diff --git a/.github/scripts/run-studio-permission-browser.sh b/.github/scripts/run-studio-permission-browser.sh
index 2007789035..e5a9a4c135 100755
--- a/.github/scripts/run-studio-permission-browser.sh
+++ b/.github/scripts/run-studio-permission-browser.sh
@@ -17,7 +17,8 @@ if [ -n "${STUDIO_PERMISSION_FRONTEND:-}" ]; then
 fi
 
 mkdir -p "$artifact_dir"
-unsloth studio reset-password
+# Wipe (not reset-password): the boot below must re-seed a fresh .bootstrap_password.
+rm -rf "$studio_home/auth"
 UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$port" "$@" \
   >"$server_log" 2>&1 &
 studio_pid=$!
diff --git a/.github/workflows/local-agent-guides-ci.yml b/.github/workflows/local-agent-guides-ci.yml
index c48328e90f..0dc0cc66d7 100644
--- a/.github/workflows/local-agent-guides-ci.yml
+++ b/.github/workflows/local-agent-guides-ci.yml
@@ -167,7 +167,9 @@ jobs:
       # ── boot the server under test (factored helper) ──────────────────
       - name: Serve unsloth run --disable-tools (gemma-4-E4B)
         run: |
-          unsloth studio reset-password
+          # Wipe, not reset-password: since #7573 the reset rotates in place and
+          # prints the new passphrase, which would land unmasked in the job log.
+          rm -rf ~/.unsloth/studio/auth
           bash .github/scripts/serve-unsloth-run.sh \
             --gguf-file "$GITHUB_WORKSPACE/gguf-cache/${GGUF_FILE}" \
             --port "$STUDIO_PORT" --log-dir logs \
@@ -371,7 +373,7 @@ jobs:
 
       - name: Serve unsloth run --disable-tools (gemma-4-E4B)
         run: |
-          unsloth studio reset-password
+          rm -rf ~/.unsloth/studio/auth
           bash .github/scripts/serve-unsloth-run.sh \
             --gguf-file "$GITHUB_WORKSPACE/gguf-cache/${GGUF_FILE}" \
             --port "$STUDIO_PORT" --log-dir logs \
@@ -554,7 +556,7 @@ jobs:
 
       - name: Serve unsloth run --disable-tools (gemma-4-E4B)
         run: |
-          unsloth studio reset-password
+          rm -rf ~/.unsloth/studio/auth
           bash .github/scripts/serve-unsloth-run.sh \
             --gguf-file "$GITHUB_WORKSPACE/gguf-cache/${GGUF_FILE}" \
             --port "$STUDIO_PORT" --log-dir logs \
@@ -718,7 +720,7 @@ jobs:
 
       - name: Serve unsloth run --disable-tools (gemma-3-270m)
         run: |
-          unsloth studio reset-password
+          rm -rf ~/.unsloth/studio/auth
           bash .github/scripts/serve-unsloth-run.sh \
             --model "$GGUF_REPO" --gguf-variant "$GGUF_VARIANT" \
             --port "$STUDIO_PORT" --log-dir logs \
diff --git a/.github/workflows/release-desktop.yml b/.github/workflows/release-desktop.yml
index 081eda4e32..0a8d71610d 100644
--- a/.github/workflows/release-desktop.yml
+++ b/.github/workflows/release-desktop.yml
@@ -766,6 +766,7 @@ jobs:
     env:
       GH_REPO: ${{ github.repository }}
       APP_VERSION: ${{ needs.prepare-version.outputs.app_version }}
+      PYPI_VERSION: ${{ needs.prepare-version.outputs.pypi_version }}
       STUDIO_VERSION: ${{ needs.prepare-version.outputs.studio_version }}
       DESKTOP_RELEASE_TAG: ${{ needs.prepare-version.outputs.desktop_release_tag }}
       DESKTOP_PRERELEASE: ${{ needs.prepare-version.outputs.prerelease }}
@@ -911,6 +912,8 @@ jobs:
           notes = pathlib.Path(os.environ['RUNNER_TEMP'], 'desktop-release-notes.md').read_text()
           metadata = {
               'version': os.environ['APP_VERSION'],
+              # App version is SemVer; CHANGELOG.md is keyed by the backend release.
+              'pypi_version': os.environ['PYPI_VERSION'],
               'notes': notes,
               'pub_date': datetime.datetime.now(datetime.timezone.utc).isoformat(timespec='milliseconds').replace('+00:00', 'Z'),
               'platforms': {
diff --git a/.github/workflows/startup-profile-ci.yml b/.github/workflows/startup-profile-ci.yml
new file mode 100644
index 0000000000..fbde99836d
--- /dev/null
+++ b/.github/workflows/startup-profile-ci.yml
@@ -0,0 +1,156 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+# Measures where Studio's startup time goes, on each platform.
+#
+# Nothing recorded a number before: main.py logs "lifespan startup completed in X ms"
+# and studio_test_kit polls /healthz, but both throw the elapsed time away. A first
+# local run (Linux, warm cache, 18-core server) put `import main` at 5.7-6.6s BEFORE
+# the server can bind, dominated by eager module-level imports pulled in by routes:
+# torch ~1.9s self, unsloth_zoo ~0.8s, routes ~0.6s, transformers ~0.5s.
+#
+# Not a gate yet: --max-healthz-seconds exists, but a budget should come from
+# observed numbers rather than a guess.
+
+name: Startup profile
+
+on:
+  pull_request:
+    paths:
+      # The measured import graph is the whole backend tree: main.py imports auth,
+      # core, hub, loggers, models, picker, routes and utils at module scope.
+      - 'studio/backend/**'
+      - '!studio/backend/tests/**'
+      # The launch phase spawns `unsloth studio --api-only`, so the CLI counts too.
+      - 'unsloth_cli/**'
+      - 'studio/src-tauri/src/preflight**'
+      # The profiler hardcodes the desktop argv that process.rs::backend_args builds,
+      # so a change there must schedule a run or the two silently diverge.
+      - 'studio/src-tauri/src/process.rs'
+      - 'scripts/profile_startup.py'
+      - '.github/workflows/startup-profile-ci.yml'
+      # The job profiles whatever `install.sh --local` built: the installers pick the
+      # venv's Python and the dependency specs, and pyproject's include list is what
+      # makes --local overlay studio.backend*.
+      - 'install.sh'
+      - 'install.ps1'
+      - 'pyproject.toml'
+      # --local also runs the checkout's setup scripts (install.sh picks
+      # $_REPO_ROOT/studio/setup.sh, the editable install resolves setup.ps1 to the
+      # repo), and both call install_python_stack.py, which picks the dependencies.
+      - 'studio/setup.sh'
+      - 'studio/setup.ps1'
+      - 'studio/install_python_stack.py'
+  workflow_dispatch:
+    inputs:
+      repeats:
+        description: 'launch repeats per OS (median reported)'
+        type: string
+        default: '3'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+permissions:
+  contents: read
+
+jobs:
+  profile:
+    name: startup ${{ matrix.os }}
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 60
+    continue-on-error: true
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-14, windows-latest]
+
+    env:
+      UNSLOTH_STUDIO_HOME: ${{ github.workspace }}/.studio-home
+      # A wildcard bind calls ifconfig.me on the startup path; loopback times our code.
+      UNSLOTH_STUDIO_DISABLE_PUBLIC_CHECK: '1'
+
+    steps:
+      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
+        with:
+          persist-credentials: false
+
+      - name: Install Studio
+        shell: bash
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          set -o pipefail
+          mkdir -p logs
+          # --local is load-bearing: it overlays the checkout, so the profiled server
+          # is this diff. Without it install.sh resolves unsloth from PyPI.
+          if [ "${{ runner.os }}" = "Windows" ]; then
+            pwsh -NoProfile -File ./install.ps1 --local 2>&1 | tee logs/install.log
+          else
+            bash install.sh --local 2>&1 | tee logs/install.log
+          fi
+
+      - name: Profile startup
+        shell: bash
+        run: |
+          BIN="$UNSLOTH_STUDIO_HOME/unsloth_studio/bin/unsloth"
+          [ -x "$BIN" ] || BIN="$UNSLOTH_STUDIO_HOME/unsloth_studio/Scripts/unsloth.exe"
+          [ -x "$BIN" ] || BIN=""
+          # Profile imports with the INSTALLED interpreter: that venv is what launches.
+          PY="$UNSLOTH_STUDIO_HOME/unsloth_studio/bin/python"
+          [ -x "$PY" ] || PY="$UNSLOTH_STUDIO_HOME/unsloth_studio/Scripts/python.exe"
+          [ -x "$PY" ] || PY="$(command -v python3 || command -v python)"
+          python3 scripts/profile_startup.py \
+            --python "$PY" \
+            ${BIN:+--bin "$BIN"} \
+            --repeats "${{ inputs.repeats || '3' }}" \
+            --json "startup-${{ matrix.os }}.json" 2>&1 | tee logs/profile.log
+
+      - name: Summary
+        if: always()
+        shell: bash
+        run: |
+          f="startup-${{ matrix.os }}.json"
+          [ -f "$f" ] || { echo "no profile produced"; exit 0; }
+          python3 - "$f" >> "$GITHUB_STEP_SUMMARY" <<'PY'
+          import json, sys
+          d = json.load(open(sys.argv[1]))
+          print(f"### {d['platform']} / {d['machine']} (py {d['python']}, {d['cpu_count']} cpu)\n")
+          imp = d.get("imports", {})
+          # Gate on ok: a failed `import main` still leaves rows, so a total can lie.
+          if imp.get("ok"):
+              print(f"**`import main`: {imp['total_seconds']}s**\n")
+              print("| package | self ms |")
+              print("|---|---:|")
+              for k, v in list(imp.get("self_by_package_ms", {}).items())[:8]:
+                  print(f"| {k} | {v} |")
+              print()
+          else:
+              print("**`import main` failed - no valid import profile**\n")
+              print("```\n" + (imp.get("error") or "")[-1500:] + "\n```\n")
+          lau = d.get("launch") or {}
+          runs = len(lau.get("runs") or [])
+          failed = lau.get("failed_runs") or 0
+          if lau.get("healthz_median_seconds") is not None:
+              # The aggregates cover only the runs that reached healthz, so flag the
+              # failures: bare numbers would read as a normal fast startup.
+              note = f" _({runs - failed} of {runs} launches; {failed} never became healthy)_" if failed else ""
+              print(f"**time to a healthy port: {lau['healthz_median_seconds']}s median, "
+                    f"{lau['healthz_max_seconds']}s max**{note}\n")
+          elif lau.get("skipped"):
+              print(f"_launch phase skipped: {lau['skipped']}_\n")
+          elif runs:
+              print(f"**no launch measurement: all {runs} launches failed to become healthy**\n")
+          PY
+
+      - name: Upload profile
+        if: always()
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
+        with:
+          name: startup-profile-${{ matrix.os }}
+          path: |
+            startup-*.json
+            logs/
+          retention-days: 14
+          if-no-files-found: warn
diff --git a/.github/workflows/studio-api-smoke.yml b/.github/workflows/studio-api-smoke.yml
index cdf1f6bf12..1cfa66fea4 100644
--- a/.github/workflows/studio-api-smoke.yml
+++ b/.github/workflows/studio-api-smoke.yml
@@ -113,7 +113,8 @@ jobs:
 
       - name: Reset auth + boot Unsloth (API-only)
         run: |
-          unsloth studio reset-password
+          # Wipe (not reset-password): the boot below must re-seed a fresh .bootstrap_password.
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
diff --git a/.github/workflows/studio-backend-ci.yml b/.github/workflows/studio-backend-ci.yml
index ec437e0c32..dd5efbb299 100644
--- a/.github/workflows/studio-backend-ci.yml
+++ b/.github/workflows/studio-backend-ci.yml
@@ -223,6 +223,16 @@ jobs:
             tests/studio/test_is_mlx_dispatch_gate.py \
             tests/studio/test_xpu_spoof_pipeline.py
 
+      - name: CLI tests (unsloth_cli)
+        # unsloth_cli/tests had no CI at all: `unsloth_cli/**` was only a paths
+        # trigger and a ruff target, so 673 tests covering the studio launcher,
+        # the pre-exposure gate and the auth secret writers ran nowhere, and
+        # four of them had been failing on main unnoticed.
+        # Own step, not folded into the tests/ discovery above: pyproject's
+        # testpaths is tests/, and this suite needs no PYTHONPATH or CUDA spoof
+        # (it self-bootstraps sys.path and imports neither unsloth nor torch).
+        run: python -m pytest unsloth_cli/tests -q --tb=short
+
       - name: Shell installer tests
         # Auto-discovered rather than allowlisted. The old hardcoded list had
         # silently fallen seven files behind tests/run_all.sh, including
diff --git a/.github/workflows/studio-inference-smoke.yml b/.github/workflows/studio-inference-smoke.yml
index c2d52eac22..c37c9555bf 100644
--- a/.github/workflows/studio-inference-smoke.yml
+++ b/.github/workflows/studio-inference-smoke.yml
@@ -127,7 +127,8 @@ jobs:
 
       - name: Reset auth + boot Unsloth (API-only)
         run: |
-          unsloth studio reset-password
+          # Wipe (not reset-password): the boot below must re-seed a fresh .bootstrap_password.
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
@@ -400,7 +401,7 @@ jobs:
         # tool_policy=None so each request's `enable_tools` field is
         # honoured.
         run: |
-          unsloth studio reset-password
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
@@ -978,7 +979,7 @@ jobs:
         # response_format requests aren't routed through the agentic
         # tool loop.
         run: |
-          unsloth studio reset-password
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
diff --git a/.github/workflows/studio-mac-api-smoke.yml b/.github/workflows/studio-mac-api-smoke.yml
index 1968885a1d..c2307f17a1 100644
--- a/.github/workflows/studio-mac-api-smoke.yml
+++ b/.github/workflows/studio-mac-api-smoke.yml
@@ -101,7 +101,8 @@ jobs:
 
       - name: Reset auth + boot Unsloth (API-only)
         run: |
-          unsloth studio reset-password
+          # Wipe (not reset-password): the boot below must re-seed a fresh .bootstrap_password.
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
diff --git a/.github/workflows/studio-mac-inference-smoke.yml b/.github/workflows/studio-mac-inference-smoke.yml
index ce15eed5c8..1dbf86ae98 100644
--- a/.github/workflows/studio-mac-inference-smoke.yml
+++ b/.github/workflows/studio-mac-inference-smoke.yml
@@ -126,7 +126,8 @@ jobs:
 
       - name: Reset auth + boot Unsloth (API-only)
         run: |
-          unsloth studio reset-password
+          # Wipe (not reset-password): the boot below must re-seed a fresh .bootstrap_password.
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
@@ -386,7 +387,7 @@ jobs:
         # tool_policy=None so each request's `enable_tools` field is
         # honoured.
         run: |
-          unsloth studio reset-password
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
@@ -831,7 +832,7 @@ jobs:
         # response_format requests aren't routed through the agentic
         # tool loop.
         run: |
-          unsloth studio reset-password
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
diff --git a/.github/workflows/studio-mac-ui-smoke.yml b/.github/workflows/studio-mac-ui-smoke.yml
index 7375e9bcbf..3bed2fcdff 100644
--- a/.github/workflows/studio-mac-ui-smoke.yml
+++ b/.github/workflows/studio-mac-ui-smoke.yml
@@ -146,7 +146,8 @@ jobs:
 
       - name: Reset auth + boot Unsloth
         run: |
-          unsloth studio reset-password
+          # Wipe (not reset-password): the boot below must re-seed a fresh .bootstrap_password.
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
@@ -190,7 +191,7 @@ jobs:
         # runner's kernel briefly runs out of socket buffers, and (3) a
         # goto 'interrupted by another navigation' when the SPA auth
         # guard redirects mid-navigation. The retry FULLY resets Unsloth
-        # (kill, reset-password, reboot, wait /api/health, re-export
+        # (kill, wipe auth, reboot, wait /api/health, re-export
         # bootstrap pw) before re-running the script. A real test failure
         # (assertion / timeout) does NOT match any pattern so it bypasses
         # retry and surfaces immediately.
@@ -213,7 +214,7 @@ jobs:
               echo "::warning::Playwright flake on attempt ${attempt}; resetting Unsloth and retrying..."
               kill "${STUDIO_PID}" 2>/dev/null || true
               sleep 2
-              unsloth studio reset-password
+              rm -rf ~/.unsloth/studio/auth
               UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
                 > "logs/studio_retry_${attempt}.log" 2>&1 &
               STUDIO_PID=$!
@@ -251,7 +252,7 @@ jobs:
 
       - name: Reset auth + boot Unsloth for extra UI tests (port 18897)
         run: |
-          unsloth studio reset-password
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p 18897 \
             > logs/studio_extra.log 2>&1 &
@@ -308,7 +309,7 @@ jobs:
               echo "::warning::Playwright flake on attempt ${attempt}; resetting Unsloth and retrying..."
               kill "${STUDIO_EXTRA_PID}" 2>/dev/null || true
               sleep 2
-              unsloth studio reset-password
+              rm -rf ~/.unsloth/studio/auth
               UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p 18897 \
                 > "logs/studio_extra_retry_${attempt}.log" 2>&1 &
               STUDIO_EXTRA_PID=$!
diff --git a/.github/workflows/studio-ui-smoke.yml b/.github/workflows/studio-ui-smoke.yml
index 97eb07b2d8..3a0713f301 100644
--- a/.github/workflows/studio-ui-smoke.yml
+++ b/.github/workflows/studio-ui-smoke.yml
@@ -115,7 +115,8 @@ jobs:
 
       - name: Reset auth + boot Unsloth
         run: |
-          unsloth studio reset-password
+          # Wipe (not reset-password): the boot below must re-seed a fresh .bootstrap_password.
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
@@ -193,7 +194,7 @@ jobs:
       # warm install we already did) so this adds little wall time.
       - name: Reset auth + boot Unsloth for extra UI tests (port 18894)
         run: |
-          unsloth studio reset-password
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p 18894 \
             > logs/studio_extra.log 2>&1 &
@@ -253,7 +254,7 @@ jobs:
       # (RAG embedder + llama.cpp probe) stay hidden from the picker.
       - name: Reset auth + boot Unsloth for model-config tests (port 18898)
         run: |
-          unsloth studio reset-password
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p 18898 \
             > logs/studio_modelcfg.log 2>&1 &
@@ -299,7 +300,7 @@ jobs:
       # earlier UI tests. No GGUF -- the bug surface is the composer.
       - name: Reset auth + boot Unsloth for IME / i18n tests (port 18896)
         run: |
-          unsloth studio reset-password
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p 18896 \
             > logs/studio_ime.log 2>&1 &
diff --git a/.github/workflows/studio-windows-api-smoke.yml b/.github/workflows/studio-windows-api-smoke.yml
index 6dbcceebbd..b328939846 100644
--- a/.github/workflows/studio-windows-api-smoke.yml
+++ b/.github/workflows/studio-windows-api-smoke.yml
@@ -179,7 +179,8 @@ jobs:
 
       - name: Reset auth + boot Unsloth (API-only)
         run: |
-          unsloth studio reset-password
+          # Wipe (not reset-password): the boot below must re-seed a fresh .bootstrap_password.
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
diff --git a/.github/workflows/studio-windows-inference-smoke.yml b/.github/workflows/studio-windows-inference-smoke.yml
index 3ebe442f52..d821664327 100644
--- a/.github/workflows/studio-windows-inference-smoke.yml
+++ b/.github/workflows/studio-windows-inference-smoke.yml
@@ -229,7 +229,8 @@ jobs:
 
       - name: Reset auth + boot Unsloth (API-only)
         run: |
-          unsloth studio reset-password
+          # Wipe (not reset-password): the boot below must re-seed a fresh .bootstrap_password.
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
@@ -573,7 +574,7 @@ jobs:
 
       - name: Reset auth + boot Unsloth (API-only, default tool policy)
         run: |
-          unsloth studio reset-password
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
@@ -1074,7 +1075,7 @@ jobs:
 
       - name: Reset auth + boot Unsloth (API-only)
         run: |
-          unsloth studio reset-password
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
@@ -1546,7 +1547,7 @@ jobs:
 
       - name: Reset auth + boot Unsloth (API-only)
         run: |
-          unsloth studio reset-password
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
@@ -1888,8 +1889,11 @@ jobs:
           # (step/substep -> Write-StudioStdoutMirror / Get-StudioAnsi).
           $script:StudioVtOk = $false
           $script:UnslothVerbose = $false
+          # Get-HostMachineArch is reached only on the absent path, where
+          # Test-VCRedistInstalled consults it before trusting the System32 DLL, so
+          # part A passes without it and only the clean-box part fails.
           foreach ($fn in @('Get-StudioAnsi', 'Write-StudioStdoutMirror', 'step', 'substep',
-                            'Invoke-SetupCommand', 'Refresh-Environment',
+                            'Invoke-SetupCommand', 'Refresh-Environment', 'Get-HostMachineArch',
                             'Test-VCRedistInstalled', 'Ensure-VCRedist')) {
               $src = Get-FunctionSource -Path $setup -Name $fn
               if (-not $src) { throw "Function '$fn' not found in setup.ps1" }
diff --git a/.github/workflows/studio-windows-ui-smoke.yml b/.github/workflows/studio-windows-ui-smoke.yml
index f401f7be44..d23cca323f 100644
--- a/.github/workflows/studio-windows-ui-smoke.yml
+++ b/.github/workflows/studio-windows-ui-smoke.yml
@@ -297,7 +297,8 @@ jobs:
 
       - name: Reset auth + boot Unsloth
         run: |
-          unsloth studio reset-password
+          # Wipe (not reset-password): the boot below must re-seed a fresh .bootstrap_password.
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
             > logs/studio.log 2>&1 &
@@ -352,7 +353,7 @@ jobs:
 
       - name: Reset auth + boot Unsloth for extra UI tests (port 18897)
         run: |
-          unsloth studio reset-password
+          rm -rf ~/.unsloth/studio/auth
           mkdir -p logs
           UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p 18897 \
             > logs/studio_extra.log 2>&1 &
diff --git a/.gitignore b/.gitignore
index fafd17aa95..fa6997cb06 100644
--- a/.gitignore
+++ b/.gitignore
@@ -208,6 +208,9 @@ tmp/
 **/node_modules/
 auth.db
 
+# Packaging snapshot of the root CHANGELOG.md (written by build.sh)
+studio/CHANGELOG.md
+
 # Tauri local build/generated output
 studio/src-tauri/target/
 studio/src-tauri/gen/
diff --git a/CHANGELOG.md b/CHANGELOG.md
new file mode 100644
index 0000000000..241e013cea
--- /dev/null
+++ b/CHANGELOG.md
@@ -0,0 +1,88 @@
+# Changelog
+
+Release notes for Unsloth and Unsloth Studio.
+
+Unsloth Studio reads this file to show release notes inside the "New Unsloth
+version" update popup. Edit it here and the popup picks the change up on the
+next update check, with no release or rebuild required.
+
+## Format
+
+Every release is a level-2 heading whose first token is the version, optionally
+followed by a date:
+
+```md
+## 2026.7.6 - 2026-07-22
+```
+
+`## [2026.7.6] - 2026-07-22` and `## v2026.7.6` also work. Everything under a
+heading, up to the next level-2 heading, is that release's notes and renders as
+Markdown in the popup.
+
+Notes are matched to one exact version. When Studio offers an update to
+`2026.7.6` it renders the `2026.7.6` section and nothing else. If that section
+is missing, the popup links out to the online changelog rather than showing
+notes from an unrelated release, so a new version needs its own section here
+before its notes can appear.
+
+Keep the newest release at the top. Lead each bullet with the change itself:
+the collapsed popup highlights the first sentence and dims the rest.
+`## Unreleased` is ignored by the popup, so it is safe to stage notes there and
+rename the heading at release time.
+
+<!-- Add new releases directly below this line. -->
+
+## Unreleased
+
+## 2026.7.5
+
+### What's Changed
+
+- AMD support is here. Train, run RL, chat with and deploy 500+ models on
+  Radeon, Instinct, Ryzen and data center GPUs across Windows, WSL and Linux,
+  up to 2x faster with 70% less VRAM and no accuracy loss.
+- Intel XPU support lands in Studio, so Arc and Data Center GPUs run chat and
+  training alongside the NVIDIA, AMD and Apple paths.
+- Local speech to text dictation runs fully offline, with slim Whisper bundles
+  and a picker for custom models.
+- DoRA training is available in Studio, selectable next to LoRA and full
+  fine-tuning in the training tab.
+- The update popup previews release notes inline, pulled from this file and
+  matched to the exact version being offered.
+
+### AMD, 23 July update
+
+Our AMD collaboration, custom Triton kernels and math algorithms bring local
+training and inference to AMD hardware. The 23 July update builds on the
+[AMD release](https://github.com/unslothai/unsloth/releases/tag/v0.1.501-beta):
+
+- RDNA2 and Gorgon Halo are supported, and the installer no longer fails to
+  detect GPUs on Strix Halo and other AMD cards.
+- RDNA4 handling is better, and HIP and ROCm failures are caught and fixed
+  automatically instead of stopping the install.
+- Unified memory safetensors loading is 2x faster, with much faster gradient
+  checkpointing on unified memory devices.
+- Voice dictation through whisper.cpp has preliminary support.
+- Rollback environments left by installs no longer eat 5GB of disk. They are
+  cleaned up automatically.
+
+Optimized ROCm builds cover GGUF and safetensors inference, and ROCm
+compatibility is improved for MI300X and MI325X. Full guide:
+[unsloth.ai/docs/basics/amd](https://unsloth.ai/docs/basics/amd).
+
+### Running larger models
+
+- Automatic GPU placement, or pick exactly which GPUs and layers to use.
+- Move MoE expert layers into system memory so larger models fit.
+- Split a model across several GPUs, or use tensor parallelism.
+- Hardware settings are saved per model and quant.
+
+### Also in this release
+
+- Remote access with `unsloth studio --secure` over free HTTPS via Cloudflare.
+- Web search reads PDF papers and manuals, and parallel tool calls, reasoning
+  output and tool retries are more reliable.
+- The model download location is configurable, so weights can live on a second
+  drive instead of the default cache.
+- Stalled Hugging Face XET downloads retry over standard HTTP, and existing
+  GGUF files are reused instead of downloaded again.
diff --git a/MANIFEST.in b/MANIFEST.in
new file mode 100644
index 0000000000..7bce036343
--- /dev/null
+++ b/MANIFEST.in
@@ -0,0 +1,2 @@
+include _changelog_build.py
+include CHANGELOG.md
diff --git a/README.md b/README.md
index 849958bfbb..e0fc8ee44c 100644
--- a/README.md
+++ b/README.md
@@ -84,8 +84,9 @@ Replace `claude` with any supported agent:
 | Hermes Agent | `unsloth start hermes` |
 | OpenClaw | `unsloth start openclaw` |
 | OpenCode | `unsloth start opencode` |
+| Pi Coding Agent | `unsloth start pi` |
 
-Claude Code, Codex and OpenCode can keep their current model and use Unsloth as a local
+Claude Code, Codex, OpenCode and Pi can keep their current model and use Unsloth as a local
 subagent:
 
 ```bash
@@ -203,7 +204,7 @@ Read our [guide](https://unsloth.ai/docs/get-started/fine-tuning-llms-guide). Ad
 ## 🦥 Unsloth News
 - **AMD training**: Train, run RL, chat and deploy on AMD GPUs across Windows, WSL and Linux. [Guide](https://unsloth.ai/docs/basics/amd)
 - **GGUF hardware controls**: Choose GPU/layer placement, offload MoE experts and use multi-GPU or Tensor Parallelism. [#6414](https://github.com/unslothai/unsloth/pull/6414)
-- **Local models for any agent**: Use `unsloth start` with Claude Code, Codex, Hermes, OpenCode, OpenClaw and more through Unsloth's OpenAI- and Anthropic-compatible APIs. [Guide](https://unsloth.ai/docs/basics/api)
+- **Local models for any agent**: Use `unsloth start` with Claude Code, Codex, Hermes, OpenCode, OpenClaw, Pi and more through Unsloth's OpenAI- and Anthropic-compatible APIs. [Guide](https://unsloth.ai/docs/basics/api)
 - **MCP control endpoint**: Let compatible clients manage models, training, recipes, checkpoints and exports. [#7191](https://github.com/unslothai/unsloth/pull/7191)
 - **Local inference reliability**: Resume long chats faster, recover stalled downloads and reuse existing GGUF files. [#7204](https://github.com/unslothai/unsloth/pull/7204) • [#6858](https://github.com/unslothai/unsloth/pull/6858) • [#7209](https://github.com/unslothai/unsloth/pull/7209)
 - **New models**: [Qwen-AgentWorld](https://huggingface.co/unsloth/Qwen-AgentWorld-35B-A3B-GGUF), [Ornith](https://huggingface.co/unsloth/models?search=ornith), [Kimi K2.7 Code](https://unsloth.ai/docs/models/kimi-k2.7-code) and [MiniMax M3](https://unsloth.ai/docs/models/minimax-m3)
diff --git a/_changelog_build.py b/_changelog_build.py
new file mode 100644
index 0000000000..f5bcf2052c
--- /dev/null
+++ b/_changelog_build.py
@@ -0,0 +1,36 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Snapshot CHANGELOG.md into the studio package at build time.
+
+CHANGELOG.md at the repo root stays the one file to edit. Copying it here,
+rather than in build.sh, means every packaging path ships it, so release notes
+still render when the popup cannot reach GitHub."""
+
+from __future__ import annotations
+
+import shutil
+from pathlib import Path
+
+from setuptools.command.build_py import build_py as _build_py
+
+ROOT = Path(__file__).resolve().parent
+SOURCE = ROOT / "CHANGELOG.md"
+SNAPSHOT = ROOT / "studio" / "CHANGELOG.md"
+
+
+class build_py(_build_py):
+    def run(self) -> None:
+        # Beside the sources only if writable (PEP 517 may build an immutable
+        # checkout); into the staging directory always.
+        if SOURCE.is_file():
+            try:
+                shutil.copyfile(SOURCE, SNAPSHOT)
+            except OSError:
+                pass
+        super().run()
+        if not SOURCE.is_file():
+            return
+        staged = Path(self.build_lib) / "studio" / "CHANGELOG.md"
+        staged.parent.mkdir(parents = True, exist_ok = True)
+        shutil.copyfile(SOURCE, staged)
diff --git a/build.sh b/build.sh
index 2a836e19d9..5b09a7791b 100644
--- a/build.sh
+++ b/build.sh
@@ -103,9 +103,13 @@ else
     STUDIO_STAMPED_VERSION="$(python scripts/stamp_studio_release.py)"
 fi
 
-# 4. Build wheel/sdist
+# 4. Build wheel/sdist. _changelog_build.py snapshots CHANGELOG.md into the studio
+# package so release notes render offline.
 python -m build
 
+# Drop the snapshot so a source checkout never serves a stale copy.
+rm -f studio/CHANGELOG.md
+
 if [ "${1:-}" = "publish" ]; then
     python scripts/stamp_studio_release.py --verify-dist dist --expected "$STUDIO_STAMPED_VERSION"
 fi
diff --git a/install.ps1 b/install.ps1
index 0b06cb3ea1..5b205df96d 100644
--- a/install.ps1
+++ b/install.ps1
@@ -57,6 +57,26 @@ function Install-UnslothStudio {
         }
     }
 
+    # Machine arch; Get-TauriDiagArch above reports the process. An emulated x64 shell on
+    # ARM64 reports AMD64, but PROCESSOR_ARCHITEW6432 is ARM64 in exactly that case.
+    function Get-HostMachineArch {
+        $osArch = ""
+        try { $osArch = [System.Runtime.InteropServices.RuntimeInformation]::OSArchitecture.ToString() } catch { $osArch = "" }
+        $signals = @([string]$env:PROCESSOR_ARCHITEW6432, [string]$env:PROCESSOR_ARCHITECTURE, $osArch)
+        foreach ($s in $signals) {
+            if ($s.ToLowerInvariant() -eq "arm64") { return "arm64" }
+        }
+        foreach ($s in $signals) {
+            if ([string]::IsNullOrWhiteSpace($s)) { continue }
+            switch ($s.ToLowerInvariant()) {
+                "amd64" { return "x86_64" }
+                "x64" { return "x86_64" }
+                "x86" { return "x86" }
+            }
+        }
+        return "unknown"
+    }
+
     function Get-TauriTorchIndexFamily {
         param([string]$TorchIndexUrl)
         if ($SkipTorch) { return "none" }
@@ -1124,10 +1144,27 @@ exit 0
         return $false
     }
 
+    # The interpreter's own arch, asked of it: win-amd64|win-arm64|win32|"".
+    function Get-PythonPlatformTag {
+        param([string]$Exe)
+        try {
+            return (& $Exe -c "import sysconfig; print(sysconfig.get_platform())" 2>$null | Out-String).Trim().ToLowerInvariant()
+        } catch { return "" }
+    }
+
     # Returns @{ Version = "3.13"; Path = "C:\...\python.exe" } or $null.
     # The resolved Path is passed to `uv venv --python` to prevent uv from
     # re-resolving the version string back to a conda interpreter.
     function Find-CompatiblePython {
+        # -X64Only: best installed x64 interpreter or $null, never ARM64. Last resort for
+        # Install-X64Python, where x64 of a lower-priority minor beats ARM64.
+        param([switch]$X64Only)
+        # Windows on ARM: prefer x64. pyarrow (via datasets) and hf-transfer ship no
+        # win_arm64 wheel, so a native ARM64 Python source-builds both and dies on CMake /
+        # Rust minutes in; x64 runs fine emulated. ARM64 is still returned when it is all
+        # there is, and the caller then bootstraps x64 or warns.
+        $preferX64 = $X64Only -or ((Get-HostMachineArch) -eq "arm64")
+        $candidates = @()
         # Try the Python Launcher first (most reliable on Windows)
         # py.exe resolves to the standard CPython install, not conda.
         # Prefer the requested $PythonVersion, then newest-first fallback.
@@ -1145,7 +1182,8 @@ exit 0
                         # Resolve the actual executable path and verify it is not conda-based
                         $resolvedExe = (& $pyLauncher.Source "-$minor" -c "import sys; print(sys.executable)" 2>$null | Out-String).Trim()
                         if ($resolvedExe -and (Test-Path $resolvedExe) -and -not (Test-IsCondaPython $resolvedExe)) {
-                            return @{ Version = $ver; Path = $resolvedExe }
+                            if (-not $preferX64) { return @{ Version = $ver; Path = $resolvedExe; Arch = "" } }
+                            $candidates += @{ Version = $ver; Path = $resolvedExe }
                         }
                     }
                 } catch {}
@@ -1166,11 +1204,53 @@ exit 0
                 try {
                     $out = & $cmd.Source --version 2>&1 | Out-String
                     if ($out -match "Python (3\.1[1-3])\.\d+") {
-                        return @{ Version = $Matches[1]; Path = $cmd.Source }
+                        if (-not $preferX64) { return @{ Version = $Matches[1]; Path = $cmd.Source; Arch = "" } }
+                        $candidates += @{ Version = $Matches[1]; Path = $cmd.Source }
                     }
                 } catch {}
             }
         }
+        # `py -3.12` runs the launcher's preferred build, normally the native ARM64 one, so
+        # a same-minor x64 install that is neither preferred nor on PATH never becomes a
+        # candidate. `-3.12-64` cannot disambiguate (deprecated, it only means "not
+        # 32-bit"), so enumerate every registration with -0p and probe each path.
+        if ($preferX64) {
+            foreach ($pyLauncher in @(Get-Command py -All -CommandType Application -ErrorAction SilentlyContinue)) {
+                if ($pyLauncher.Source -match $script:CondaSkipPattern) { continue }
+                $listed = @()
+                try { $listed = @(& $pyLauncher.Source "-0p" 2>$null) } catch {}
+                foreach ($line in $listed) {
+                    # " -V:3.12 *   C:\...\python.exe": tag, optional default marker, path.
+                    $m = [regex]::Match([string]$line, '(?i)^\s*-\S+\s+\*?\s*"?(?<p>\S.*?\.exe)"?\s*$')
+                    if (-not $m.Success) { continue }
+                    $exe = $m.Groups['p'].Value.Trim()
+                    if ($candidates | Where-Object { $_.Path -eq $exe }) { continue }
+                    if (-not (Test-Path -LiteralPath $exe)) { continue }
+                    if (Test-IsCondaPython $exe) { continue }
+                    try {
+                        $out = & $exe --version 2>&1 | Out-String
+                        if ($out -match "Python (3\.1[1-3])\.\d+") {
+                            $candidates += @{ Version = $Matches[1]; Path = $exe }
+                        }
+                    } catch {}
+                }
+            }
+        }
+        # Prefer x64, but only within one minor: $minors is the caller's version preference,
+        # so ranking on arch alone would answer UNSLOTH_PYTHON=3.12 with an x64 3.13 and
+        # never bootstrap x64 3.12. Probing costs a subprocess, so non-ARM returned above.
+        foreach ($c in $candidates) {
+            $tag = Get-PythonPlatformTag $c.Path
+            $c.Arch = if ($tag -eq "win-amd64") { "x86_64" } elseif ($tag -eq "win-arm64") { "arm64" } else { "unknown" }
+        }
+        foreach ($minor in $minors) {
+            $sameMinor = @($candidates | Where-Object { $_.Version -eq $minor })
+            if ($sameMinor.Count -eq 0) { continue }
+            $x64 = $sameMinor | Where-Object { $_.Arch -eq "x86_64" } | Select-Object -First 1
+            if ($x64) { return $x64 }
+            if (-not $X64Only) { return $sameMinor[0] }
+        }
+        if (-not $X64Only -and $candidates.Count -gt 0) { return $candidates[0] }
         return $null
     }
 
@@ -1181,8 +1261,11 @@ exit 0
     # (no UAC), putting python.exe + the py launcher on PATH. Mirrors the uv ->
     # astral.sh fallback below. Returns @{ Version; Path } or $null.
     function Install-PythonFromPythonOrg {
+        # $Arch overrides the host arch, to pull x64 onto an ARM64 box.
+        param([string]$Arch = "")
         # python.org ships one installer per architecture.
-        $archSuffix = switch (Get-TauriDiagArch) {
+        $targetArch = if ($Arch) { $Arch } else { Get-TauriDiagArch }
+        $archSuffix = switch ($targetArch) {
             "x86_64" { "-amd64" }
             "arm64"  { "-arm64" }
             "x86"    { "" }
@@ -1247,6 +1330,28 @@ exit 0
         return (Find-CompatiblePython)
     }
 
+    # ── Windows on ARM: get an x64 CPython ──
+    # --architecture x64 forces winget off the ARM64 build; python.org takes the same override.
+    function Install-X64Python {
+        if ($script:WingetAvailable) {
+            $prevEAP = $ErrorActionPreference
+            $ErrorActionPreference = "Continue"
+            try {
+                winget install -e --id "Python.Python.$PythonVersion" --source winget --architecture x64 --accept-package-agreements --accept-source-agreements
+            } catch { }
+            $ErrorActionPreference = $prevEAP
+            Refresh-SessionPath
+            $found = Find-CompatiblePython
+            if ($found -and $found.Arch -eq "x86_64") { return $found }
+            substep "winget could not provide an x64 Python -- trying python.org..." "Yellow"
+        }
+        $found = Install-PythonFromPythonOrg -Arch "x86_64"
+        if ($found -and $found.Arch -eq "x86_64") { return $found }
+        # Nothing installable (offline / no winget): an x64 build of another supported minor
+        # still runs the wheels ARM64 cannot, so take it over the native interpreter.
+        return (Find-CompatiblePython -X64Only)
+    }
+
     # ── Install Python if no compatible version (3.11-3.13) found ──
     # Find-CompatiblePython returns @{ Version = "3.13"; Path = "C:\...\python.exe" } or $null.
     Write-TauriLog "STEP" "Installing Python"
@@ -1318,6 +1423,26 @@ exit 0
             return (Exit-InstallFailure "Python installation failed")
         }
     }
+    # ── Windows on ARM: swap a native ARM64 interpreter for x64 ──
+    # pyarrow and hf-transfer publish no win_arm64 wheel, so an ARM64 Python source-builds
+    # both and fails deep into the run. Warn up front if x64 is unobtainable.
+    if ($DetectedPython -and (Get-HostMachineArch) -eq "arm64" -and $DetectedPython.Arch -ne "x86_64") {
+        substep "windows on arm: only a native ARM64 Python $($DetectedPython.Version) was found." "Yellow"
+        substep "pyarrow and hf-transfer publish no win_arm64 wheels, so installing x64 Python..." "Yellow"
+        $X64Python = Install-X64Python
+        if ($X64Python) {
+            $DetectedPython = $X64Python
+            step "python" "using x64 Python $($DetectedPython.Version) under emulation"
+        } else {
+            Write-Host "[WARN] Could not install an x64 Python on this ARM64 machine." -ForegroundColor Yellow
+            Write-Host "       Continuing with ARM64 Python $($DetectedPython.Version), but the install is likely to fail:" -ForegroundColor Yellow
+            Write-Host "       pyarrow (via datasets) and hf-transfer ship no win_arm64 wheels and will be" -ForegroundColor Yellow
+            Write-Host "       built from source, which needs CMake plus the MSVC and Rust toolchains." -ForegroundColor Yellow
+            Write-Host "       Fix: install x64 Python from https://www.python.org/downloads/windows/" -ForegroundColor Yellow
+            Write-Host "       (choose 'Windows installer (64-bit)', not ARM64), then re-run this installer." -ForegroundColor Yellow
+        }
+    }
+
     $DiagPythonVersion = $PythonVersion
     if ($DetectedPython) { $DiagPythonVersion = $DetectedPython.Version }
     $InitialGpuBranch = "unknown"
@@ -2438,6 +2563,13 @@ exit 0
             }
         } else {
             Write-TauriLog "STEP" "Installing PyTorch"
+            # Windows on ARM lacks only torchaudio (whl/cpu win_arm64: torch 42,
+            # torchvision 60, torchaudio 0), so drop that pin instead of aborting. Ask the
+            # interpreter, not PROCESSOR_ARCHITECTURE; reached when no x64 Python exists.
+            $VenvPlatform = ""
+            try {
+                $VenvPlatform = (& $VenvPython -c "import sysconfig; print(sysconfig.get_platform())" 2>$null | Out-String).Trim().ToLowerInvariant()
+            } catch { $VenvPlatform = "" }
             substep "installing PyTorch ($(Remove-IndexUrlCredentials $TorchIndexUrl))..."
             # Bound the companions to the capped torch on EVERY index, cu<digits>
             # families included: torchaudio 2.11 dropped its exact torch pin from
@@ -2445,7 +2577,13 @@ exit 0
             # resolve a mismatched 2.11.0 build. Mirrors install.sh.
             $_pinVisionSpec = "torchvision>=0.19,<0.26.0"
             $_pinAudioSpec = "torchaudio>=2.4,<2.11.0"
-            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch" { uv pip install --python $VenvPython "torch>=2.4,<2.11.0" $_pinVisionSpec $_pinAudioSpec --default-index $TorchIndexUrl }
+            $_torchSpecs = @("torch>=2.4,<2.11.0", $_pinVisionSpec, $_pinAudioSpec)
+            if ($VenvPlatform -eq "win-arm64") {
+                substep "windows on arm: skipping torchaudio (upstream publishes no"
+                substep "win_arm64 wheel); torch and torchvision install normally."
+                $_torchSpecs = @("torch>=2.4,<2.11.0", $_pinVisionSpec)
+            }
+            $torchInstallExit = Invoke-InstallCommandRetry -Label "install PyTorch" { uv pip install --python $VenvPython @_torchSpecs --default-index $TorchIndexUrl }
             if ($torchInstallExit -ne 0) {
                 Write-Host "[ERROR] Failed to install PyTorch (exit code $torchInstallExit)" -ForegroundColor Red
                 return (Exit-InstallFailure "Failed to install PyTorch (exit code $torchInstallExit)" $torchInstallExit)
diff --git a/install.sh b/install.sh
index fc9aa0a431..166beeb52c 100755
--- a/install.sh
+++ b/install.sh
@@ -19,6 +19,17 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 set -e
+# ── Why the installer lives in a function ──
+# Under `curl ... | sh`, sh is the pipe READER. This file is ~150KB, so a top-level
+# `exit` left most of it unread, the write end failed, and curl tacked
+# "(56) Failure writing output to destination" onto our own error message. Wrapping
+# the body forces sh to parse to the closing brace first, so the pipe always drains
+# (install.ps1 has always had this shape).
+#
+# Body is deliberately NOT reindented: reflowing 4000+ lines would bury the change,
+# and `exit` still exits the shell from inside a function. Do not add
+# `exec < /dev/null`: for a piped shell that closes the script's own source.
+_unsloth_main() {
 
 # ── Output style (aligned with studio/setup.sh) ──
 RULE=""
@@ -4447,3 +4458,8 @@ else
     substep "(add -H 0.0.0.0 --cloudflare for a public Cloudflare HTTPS link, or --secure to keep the raw port private; anyone with the API key can run code)"
     echo ""
 fi
+
+}
+
+# Every byte above is parsed before this line runs, which is the point.
+_unsloth_main "$@"
diff --git a/pyproject.toml b/pyproject.toml
index 7359a51fa6..8895bf0686 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -47,9 +47,14 @@ version = {attr = "unsloth.models._utils.__version__"}
 [tool.setuptools]
 include-package-data = true
 
+[tool.setuptools.cmdclass]
+# Snapshots CHANGELOG.md into studio/ so every build path ships it.
+build_py = "_changelog_build.build_py"
+
 [tool.setuptools.package-data]
 unsloth_cli = ["codex_fallback_prompt.md", "pi_subagent.ts"]
 studio = [
+    "CHANGELOG.md",
     "*.sh",
     "*.ps1",
     "*.bat",
@@ -128,14 +133,19 @@ huggingfacenotorch = [
 ]
 # torchcodec backend for Gemma audio / datasets>=4 (#7225).
 # Pick the audio-torch* pin matching your torch minor (see TORCH_TORCHCODEC).
+# torchcodec publishes no sdist and only manylinux_2_28_x86_64, macosx_*_arm64
+# and win_amd64 wheels, so Linux aarch64, Windows ARM64 and Intel Mac have
+# nothing to resolve and pip fails the whole install rather than skipping audio.
+# Gate on the platforms that have a wheel, matching
+# PLATFORM_LACKS_TORCHCODEC_WHEEL in studio/install_python_stack.py.
 audio-torch210 = [
-    "torchcodec>=0.10.0,<0.11.0 ; python_version >= '3.10'",
+    "torchcodec>=0.10.0,<0.11.0 ; python_version >= '3.10' and (((sys_platform == 'linux' or sys_platform == 'win32') and (platform_machine == 'x86_64' or platform_machine == 'AMD64')) or (sys_platform == 'darwin' and platform_machine == 'arm64'))",
 ]
 audio-torch290 = [
-    "torchcodec>=0.8.0,<0.10.0 ; python_version >= '3.10'",
+    "torchcodec>=0.8.0,<0.10.0 ; python_version >= '3.10' and (((sys_platform == 'linux' or sys_platform == 'win32') and (platform_machine == 'x86_64' or platform_machine == 'AMD64')) or (sys_platform == 'darwin' and platform_machine == 'arm64'))",
 ]
 audio-torch280 = [
-    "torchcodec>=0.6.0,<0.8.0 ; python_version >= '3.9'",
+    "torchcodec>=0.6.0,<0.8.0 ; python_version >= '3.9' and (((sys_platform == 'linux' or sys_platform == 'win32') and (platform_machine == 'x86_64' or platform_machine == 'AMD64')) or (sys_platform == 'darwin' and platform_machine == 'arm64'))",
 ]
 huggingface = [
     "unsloth[huggingfacenotorch]",
diff --git a/scripts/profile_startup.py b/scripts/profile_startup.py
new file mode 100644
index 0000000000..937d007ac1
--- /dev/null
+++ b/scripts/profile_startup.py
@@ -0,0 +1,377 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Measure where Unsloth Studio's startup time goes, per platform.
+
+Nothing measured this before: the backend logs "lifespan startup completed in X ms"
+but no test or CI job asserted a budget, and studio_test_kit discards the elapsed
+time of its /healthz poll. A first local run (Linux, warm cache, fast server CPU)
+found `import main` alone costs 6.6s before the server can bind, dominated by eager
+module-level imports pulled in by the `routes` package:
+
+    torch          1930 ms self
+    unsloth_zoo     914 ms self
+    routes          779 ms self
+    transformers    524 ms self
+
+Phases measured:
+  import   `python -X importtime -c "import main"`, top cumulative + per-package self
+  spawn    process start -> first byte on stdout
+  healthz  process start -> /api/health (or /healthz) answers 200
+  lifespan the backend's own "lifespan startup completed in X ms" log line
+
+Usage:
+    python scripts/profile_startup.py --repeats 3 --json out.json
+    python scripts/profile_startup.py --import-only     # no server, no port needed
+
+Exit code is 0 unless --max-healthz-seconds is given and exceeded.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import os
+import platform
+import re
+import shutil
+import socket
+import statistics
+import subprocess
+import sys
+import threading
+import time
+import urllib.error
+import urllib.request
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+BACKEND = REPO_ROOT / "studio" / "backend"
+
+_IMPORTTIME_RE = re.compile(r"import time:\s+(\d+)\s+\|\s+(\d+)\s+\|(\s*)(\S.*)")
+
+
+def _free_port() -> int:
+    with socket.socket() as s:
+        s.bind(("127.0.0.1", 0))
+        return int(s.getsockname()[1])
+
+
+def profile_imports(python: str, top: int = 15) -> dict:
+    """Cumulative and self import cost for the backend's module graph.
+
+    Run in a subprocess with -X importtime: the numbers are only meaningful for a
+    cold interpreter, and importing in-process would measure a warm sys.modules.
+    """
+    proc = subprocess.run(
+        [python, "-X", "importtime", "-c", "import sys; sys.path.insert(0, '.'); import main"],
+        cwd = BACKEND,
+        capture_output = True,
+        text = True,
+        timeout = 900,
+    )
+    rows = []
+    for line in proc.stderr.splitlines():
+        m = _IMPORTTIME_RE.match(line)
+        if m:
+            rows.append((int(m.group(1)), int(m.group(2)), m.group(4).strip()))
+    if not rows:
+        return {"ok": False, "error": (proc.stderr or proc.stdout)[-2000:]}
+    if proc.returncode != 0:
+        # Rows survive up to the failure, so any total from a partial graph is wrong.
+        return {
+            "ok": False,
+            "error": (proc.stderr or proc.stdout)[-2000:],
+            "partial_rows": len(rows),
+        }
+
+    by_cum = sorted(rows, key = lambda r: -r[1])
+    # Total comes from the `main` row, not by_cum[0]: -X importtime also prints the
+    # interpreter's own startup graph (`site`), which can outrank a trivial main.
+    main_row = next((r for r in reversed(rows) if r[2] == "main"), None)
+    if main_row is None:
+        return {
+            "ok": False,
+            "error": "no `import main` row in -X importtime output\n"
+            + (proc.stderr or proc.stdout)[-2000:],
+        }
+    self_by_pkg: dict[str, int] = {}
+    for self_us, _cum, name in rows:
+        pkg = name.split(".")[0]
+        self_by_pkg[pkg] = self_by_pkg.get(pkg, 0) + self_us
+
+    return {
+        "ok": True,
+        "total_seconds": round(main_row[1] / 1e6, 3),
+        "top_cumulative": [
+            {"module": n, "seconds": round(c / 1e6, 3)} for _s, c, n in by_cum[:top]
+        ],
+        "self_by_package_ms": {
+            k: round(v / 1000) for k, v in sorted(self_by_pkg.items(), key = lambda x: -x[1])[:top]
+        },
+    }
+
+
+def _terminate_tree(proc: subprocess.Popen) -> None:
+    """Stop the server AND its children, which on Windows are a separate process.
+
+    CI profiles `Scripts/unsloth.exe`, a distlib launcher stub that CreateProcess's
+    the venv python and waits, so terminate() reaps the stub only: the real backend
+    keeps the inherited stdout handle, the reader thread never sees EOF, and
+    --repeats strands one server per iteration on the shared UNSLOTH_STUDIO_HOME.
+    taskkill /T walks the tree, as unsloth_cli/commands/start.py already does.
+    """
+    if proc.poll() is not None:
+        return
+    if os.name == "nt":
+        try:
+            killed = subprocess.run(
+                ["taskkill", "/PID", str(proc.pid), "/T", "/F"],
+                capture_output = True,
+                timeout = 30,
+                check = False,
+            )
+            if killed.returncode == 0:
+                return
+        except Exception:
+            # taskkill missing or timed out; fall through so the stub still dies.
+            pass
+        # check=False: a nonzero taskkill does not raise, so fall through as well.
+    proc.terminate()
+
+
+def profile_launch(
+    bin_path: str,
+    port: int,
+    timeout_s: int = 300,
+) -> dict:
+    """Spawn the backend the way the desktop app does and time it to first 200."""
+    log_lines: list[str] = []
+    first_byte: list[float] = []
+    t0 = time.perf_counter()
+    proc = subprocess.Popen(
+        [bin_path, "studio", "--api-only", "-H", "127.0.0.1", "-p", str(port)],
+        cwd = REPO_ROOT,
+        stdout = subprocess.PIPE,
+        stderr = subprocess.STDOUT,
+        text = True,
+        bufsize = 1,
+    )
+
+    def _drain() -> None:
+        # Runs alongside the health polling: the first read timestamps the spawn
+        # phase, and an undrained pipe blocks the backend before it binds.
+        for line in proc.stdout:
+            if not first_byte:
+                first_byte.append(time.perf_counter() - t0)
+            log_lines.append(line.rstrip("\n"))
+
+    reader = threading.Thread(target = _drain, daemon = True)
+    reader.start()
+
+    t_healthz = None
+    deadline = t0 + timeout_s
+    try:
+        while time.perf_counter() < deadline:
+            if proc.poll() is not None:
+                break
+            if t_healthz is None:
+                for url in (
+                    f"http://127.0.0.1:{port}/api/health",
+                    f"http://127.0.0.1:{port}/healthz",
+                ):
+                    try:
+                        with urllib.request.urlopen(url, timeout = 2) as r:
+                            if r.status == 200:
+                                t_healthz = time.perf_counter() - t0
+                                break
+                    except (urllib.error.URLError, OSError, TimeoutError):
+                        pass
+            if t_healthz is not None:
+                break
+            time.sleep(0.25)
+    finally:
+        _terminate_tree(proc)
+        try:
+            # Safe: the reader drains the pipe, so the child cannot block on write().
+            proc.wait(timeout = 30)
+        except subprocess.TimeoutExpired:
+            proc.kill()
+            proc.wait()
+        reader.join(timeout = 10)
+
+    t_first_byte = first_byte[0] if first_byte else None
+    lifespan_ms = None
+    for line in log_lines:
+        m = re.search(r"lifespan startup completed in ([\d.]+)ms", line)
+        if m:
+            lifespan_ms = float(m.group(1))
+    return {
+        "spawn_seconds": round(t_first_byte, 3) if t_first_byte is not None else None,
+        "healthz_seconds": round(t_healthz, 3) if t_healthz is not None else None,
+        "lifespan_ms": lifespan_ms,
+        "reached_healthz": t_healthz is not None,
+        "log_tail": log_lines[-25:],
+    }
+
+
+def python_version_of(python: str) -> str:
+    """Version of the interpreter that runs the imports, not the one running us.
+
+    --python points at the installed Studio venv while this script runs under the
+    runner's system python, so platform.python_version() would label it wrong.
+    """
+    if python == sys.executable:
+        return platform.python_version()
+    try:
+        proc = subprocess.run(
+            [python, "-c", "import platform; print(platform.python_version())"],
+            capture_output = True,
+            text = True,
+            timeout = 60,
+        )
+        if proc.returncode == 0 and proc.stdout.strip():
+            return proc.stdout.strip()
+    except (OSError, subprocess.SubprocessError):
+        pass
+    return "unknown"
+
+
+def find_bin() -> str | None:
+    home = os.environ.get("UNSLOTH_STUDIO_HOME") or str(Path.home() / ".unsloth" / "studio")
+    names = ["unsloth.exe", "unsloth"] if platform.system() == "Windows" else ["unsloth"]
+    subdirs = ["unsloth_studio/Scripts", "unsloth_studio/bin", "bin", "Scripts"]
+    for sd in subdirs:
+        for n in names:
+            p = Path(home) / sd / n
+            if p.exists():
+                return str(p)
+    return shutil.which("unsloth")
+
+
+def main(argv: list[str]) -> int:
+    ap = argparse.ArgumentParser(
+        description = __doc__, formatter_class = argparse.RawDescriptionHelpFormatter
+    )
+    ap.add_argument(
+        "--repeats",
+        type = int,
+        default = 1,
+        help = "launch repeats; the median is reported (imports are measured once)",
+    )
+    ap.add_argument(
+        "--python",
+        default = sys.executable,
+        help = "interpreter used for the import profile (default: this one)",
+    )
+    ap.add_argument("--bin", help = "path to the unsloth CLI (default: autodetect)")
+    ap.add_argument(
+        "--import-only",
+        action = "store_true",
+        help = "skip the server phases (no install needed beyond the deps)",
+    )
+    ap.add_argument(
+        "--max-healthz-seconds",
+        type = float,
+        help = "fail if the median time to a healthy port exceeds this",
+    )
+    ap.add_argument("--json", help = "write the full report here")
+    a = ap.parse_args(argv)
+    # range(0) launches nothing, leaving the budget check with nothing to fail on.
+    if a.repeats < 1:
+        ap.error("--repeats must be at least 1")
+    # Same reason: --import-only never launches anything.
+    if a.import_only and a.max_healthz_seconds is not None:
+        ap.error("--max-healthz-seconds cannot be combined with --import-only")
+    # nan and inf parse fine as floats but `med > budget` is then always False,
+    # so the gate would report success without ever bounding anything.
+    if a.max_healthz_seconds is not None and not math.isfinite(a.max_healthz_seconds):
+        ap.error("--max-healthz-seconds must be a finite number")
+
+    report: dict = {
+        "platform": platform.system().lower(),
+        "machine": platform.machine(),
+        "python": python_version_of(a.python),
+        "cpu_count": os.cpu_count(),
+    }
+
+    print("== import graph ==")
+    report["imports"] = profile_imports(a.python)
+    imp = report["imports"]
+    if imp.get("ok"):
+        print(f"  import main: {imp['total_seconds']}s")
+        for row in imp["top_cumulative"][:8]:
+            print(f"    {row['seconds']:7.3f}s  {row['module']}")
+        print("  self time by package (ms):")
+        for k, v in list(imp["self_by_package_ms"].items())[:8]:
+            print(f"    {v:8} ms  {k}")
+    else:
+        print(f"  FAILED: {imp.get('error', '')[:400]}")
+
+    if not a.import_only:
+        bin_path = a.bin or find_bin()
+        if not bin_path:
+            print(
+                "== launch == skipped: no unsloth CLI found "
+                "(set UNSLOTH_STUDIO_HOME or pass --bin)"
+            )
+            report["launch"] = {"skipped": "no unsloth CLI found"}
+        else:
+            print(f"== launch == {bin_path}")
+            runs = []
+            for i in range(a.repeats):
+                r = profile_launch(bin_path, _free_port())
+                runs.append(r)
+                print(
+                    f"  run {i + 1}: healthz={r['healthz_seconds']}s "
+                    f"lifespan={r['lifespan_ms']}ms reached={r['reached_healthz']}"
+                )
+            got = [r["healthz_seconds"] for r in runs if r["healthz_seconds"] is not None]
+            report["launch"] = {
+                "runs": runs,
+                "failed_runs": sum(1 for r in runs if not r["reached_healthz"]),
+                "healthz_median_seconds": round(statistics.median(got), 3) if got else None,
+                "healthz_max_seconds": round(max(got), 3) if got else None,
+            }
+            if got:
+                print(
+                    f"  median time to healthy port: {report['launch']['healthz_median_seconds']}s"
+                )
+
+    if a.json:
+        Path(a.json).write_text(json.dumps(report, indent = 2), encoding = "utf-8")
+        print(f"\nwrote {a.json}")
+
+    if a.max_healthz_seconds is not None:
+        launch = report.get("launch") or {}
+        med = launch.get("healthz_median_seconds")
+        failed = launch.get("failed_runs") or 0
+        if failed:
+            # Failed launches fail the budget; dropping them would keep only the fast ones.
+            print(
+                f"::error::startup regression: {failed} of {len(launch.get('runs') or [])} "
+                f"launches never became healthy within the timeout"
+            )
+            return 1
+        if med is None:
+            # Nothing measured: exiting 0 would pass a requested budget without a
+            # single health request, so fail closed.
+            print(
+                "::error::startup regression: no healthz measurement, so the "
+                f"{a.max_healthz_seconds}s budget was never checked "
+                f"({launch.get('skipped') or 'launch phase produced no runs'})"
+            )
+            return 1
+        elif med > a.max_healthz_seconds:
+            print(
+                f"::error::startup regression: {med}s median to a healthy port "
+                f"exceeds the {a.max_healthz_seconds}s budget"
+            )
+            return 1
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main(sys.argv[1:]))
diff --git a/studio/backend/auth/authentication.py b/studio/backend/auth/authentication.py
index 94df994928..2e9520827e 100644
--- a/studio/backend/auth/authentication.py
+++ b/studio/backend/auth/authentication.py
@@ -11,11 +11,12 @@ import jwt
 
 from .storage import (
     API_KEY_PREFIX,
+    credential_generation,
     get_jwt_secret,
     get_user_and_secret,
     load_jwt_secret,
     save_refresh_token,
-    validate_api_key,
+    validate_api_key_with_credential,
     verify_refresh_token,
 )
 
@@ -54,11 +55,14 @@ def create_access_token(
     expires_delta: Optional[timedelta] = None,
     *,
     desktop: bool = False,
+    secret: Optional[str] = None,
 ) -> str:
     """
     Create a signed JWT for the given subject (e.g. username).
 
-    Valid across restarts: the signing secret is stored in SQLite.
+    Valid across restarts: the signing secret is stored in SQLite. Callers that
+    already verified a credential pass ``secret`` so a rotation landing mid-request
+    cannot sign the token with the credential that just replaced it.
     """
     to_encode = {"sub": subject}
     if desktop:
@@ -69,7 +73,7 @@ def create_access_token(
     to_encode.update({"exp": expire})
     return jwt.encode(
         to_encode,
-        _get_secret_for_subject(subject),
+        secret if secret is not None else _get_secret_for_subject(subject),
         algorithm = ALGORITHM,
     )
 
@@ -96,15 +100,28 @@ def is_desktop_access_token(token: str) -> bool:
     return payload.get("sub") == subject and payload.get("desktop") is True
 
 
-def create_refresh_token(subject: str, *, desktop: bool = False) -> str:
+def create_refresh_token(
+    subject: str,
+    *,
+    desktop: bool = False,
+    secret: Optional[str] = None,
+) -> str:
     """
     Create a random refresh token, store its hash in SQLite, and return it.
 
     Refresh tokens are opaque (not JWTs); expire after REFRESH_TOKEN_EXPIRE_DAYS.
+    ``secret`` stamps the token with the credential version the caller verified,
+    so a rotation cannot leave a token minted from the replaced credential valid.
     """
     token = secrets.token_urlsafe(48)
     expires_at = datetime.now(timezone.utc) + timedelta(days = REFRESH_TOKEN_EXPIRE_DAYS)
-    save_refresh_token(token, subject, expires_at.isoformat(), is_desktop = desktop)
+    save_refresh_token(
+        token,
+        subject,
+        expires_at.isoformat(),
+        is_desktop = desktop,
+        secret_gen = credential_generation(secret) if secret is not None else None,
+    )
     return token
 
 
@@ -137,7 +154,22 @@ def reload_secret() -> None:
 
 async def get_current_subject(credentials: HTTPAuthorizationCredentials = Depends(security)) -> str:
     """Validate JWT and require the password-change flow to be completed."""
-    return await _get_current_subject(
+    subject, _generation = await _get_current_credential(
+        credentials,
+        allow_password_change = False,
+    )
+    return subject
+
+
+async def get_current_credential(
+    credentials: HTTPAuthorizationCredentials = Depends(security),
+) -> Tuple[str, Optional[str]]:
+    """As get_current_subject, but also returns the credential generation.
+
+    For routes that persist a new credential and must not do so on behalf of one
+    a concurrent reset has revoked.
+    """
+    return await _get_current_credential(
         credentials,
         allow_password_change = False,
     )
@@ -158,10 +190,11 @@ async def get_current_subject_allow_password_change(
     credentials: HTTPAuthorizationCredentials = Depends(security),
 ) -> str:
     """Validate JWT but allow access to the password-change endpoint."""
-    return await _get_current_subject(
+    subject, _generation = await _get_current_credential(
         credentials,
         allow_password_change = True,
     )
+    return subject
 
 
 # The literal the examples ship with; pasted unedited more often than a revoked key.
@@ -179,21 +212,27 @@ def _invalid_api_key_detail(token: str) -> str:
     return "Invalid or expired API key"
 
 
-async def _get_current_subject(
+async def _get_current_credential(
     credentials: HTTPAuthorizationCredentials, *, allow_password_change: bool
-) -> str:
-    """FastAPI dependency: validate the JWT and return the subject. Use on protected routes."""
+) -> Tuple[str, Optional[str]]:
+    """Validate the bearer and return ``(subject, credential generation)``.
+
+    The generation is the credential version this request actually authenticated
+    against. Routes that persist new credentials must bind their write to it, or
+    a reset landing mid-request would bless what it just revoked.
+    """
     token = credentials.credentials
 
     # --- API key path (sk-unsloth-...) ---
     if token.startswith(API_KEY_PREFIX):
-        username = validate_api_key(token)
-        if username is None:
+        verified = validate_api_key_with_credential(token)
+        if verified is None:
             raise HTTPException(
                 status_code = status.HTTP_401_UNAUTHORIZED,
                 detail = _invalid_api_key_detail(token),
             )
-        return username
+        username, secret = verified
+        return username, credential_generation(secret)
 
     # --- JWT path ---
     subject = _decode_subject_without_verification(token)
@@ -224,7 +263,7 @@ async def _get_current_subject(
                 status_code = status.HTTP_403_FORBIDDEN,
                 detail = "Password change required",
             )
-        return subject
+        return subject, credential_generation(jwt_secret)
     except jwt.InvalidTokenError:
         raise HTTPException(
             status_code = status.HTTP_401_UNAUTHORIZED,
diff --git a/studio/backend/auth/storage.py b/studio/backend/auth/storage.py
index 5f80ad89a3..6cf4d44834 100644
--- a/studio/backend/auth/storage.py
+++ b/studio/backend/auth/storage.py
@@ -9,6 +9,7 @@ import ipaddress
 import os
 import secrets
 import sqlite3
+import tempfile
 import threading
 from datetime import datetime, timezone
 from typing import Optional, Tuple
@@ -30,6 +31,97 @@ _BOOTSTRAP_PW_PATH = DB_PATH.parent / ".bootstrap_password"
 _bootstrap_password: Optional[str] = None
 
 
+def _bootstrap_file_bytes(password: str) -> bytes:
+    """Exact on-disk form: the secret plus one LF.
+
+    Bytes, not text: text mode writes CRLF on Windows, and `$(cat ...)` strips
+    the LF but leaves the CR attached to the credential.
+    """
+    return (password + "\n").encode("utf-8")
+
+
+def _persist_bootstrap_password(password: str) -> None:
+    """Atomically write the bootstrap password 0600, LF terminated on every OS.
+
+    A partial write would destroy the only plaintext recovery credential.
+    """
+    fd, tmp_name = tempfile.mkstemp(
+        prefix = f".{_BOOTSTRAP_PW_PATH.name}.", dir = _BOOTSTRAP_PW_PATH.parent
+    )
+    try:
+        with os.fdopen(fd, "wb") as f:
+            f.write(_bootstrap_file_bytes(password))
+        try:
+            os.chmod(tmp_name, 0o600)
+        except OSError:
+            pass
+        os.replace(tmp_name, _BOOTSTRAP_PW_PATH)
+    except BaseException:
+        try:
+            os.unlink(tmp_name)
+        except OSError:
+            pass
+        raise
+
+
+def _normalise_bootstrap_file(raw: bytes, password: str) -> None:
+    """Append the LF a pre-newline release left off.
+
+    Append-only, and only when the file is exactly the credential:
+    clear_bootstrap_password() may unlink or (when unlink fails, notably on
+    Windows while this descriptor is open) truncate through another descriptor
+    after we read, so a rewrite could restore revoked plaintext. An append
+    cannot: worst case is a lone "\\n" over a cleared file, which strips back to
+    no bootstrap password. Pre-newline releases wrote no terminator at all, so
+    that is the only shape in the wild; anything else reads fine, since every
+    reader strips, and is left alone.
+    """
+    if raw != password.encode("utf-8"):
+        return
+
+    # O_BINARY: without it Windows opens in text mode and turns the LF straight
+    # back into CRLF, the bug being fixed.
+    fd = os.open(
+        _BOOTSTRAP_PW_PATH,
+        os.O_WRONLY | os.O_APPEND | getattr(os, "O_BINARY", 0),
+    )
+    try:
+        os.write(fd, b"\n")
+        try:
+            os.fchmod(fd, 0o600)
+        except (AttributeError, OSError):
+            # fchmod only reached Windows in 3.13.
+            pass
+    finally:
+        os.close(fd)
+
+
+def _read_persisted_bootstrap_password() -> Optional[str]:
+    """Read the persisted password, normalising the file if it is malformed."""
+    if not _BOOTSTRAP_PW_PATH.is_file():
+        return None
+
+    # No caller handles a raise, so an unreadable file has to mean "no bootstrap
+    # password", not a dead backend. We write UTF-8, so undecodable bytes are
+    # damage whose plaintext is worthless anyway.
+    try:
+        raw = _BOOTSTRAP_PW_PATH.read_bytes()
+        password = raw.decode("utf-8").strip()
+    except (OSError, UnicodeDecodeError):
+        return None
+    if not password:
+        return None
+
+    # Older releases wrote no terminator; best-effort, a read-only auth dir must
+    # not fail startup.
+    if raw != _bootstrap_file_bytes(password):
+        try:
+            _normalise_bootstrap_file(raw, password)
+        except OSError:
+            pass
+    return password
+
+
 def generate_bootstrap_password() -> str:
     """Generate a 4-word diceware passphrase and persist it to disk.
 
@@ -43,10 +135,10 @@ def generate_bootstrap_password() -> str:
         return _bootstrap_password
 
     # Persisted from a previous run?
-    if _BOOTSTRAP_PW_PATH.is_file():
-        _bootstrap_password = _BOOTSTRAP_PW_PATH.read_text(encoding = "utf-8").strip()
-        if _bootstrap_password:
-            return _bootstrap_password
+    persisted = _read_persisted_bootstrap_password()
+    if persisted:
+        _bootstrap_password = persisted
+        return _bootstrap_password
 
     # First startup: generate a fresh passphrase.
     import diceware
@@ -57,11 +149,7 @@ def generate_bootstrap_password() -> str:
 
     # Persist so the same passphrase survives restarts until password change.
     ensure_dir(_BOOTSTRAP_PW_PATH.parent)
-    _BOOTSTRAP_PW_PATH.write_text(_bootstrap_password, encoding = "utf-8")
-    try:
-        os.chmod(_BOOTSTRAP_PW_PATH, 0o600)
-    except OSError:
-        pass
+    _persist_bootstrap_password(_bootstrap_password)
 
     return _bootstrap_password
 
@@ -72,13 +160,14 @@ def get_bootstrap_password() -> Optional[str]:
 
 
 def _load_bootstrap_password() -> Optional[str]:
-    """Load an existing bootstrap password without creating one."""
+    """Load an existing bootstrap password without creating one.
+
+    Upgrades take this path, not generate_bootstrap_password()
+    (ensure_default_admin short-circuits once the admin row exists), so it has
+    to normalise too.
+    """
     global _bootstrap_password
-    _bootstrap_password = None
-    if _BOOTSTRAP_PW_PATH.is_file():
-        bootstrap_password = _BOOTSTRAP_PW_PATH.read_text(encoding = "utf-8").strip()
-        if bootstrap_password:
-            _bootstrap_password = bootstrap_password
+    _bootstrap_password = _read_persisted_bootstrap_password()
     return _bootstrap_password
 
 
@@ -97,7 +186,7 @@ def clear_bootstrap_password() -> None:
             # Removal failed (Windows AV, read-only auth dir). The hash is already
             # committed, so don't fail the change -- but truncate the file so its
             # stale plaintext can't be re-seeded by generate_bootstrap_password()
-            # if a later reset-password deletes auth.db and re-validates it.
+            # if auth.db is ever recreated.
             try:
                 _BOOTSTRAP_PW_PATH.write_text("", encoding = "utf-8")
                 cleared = True
@@ -132,6 +221,31 @@ def _hash_token(token: str) -> str:
     return hashlib.sha256(token.encode("utf-8")).hexdigest()
 
 
+class CredentialRotated(Exception):
+    """A password reset revoked the credential this request authenticated with."""
+
+
+def credential_generation(jwt_secret: str) -> str:
+    """Marker for the credential version a refresh token was issued under.
+
+    Every password change rotates ``jwt_secret``, so a token stamped with the
+    previous one is rejected even if it was inserted after the revoking DELETE.
+    """
+    return hashlib.sha256(jwt_secret.encode("utf-8")).hexdigest()
+
+
+def _current_secret(conn: sqlite3.Connection, username: str) -> Optional[str]:
+    row = conn.execute(
+        "SELECT jwt_secret FROM auth_user WHERE username = ?", (username,)
+    ).fetchone()
+    return row["jwt_secret"] if row else None
+
+
+def _current_generation(conn: sqlite3.Connection, username: str) -> Optional[str]:
+    secret = _current_secret(conn, username)
+    return credential_generation(secret) if secret is not None else None
+
+
 def get_connection() -> sqlite3.Connection:
     """Get a connection to the auth database, creating tables if needed."""
     ensure_dir(DB_PATH.parent)
@@ -175,7 +289,8 @@ def get_connection() -> sqlite3.Connection:
             token_hash TEXT NOT NULL,
             username TEXT NOT NULL,
             expires_at TEXT NOT NULL,
-            is_desktop INTEGER NOT NULL DEFAULT 0
+            is_desktop INTEGER NOT NULL DEFAULT 0,
+            secret_gen TEXT
         );
         """
     )
@@ -214,6 +329,8 @@ def get_connection() -> sqlite3.Connection:
     refresh_columns = {row["name"] for row in conn.execute("PRAGMA table_info(refresh_tokens)")}
     if "is_desktop" not in refresh_columns:
         conn.execute("ALTER TABLE refresh_tokens ADD COLUMN is_desktop INTEGER NOT NULL DEFAULT 0")
+    if "secret_gen" not in refresh_columns:
+        conn.execute("ALTER TABLE refresh_tokens ADD COLUMN secret_gen TEXT")
     conn.commit()
     return conn
 
@@ -587,12 +704,22 @@ def update_password(
     new_password: str,
     *,
     revoke_refresh_tokens: bool = False,
-) -> bool:
+    expect_password_hash: Optional[str] = None,
+) -> Optional[str]:
     """Update password, clear first-login requirement, rotate JWT secret.
 
+    Returns the new JWT secret, or None when nothing was updated. Callers that
+    mint tokens for the caller must sign with the returned secret: re-reading it
+    would pick up a reset that landed between this commit and the mint.
+
     ``revoke_refresh_tokens`` deletes the user's refresh tokens in the SAME
     transaction: a separate delete could fail after the password commit and
     leave a pre-change token still able to mint access tokens.
+
+    ``expect_password_hash`` makes the write conditional on the credential the
+    caller verified still being current, so a request that checked the old
+    password cannot overwrite a reset that landed while it was in flight.
+    Returns False when the credential moved underneath it.
     """
     from .hashing import hash_password
 
@@ -600,21 +727,32 @@ def update_password(
     jwt_secret = secrets.token_urlsafe(64)
     conn = get_connection()
     try:
-        cursor = conn.execute(
-            """
-            UPDATE auth_user
-            SET password_salt = ?, password_hash = ?, jwt_secret = ?, must_change_password = 0
-            WHERE username = ?
-            """,
-            (salt, pwd_hash, jwt_secret, username),
-        )
+        if expect_password_hash is None:
+            cursor = conn.execute(
+                """
+                UPDATE auth_user
+                SET password_salt = ?, password_hash = ?, jwt_secret = ?, must_change_password = 0
+                WHERE username = ?
+                """,
+                (salt, pwd_hash, jwt_secret, username),
+            )
+        else:
+            cursor = conn.execute(
+                """
+                UPDATE auth_user
+                SET password_salt = ?, password_hash = ?, jwt_secret = ?, must_change_password = 0
+                WHERE username = ? AND password_hash = ?
+                """,
+                (salt, pwd_hash, jwt_secret, username, expect_password_hash),
+            )
         if revoke_refresh_tokens and cursor.rowcount > 0:
             conn.execute("DELETE FROM refresh_tokens WHERE username = ?", (username,))
         conn.commit()
         if cursor.rowcount > 0:
             clear_bootstrap_password()
             clear_desktop_secret()
-        return cursor.rowcount > 0
+            return jwt_secret
+        return None
     finally:
         conn.close()
 
@@ -625,35 +763,49 @@ def save_refresh_token(
     expires_at: str,
     *,
     is_desktop: bool = False,
+    secret_gen: Optional[str] = None,
 ) -> None:
     """
     Store a hashed refresh token with its associated username and expiry.
+
+    ``secret_gen`` binds the token to a credential version; it defaults to the
+    current one, and callers that already verified a credential must pass the
+    version they verified rather than let this re-read a rotated one.
     """
     token_hash = _hash_token(token)
     conn = get_connection()
     try:
+        if secret_gen is None:
+            secret_gen = _current_generation(conn, username)
         conn.execute(
             """
-            INSERT INTO refresh_tokens (token_hash, username, expires_at, is_desktop)
-            VALUES (?, ?, ?, ?)
+            INSERT INTO refresh_tokens (token_hash, username, expires_at, is_desktop, secret_gen)
+            VALUES (?, ?, ?, ?, ?)
             """,
-            (token_hash, username, expires_at, int(is_desktop)),
+            (token_hash, username, expires_at, int(is_desktop), secret_gen),
         )
         conn.commit()
     finally:
         conn.close()
 
 
-def consume_refresh_token(token: str) -> Optional[Tuple[str, bool]]:
+def consume_refresh_token(token: str) -> Optional[Tuple[str, bool, str]]:
     """Atomically validate-and-delete a refresh token for single-use rotation.
 
     DELETE RETURNING fuses validate and delete into one statement so two
-    concurrent refresh requests cannot both consume the same token.
+    concurrent refresh requests cannot both consume the same token. Returns
+    ``(username, is_desktop, jwt_secret)``; the caller must mint the replacement
+    tokens against that secret so a rotation landing mid-refresh cannot issue a
+    post-rotation session from a pre-rotation token.
     """
     token_hash = _hash_token(token)
     now = datetime.now(timezone.utc).isoformat()
     conn = get_connection()
     try:
+        # One transaction with the delete: an unstamped legacy row has no
+        # generation to compare, so reading the credential after committing would
+        # hand a reset's new secret to a token issued before it.
+        conn.execute("BEGIN IMMEDIATE")
         conn.execute(
             "DELETE FROM refresh_tokens WHERE expires_at < ?",
             (now,),
@@ -662,15 +814,21 @@ def consume_refresh_token(token: str) -> Optional[Tuple[str, bool]]:
             """
             DELETE FROM refresh_tokens
             WHERE token_hash = ? AND expires_at >= ?
-            RETURNING username, is_desktop
+            RETURNING username, is_desktop, secret_gen
             """,
             (token_hash, now),
         )
         row = cur.fetchone()
-        conn.commit()
         if row is None:
+            conn.commit()
             return None
-        return row["username"], bool(row["is_desktop"])
+        secret = _current_secret(conn, row["username"])
+        conn.commit()
+        if secret is None:
+            return None
+        if row["secret_gen"] is not None and row["secret_gen"] != credential_generation(secret):
+            return None
+        return row["username"], bool(row["is_desktop"]), secret
     finally:
         conn.close()
 
@@ -694,7 +852,7 @@ def verify_refresh_token(token: str) -> Optional[Tuple[str, bool]]:
 
         cur = conn.execute(
             """
-            SELECT id, username, expires_at, is_desktop FROM refresh_tokens
+            SELECT id, username, expires_at, is_desktop, secret_gen FROM refresh_tokens
             WHERE token_hash = ?
             """,
             (token_hash,),
@@ -703,6 +861,13 @@ def verify_refresh_token(token: str) -> Optional[Tuple[str, bool]]:
         if row is None:
             return None
 
+        if row["secret_gen"] is not None and row["secret_gen"] != _current_generation(
+            conn, row["username"]
+        ):
+            conn.execute("DELETE FROM refresh_tokens WHERE id = ?", (row["id"],))
+            conn.commit()
+            return None
+
         # Check expiry
         expires_at = datetime.fromisoformat(row["expires_at"])
         if datetime.now(timezone.utc) > expires_at:
@@ -747,30 +912,41 @@ def create_desktop_secret() -> str:
         conn.close()
 
 
-def validate_desktop_secret(raw_secret: str) -> Optional[str]:
-    """Return the real admin username when the desktop secret matches."""
+def validate_desktop_secret_with_credential(raw_secret: str) -> Optional[Tuple[str, str]]:
+    """Validate the desktop secret and return ``(username, jwt_secret)``.
+
+    Both reads share one transaction so the returned secret is the credential
+    version the desktop secret was checked against; a reset landing mid-request
+    then invalidates the tokens minted from it rather than blessing them.
+    """
     if not raw_secret.startswith(DESKTOP_SECRET_PREFIX):
         return None
-    if get_user_and_secret(DEFAULT_ADMIN_USERNAME) is None:
-        return None
 
     secret_hash = _pbkdf2_desktop_secret(raw_secret)
     conn = get_connection()
     try:
-        cur = conn.execute(
+        conn.execute("BEGIN")
+        row = conn.execute(
             "SELECT value FROM app_secrets WHERE key = ?",
             (_DESKTOP_SECRET_HASH_KEY,),
-        )
-        row = cur.fetchone()
-        if row is None:
+        ).fetchone()
+        if row is None or not secrets.compare_digest(row["value"], secret_hash):
             return None
-        if not secrets.compare_digest(row["value"], secret_hash):
+        jwt_secret = _current_secret(conn, DEFAULT_ADMIN_USERNAME)
+        if jwt_secret is None:
             return None
-        return DEFAULT_ADMIN_USERNAME
+        return DEFAULT_ADMIN_USERNAME, jwt_secret
     finally:
+        conn.rollback()
         conn.close()
 
 
+def validate_desktop_secret(raw_secret: str) -> Optional[str]:
+    """Return the real admin username when the desktop secret matches."""
+    verified = validate_desktop_secret_with_credential(raw_secret)
+    return verified[0] if verified else None
+
+
 def clear_desktop_secret() -> None:
     """Remove backend-side desktop auth state."""
     conn = get_connection()
@@ -796,6 +972,7 @@ def create_api_key(
     name: str,
     expires_at: Optional[str] = None,
     internal: bool = False,
+    expect_gen: Optional[str] = None,
 ) -> Tuple[str, dict]:
     """Create a new API key for *username*.
 
@@ -804,6 +981,10 @@ def create_api_key(
 
     Pass ``internal=True`` for keys minted by workflows (e.g. data-recipe
     runs) that should not appear in user-facing key listings.
+
+    ``expect_gen`` ties the insert to the credential generation the request
+    authenticated under, so a session revoked by a concurrent password reset
+    cannot mint a key that outlives it. Raises ``CredentialRotated`` if it moved.
     """
     raw_key = API_KEY_PREFIX + secrets.token_hex(16)
     key_hash = _pbkdf2_api_key(raw_key)
@@ -812,6 +993,12 @@ def create_api_key(
 
     conn = get_connection()
     try:
+        if expect_gen is not None:
+            conn.execute("BEGIN IMMEDIATE")
+            if _current_generation(conn, username) != expect_gen:
+                raise CredentialRotated(
+                    "The credential this request authenticated with was revoked."
+                )
         conn.execute(
             """
             INSERT INTO api_keys (username, key_prefix, key_hash, name, created_at, expires_at, is_internal)
@@ -900,15 +1087,25 @@ def revoke_internal_api_key(key_id: int) -> bool:
 
 
 def validate_api_key(raw_key: str) -> Optional[str]:
-    """Validate *raw_key* and return the owning username, or ``None``.
+    """Validate *raw_key* and return the owning username, or ``None``."""
+    verified = validate_api_key_with_credential(raw_key)
+    return verified[0] if verified else None
 
-    Also updates ``last_used_at`` on success.
+
+def validate_api_key_with_credential(raw_key: str) -> Optional[Tuple[str, str]]:
+    """Validate *raw_key* and return ``(username, jwt_secret)``, or ``None``.
+
+    Also updates ``last_used_at`` on success. The key check and the credential
+    read share one write transaction, so the returned version is the one the key
+    was actually valid under: a reset committing right after cannot have its new
+    generation handed to a request the key it revoked authenticated.
     """
     cache_id = _api_key_cache_id(raw_key)
     cached_hash = _api_key_hash_cache.get(cache_id)
     key_hash = cached_hash if cached_hash is not None else _pbkdf2_api_key(raw_key)
     conn = get_connection()
     try:
+        conn.execute("BEGIN IMMEDIATE")
         cur = conn.execute(
             "SELECT id, username, is_active, expires_at FROM api_keys WHERE key_hash = ?",
             (key_hash,),
@@ -928,11 +1125,15 @@ def validate_api_key(raw_key: str) -> Optional[str]:
             expires = datetime.fromisoformat(row["expires_at"])
             if datetime.now(timezone.utc) > expires:
                 return None
+        secret = _current_secret(conn, row["username"])
+        if secret is None:
+            return None
         conn.execute(
             "UPDATE api_keys SET last_used_at = ? WHERE id = ?",
             (datetime.now(timezone.utc).isoformat(), row["id"]),
         )
         conn.commit()
-        return row["username"]
+        return row["username"], secret
     finally:
+        conn.rollback()
         conn.close()
diff --git a/studio/backend/cloudflare_tunnel.py b/studio/backend/cloudflare_tunnel.py
index 78fce0c70a..f7967e2faa 100644
--- a/studio/backend/cloudflare_tunnel.py
+++ b/studio/backend/cloudflare_tunnel.py
@@ -310,6 +310,7 @@ class CloudflareTunnel:
                 stderr = subprocess.STDOUT,
                 stdin = subprocess.DEVNULL,
                 text = True,
+                encoding = "utf-8",
                 errors = "replace",
                 bufsize = 1,
                 **_windows_hidden_kwargs(),
diff --git a/studio/backend/core/data_recipe/local_callable_validators.py b/studio/backend/core/data_recipe/local_callable_validators.py
index ffc81669ae..143895d781 100644
--- a/studio/backend/core/data_recipe/local_callable_validators.py
+++ b/studio/backend/core/data_recipe/local_callable_validators.py
@@ -257,6 +257,8 @@ def _run_oxc_batch(
             cwd = str(_OXC_TOOL_DIR),
             input = json.dumps(payload),
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             capture_output = True,
             check = False,
             env = env,
diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index 0af37e627f..e78bf1be8d 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -567,7 +567,7 @@ class InferenceBackend:
                         _meta_path = Path(config.path) / "export_metadata.json"
                         try:
                             if _meta_path.exists():
-                                _meta = json.loads(_meta_path.read_text(encoding = "utf-8"))
+                                _meta = json.loads(_meta_path.read_text(encoding = "utf-8-sig"))
                                 if _meta.get("base_model"):
                                     processor_source = _meta["base_model"]
                         except Exception:
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 144aa1fd37..712caf43e5 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -85,6 +85,7 @@ from core.tool_healing import (
     strip_outside_think,
 )
 from utils.native_path_leases import child_env_without_native_path_secret
+from utils.child_stdio import utf8_child_env
 from utils.hf_xet_fallback import hf_hub_download_with_xet_fallback
 from utils.subprocess_compat import (
     windows_hidden_subprocess_kwargs as _windows_hidden_subprocess_kwargs,
@@ -94,6 +95,8 @@ from core.inference.tool_call_parser import (
     MAX_ACT_REPROMPTS as _MAX_REPROMPTS,
     NUDGE_TOOL_CALLS_STATUS as _NUDGE_TOOL_CALLS_STATUS,
     REPROMPT_MAX_CHARS as _REPROMPT_MAX_CHARS,
+    is_reprompt_repeat as _is_reprompt_repeat,
+    is_reprompt_restatement as _is_reprompt_restatement,
     is_short_intent_without_action as _is_short_intent_without_action,
     reprompt_to_act_message as _reprompt_to_act_message,
 )
@@ -360,12 +363,32 @@ _DEFAULT_STREAM_STALL_TIMEOUT_S = 120.0  # 2 min
 # loop). Structured delta.tool_calls are grammar-bounded by llama-server; text
 # parsed from content is not, so one runaway turn could fan out unbounded.
 _MAX_TOOL_CALLS_PER_TURN = 8
-_FORCED_REPEAT_PLAN_SIGNAL = re.compile(
-    r"\b(?:i\s+will|i'll|let\s+me|going\s+to|need\s+to|call|use|run|search|fetch|render)\b",
+# Obligation phrasing INTENT_SIGNAL leaves alone ("I need to call ..."), paired with
+# an action verb. Sentence-anchored: mid-sentence the same words are prose that names
+# a tool ("The API I should invoke is foo() because ..."), and suppressing that loses
+# a real answer. "should"/"must" sit outside the need|have|ought group because they
+# take a bare infinitive. "invoke"/"query" stay out of the verb list: they read as
+# technical prose far more often than as a stall.
+_FORCED_PLAN_INTENT = re.compile(
+    r"(?:^|[.!?]\s+)\s*"
+    r"(?:i\s+(?:(?:need|have|ought)\s+to|should|must)|need\s+to|going\s+to|must|should)"
+    r"\s+(?:\w+\s+){0,2}?(?:call|use|run|search|fetch|render)\b",
+    re.I | re.M,
+)
+# "the answer is not in the context" announces a *missing* answer, so the negated
+# forms are excluded or the plan behind them would ship as the final response.
+_FINAL_ANSWER_SIGNAL = re.compile(
+    r"\b(?:final\s+answer|answer\s*:|here\s+is|here's|in\s+summary|result\s*:"
+    r"|(?:the\s+)?answer\s+is(?!\s+(?:not|unavailable|unknown|unclear|missing)\b))\b",
     re.I,
 )
-_FINAL_ANSWER_SIGNAL = re.compile(
-    r"\b(?:final\s+answer|answer\s*:|here\s+is|here's|in\s+summary|result\s*:)\b",
+# A plan that pivots ("I should call web_search, but Tokyo is the capital") has an
+# answer attached, so the turn must survive. Leaking a plan sentence is cosmetic;
+# dropping an answer is not, so the doubtful case keeps the output. The pivot has to
+# carry text of its own: "I should call web_search, though." answers nothing.
+_ANSWER_PIVOT = re.compile(
+    r"\b(?:but|however|although|though|that\s+said|in\s+the\s+meantime|meanwhile)\b"
+    r"[\W_]*(?:\w+[\W_]+){1,}\w",
     re.I,
 )
 
@@ -457,14 +480,28 @@ def _held_rehearsal_tail_len(text: str, active_tools: list[dict]) -> int:
     return len(tail) if tail and _is_rehearsal_prefix(tail, active_tools) else 0
 
 
-def _should_suppress_forced_no_tool_output(text: str) -> bool:
-    """Suppress only repeated forced-turn planning text, not final answers."""
+def _should_suppress_forced_no_tool_output(text: str, previous: str = "") -> bool:
+    """Suppress only repeated forced-turn planning text, not final answers.
+
+    ``previous`` is the stall text that triggered the nudge, so a retry that
+    moved on can be told from one that just said the same thing again.
+    """
     stripped = text.strip()
     if not stripped or len(stripped) >= _REPROMPT_MAX_CHARS:
         return False
     if _FINAL_ANSWER_SIGNAL.search(stripped):
         return False
-    return _FORCED_REPEAT_PLAN_SIGNAL.search(stripped) is not None
+    plan = _FORCED_PLAN_INTENT.search(stripped)
+    if plan is not None:
+        # Only the plan itself is safe to drop; anything the turn pivots to after it
+        # is the answer the user is waiting for.
+        return _ANSWER_PIVOT.search(stripped[plan.end() :]) is None
+    if not _is_short_intent_without_action(stripped):
+        return False
+    # INTENT_SIGNAL also fires on lead-ins to a real answer ("Now I have the results.
+    # The capital is Tokyo."), so a bare intent match is a stall only when the retry
+    # adds nothing. No ``previous`` keeps the standalone "is this a stall?" contract.
+    return not previous or _is_reprompt_restatement(stripped, previous)
 
 
 # ── Pre-compiled patterns for GGUF shard detection ───────────
@@ -581,7 +618,7 @@ def _load_swa_cache() -> dict:
         if _SWA_CACHE is not None:
             return _SWA_CACHE
         try:
-            with open(_swa_cache_path(), encoding = "utf-8") as f:
+            with open(_swa_cache_path(), encoding = "utf-8-sig") as f:
                 _SWA_CACHE = json.load(f)
                 if not isinstance(_SWA_CACHE, dict):
                     _SWA_CACHE = {}
@@ -632,7 +669,7 @@ def _fetch_swa_entry_from_hf(repo_id: str) -> Optional[object]:
             repo_type = "model",
             cache_dir = active_hf_hub_cache(),
         )
-        with open(cfg_path, encoding = "utf-8") as f:
+        with open(cfg_path, encoding = "utf-8-sig") as f:
             cfg = json.load(f)
     except Exception:
         return None
@@ -3046,6 +3083,7 @@ class LlamaCppBackend:
                 [bin_path, "--help"],
                 capture_output = True,
                 text = True,
+                encoding = "utf-8",
                 errors = "replace",
                 timeout = 10,
                 check = False,
@@ -3618,6 +3656,8 @@ class LlamaCppBackend:
                 ],
                 capture_output = True,
                 text = True,
+                encoding = "utf-8",
+                errors = "replace",
                 timeout = 10,
                 env = child_env_without_native_path_secret(),
                 **_windows_hidden_subprocess_kwargs(),
@@ -3732,7 +3772,7 @@ class LlamaCppBackend:
                 encoding = "utf-8",
                 errors = "replace",
                 timeout = 15,
-                env = env,
+                env = utf8_child_env(env),
                 **_windows_hidden_subprocess_kwargs(),
             )
             if result.returncode != 0:
@@ -5482,7 +5522,9 @@ class LlamaCppBackend:
             stdout = subprocess.PIPE,
             stderr = subprocess.STDOUT,
             text = True,
-            env = env,
+            encoding = "utf-8",
+            errors = "replace",
+            env = utf8_child_env(env),
             **_windows_hidden_subprocess_kwargs(),
             **_child_popen_kwargs(),
         )
@@ -6696,6 +6738,8 @@ class LlamaCppBackend:
             stdout = subprocess.PIPE,
             stderr = subprocess.STDOUT,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             env = env,
             **_windows_hidden_subprocess_kwargs(),
             **_child_popen_kwargs(),
@@ -8712,6 +8756,8 @@ class LlamaCppBackend:
                             stdout = subprocess.PIPE,
                             stderr = subprocess.STDOUT,
                             text = True,
+                            encoding = "utf-8",
+                            errors = "replace",
                             env = env,
                             **_windows_hidden_subprocess_kwargs(),
                             **_child_popen_kwargs(),
@@ -10214,6 +10260,8 @@ class LlamaCppBackend:
                     ["pgrep", "-a", "-f", "llama-server"],
                     capture_output = True,
                     text = True,
+                    encoding = "utf-8",
+                    errors = "replace",
                     timeout = 5,
                     env = child_env_without_native_path_secret(),
                 )
@@ -11712,6 +11760,10 @@ class LlamaCppBackend:
         # direct answer ("4", "Hello!") won't match. Pattern shared with the
         # safetensors loop (tool_call_parser.INTENT_SIGNAL).
         _reprompt_count = 0
+        # Budgeted apart from _reprompt_count so a pre-tool nudge can't spend it.
+        _post_tool_reprompts = 0
+        # Text that triggered the last nudge; if the retry restates it, stop.
+        _last_reprompt_text = ""
         # Gates ``max_tool_iterations`` on real tool turns (not the enlarged range) so reserved
         # re-prompt slots don't extend the budget. Mirrors the safetensors guard.
         _tool_iters_done = 0
@@ -11719,7 +11771,7 @@ class LlamaCppBackend:
 
         # Reserve extra iterations for re-prompts so they don't consume the
         # caller's tool-call budget; only when tool iterations are allowed.
-        _extra = _MAX_REPROMPTS if max_tool_iterations > 0 else 0
+        _extra = _MAX_REPROMPTS + 1 if max_tool_iterations > 0 else 0
         for iteration in range(max_tool_iterations + _extra):
             if cancel_event is not None and cancel_event.is_set():
                 return
@@ -12364,12 +12416,10 @@ class LlamaCppBackend:
                     )
                     if not _safety_tc:
                         # ── Re-prompt on plan-without-action ──
-                        # If the model described its intent (forward-looking
-                        # language) without calling a tool, nudge it to act.
-                        # Fires at most once per request, only on short
-                        # responses with intent signals -- "4" or "Hello!"
-                        # won't trigger it. Use content if available, else
-                        # fall back to reasoning text (reasoning-only stalls).
+                        # Intent described without a tool call: nudge it to act. Up
+                        # to _MAX_REPROMPTS times, only on short responses with intent
+                        # signals -- "4" or "Hello!" won't trigger it. Uses content,
+                        # else reasoning text (reasoning-only stalls).
                         _stripped = content_accum.strip()
                         if not _stripped:
                             _stripped = reasoning_accum.strip()
@@ -12379,18 +12429,33 @@ class LlamaCppBackend:
                             r"(?i)\brender[_\s-]?html\b",
                             _stripped,
                         )
+                        # A post-tool stall still deserves a nudge, but each retry
+                        # re-runs tools, so allow only one. RAG autoinject never lands
+                        # in history, so _auto keeps a doc-grounded turn from reading
+                        # as pre-tool (mirrors safetensors rag_autoinjected).
+                        _already_acted = bool(_auto) or any(
+                            record.executed for record in tool_controller.history
+                        )
+                        if _already_acted:
+                            _reprompt_used, _reprompt_cap = _post_tool_reprompts, 1
+                        else:
+                            _reprompt_used, _reprompt_cap = _reprompt_count, _MAX_REPROMPTS
                         # None keeps the default-on re-prompt; False disables it.
                         if (
                             auto_heal_tool_calls
                             and (nudge_tool_calls is None or nudge_tool_calls)
                             and active_tools
                             and not _render_html_already_done_intent
-                            and _reprompt_count < _MAX_REPROMPTS
+                            and _reprompt_used < _reprompt_cap
+                            and not _is_reprompt_repeat(_stripped, _last_reprompt_text)
                             and _is_short_intent_without_action(_stripped)
                         ):
                             _reprompt_count += 1
+                            if _already_acted:
+                                _post_tool_reprompts += 1
+                            _last_reprompt_text = _stripped
                             logger.info(
-                                f"Re-prompt {_reprompt_count}/{_MAX_REPROMPTS}: "
+                                f"Re-prompt {_reprompt_used + 1}/{_reprompt_cap}: "
                                 f"model responded without calling tools "
                                 f"({len(_stripped)} chars)"
                             )
@@ -12428,7 +12493,10 @@ class LlamaCppBackend:
 
                         if _forced_tool_call_pending:
                             _forced_tool_call_pending = False
-                            if not _should_suppress_forced_no_tool_output(_stripped):
+                            if not _should_suppress_forced_no_tool_output(
+                                _stripped,
+                                _last_reprompt_text,
+                            ):
                                 if cumulative_display:
                                     forced_visible_text = _strip_tool_markup(
                                         cumulative_display,
@@ -12758,6 +12826,10 @@ class LlamaCppBackend:
                             _kb_search_count += 1
                     completion = tool_controller.record_result(decision, result)
                     resolved_provisional_tool_call_ids.add(decision.tool_call_id)
+                    # A real execution opens the post-tool phase; carrying the pre-tool
+                    # stall text over would read the same sentence as a repeat and
+                    # swallow the one post-tool nudge.
+                    _last_reprompt_text = ""
                     # A tool ran this turn, so it counts against the caller's budget.
                     _turn_executed_real_tool = True
                     yield completion.tool_end_event()
diff --git a/studio/backend/core/inference/safetensors_agentic.py b/studio/backend/core/inference/safetensors_agentic.py
index 3057f7c2ac..3b733a85be 100644
--- a/studio/backend/core/inference/safetensors_agentic.py
+++ b/studio/backend/core/inference/safetensors_agentic.py
@@ -39,6 +39,7 @@ from core.inference.tool_call_parser import (
     RAG_MAX_SEARCHES_PER_TURN,
     RAG_SEARCH_CAP_NUDGE,
     TOOL_XML_SIGNALS,
+    is_reprompt_repeat,
     is_short_intent_without_action,
     parse_tool_calls_from_text,
     reprompt_to_act_message,
@@ -565,6 +566,8 @@ def run_safetensors_tool_loop(
     final_attempt_done = False
     next_call_id = 0
     reprompt_count = 0
+    # Text that triggered the last nudge; if the retry restates it, stop (GGUF parity).
+    last_reprompt_text = ""
     # A denied tool confirmation must not be answered with a plan-without-action
     # re-prompt (which would raise the confirmation gate again).
     tool_denied = False
@@ -1015,9 +1018,11 @@ def run_safetensors_tool_loop(
                     and not rag_autoinjected
                     and not tool_denied
                     and not any(record.executed for record in tool_controller.history)
+                    and not is_reprompt_repeat(intent_text, last_reprompt_text)
                     and is_short_intent_without_action(intent_text)
                 ):
                     reprompt_count += 1
+                    last_reprompt_text = intent_text
                     logger.info(
                         "Safetensors re-prompt %d/%d: model responded without "
                         "calling tools (%d chars)",
diff --git a/studio/backend/core/inference/tool_call_parser.py b/studio/backend/core/inference/tool_call_parser.py
index 4c3fe234ae..28b544303d 100644
--- a/studio/backend/core/inference/tool_call_parser.py
+++ b/studio/backend/core/inference/tool_call_parser.py
@@ -166,15 +166,40 @@ RAG_SEARCH_CAP_NUDGE = (
 
 
 # ── Plan-without-action re-prompt (shared by the GGUF and safetensors loops) ──
+# Verbs naming work this turn. Narrow on purpose: "install"/"add"/"open" belong to
+# advice for the user, which must not be re-prompted.
+_ACTION_VERB = (
+    r"(?:search|check|look|find|fetch|get|call|use|run|query|invoke|analy[sz]e"
+    r"|review|inspect|read|gather|examine|retrieve|browse|consult|verify"
+    r"|confirm|compute|calculate|determine|identify|render)"
+)
+# Offering to help hands control back exactly like "let me know": measured on real
+# turns, "I'll do my best to help" and "allow me to assist" close a clarification
+# request and never precede a tool call. "help you" keeps its plan reading when an
+# action follows it ("I'll help you search the web").
+_HELP_OFFER = (
+    r"(?:do(?:ing)?\s+my\s+best|try\s+my\s+best|be\s+(?:able|happy|glad)\s+to\b"
+    r"|assist\b|help\s+you\b(?!\s+" + _ACTION_VERB + r")|give\s+you\s+accurate\b)"
+)
 # Forward-looking intent: the model says what it *will* do, not a final answer.
 INTENT_SIGNAL = re.compile(
-    r"(?i)("
-    # Direct intent ("I'll", "Let me"); lookahead drops negated forms
-    # ("I will not") so a refusal does not re-prompt.
-    r"\b(i['\u2019](ll|m going to|m gonna)|i am (going to|gonna)|i will|i shall|let me|allow me)\b(?!\s+(?:not|never)\b)"
+    r"(?im)("
+    # Direct intent ("I'll"); lookahead drops negated forms ("I will not").
+    r"\b(i['\u2019](ll|m going to|m gonna)|i am (going to|gonna)|i will|i shall)\b"
+    r"(?!\s+(?:not|never)\b)(?!\s+" + _HELP_OFFER + r")"
     r"|"
-    # Step/plan framing: "First ...", "Step 1:", "Here's my plan"
-    r"\b(?:first\b|step \d+:?|here['\u2019]?s (?:my |the |a )?(?:plan|approach))"
+    # "let me know" hands control back rather than announcing an action.
+    r"\b(?:let me|allow me)\b(?!\s+(?:not|never|know)\b)(?!\s+to\s+" + _HELP_OFFER + r")"
+    r"|"
+    # Step/plan framing. "first" must open a sentence and be followed by a plan
+    # (pronoun, "my/our plan", or an action verb); otherwise it is prose ("The
+    # first line is blank.", "First place went to Alice") or advice to the user.
+    r"(?:^|[.!?]\s+)\s*(?:the\s+)?first\s+step\b"
+    r"|(?:^|[.!?]\s+)\s*first\s*[,:–—-]?\s+(?:my|our)\s+(?:plan|approach|step)\b"
+    r"|(?:^|[.!?]\s+)\s*first\s*[,:–—-]?\s+(?:i|we|let['’]?s|let us)\b"
+    r"|(?:^|[.!?]\s+)\s*first\s*[,:–—-]?\s+" + _ACTION_VERB + r"\b"
+    r"|"
+    r"\b(?:step \d+:?|here['\u2019]?s (?:my |the |a )?(?:plan|approach))"
     r"|"
     r"\b(?:now i|next i)\b"
     r")"
@@ -193,6 +218,41 @@ def is_short_intent_without_action(text: str) -> bool:
     return 0 < len(stripped) < REPROMPT_MAX_CHARS and INTENT_SIGNAL.search(stripped) is not None
 
 
+# Leading marks are kept unless they are quotes or brackets, so ".NET" survives;
+# stripping all non-word chars would collapse "C++" and "C#" to the same token.
+_REPEAT_TRAIL_PUNCT = ".,;:!?\"'`()[]{}<>‘’“”"
+_REPEAT_LEAD_PUNCT = "\"'`([{‘“"
+
+
+def _normalize_for_repeat(text: str) -> str:
+    words = []
+    for word in text.lower().split():
+        stripped = word.rstrip(_REPEAT_TRAIL_PUNCT).lstrip(_REPEAT_LEAD_PUNCT)
+        # Keep marks-only tokens: "value is 5" and "value is < 5" differ, and
+        # dropping the "<" threw the corrected attempt away.
+        words.append(stripped or word)
+    return " ".join(words)
+
+
+# A nudge that just gets the same answer back has not worked, so stop there.
+# Exact after normalisation, deliberately. Every relaxation tried here lost a real
+# correction: a similarity ratio is length dependent (one changed token in a 50-word
+# plan still scored 0.98), a set ignores order ("cats not dogs"), and ignoring filler
+# words eats the target itself ("The Who", "OK Go"). A missed repeat costs one nudge
+# out of MAX_ACT_REPROMPTS; a false one strands the plan unexecuted.
+def is_reprompt_repeat(text: str, previous: str) -> bool:
+    return is_reprompt_restatement(text, previous)
+
+
+# Same comparison, different decision: this one discards the turn. An appended answer
+# must not match, and deletions flip meaning ("is not supported" -> "is supported").
+def is_reprompt_restatement(text: str, previous: str) -> bool:
+    if not previous:
+        return False
+    a, b = _normalize_for_repeat(text), _normalize_for_repeat(previous)
+    return bool(a) and a == b
+
+
 def reprompt_to_act_message(tool_hint: str) -> str:
     """The user message appended when re-prompting a plan-without-action turn."""
     return (
diff --git a/studio/backend/core/inference/worker.py b/studio/backend/core/inference/worker.py
index 3f32b3bd57..f208183300 100644
--- a/studio/backend/core/inference/worker.py
+++ b/studio/backend/core/inference/worker.py
@@ -151,7 +151,7 @@ def _resolve_lora_4bit(mc, load_in_4bit: bool) -> bool:
     import json
 
     try:
-        with open(adapter_cfg_path, encoding = "utf-8") as f:
+        with open(adapter_cfg_path, encoding = "utf-8-sig") as f:
             adapter_cfg = json.load(f)
         training_method = adapter_cfg.get("unsloth_training_method")
         if training_method == "lora" and load_in_4bit:
@@ -963,7 +963,7 @@ def run_inference_process(
     if _local_adapter_cfg.is_file():
         try:
             _lora_base = (
-                _json.loads(_local_adapter_cfg.read_text(encoding = "utf-8")).get(
+                _json.loads(_local_adapter_cfg.read_text(encoding = "utf-8-sig")).get(
                     "base_model_name_or_path"
                 )
                 or None
diff --git a/studio/backend/core/rag/embed_llama_server.py b/studio/backend/core/rag/embed_llama_server.py
index facd989b27..b3ac62e520 100644
--- a/studio/backend/core/rag/embed_llama_server.py
+++ b/studio/backend/core/rag/embed_llama_server.py
@@ -103,6 +103,8 @@ class LlamaServerBackend:
                 [binary, "--help"],
                 capture_output = True,
                 text = True,
+                encoding = "utf-8",
+                errors = "replace",
                 timeout = 30,
                 **windows_hidden_subprocess_kwargs(),
             )
@@ -331,6 +333,8 @@ class LlamaServerBackend:
             stdout = subprocess.PIPE,
             stderr = subprocess.STDOUT,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             env = env,
             **windows_hidden_subprocess_kwargs(),
             **child_popen_kwargs(),
diff --git a/studio/backend/core/rag/embeddings.py b/studio/backend/core/rag/embeddings.py
index c86c0d3c51..95b8a866b2 100644
--- a/studio/backend/core/rag/embeddings.py
+++ b/studio/backend/core/rag/embeddings.py
@@ -100,7 +100,7 @@ def _st_module_subdirs(name: str, token: str | None) -> tuple[str, ...]:
             path = Path(normalize_path(name)).expanduser() / "modules.json"
             if not path.is_file():
                 return ()
-            data = json.loads(path.read_text(encoding = "utf-8"))
+            data = json.loads(path.read_text(encoding = "utf-8-sig"))
         else:
             from huggingface_hub import hf_hub_download
             from huggingface_hub.utils import EntryNotFoundError
@@ -115,7 +115,7 @@ def _st_module_subdirs(name: str, token: str | None) -> tuple[str, ...]:
                 )
             except EntryNotFoundError:
                 return ()
-            data = json.loads(open(local, encoding = "utf-8").read())
+            data = json.loads(open(local, encoding = "utf-8-sig").read())
         subdirs = []
         for module in data or ():
             sub = str((module or {}).get("path", "")).strip().strip("/")
diff --git a/studio/backend/core/training/worker.py b/studio/backend/core/training/worker.py
index baf6329dae..b5fb5d224e 100644
--- a/studio/backend/core/training/worker.py
+++ b/studio/backend/core/training/worker.py
@@ -43,6 +43,7 @@ if sys.platform.startswith("linux") and "HSA_ENABLE_DXG_DETECTION" not in os.env
         pass
 
 logger = get_logger(__name__)
+from utils.child_stdio import utf8_child_env
 from utils.hardware import apply_gpu_ids
 from utils.training_runs import build_default_output_dir_name
 from utils.wheel_utils import (
@@ -385,6 +386,10 @@ def _install_package_wheel_first(
         "stdout": _sp.PIPE,
         "stderr": _sp.STDOUT,
         "text": True,
+        "encoding": "utf-8",
+        "errors": "replace",
+        # Make the Python child emit the UTF-8 we decode above.
+        "env": utf8_child_env(),
     }
     if is_hip:
         _run_kwargs["timeout"] = 1800
@@ -606,6 +611,9 @@ def _ensure_flash_linear_attention_unconditional(event_queue: Any) -> bool:
             stdout = _sp.PIPE,
             stderr = _sp.STDOUT,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
+            env = utf8_child_env(),
             timeout = _TILELANG_INSTALL_TIMEOUT_S,
         )
     except _sp.TimeoutExpired:
@@ -849,6 +857,9 @@ def _run_pip(cmd: list[str], event_queue: Any, label: str) -> bool:
             stdout = _sp.PIPE,
             stderr = _sp.STDOUT,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
+            env = utf8_child_env(),
             timeout = _TILELANG_INSTALL_TIMEOUT_S,
         )
     except _sp.TimeoutExpired:
diff --git a/studio/backend/hub/services/models/ollama.py b/studio/backend/hub/services/models/ollama.py
index 56275c22a9..da30f7e98c 100644
--- a/studio/backend/hub/services/models/ollama.py
+++ b/studio/backend/hub/services/models/ollama.py
@@ -215,7 +215,7 @@ def _ollama_model_info_from_manifest(
         return None
 
     try:
-        manifest = json.loads(tag_file.read_text(encoding = "utf-8"))
+        manifest = json.loads(tag_file.read_text(encoding = "utf-8-sig"))
     except (json.JSONDecodeError, OSError, UnicodeDecodeError) as e:
         logger.debug("Skipping unreadable/invalid Ollama manifest %s: %s", tag_file, e)
         return None
@@ -228,7 +228,7 @@ def _ollama_model_info_from_manifest(
         config_blob = _ollama_blob_path(blobs_dir, config_digest)
         if config_blob is not None and _safe_is_file(config_blob):
             try:
-                cfg = json.loads(config_blob.read_text(encoding = "utf-8"))
+                cfg = json.loads(config_blob.read_text(encoding = "utf-8-sig"))
                 model_type = cfg.get("model_type", "")
                 file_type = cfg.get("file_type", "")
             except (json.JSONDecodeError, OSError, UnicodeDecodeError) as e:
diff --git a/studio/backend/hub/utils/download_registry.py b/studio/backend/hub/utils/download_registry.py
index 39c27208b1..760ef6b01c 100644
--- a/studio/backend/hub/utils/download_registry.py
+++ b/studio/backend/hub/utils/download_registry.py
@@ -464,6 +464,8 @@ def _read_marker_value(marker: Path) -> Optional[str]:
             return None
         value = marker.read_text(encoding = "utf-8").strip()
     except (OSError, UnicodeDecodeError):
+        # UnicodeDecodeError is a ValueError, so it would escape and abort
+        # prepare_cache_for_transport. An unknown value just purges and restarts.
         return None
     return value if value in VALID_TRANSPORTS else None
 
diff --git a/studio/backend/loggers/config.py b/studio/backend/loggers/config.py
index 688d3c7ebe..57cf7cecd6 100644
--- a/studio/backend/loggers/config.py
+++ b/studio/backend/loggers/config.py
@@ -42,8 +42,12 @@ class LogConfig:
         log_level_name = os.getenv("LOG_LEVEL", "INFO").upper()
         log_level = getattr(logging, log_level_name, logging.INFO)
 
-        if sys.platform == "win32":
-            for stream in (sys.stdout, sys.stderr):
+        # Non-ASCII on a non-UTF-8 stream raises UnicodeEncodeError (Windows,
+        # LANG=C), so key off the stream, not the platform.
+        for stream in (sys.stdout, sys.stderr):
+            if getattr(stream, "encoding", "") and not str(stream.encoding).lower().replace(
+                "-", ""
+            ).startswith("utf8"):
                 if hasattr(stream, "reconfigure"):
                     try:
                         stream.reconfigure(encoding = "utf-8", errors = "replace")
diff --git a/studio/backend/main.py b/studio/backend/main.py
index 02f5a20106..9a2e598314 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -347,6 +347,7 @@ from utils.update_status import (
     get_studio_install_source_status,
     get_studio_update_status,
 )
+from utils.changelog import get_release_notes, is_supported_version_query
 from utils.studio_version import get_studio_version
 from utils.api_errors import install_api_error_handlers
 
@@ -1154,6 +1155,18 @@ def studio_update_status(_current_subject: str = Depends(get_current_subject)):
     return get_studio_update_status(UNSLOTH_VERSION)
 
 
+@app.get("/api/studio/release-notes")
+def studio_release_notes(
+    version: str = Query(..., max_length = 64),
+    refresh: bool = Query(False),
+    _current_subject: str = Depends(get_current_subject),
+):
+    """Return CHANGELOG.md notes for exactly `version` (never a nearby one)."""
+    if not is_supported_version_query(version):
+        raise HTTPException(status_code = 422, detail = "Invalid version.")
+    return get_release_notes(version, refresh = refresh)
+
+
 @app.get(
     "/api/studio/download-transport-capabilities",
     response_model = TransportCapabilities,
diff --git a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper_impl/state_store.py b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper_impl/state_store.py
index b4c226136b..b059fad7ff 100644
--- a/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper_impl/state_store.py
+++ b/studio/backend/plugins/data-designer-github-repo-seed/src/data_designer_github_repo_seed/scraper_impl/state_store.py
@@ -6,10 +6,93 @@
 from __future__ import annotations
 
 import json
+import locale
 import os
 import threading
 from pathlib import Path
-from typing import Any, Dict
+from typing import Any, Dict, NamedTuple
+
+
+def _locale_encoding() -> str:
+    """The codepage a pre-UTF-8 release here would have written, or "".
+
+    Empty on a UTF-8 host, where there is no codepage to attribute the file to.
+    """
+    try:
+        preferred = locale.getencoding()
+    except AttributeError:  # Python < 3.11
+        preferred = locale.getpreferredencoding(False)
+    if preferred.lower().replace("-", "").replace("_", "") == "utf8":
+        return ""
+    return preferred
+
+
+# Trail bytes can land on JSON punctuation, so a single-byte fallback misreads these.
+_DOUBLE_BYTE_ENCODINGS = ("cp932", "cp936", "cp949", "cp950")
+
+
+def _parse(raw: bytes, encoding: str) -> Any:
+    """Parse one JSON document under *encoding*, or None if it does not.
+
+    RecursionError is a RuntimeError, so nesting json.loads will not descend is
+    the one parse failure the other three miss. Both callers run this outside
+    any further handler, so it has to answer None here or a single damaged
+    record aborts the scraper at startup instead of being skipped.
+    """
+    try:
+        return json.loads(raw.decode(encoding))
+    except (UnicodeDecodeError, LookupError, ValueError, RecursionError):
+        return None
+
+
+class _Reading(NamedTuple):
+    as_utf8: Any
+    as_legacy: Any
+
+
+def _read_line(raw: bytes, codepage: str) -> _Reading:
+    """Read one line as UTF-8 and as a codepage, for dedup keys only.
+
+    Requiring valid JSON, not merely a successful decode, is what separates a
+    genuine legacy record from a half-written UTF-8 one: a torn multibyte
+    character decodes under cp1252 but leaves the JSON unterminated. Some byte
+    strings parse both ways, e.g. cp1251 ``Р°`` is ``D0 B0``, which is also
+    UTF-8 ``а``.
+
+    The codepage reading is never authoritative, because the file's own encoding
+    cannot be recovered from its bytes. Reading a cp1251 shard on a cp1252
+    machine turns ``Привет`` into ``Ïðèâåò`` and every byte of it decodes
+    cleanly, so a successful decode proves nothing about who wrote it. It is
+    used only to recover the dedup keys, which are ASCII ids and come back the
+    same under any of these, so the first reading that parses will do.
+
+    That is also why several are tried. latin-1 alone mangles the double-byte
+    codepages: cp932 ``表`` is ``95 5C``, and latin-1 turns the trail byte into
+    a JSON backslash, so the record fails to parse and its id is forgotten.
+    """
+    as_utf8 = _parse(raw, "utf-8")
+    # A record that reads as UTF-8 needs no second reading: re-parsing cost 2.8x on a
+    # 76 MB shard, and these reach gigabytes. Only a dict, since key lookup falls
+    # through to the codepage when UTF-8 yields none.
+    if isinstance(as_utf8, dict):
+        return _Reading(as_utf8, None)
+    for encoding in (codepage, "latin-1", *_DOUBLE_BYTE_ENCODINGS):
+        if not encoding:
+            continue
+        as_legacy = _parse(raw, encoding)
+        if as_legacy is not None:
+            return _Reading(as_utf8, as_legacy)
+    return _Reading(as_utf8, None)
+
+
+class _Scan(NamedTuple):
+    """What a pass over an existing shard established about it."""
+
+    legacy: bool  # enough evidence to trust the codepage reading's keys
+    readable: bool
+    saw_non_ascii: bool  # some line's meaning depends on the encoding
+    utf8_keys: set  # keys from lines UTF-8 could read
+    legacy_keys: set  # keys only the codepage reading yields
 
 
 class StateStore:
@@ -18,12 +101,19 @@ class StateStore:
         self.path.parent.mkdir(parents = True, exist_ok = True)
         self._lock = threading.Lock()
         self._data: Dict[str, Any] = {}
+        # Read whole, and UTF-8 only unlike the shards below: a checkpoint holds
+        # nothing but base64 cursors and booleans, so a codepage retry could only ever
+        # add non-ASCII. That would resume on a mojibaked cursor, which GitHub rejects
+        # with INVALID_CURSOR_ARGUMENTS, and the empty page it returns marks the stream
+        # done and skips the rest for good. Dropping a damaged checkpoint re-scrapes
+        # from the first page, which the writers dedup.
         if self.path.exists():
             try:
-                with self.path.open(encoding = "utf-8") as f:
-                    self._data = json.load(f)
-            except Exception:
-                self._data = {}
+                raw = self.path.read_bytes()
+            except OSError:
+                raw = b""
+            data = _parse(raw, "utf-8")
+            self._data = data if isinstance(data, dict) else {}
 
     def get(
         self,
@@ -63,24 +153,83 @@ class JsonlWriter:
         self.path = Path(path)
         self.path.parent.mkdir(parents = True, exist_ok = True)
         self._lock = threading.Lock()
-        self._fh = self.path.open("a", buffering = 1, encoding = "utf-8")
         self._count_seen_keys: set[str] = set()
-        # Preload seen keys for dedup across resumes
+        self._codepage = _locale_encoding()
+        self._ensure_ascii = False
+        encoding = "utf-8"
         if self.path.exists() and self.path.stat().st_size > 0:
-            try:
-                # No guess is safe for a file an older build wrote in the
-                # operator's locale, so read past whatever will not decode.
-                with self.path.open(encoding = "utf-8", errors = "replace") as f:
-                    for line in f:
-                        try:
-                            obj = json.loads(line)
-                            k = self._key(obj)
-                            if k is not None:
-                                self._count_seen_keys.add(k)
-                        except Exception:
-                            pass
-            except Exception:
-                pass
+            scan = self._scan_existing()
+            self._count_seen_keys = scan.utf8_keys
+            if scan.legacy:
+                self._count_seen_keys |= scan.legacy_keys
+            if scan.saw_non_ascii or not scan.readable:
+                # Never convert: the writing encoding is unrecoverable and guessing
+                # mojibakes the records. Pure ASCII appends store identically under
+                # every codepage, and json.loads turns the \uXXXX escapes back.
+                encoding = "ascii"
+                self._ensure_ascii = True
+        self._fh = self.path.open("a", buffering = 1, encoding = encoding, errors = "strict")
+
+    def _scan_existing(self) -> _Scan:
+        """Read the shard once to recover dedup keys and judge its encoding.
+
+        Line by line: these shards reach gigabytes on a large scrape, so neither
+        the bytes nor the decoded text are held whole.
+
+        The verdict weighs the whole file. Each line with non-ASCII bytes votes:
+        one that parses only under the codepage is evidence of a legacy shard,
+        one that parses as UTF-8 is evidence against, since arbitrary codepage
+        text almost never forms valid multibyte UTF-8. A single corrupt byte in
+        a healthy shard therefore cannot outvote the records around it, and a
+        genuinely legacy shard has a legacy vote on every line that carries an
+        umlaut.
+
+        More than one such line is required, because a single one is genuinely
+        undecidable: a legacy record holding one accented character and an ASCII
+        record holding one stray byte are the same shape. Reading it as damage
+        risks a duplicate; reading it as legacy marks an unreadable record seen
+        and blocks the retry that would replace it, losing it for good. Only one
+        of those is recoverable.
+
+        The verdict only picks which reading supplies the dedup keys. The file
+        itself is never rewritten either way, so a wrong answer costs at most a
+        duplicate, never a corrupted record.
+        """
+        legacy_votes = 0
+        utf8_votes = 0
+        saw_non_ascii = False
+        utf8_keys: set[str] = set()
+        legacy_keys: set[str] = set()
+        try:
+            with self.path.open("rb") as handle:
+                for raw in handle:
+                    line = raw.strip()
+                    reading = _read_line(line, self._codepage)
+                    # ASCII reads the same everywhere: no vote, no constraint.
+                    if not line.isascii():
+                        saw_non_ascii = True
+                        if reading.as_utf8 is None and reading.as_legacy is not None:
+                            legacy_votes += 1
+                        elif reading.as_utf8 is not None:
+                            utf8_votes += 1
+                    # Kept apart so a damaged line does not block its own retry.
+                    if isinstance(reading.as_utf8, dict):
+                        key = self._key(reading.as_utf8)
+                        if key is not None:
+                            utf8_keys.add(key)
+                    elif isinstance(reading.as_legacy, dict):
+                        key = self._key(reading.as_legacy)
+                        if key is not None:
+                            legacy_keys.add(key)
+        except OSError:
+            return _Scan(False, False, False, utf8_keys, legacy_keys)
+        return _Scan(
+            legacy_votes > 1 and legacy_votes > utf8_votes,
+            True,
+            saw_non_ascii,
+            utf8_keys,
+            legacy_keys,
+        )
 
     def _key(self, obj: dict) -> str | None:
         for k in ("id", "node_id", "number", "sha", "url"):
@@ -99,7 +248,7 @@ class JsonlWriter:
                 return False
             if k is not None:
                 self._count_seen_keys.add(k)
-            self._fh.write(json.dumps(obj, default = str, ensure_ascii = False))
+            self._fh.write(json.dumps(obj, default = str, ensure_ascii = self._ensure_ascii))
             self._fh.write("\n")
             self._fh.flush()
         return True
diff --git a/studio/backend/plugins/data-designer-unstructured-seed/src/data_designer_unstructured_seed/impl.py b/studio/backend/plugins/data-designer-unstructured-seed/src/data_designer_unstructured_seed/impl.py
index ce0c88e5bf..825b050e07 100644
--- a/studio/backend/plugins/data-designer-unstructured-seed/src/data_designer_unstructured_seed/impl.py
+++ b/studio/backend/plugins/data-designer-unstructured-seed/src/data_designer_unstructured_seed/impl.py
@@ -30,6 +30,8 @@ class UnstructuredSeedReader(SeedReader[UnstructuredSeedSource]):
                     meta = json_mod.loads(meta_path.read_text(encoding = "utf-8"))
                     orig_name = meta.get("original_filename", path_obj.name)
                 except (json_mod.JSONDecodeError, OSError, UnicodeDecodeError):
+                    # Undecodable metadata is as malformed as invalid JSON, so
+                    # fall back to the file's own name rather than abort the seed.
                     pass
             file_entries.append((path_obj, orig_name))
 
diff --git a/studio/backend/routes/auth.py b/studio/backend/routes/auth.py
index 1acc48e3a3..fe2f09fcd9 100644
--- a/studio/backend/routes/auth.py
+++ b/studio/backend/routes/auth.py
@@ -31,6 +31,7 @@ from auth import storage, hashing
 from auth.authentication import (
     create_access_token,
     create_refresh_token,
+    get_current_credential,
     get_current_subject,
     get_current_subject_allow_password_change,
     refresh_access_token,
@@ -399,7 +400,7 @@ async def login(payload: AuthLoginRequest, request: Request) -> Token:
             detail = f"Incorrect password. To reset it, run this in your terminal: {_reset_password_command()}",
         )
 
-    salt, pwd_hash, _jwt_secret, must_change_password = record
+    salt, pwd_hash, jwt_secret, must_change_password = record
     if not hashing.verify_password(payload.password, salt, pwd_hash):
         _record_login_failure(key)
         raise HTTPException(
@@ -409,8 +410,10 @@ async def login(payload: AuthLoginRequest, request: Request) -> Token:
 
     _clear_login_bucket(key)
     _clear_login_bucket(unknown_key)
-    access_token = create_access_token(subject = payload.username)
-    refresh_token = create_refresh_token(subject = payload.username)
+    # Issue against the credential version just verified, not whatever is in the DB
+    # now: a concurrent reset-password must not hand this login a post-reset session.
+    access_token = create_access_token(subject = payload.username, secret = jwt_secret)
+    refresh_token = create_refresh_token(subject = payload.username, secret = jwt_secret)
     return Token(
         access_token = access_token,
         refresh_token = refresh_token,
@@ -438,16 +441,17 @@ async def logout(
 @router.post("/desktop-login", response_model = Token)
 async def desktop_login(payload: DesktopLoginRequest) -> Token:
     """Exchange a local desktop secret for normal admin-subject tokens."""
-    username = storage.validate_desktop_secret(payload.secret)
-    if username is None:
+    verified = storage.validate_desktop_secret_with_credential(payload.secret)
+    if verified is None:
         raise HTTPException(
             status_code = status.HTTP_401_UNAUTHORIZED,
             detail = "Desktop authentication failed",
         )
+    username, jwt_secret = verified
 
     return Token(
-        access_token = create_access_token(subject = username, desktop = True),
-        refresh_token = create_refresh_token(subject = username, desktop = True),
+        access_token = create_access_token(subject = username, desktop = True, secret = jwt_secret),
+        refresh_token = create_refresh_token(subject = username, desktop = True, secret = jwt_secret),
         token_type = "bearer",
         must_change_password = False,
     )
@@ -462,9 +466,11 @@ async def refresh(payload: RefreshTokenRequest) -> Token:
             status_code = status.HTTP_401_UNAUTHORIZED,
             detail = "Invalid or expired refresh token",
         )
-    username, is_desktop = consumed
-    new_access_token = create_access_token(subject = username, desktop = is_desktop)
-    new_refresh_token = create_refresh_token(subject = username, desktop = is_desktop)
+    username, is_desktop, jwt_secret = consumed
+    new_access_token = create_access_token(subject = username, desktop = is_desktop, secret = jwt_secret)
+    new_refresh_token = create_refresh_token(
+        subject = username, desktop = is_desktop, secret = jwt_secret
+    )
 
     return Token(
         access_token = new_access_token,
@@ -507,13 +513,25 @@ async def change_password(
 
     # Single transaction: a separate refresh-token purge could fail after the
     # password commit, leaving pre-change tokens able to mint access tokens.
-    storage.update_password(current_subject, payload.new_password, revoke_refresh_tokens = True)
+    # Conditional on the hash just verified: a reset-password that landed while
+    # this request was in flight must not be overwritten by it.
+    new_secret = storage.update_password(
+        current_subject,
+        payload.new_password,
+        revoke_refresh_tokens = True,
+        expect_password_hash = pwd_hash,
+    )
+    if new_secret is None:
+        raise HTTPException(
+            status_code = status.HTTP_409_CONFLICT,
+            detail = "The password changed while this request was in flight. Sign in again.",
+        )
     try:
         request.app.state.bootstrap_password = None
     except AttributeError:
         pass
-    access_token = create_access_token(subject = current_subject)
-    refresh_token = create_refresh_token(subject = current_subject)
+    access_token = create_access_token(subject = current_subject, secret = new_secret)
+    refresh_token = create_refresh_token(subject = current_subject, secret = new_secret)
     return Token(
         access_token = access_token,
         refresh_token = refresh_token,
@@ -541,20 +559,28 @@ def _row_to_api_key_response(row: dict) -> ApiKeyResponse:
 
 @router.post("/api-keys", response_model = CreateApiKeyResponse)
 async def create_api_key(
-    payload: CreateApiKeyRequest, current_subject: str = Depends(get_current_subject)
+    payload: CreateApiKeyRequest, credential: tuple = Depends(get_current_credential)
 ) -> CreateApiKeyResponse:
     """Create a new API key. The raw key is returned once and cannot be retrieved later."""
+    current_subject, generation = credential
     expires_at = None
     if payload.expires_in_days is not None:
         expires_at = (
             datetime.now(timezone.utc) + timedelta(days = payload.expires_in_days)
         ).isoformat()
 
-    raw_key, row = storage.create_api_key(
-        username = current_subject,
-        name = payload.name,
-        expires_at = expires_at,
-    )
+    try:
+        raw_key, row = storage.create_api_key(
+            username = current_subject,
+            name = payload.name,
+            expires_at = expires_at,
+            expect_gen = generation,
+        )
+    except storage.CredentialRotated:
+        raise HTTPException(
+            status_code = status.HTTP_401_UNAUTHORIZED,
+            detail = "Invalid or expired token",
+        )
     return CreateApiKeyResponse(
         key = raw_key,
         api_key = _row_to_api_key_response(row),
diff --git a/studio/backend/routes/data_recipe/jobs.py b/studio/backend/routes/data_recipe/jobs.py
index e870e8855e..7fdf0abada 100644
--- a/studio/backend/routes/data_recipe/jobs.py
+++ b/studio/backend/routes/data_recipe/jobs.py
@@ -10,7 +10,10 @@ from datetime import datetime, timedelta, timezone
 from typing import Any, Optional
 from urllib.parse import urlparse
 
-from fastapi import APIRouter, HTTPException, Query, Request
+from fastapi import APIRouter, Depends, HTTPException, Query, Request
+
+from auth.authentication import get_current_credential
+from auth.storage import CredentialRotated
 from fastapi.responses import JSONResponse, StreamingResponse
 from pydantic import ValidationError
 
@@ -257,7 +260,11 @@ def _inject_local_structured_response_format(
         model_configs.extend(new_configs)
 
 
-def _inject_local_providers(recipe: dict[str, Any], request: Request) -> Optional[int]:
+def _inject_local_providers(
+    recipe: dict[str, Any],
+    request: Request,
+    expect_gen: Optional[str] = None,
+) -> Optional[int]:
     """Mutate recipe in-place: point is_local providers at this server and mint
     a short-lived internal sk-unsloth-* key for workflow auth.
 
@@ -313,6 +320,7 @@ def _inject_local_providers(recipe: dict[str, Any], request: Request) -> Optiona
             name = "data-recipe workflow",
             expires_at = expires_at,
             internal = True,
+            expect_gen = expect_gen,
         )
         internal_key_id = int(row["id"])
 
@@ -375,7 +383,11 @@ def _normalize_run_name(value: Any) -> str | None:
 
 
 @router.post("/jobs", response_class = JSONResponse, response_model = JobCreateResponse)
-def create_job(payload: RecipePayload, request: Request):
+def create_job(
+    payload: RecipePayload,
+    request: Request,
+    credential: tuple = Depends(get_current_credential),
+):
     recipe = payload.recipe
     if not recipe.get("columns"):
         raise HTTPException(status_code = 400, detail = "Recipe must include columns.")
@@ -406,7 +418,11 @@ def create_job(payload: RecipePayload, request: Request):
             ) from exc
 
     try:
-        internal_api_key_id = _inject_local_providers(recipe, request)
+        internal_api_key_id = _inject_local_providers(recipe, request, credential[1])
+    except CredentialRotated as exc:
+        # A reset-password landed after this request authenticated; the workflow key
+        # is refused, so answer like any other revoked credential rather than 500.
+        raise HTTPException(status_code = 401, detail = "Invalid or expired token") from exc
     except ValueError as exc:
         raise log_and_http_error(
             exc,
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index d0a2d97f74..20a5af1409 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -4434,7 +4434,7 @@ def _effective_load_in_4bit(config: ModelConfig, requested: bool) -> bool:
     if not adapter_cfg_path.exists():
         return load_in_4bit
     try:
-        with open(adapter_cfg_path, encoding = "utf-8") as f:
+        with open(adapter_cfg_path, encoding = "utf-8-sig") as f:
             adapter_cfg = json.load(f)
         if not isinstance(adapter_cfg, dict):  # malformed -> keep requested
             return load_in_4bit
diff --git a/studio/backend/routes/models.py b/studio/backend/routes/models.py
index 96c5b96d73..6e587c18e8 100644
--- a/studio/backend/routes/models.py
+++ b/studio/backend/routes/models.py
@@ -722,7 +722,7 @@ def _scan_ollama_dir(ollama_dir: Path, limit: Optional[int] = None) -> List[Loca
             stem_hash = hashlib.sha256(manifest_key.encode()).hexdigest()[:10]
 
             try:
-                manifest = json.loads(tag_file.read_text(encoding = "utf-8"))
+                manifest = json.loads(tag_file.read_text(encoding = "utf-8-sig"))
             except (json.JSONDecodeError, OSError, UnicodeDecodeError) as e:
                 logger.debug(
                     "Skipping unreadable/invalid Ollama manifest %s: %s",
@@ -738,7 +738,7 @@ def _scan_ollama_dir(ollama_dir: Path, limit: Optional[int] = None) -> List[Loca
                 config_blob = blobs_dir / config_digest.replace(":", "-")
                 if config_blob.is_file():
                     try:
-                        cfg = json.loads(config_blob.read_text(encoding = "utf-8"))
+                        cfg = json.loads(config_blob.read_text(encoding = "utf-8-sig"))
                         model_type = cfg.get("model_type", "")
                         file_type = cfg.get("file_type", "")
                     except (json.JSONDecodeError, OSError, UnicodeDecodeError) as e:
@@ -1042,7 +1042,7 @@ def _dir_has_downloaded_model(directory: Path, max_entries: int = 4000) -> bool:
                 if not m.is_file():
                     continue
                 try:
-                    manifest = json.loads(m.read_text(encoding = "utf-8"))
+                    manifest = json.loads(m.read_text(encoding = "utf-8-sig"))
                 except (json.JSONDecodeError, OSError, ValueError):
                     continue
                 for layer in manifest.get("layers") or []:
@@ -3360,6 +3360,8 @@ def _wsl_reveal_in_explorer(path: Path) -> bool:
             ["wslpath", "-w", str(path)],
             capture_output = True,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             check = True,
             timeout = 10,
         ).stdout.strip()
diff --git a/studio/backend/run.py b/studio/backend/run.py
index 08d1c5299e..2d9e714d90 100644
--- a/studio/backend/run.py
+++ b/studio/backend/run.py
@@ -10,7 +10,7 @@ import os
 import sys
 import time
 from pathlib import Path
-from typing import Optional, Tuple
+from typing import NoReturn, Optional, Sequence, Tuple
 
 
 def _fix_torch_cuda_ld_path():
@@ -689,6 +689,33 @@ def _get_pid_on_port(port: int) -> "tuple[int, str] | None":
     return None
 
 
+def _bind_addresses(host: str, port: int) -> "set[str]":
+    """Every address *host* resolves to. `localhost` is both 127.0.0.1 and ::1, and
+    recording only the first lets a later launch on the other one miss us."""
+    import socket
+
+    try:
+        infos = socket.getaddrinfo(host, port, socket.AF_UNSPEC, socket.SOCK_STREAM)
+    except OSError:
+        return {host}
+    return {info[4][0] for info in infos} or {host}
+
+
+def _addresses_collide(recorded: "str | None", host: str, port: int) -> bool:
+    """Would a server bound to *recorded* block a bind to *host*?
+
+    *recorded* may list several addresses. Unknown or wildcard on either side
+    collides: refusing with a clear message beats silently starting a duplicate.
+    """
+    wildcards = ("0.0.0.0", "::", "")
+    if not recorded or host in wildcards:
+        return True
+    listed = {a.strip() for a in recorded.split(",") if a.strip()}
+    if not listed or listed & set(wildcards):
+        return True
+    return bool(listed & _bind_addresses(host, port))
+
+
 def _is_port_free(host: str, port: int) -> bool:
     """Check if a port is available for binding.
 
@@ -733,18 +760,213 @@ def _find_free_port(
     host: str,
     start: int,
     max_attempts: int = 20,
+    avoid_own_studio: bool = False,
 ) -> int:
-    """Find a free port from `start`, trying up to max_attempts ports."""
+    """Find a free port from `start`, trying up to max_attempts ports.
+
+    ``avoid_own_studio`` aborts rather than skipping past one of our own servers
+    in the fallback range, which would start a duplicate on a later port.
+    """
     for offset in range(max_attempts):
         candidate = start + offset
         if _is_port_free(host, candidate):
             return candidate
+        if avoid_own_studio:
+            own = _own_studio_on_port(candidate, host)
+            if own is not None:
+                _abort_already_running(own, candidate)
     raise RuntimeError(f"Could not find a free port in range {start}-{start + max_attempts - 1}")
 
 
 from utils.paths.storage_roots import studio_root as _studio_root
 
+# Legacy single-instance file; still read so `stop` finds an older build's server.
 _PID_FILE = _studio_root() / "studio.pid"
+PID_FILE_GLOB = "studio-*.pid"
+
+
+def _pid_file_for_port(port: int) -> Path:
+    # PID in the name: 127.0.0.1 and ::1 can share a port, and one file per port
+    # would let the second bind overwrite the first.
+    return _studio_root() / f"studio-{port}-{os.getpid()}.pid"
+
+
+def _pid_alive(pid: int) -> bool:
+    try:
+        import psutil
+        return psutil.pid_exists(pid)
+    except ImportError:
+        pass
+    if sys.platform == "win32":
+        # os.kill(pid, 0) raises OSError for every pid on Windows, so tasklist is
+        # the only usable probe here.
+        import subprocess
+        try:
+            out = subprocess.run(
+                ["tasklist", "/FI", f"PID eq {int(pid)}", "/NH", "/FO", "CSV"],
+                capture_output = True,
+                text = True,
+                timeout = 10,
+            ).stdout
+        except Exception:
+            # Unconfirmed means keep, matching the CLI's _pid_alive. Pruning a
+            # live server's record is what lets the next launch fall back past it
+            # and strand it, which is the bug this file exists to fix. A stale
+            # record instead costs one clear "already running" message.
+            return True
+        return f'"{int(pid)}"' in out
+    try:
+        os.kill(pid, 0)
+    except ProcessLookupError:
+        return False
+    except OSError:
+        return True
+    return True
+
+
+def _process_create_time(pid: int) -> "float | None":
+    try:
+        import psutil
+        return psutil.Process(pid).create_time()
+    except Exception:
+        return None
+
+
+def _read_pid_record(path: Path) -> "tuple[int, float | None, str | None] | None":
+    """Parse ``pid`` / optional ``create_time`` / optional bind address."""
+    try:
+        lines = path.read_text(encoding = "utf-8").splitlines()
+    except (OSError, UnicodeDecodeError):
+        return None
+    if not lines or not lines[0].strip().isdigit():
+        return None
+    try:
+        # isdigit() is not enough: a superscript two passes it but int() rejects it.
+        pid = int(lines[0].strip())
+    except ValueError:
+        return None
+    # kill(0) signals our whole process group; kill(1) is init. Never either.
+    if pid < 2:
+        return None
+    created = None
+    if len(lines) > 1:
+        try:
+            created = float(lines[1].strip())
+        except ValueError:
+            created = None
+    address = lines[2].strip() if len(lines) > 2 and lines[2].strip() else None
+    return pid, created, address
+
+
+def _pid_is_studio_backend(pid: int, created_times: "Sequence[float | None]" = ()) -> bool:
+    """False only when a recorded start time proves this PID is a different process.
+
+    Any recorded time matching is enough -- a stale record must not veto a live
+    server that reused the PID. Untimed records cannot be checked at all, so they
+    are trusted: a legacy `python run.py` has no telltale argv, and guessing from
+    the command line rejected real servers.
+    """
+    known = [c for c in created_times if c is not None]
+    if not known:
+        return True
+    actual = _process_create_time(pid)
+    if actual is None:
+        return True
+    return any(abs(actual - c) < 1.0 for c in known)
+
+
+def _own_studio_on_port(port: int, host: str) -> "int | None":
+    """PID of one of our own servers already bound to *port* for *host*.
+
+    Reads our own records rather than enumerating listeners: psutil is optional,
+    and without it a listener scan finds nothing and we silently start a duplicate.
+    """
+    try:
+        paths = list(_studio_root().glob(f"studio-{port}-*.pid"))
+    except OSError:
+        return None
+    for path in paths:
+        record = _read_pid_record(path)
+        if record is None:
+            continue
+        pid, created, address = record
+        if not _pid_alive(pid):
+            # Pruning is a courtesy; an undeletable record must not abort startup.
+            try:
+                path.unlink(missing_ok = True)
+            except OSError:
+                pass
+            continue
+        if not _addresses_collide(address, host, port):
+            continue
+        if _pid_is_studio_backend(pid, [created]):
+            return pid
+    return _legacy_studio_on_port(port)
+
+
+def _legacy_studio_on_port(port: int) -> "int | None":
+    """A pre-upgrade server recorded only its PID, so match it to the listener.
+
+    Falling back past one leaves it running while `_write_pid_file` overwrites the
+    only record of it. When the listener is unknowable, assume it is ours.
+    """
+    record = _read_pid_record(_PID_FILE)
+    if record is None:
+        return None
+    pid, created, _address = record
+    if not _pid_alive(pid):
+        return None
+    # A current build writes a per-port file too, so its port is already known --
+    # and this port's records were just checked. Only count a record that still
+    # matches the live process: a stale one may just share a reused PID.
+    for other in _per_port_records():
+        if other and other[0] == pid and _pid_is_studio_backend(pid, [other[1]]):
+            return None
+    blocker = _get_pid_on_port(port)
+    if blocker is not None and blocker[0] != pid:
+        return None
+    if not _pid_is_studio_backend(pid, [created]):
+        return None
+    return pid
+
+
+def _per_port_records() -> "list[tuple[int, float | None, str | None] | None]":
+    try:
+        return [_read_pid_record(p) for p in _studio_root().glob(PID_FILE_GLOB)]
+    except OSError:
+        return []
+
+
+def _resolve_port(
+    host: str,
+    port: int,
+    avoid_own_studio: bool = True,
+) -> int:
+    """The requested port, or the next free one.
+
+    With ``avoid_own_studio`` this aborts rather than falling back past one of our
+    own servers, on *port* itself or anywhere in the fallback range: skipping one
+    is what strands it. Callers that read the bound port back pass False and keep
+    the plain fallback.
+    """
+    if _is_port_free(host, port):
+        return port
+    if avoid_own_studio:
+        own = _own_studio_on_port(port, host)
+        if own is not None:
+            _abort_already_running(own, port)
+    return _find_free_port(host, port + 1, avoid_own_studio = avoid_own_studio)
+
+
+def _abort_already_running(pid: int, port: int) -> "NoReturn":
+    print(
+        f"Error: Unsloth Studio is already running on port {port} (PID {pid}). Run "
+        "`unsloth studio stop` first, or start this one on a different --port.",
+        file = sys.stderr,
+        flush = True,
+    )
+    sys.exit(1)
+
 
 # Direct backend launches bypass the CLI's env re-export; do it here for
 # real custom roots so unsloth-zoo's import-time LLAMA_CPP_DEFAULT_DIR
@@ -770,23 +992,101 @@ if _STUDIO_ROOT_RESOLVED != _LEGACY_STUDIO_ROOT:
 os.environ.setdefault("UNSLOTH_IS_PRESENT", "1")
 
 
-def _write_pid_file():
-    """Write the current process PID to the studio PID file."""
+_OWN_PID_FILE: "Path | None" = None
+
+
+def _write_pid_file(port: int, host: str = ""):
+    """Record this PID under its own port so `stop` can find every server."""
+    global _OWN_PID_FILE
+    path = _pid_file_for_port(port)
     try:
-        _PID_FILE.parent.mkdir(parents = True, exist_ok = True)
-        _PID_FILE.write_text(str(os.getpid()), encoding = "utf-8")
+        path.parent.mkdir(parents = True, exist_ok = True)
+    except OSError:
+        pass
+    try:
+        # Start time pins the record to this process; the bind address tells a
+        # later launch whether this server would actually block it.
+        created = _process_create_time(os.getpid())
+        address = ",".join(sorted(_bind_addresses(host, port))) if host else ""
+        body = f"{os.getpid()}\n{'' if created is None else repr(created)}\n{address}"
+        # Write-then-rename: `stop` reads these concurrently, and a reader that
+        # catches the truncate window sees a corrupt record and deletes it.
+        tmp = path.with_name(path.name + ".tmp")
+        try:
+            tmp.write_text(body, encoding = "utf-8")
+            os.replace(tmp, path)
+        finally:
+            # A failed replace would otherwise leave the scratch file behind. It
+            # does not end in .pid, so no glob picks it up either way.
+            tmp.unlink(missing_ok = True)
+    except OSError:
+        pass
+    else:
+        _OWN_PID_FILE = path
+    # An older CLI's `stop` only reads this one, and expects a bare PID. Written
+    # independently of the per-port record: if that one failed, this is the only
+    # thing keeping the server stoppable at all.
+    try:
+        # Never take it from a server that is still running. A pre-upgrade server
+        # is recorded here and nowhere else, so overwriting its entry is exactly
+        # what strands it -- the orphan this file exists to prevent.
+        prior = _read_pid_record(_PID_FILE) if _PID_FILE.is_file() else None
+        if prior is None or prior[0] == os.getpid() or not _pid_alive(prior[0]):
+            _PID_FILE.write_text(str(os.getpid()), encoding = "utf-8")
     except OSError:
         pass
 
 
-def _remove_pid_file():
-    """Remove the PID file if it belongs to this process."""
+def _legacy_heir() -> "int | None":
+    """Another live server's PID, to hand the legacy studio.pid over to.
+
+    Only one server owns studio.pid at a time, so its exit would otherwise drop
+    the single record an older CLI can read, stranding any sibling that is still
+    serving.
+    """
     try:
-        if _PID_FILE.is_file():
-            stored = _PID_FILE.read_text(encoding = "utf-8").strip()
-            if stored == str(os.getpid()):
+        paths = sorted(_studio_root().glob(PID_FILE_GLOB))
+    except OSError:
+        return None
+    for path in paths:
+        if _OWN_PID_FILE is not None and path == _OWN_PID_FILE:
+            continue
+        record = _read_pid_record(path)
+        if record is None or record[0] == os.getpid():
+            continue
+        if _pid_alive(record[0]) and _pid_is_studio_backend(record[0], [record[1]]):
+            return record[0]
+    return None
+
+
+def _remove_pid_file():
+    """Remove the PID files that belong to this process.
+
+    _PID_FILE is checked even when the per-port record was never written, since
+    _write_pid_file writes the two independently.
+    """
+    # Nothing here may raise: _graceful_shutdown calls this at the end, and an
+    # unreadable or undeletable record must not abandon the rest of the exit
+    # path. _read_pid_record already swallows OSError/UnicodeDecodeError.
+    if _OWN_PID_FILE is not None:
+        try:
+            record = _read_pid_record(_OWN_PID_FILE) if _OWN_PID_FILE.is_file() else None
+            if record is not None and record[0] == os.getpid():
+                _OWN_PID_FILE.unlink(missing_ok = True)
+        except OSError:
+            pass
+    try:
+        record = _read_pid_record(_PID_FILE) if _PID_FILE.is_file() else None
+        if record is not None and record[0] == os.getpid():
+            # Hand the pointer to a live sibling rather than deleting it. An
+            # older CLI reads only this file, so dropping it while another
+            # server is still up leaves that server unstoppable.
+            heir = _legacy_heir()
+            if heir is None:
                 _PID_FILE.unlink(missing_ok = True)
-    except (OSError, UnicodeDecodeError):
+            else:
+                _PID_FILE.write_text(str(heir), encoding = "utf-8")
+    except OSError:
         pass
 
 
@@ -796,7 +1096,6 @@ def _graceful_shutdown(server = None):
     Called from signal handlers to clean up children before exit. Critical on
     Windows where atexit handlers are unreliable after Ctrl+C.
     """
-    _remove_pid_file()
     logger.info("Graceful shutdown initiated -- cleaning up subprocesses...")
 
     # 1. Shut down uvicorn (releases the listening socket).
@@ -849,6 +1148,9 @@ def _graceful_shutdown(server = None):
     except Exception as e:
         logger.warning("Error in process-lifetime sweep: %s", e)
 
+    # Last: while cleanup runs the server is still alive, and dropping the record
+    # early leaves a retried `stop` or a new launch unable to find it.
+    _remove_pid_file()
     logger.info("All subprocesses cleaned up")
 
 
@@ -1326,7 +1628,8 @@ def _apply_supplied_password(password_value: "Optional[str]") -> None:
     if not _auth_storage.requires_password_change(_admin):
         print(
             "Error: an Unsloth admin password is already set; --password only sets "
-            "the initial password. Run `unsloth studio reset-password` first.",
+            "the initial password. Change it in the UI, or run `unsloth studio "
+            "reset-password` for a new one.",
             file = sys.stderr,
             flush = True,
         )
@@ -1397,6 +1700,7 @@ def run_server(
     enable_tools: "Optional[bool]" = None,
     password: "Optional[str]" = None,
     emit_tauri_port: bool = True,
+    abort_if_own_studio: "Optional[bool]" = None,
 ):
     """
     Start the FastAPI server.
@@ -1530,10 +1834,16 @@ def run_server(
     )
 
     # Auto-find a free port if the requested one is in use.
-    if not _is_port_free(host, port):
-        original_port = port
-        blocker = _get_pid_on_port(port)
-        port = _find_free_port(host, port + 1)
+    original_port = port
+    # Refusing rather than falling back is for callers that cannot follow us to
+    # the new port. `studio run` reads app.state.server_port back and the desktop
+    # app reads TAURI_PORT, so both should keep the plain fallback; only the
+    # bare launch, which has nothing but the banner, benefits from the refusal.
+    if abort_if_own_studio is None:
+        abort_if_own_studio = not api_only
+    port = _resolve_port(host, port, avoid_own_studio = abort_if_own_studio)
+    if port != original_port:
+        blocker = _get_pid_on_port(original_port)
         if not silent:
             print("")
             print("=" * 50)
@@ -1731,7 +2041,7 @@ def run_server(
         (time.perf_counter() - boot_started) * 1000,
     )
 
-    _write_pid_file()
+    _write_pid_file(port, host)
     import atexit
 
     atexit.register(_remove_pid_file)
diff --git a/studio/backend/tests/data/plan_vs_answer.jsonl b/studio/backend/tests/data/plan_vs_answer.jsonl
new file mode 100644
index 0000000000..7397c8da7e
--- /dev/null
+++ b/studio/backend/tests/data/plan_vs_answer.jsonl
@@ -0,0 +1,300 @@
+{"id": "7cdc94d4ad818e68", "model": "llama3.2-1b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "import statistics\n# Define the list of numbers\nnumbers = [8, 3, 91, 12, 7, 44, 2]\n# Sort the list in ascending order\nsorted_numbers = sorted(numbers)\n# Calculate the median\nmedian = statistics.median(sorted_numbers)\n# Return the result in a JSON format\nreturn", "retry_text": "", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "300cc2c42f91bf35", "model": "llama3.2-1b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "import json\nimport statistics\n\n# Define the list of numbers\nnumbers = [8, 3, 91, 12, 7, 44, 2]\n\n# Sort the list in ascending order\nnumbers.sort()\n\n# Calculate the median\nmedian = statistics.median(numbers)\n\n# Convert the result to a JSON string\nresult = json.dumps(", "retry_text": "web_search(query=\"statistics.median\", url=\"https://www.example.com\")", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "d35c286e3b43a3bb", "model": "llama3.2-1b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "import json\nimport statistics\n\n# Define the function calls\nfunction_calls = [", "retry_text": "import json\n\ndef web_search(query, url):\n    # This is a mock web_search function\n    print(f\"Searching for", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "31ea3bc4cfb9df5d", "model": "llama3.2-1b", "prompt_class": "no_tool", "prompt": "What is the difference between a list and a tuple in Python?", "text": "import json\n\n# List and tuple are both mutable, meaning they can be changed after creation.\n# They are also both ordered collections, meaning they maintain the order in which elements were added.\n\n# Create a list\nmy_list = [1, 2, 3, 4, 5]\n\n# Create a tuple\nmy_tuple = (1, 2, 3, 4, 5)\n\n# Modifying a list\nmy_list.append(6)\nprint(my_list)  # Output: [1, 2, 3, 4, 5, 6]\n\n# Modifying a tuple\nmy_tuple = (1, 2, 3, 4, 5)\nmy_tuple.append(6)\nprint(my_tuple)  # Output: (1, 2, 3, 4, 5, 6)\n\n# Accessing elements in a list\nprint(my_list[0])  # Output: 1\n\n# Accessing elements in a tuple\nprint(my_tuple[0])  # Output: 1\n\n# Modifying a list\nmy_list.insert(1, 7)\nprint(my_list)  # Output: [1, 7, 2, 3, 4, 5]\n\n# Modifying a tuple\nmy_tuple.insert(1, 7)\nprint(my_tuple)  # Output: (1, 7, 2, 3, 4, 5)\n\n# Creating a list and a tuple\nmy_list2 = []\nmy_tuple2 = ()\n\n# Modifying a list\nmy_list2.append(8)\nprint(my_list2)  # Output: [1, 2, 3, 4, 5, 6, 8]\n\n# Modifying a tuple\nmy_tuple2 = (1, 2, 3, 4, 5)\nmy_tuple2.append(8)\nprint(my_tuple2)  # Output: (1, 2, 3, 4, 5, 8)", "retry_text": "", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "2c701ec6cbbad60f", "model": "llama3.2-1b", "prompt_class": "format", "prompt": "Describe your approach to reviewing a pull request, as a list.", "text": "import json\n\n# Define the functions as described in the prompt\ndef web_search(query, url):\n    return", "retry_text": "import web_search\n\nweb_search(\"example query\", \"https://www.example.com\")", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "17930edef7c0a391", "model": "llama3.2-1b", "prompt_class": "non_english", "prompt": "Explica en una frase que es el sobreajuste.", "text": "El sobreajuste se refiere a un problema en el que dos o más variables tienen valores equivalentes o cercanos, lo que puede provocar errores en el ejecución de una programación o comprensión de un lenguaje. Este sobreajuste puede ocurrir en diversas situaciones, como en la programación con variables dinámicas, en la resolución de ecuaciones cuadráticas, o en la interpretación de lenguajes de programación.", "retry_text": "", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "4fdf373b03b9674c", "model": "qwen3-0.6b", "prompt_class": "needs_tool_web", "prompt": "Who won the most recent Formula 1 race?", "text": "The most recent Formula 1 race was held in **Monaco** in the 2023 season, and the winner was **Lohman**.", "retry_text": "The most recent Formula 1 race was held in **Monaco** in the 2023 season, and the winner was **Lohman**.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "2f93cfb96347d03b", "model": "qwen3-0.6b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "Sorted list: [2, 3, 7, 8, 12, 44, 91]  \n**Median**: 8", "retry_text": "Sorted list: [2, 3, 7, 8, 12, 44, 91]  \n**Median**: 8", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ffe902b3a77da3a3", "model": "qwen3-0.6b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "Sorted list: [2, 3, 7, 8, 12, 44, 91]  \nMedian: 8  \n\nThe median is the middle number when the list has an odd number of elements.", "retry_text": "Sorted list: [2, 3, 7, 8, 12, 44, 91]  \nMedian: 8  \n\nThe median is the middle number when the list has an odd number of elements.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "43700884f9e483bb", "model": "qwen3-0.6b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "The sorted list is [2, 3, 7, 8, 12, 44, 91]. The median is 8.", "retry_text": "The sorted list is [2, 3, 7, 8, 12, 44, 91]. The median is 8.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "bb267a38ffd74bb8", "model": "qwen3-0.6b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "The sorted list is [2, 3, 7, 8, 12, 44, 91], and the median is 12.", "retry_text": "The sorted list is [2, 3, 7, 8, 12, 44, 91], and the median is 8.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "b8428bd477953912", "model": "qwen3-0.6b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "Sorted list: [2, 3, 7, 8, 12, 44, 91]  \nMedian: 7", "retry_text": "Sorted list: [2, 3, 7, 8, 12, 44, 91]  \nMedian: 8", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a6257c91990092ea", "model": "qwen3-0.6b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "Sorted list: 2, 3, 7, 8, 12, 44, 91  \nMedian: 8 (the middle value when the list has an odd number of elements)", "retry_text": "Sorted list: 2, 3, 7, 8, 12, 44, 91  \nMedian: 8", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "8ca73b260971a970", "model": "qwen3-0.6b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "Sorted list: 2, 3, 7, 8, 12, 44, 91  \nMedian: 8", "retry_text": "Sorted list: 2, 3, 7, 8, 12, 44, 91  \nMedian: 8", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "2cf3c930c6256595", "model": "qwen3-0.6b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "Sorted list: [2, 3, 7, 8, 12, 44, 91]  \nMedian: 12  \n\nThe median is 12.", "retry_text": "Sorted list: [2, 3, 7, 8, 12, 44, 91]  \nMedian: 12  \n\nThe median is 12.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "c094102b1a6f9645", "model": "qwen3-0.6b", "prompt_class": "needs_tool_python", "prompt": "Calculate the standard deviation of 4, 8, 15, 16, 23, 42.", "text": "The standard deviation of the dataset 4, 8, 15, 16, 23, and 42 is approximately 12.32. \n\n**Step-by-Step Explanation:**\n1. **Calculate the Mean:**  \n   Sum of the numbers = 108, divided by 6 (count) = **18**.  \n2. **Calculate Squared Differences:**  \n   (4-18)² = 196, (8-18)² = 100, (15-18)² = 9, (16-18)² = 4, (23-18)² = 25, (42-18)² = 576.  \n   Sum of squared differences = **910**.  \n3. **Compute Variance:**  \n   Variance = 910 / 6 ≈ **151.666...**.  \n4. **Take Square Root:**  \n   Standard Deviation ≈ √151.666... ≈ **12.32**.  \n\n**Answer:** The standard deviation is approximately **12.32**.", "retry_text": "The standard deviation of the dataset 4, 8, 15, 16, 23, and 42 is approximately **12.32**. \n\n**Step-by-Step Explanation:**\n1. **Calculate the Mean:**  \n   Sum of the numbers = 108, divided by 6 = **18**.  \n2. **Calculate Squared Differences:**  \n   (4-18)² = 196, (8-18)² = 100, (15-18)² = 9, (16-18)² = 4, (23-18)² = 25, (42-18)² = 576.  \n   Sum of squared differences = **910**.  \n3. **Compute Variance:**  \n   Variance = 910 / 6 ≈ **151.666...**.  \n4. **Take Square Root:**  \n   Standard Deviation ≈ √151.666... ≈ **12.32**.  \n\n**Answer:** The standard deviation is approximately **12.32**.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "82ad4f08c9443c13", "model": "qwen3-0.6b", "prompt_class": "needs_tool_python", "prompt": "Calculate the standard deviation of 4, 8, 15, 16, 23, 42.", "text": "The standard deviation of the numbers 4, 8, 15, 16, 23, and 42 is approximately 12.32.\n\n**Step-by-Step Explanation:**\n1. **Calculate the mean:**  \n   $ \\frac{4 + 8 + 15 + 16 + 23 + 42}{6} = \\frac{108}{6} = 18 $.\n\n2. **Find squared deviations from the mean:**  \n   $ (4 - 18)^2 = 196 $, $ (8 - 18)^2 = 100 $, $ (15 - 18)^2 = 9 $, $ (16 - 18)^2 = 4 $, $ (23 - 18)^2 = 25 $, $ (42 - 18)^2 = 576 $.\n\n3. **Sum squared deviations:**  \n   $ 196 + 100 + 9 + 4 + 25 + 576 = 910 $.\n\n4. **Calculate variance:**  \n   $ \\frac{910}{6} \\approx 151.67 $.\n\n5. **Take the square root of variance:**  \n   $ \\sqrt{151.67} \\approx 12.32 $.\n\n**Answer:** The standard deviation is approximately 12.32.", "retry_text": "The standard deviation of the numbers 4, 8, 15, 16, 23, and 42 is approximately 12.32.\n\n**Step-by-Step Explanation:**\n1. **Calculate the mean:**  \n   $ \\frac{4 + 8 + 15 + 16 + 23 + 42}{6} = \\frac{108}{6} = 18 $.\n\n2. **Find squared deviations from the mean:**  \n   $ (4 - 18)^2 = 196 $, $ (8 - 18)^2 = 100 $, $ (15 - 18)^2 = 9 $, $ (16 - 18)^2 = 4 $, $ (23 - 18)^2 = 25 $, $ (42 - 18)^2 = 576 $.\n\n3. **Sum squared deviations:**  \n   $ 196 + 100 + 9 + 4 + 25 + 576 = 910 $.\n\n4. **Calculate variance:**  \n   $ \\frac{910}{6} \\approx 151.67 $.\n\n5. **Take the square root of variance:**  \n   $ \\sqrt{151.67} \\approx 12.32 $.\n\n**Answer:** The standard deviation is approximately 12.32.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "434d3a6198ee764a", "model": "qwen3-0.6b", "prompt_class": "needs_tool_python", "prompt": "Calculate the standard deviation of 4, 8, 15, 16, 23, 42.", "text": "The standard deviation of the dataset 4, 8, 15, 16, 23, 42 is approximately 12.32. \n\n**Answer:** The standard deviation is approximately 12.32.", "retry_text": "The standard deviation of the dataset 4, 8, 15, 16, 23, and 42 is approximately **12.32**. \n\n**Answer:** The standard deviation is approximately 12.32.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "9f22d6beb4d34f32", "model": "qwen3-0.6b", "prompt_class": "needs_tool_python", "prompt": "Calculate the standard deviation of 4, 8, 15, 16, 23, 42.", "text": "The standard deviation of the dataset [4, 8, 15, 16, 23, 42] is approximately 12.32. \n\n**Steps:**\n1. **Calculate the mean**: (4 + 8 + 15 + 16 + 23 + 42) / 6 = 108 / 6 = 18.\n2. **Compute squared differences from the mean**: \n   - (4-18)² = 196, (8-18)² = 100, (15-18)² = 9, (16-18)² = 4, (23-18)² = 25, (42-18)² = 576.\n3. **Sum squared differences**: 196 + 100 + 9 + 4 + 25 + 576 = 910.\n4. **Divide by count (6)**: 910 / 6 ≈ 151.666.\n5. **Take square root**: √(151.666...) ≈ 12.32.", "retry_text": "The standard deviation of the dataset [4, 8, 15, 16, 23, 42] is approximately **12.32**.\n\n**Steps:**\n1. **Mean**: (4 + 8 + 15 + 16 + 23 + 42) / 6 = 108 / 6 = 18.\n2. **Squared differences**: (4-18)² = 196, (8-18)² = 100, (15-18)² = 9, (16-18)² = 4, (23-18)² = 25, (42-18)² = 576.\n3. **Sum of squared differences**: 196 + 100 + 9 + 4 + 25 + 576 = 910.\n4. **Divide by count**: 910 / 6 ≈ 151.666.\n5. **Square root**: √(151.666...) ≈ 12.32.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "b612ea05e2b54573", "model": "qwen3-0.6b", "prompt_class": "needs_tool_python", "prompt": "How many primes are there below 10000?", "text": "The number of primes below 10000 is **1684**. This value is derived using the Sieve of Eratosthenes algorithm, which efficiently marks non-prime numbers by eliminating multiples of prime numbers, ensuring accuracy without unnecessary computations.", "retry_text": "The number of primes below 10,000 is **1684**. This value is derived using the Sieve of Eratosthenes algorithm, which efficiently marks non-prime numbers by eliminating their multiples, ensuring accuracy without unnecessary computations.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "eead70908f20af93", "model": "qwen3-0.6b", "prompt_class": "needs_tool_python", "prompt": "Work out the compound interest on 5000 at 3.5% for 12 years.", "text": "The compound interest on 5000 at 3.5% for 12 years can be calculated using the formula:\n\n$$ A = P \\times (1 + r)^t $$\n\nWhere:\n- $ P = 5000 $ (principal)\n- $ r = 3.5\\% = 0.035 $\n- $ t = 12 $ (time in years)\n\nFirst, compute $ (1 + 0.035)^{12} \\approx 1.4533 $. Then, the total amount $ A $ is:\n\n$$ 5000 \\times 1.4533 \\approx 7266.5 $$\n\nThe compound interest is:\n\n$$ 7266.5 - 5000 = 2266.5 $$\n\n**Answer:** The compound interest is approximately **$2266.50**.", "retry_text": "The compound interest on $5000 at 3.5% for 12 years is approximately **$2266.50**.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "98e626024169106e", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the capital of Japan?", "text": "The capital of Japan is Tokyo.", "retry_text": "The capital of Japan is Tokyo.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "5d703c2433475f71", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the capital of Japan?", "text": "The capital of Japan is Tokyo.", "retry_text": "The capital of Japan is Tokyo.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "90ad72f79996edab", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the capital of Japan?", "text": "The capital of Japan is Tokyo.", "retry_text": "The capital of Japan is Tokyo.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "114e4cb212318cc7", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the capital of Japan?", "text": "The capital of Japan is Tokyo.", "retry_text": "The capital of Japan is Tokyo.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "e1d5199f79924aa7", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the capital of Japan?", "text": "The capital of Japan is Tokyo.", "retry_text": "The capital of Japan is Tokyo.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "70d28a74a973cccc", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the capital of Japan?", "text": "The capital of Japan is Tokyo.", "retry_text": "The capital of Japan is Tokyo.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "455e9970d6a3ec42", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the capital of Japan?", "text": "The capital of Japan is Tokyo.", "retry_text": "The capital of Japan is Tokyo.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "82f8b8f25827ef49", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the capital of Japan?", "text": "The capital of Japan is Tokyo.", "retry_text": "The capital of Japan is Tokyo.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "8806bfc3a37d5c41", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Explain what a mutex is in two sentences.", "text": "A mutex (mutual exclusion object) is a synchronization mechanism that ensures only one thread can access a shared resource at a time, preventing race conditions. It works by locking the resource when a thread accesses it and unlocking it when another thread locks it, thus controlling concurrent access.", "retry_text": "A mutex (mutual exclusion object) is a synchronization mechanism that ensures only one thread can access a shared resource at a time, preventing race conditions by locking the resource when a thread accesses it and unlocking it when another thread locks it.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "0259d6328e80d33e", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Explain what a mutex is in two sentences.", "text": "A mutex (mutual exclusion object) is a synchronization primitive that ensures only one thread can access a shared resource at a time, preventing race conditions. It works by blocking other threads from accessing the resource until one is signaled, ensuring thread safety in concurrent programming.", "retry_text": "A mutex (mutual exclusion object) is a synchronization primitive that ensures only one thread can access a shared resource at a time, preventing race conditions. It works by blocking other threads from accessing the resource until one is signaled, ensuring thread safety in concurrent programming.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ee78b728ee51cc8c", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Explain what a mutex is in two sentences.", "text": "A mutex is a synchronization primitive used to protect shared resources from concurrent access. It ensures that only one thread at a time can access the resource, preventing race conditions.", "retry_text": "A mutex is a synchronization primitive used to protect shared resources from concurrent access. It ensures that only one thread at a time can access the resource, preventing race conditions.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "c931946998269339", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Explain what a mutex is in two sentences.", "text": "A mutex is a synchronization primitive used to protect shared resources from concurrent access, ensuring that only one thread can access them at a time. It achieves this by locking the resource when a thread needs to access it and unlocking it when the resource is free, preventing race conditions.", "retry_text": "A mutex is a synchronization primitive used to protect shared resources from concurrent access, ensuring that only one thread can access them at a time. It achieves this by locking the resource when a thread needs to access it and unlocking it when the resource is free, preventing race conditions.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "9a5a6581c9019f7f", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Explain what a mutex is in two sentences.", "text": "A mutex (mutual exclusive lock) is a synchronization primitive used to control shared resources, ensuring that multiple threads or processes can access them without interference. It prevents race conditions by enforcing exclusive access to the resource at a time.", "retry_text": "A mutex (mutual exclusive lock) is a synchronization primitive used to control shared resources, ensuring that multiple threads or processes can access them without interference. It prevents race conditions by enforcing exclusive access to the resource at a time.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a12923f96ffafa6e", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Explain what a mutex is in two sentences.", "text": "A mutex (mutual exclusion object) is a synchronization primitive used to protect shared resources by ensuring that only one thread can access them at a time. It prevents race conditions by allowing one thread to acquire the lock, and another thread must relinquish the lock before it can proceed.  \n\nIn two sentences:  \nA mutex is a synchronization object used to protect shared resources, ensuring that only one thread can access them at a time to avoid race conditions.", "retry_text": "A mutex (mutual exclusion object) is a synchronization primitive used to protect shared resources, ensuring that only one thread can access them at a time to avoid race conditions.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "e246df27c5212f5b", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Explain what a mutex is in two sentences.", "text": "A mutex (mutual exclusive lock) is a synchronization mechanism that protects shared resources from concurrent access by ensuring only one thread can access them at a time. It prevents race conditions by allowing only one thread to acquire the lock before others can do so.  \n\nAnother way: A mutex is a tool that ensures thread safety when multiple threads access shared resources, preventing interference and ensuring data consistency.", "retry_text": "A mutex (mutual exclusive lock) is a synchronization mechanism that protects shared resources from concurrent access by ensuring only one thread can access them at a time. It prevents race conditions by allowing only one thread to acquire the lock before others can do so.  \n\nAnother way: A mutex is a tool that ensures thread safety when multiple threads access shared resources, preventing interference and ensuring data consistency.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "55adde375e13956d", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Explain what a mutex is in two sentences.", "text": "A mutex is a synchronization primitive used to protect shared resources from concurrent access. It ensures that only one thread can access the resource at a time, preventing race conditions by controlling access to shared data.", "retry_text": "A mutex is a synchronization primitive used to protect shared resources from concurrent access. It ensures that only one thread can access the resource at a time, preventing race conditions by controlling access to shared data.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "f3ae8b3e9485a3f7", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "In plain English, what is a transformer model?", "text": "A transformer model is a type of neural network used in machine learning and natural language processing to handle long-range dependencies in sequences. Unlike traditional models like recurrent or RNNs, transformers use self-attention mechanisms to process the input in a way that allows the model to focus on relevant parts of the sequence, making them more efficient for tasks like language modeling and text generation.", "retry_text": "A transformer model is a type of neural network used in machine learning and natural language processing to handle long-range dependencies in sequences. Unlike traditional models like recurrent or RNNs, transformers use self-attention mechanisms to focus on relevant parts of the input, making them more efficient for tasks like language modeling and text generation.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "7501627df64f5901", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "In plain English, what is a transformer model?", "text": "A transformer model is a type of neural network designed to process long sequences of text, such as sentences or paragraphs, more efficiently than traditional models like RNNs or LSTMs. Here's a simple explanation:\n\n1. **Core Idea**: Transformers use self-attention mechanisms to focus on specific parts of the input at different times. Unlike traditional models, which process information sequentially, transformers can handle complex, long-term dependencies in text.\n\n2. **Key Advancement**: This allows models to understand and generate text even when the input is very long or contains complex relationships between words.\n\n3. **Common Use Cases**: Transformer models are widely used in tasks like language modeling (text generation, translation), summarization, and summarizing long documents.\n\nIn plain English, transformers enable powerful processing of text with greater efficiency and flexibility.", "retry_text": "A transformer model is a type of neural network designed to process long sequences of text, like sentences or paragraphs, more efficiently than traditional models like RNNs or LSTMs. It uses self-attention mechanisms to focus on specific parts of the input at different times, allowing it to handle complex, long-term dependencies in text.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "91e15fc0eb0e2627", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "In plain English, what is a transformer model?", "text": "A transformer model is a type of machine learning model used in **natural language processing (NLP)** to handle complex text and sequences. Here's a simple explanation:\n\n- **Purpose**: It's designed to process long sequences of text (like paragraphs or sentences) and understand context, which is useful for tasks like language translation, text generation, or summarization.\n- **Key Mechanism**: Unlike traditional models like RNNs or LSTMs, transformers use **self-attention** to dynamically determine which parts of the input to consider relevantly. This allows them to focus on the context and not just memorize the data.\n- **Comparison**: Unlike RNNs or LSTMs, which process data sequentially, transformers can handle long sequences more efficiently.\n\nIn short, a transformer model helps machines understand and generate text more effectively by focusing on context and long-term dependencies.", "retry_text": "A transformer model is a type of machine learning model used in **natural language processing (NLP)** to handle complex text and sequences. Here's a simple explanation:\n\n- **Purpose**: It's designed to process long sequences of text (like paragraphs or sentences) and understand context, which is useful for tasks like language translation, text generation, or summarization.  \n- **Key Mechanism**: Unlike traditional models like RNNs or LSTMs, transformers use **self-attention** to dynamically determine which parts of the input to consider relevantly. This allows them to focus on context and not just memorize the data.  \n- **Comparison**: Unlike RNNs or LSTMs, which process data sequentially, transformers can handle long sequences more efficiently.  \n\nIn short, a transformer model helps machines understand and generate text more effectively by focusing on context and long-term dependencies.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "5c344eaff3a31f90", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "In plain English, what is a transformer model?", "text": "A transformer model is a type of deep learning model used to process and generate text or other sequential data. Unlike traditional neural networks that use recurrent units (like RNNs), transformers use self-attention mechanisms to focus on relevant parts of the input. This makes them highly efficient for handling long, complex sequences of text, such as in machine translation or text generation tasks.", "retry_text": "A transformer model is a type of deep learning model used to process and generate text or other sequential data. Unlike traditional neural networks that use recurrent units (like RNNs), transformers use self-attention mechanisms to focus on relevant parts of the input. This makes them highly efficient for handling long, complex sequences of text, such as in machine translation or text generation.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "70de4d7453ea6e5f", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "In plain English, what is a transformer model?", "text": "A transformer model is a type of neural network used to process and understand text, like how humans do. It works by breaking down long sequences of text into smaller parts. Transformers use self-attention to focus on different parts of the text, allowing them to process information efficiently. This makes them great for tasks like language translation, text generation, and speech recognition.", "retry_text": "A transformer model is a type of neural network used in natural language processing to process and understand text, like how humans do. It works by breaking down long sequences of text into smaller parts, using self-attention to focus on different parts of the text, and allowing efficient processing of information. This makes it great for tasks like language translation, text generation, and speech recognition.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "fec9676d5b701fb7", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "In plain English, what is a transformer model?", "text": "A transformer model is a type of neural network used in **natural language processing** to help with tasks like language modeling, text generation, and machine translation. Here's a simple breakdown in plain English:\n\n- **Purpose**: Processes long sequences of text (like sentences or paragraphs) to understand and generate human-like content.\n- **Key Components**:  \n  - **Self-Attention**: Allows the model to weigh the importance of different parts of the input sequence.  \n  - **Encoder-Decoder Architecture**: Breaks the problem into encoding the input text and decoding it into output text.  \n- **Example**: Uses a transformer to translate text from one language to another, or generate text that aligns with a given prompt.\n\nIn short, a transformer model is a powerful tool for handling complex patterns in sequential data.", "retry_text": "A transformer model is a type of neural network used in natural language processing (NLP) to help with tasks like language modeling, text generation, and machine translation. In plain English:\n\n- It processes long sequences of text (e.g., sentences or paragraphs) to understand and generate human-like content.  \n- It uses **self-attention** to weigh the importance of different parts of the input, allowing it to learn patterns and dependencies between words.  \n- The model is structured as an **encoder-decoder** system, breaking the problem into encoding the input and decoding it into output.  \n- Examples include translating text between languages or generating creative text.  \n\nIn summary, a transformer model helps machines understand complex patterns in text by focusing on relationships between words rather than just individual elements.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "2d91756e22f14192", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "In plain English, what is a transformer model?", "text": "A transformer model is a type of neural network designed to process and understand long sequences of text, such as in natural language processing tasks. Unlike traditional RNNs or LSTMs, which process information sequentially, transformers use **self-attention** to dynamically highlight and focus on specific parts of the input, allowing them to handle complex, long-term dependencies in text more effectively. This makes them powerful for tasks like machine translation and text generation.", "retry_text": "A transformer model is a type of neural network used in natural language processing to handle long sequences of text. Unlike traditional RNNs or LSTMs (which process information sequentially), transformers use **self-attention** to dynamically highlight and focus on specific parts of the input, enabling them to process complex, long-term dependencies more effectively. This makes them powerful for tasks like machine translation and text generation.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "62f5e4af91d404bc", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "In plain English, what is a transformer model?", "text": "A transformer model is a type of neural network used in machine learning to process and transform input data efficiently. Unlike traditional models like RNNs or LSTMs, which process information sequentially, transformer models use **self-attention** to focus on specific parts of the input at different times. This allows them to learn complex patterns and dependencies in the data more effectively. The core idea is that the model can dynamically \"look\" at multiple parts of the input, enabling it to capture context and relationships in a way that traditional models might not.", "retry_text": "A transformer model is a type of neural network used in machine learning to process and transform input data efficiently. Unlike traditional models like RNNs or LSTMs, which process information sequentially, transformer models use **self-attention** to focus on specific parts of the input at different times. This allows them to learn complex patterns and dependencies in the data more effectively.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "29d4bd9e22b119a2", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is 2 + 2?", "text": "2 + 2 equals 4.", "retry_text": "2 + 2 equals 4.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "fccda71d724525a4", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is 2 + 2?", "text": "2 + 2 equals 4.", "retry_text": "2 + 2 equals 4.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "4368361fe083e78b", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is 2 + 2?", "text": "The result of 2 + 2 is 4.", "retry_text": "2 + 2 equals 4.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "d1d0c89509f6e2b1", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is 2 + 2?", "text": "2 + 2 equals 4.", "retry_text": "2 + 2 equals 4.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "553666af55035ee6", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is 2 + 2?", "text": "2 + 2 equals 4.", "retry_text": "2 + 2 equals 4.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "09332e87b52a264c", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is 2 + 2?", "text": "2 + 2 equals 4.", "retry_text": "2 + 2 equals 4.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "02910af1afb0fdeb", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is 2 + 2?", "text": "2 + 2 equals 4.", "retry_text": "2 + 2 equals 4.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "2ae2f8f18e6fb620", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is 2 + 2?", "text": "2 + 2 equals 4.", "retry_text": "2 + 2 equals 4.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "5ebee57bef9e7cc3", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Give me a short definition of overfitting.", "text": "Overfitting occurs when a model learns too much from the training data, leading to poor generalization and high error rates.", "retry_text": "Overfitting occurs when a model learns too much from the training data, leading to poor generalization and high error rates.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "653926096e613ab4", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Give me a short definition of overfitting.", "text": "Overfitting occurs when a model's complexity increases beyond the actual data, causing it to perform poorly on new, unseen data.", "retry_text": "Overfitting occurs when a model's complexity increases beyond the actual data, causing it to perform poorly on new, unseen data.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "55755ab88045d4cf", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Give me a short definition of overfitting.", "text": "Overfitting occurs when a model's training data is too similar to the test data, causing the model to learn the patterns present in the training data instead of the underlying generalization.", "retry_text": "Overfitting occurs when a model learns too much from the training data, making it perform well on the training set but poorly on the test set.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "407cb9a2b45d875e", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Give me a short definition of overfitting.", "text": "Overfitting occurs when a model learns the training data too well, leading to poor performance on new, unseen data.", "retry_text": "Overfitting occurs when a model learns the training data too well, leading to poor performance on new, unseen data.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "52fadbaf0c564664", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Give me a short definition of overfitting.", "text": "Overfitting occurs when a model learns too much from the training data, leading to poor generalization to new, unseen data.", "retry_text": "Overfitting occurs when a model learns too much from the training data, leading to poor generalization to new, unseen data.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "b3ea25024fa7c2e5", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Give me a short definition of overfitting.", "text": "Overfitting occurs when a model learns too much from the training data, leading to poor generalization and poor performance on new, unseen data.", "retry_text": "Overfitting occurs when a model learns too much from the training data, leading to poor generalization and poor performance on new, unseen data.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "e253899c9a49119c", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Give me a short definition of overfitting.", "text": "Overfitting refers to a situation in machine learning where a model learns the training data too well, making it perform poorly on new, unseen data.", "retry_text": "Overfitting refers to a situation in machine learning where a model learns the training data too well, making it perform poorly on new, unseen data.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "cff4f8216561e80a", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Give me a short definition of overfitting.", "text": "Overfitting occurs when a model's parameters become too specific to the training data, leading it to perform well on the training set but poorly on new, unseen data.", "retry_text": "Overfitting occurs when a model's parameters become too specific to the training data, leading it to perform well on the training set but poorly on new, unseen data.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "5f3fda191881eae4", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Who wrote Pride and Prejudice?", "text": "The novel *Pride and Prejudice* was written by **Jane Austen**. It was published in 1813.", "retry_text": "Jane Austen.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "53710dec1aba5d15", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the difference between a list and a tuple in Python?", "text": "A list and a tuple in Python are both used to store collections of elements, but they have key differences:\n\n### **Difference Between List and Tuple:**\n1. **Mutability:**\n   - **List:** Mutable, allows elements to be added, removed, or modified in-place.\n   - **Tuple:** Immutable, cannot be changed once created.\n\n2. **Fixed Length (Tuples):**\n   - **Tuple:** Has a fixed length, and all elements must be the same type.\n   - **List:** Can have variable lengths and elements of different types.\n\n3. **Uniqueness:**\n   - **Tuple:** Can be initialized with all elements, but lists can have any elements.\n   - **List:** Can be initialized with any elements, not just a single element.\n\n### **Example:**\n```python\n# List\nmy_list = [1, 2, 3]\nmy_list.append(4)  # List is mutable\nprint(my_list)   # Output: [1, 2, 3, 4]\n\n# Tuple\nmy_tuple = (1, 2, 3)\n# Cannot modify my_tuple\nprint(my_tuple)   # Output: (1, 2, 3)\n```\n\nThese differences make lists suitable for dynamic data and tuples for static, unchangeable data.", "retry_text": "A list and a tuple in Python differ in several key aspects:\n\n1. **Mutability:**  \n   - **List:** Mutable, allows elements to be added, removed, or modified in-place.  \n   - **Tuple:** Immutable, cannot be changed once created.\n\n2. **Fixed Length:**  \n   - **Tuple:** Has a fixed length and all elements must be of the same type.  \n   - **List:** Can have variable lengths and elements of different types.\n\n3. **Uniqueness:**  \n   - **Tuple:** Can be initialized with all elements, but lists can have any elements.  \n   - **List:** Can be initialized with any elements, not just a single element.\n\n**Example:**  \n```python\n# List\nmy_list = [1, 2, 3]\nmy_list.append(4)  # List is mutable\nprint(my_list)   # Output: [1, 2, 3, 4]\n\n# Tuple\nmy_tuple = (1, 2, 3)\n# Cannot modify my_tuple\nprint(my_tuple)   # Output: (1, 2, 3)\n```", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "b681f92c716c4f2a", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the difference between a list and a tuple in Python?", "text": "In Python, a **list** and a **tuple** are both used to store collections of elements, but they differ in several key aspects:\n\n### 1. **Immutable vs. Mutable**  \n- **Tuple**: Once created, its elements **cannot be changed**.  \n  Example: `(1, 2, 3)` → changes to `2, 3, 4` would break the tuple.  \n- **List**: Elements can be **modified** as long as you don't change them.  \n  Example: `[1, 2, 3]` → modifying `3` to `2` would update the list.\n\n### 2. **Element Type Consistency**  \n- **Tuple**: All elements must be of the **same type** (e.g., integers, strings).  \n- **List**: Elements can be of any type (e.g., integers, strings, etc).\n\n### 3. **Use Cases**  \n- **Tuple**: Best when the data **remains constant** or **cannot be modified**.  \n- **List**: Best when the data **can be updated or modified**.\n\n### Example\n```python\n# Tuple (immutable)  \nt1 = (1, 2, 3)  \nt2 = t1 + (4, 5)  # t2 becomes (1, 2, 3, 4, 5)\n\n# List (mutable)  \nl1 = [1, 2, 3]  \nl2 = l1 + [4, 5]  # l2 becomes (1, 2, 3, 4, 5)\n```\n\n### Summary\n- **Tuple**: Best for data that won’t change, or for immutable data.  \n- **List**: Best for dynamic, mutable data.", "retry_text": "In Python, a **list** and a **tuple** are both used to store collections of elements, but they differ in several key aspects:\n\n### 1. **Immutable vs. Mutable**  \n- **Tuple**: Once created, its elements **cannot be changed**.  \n  Example: `(1, 2, 3)` → changes to `2, 3, 4` would break the tuple.  \n- **List**: Elements can be **modified** as long as you don't change them.  \n  Example: `[1, 2, 3]` → modifying `3` to `2` would update the list.\n\n### 2. **Element Type Consistency**  \n- **Tuple**: All elements must be of the **same type** (e.g., integers, strings).  \n- **List**: Elements can be of any type (e.g., integers, strings, etc).\n\n### 3. **Use Cases**  \n- **Tuple**: Best for immutable data or when the data **remains constant**.  \n- **List**: Best for dynamic, mutable data where elements can be updated or modified.\n\n### Example\n```python\n# Tuple (immutable)  \nt1 = (1, 2, 3)  \nt2 = t1 + (4, 5)  # t2 becomes (1, 2, 3, 4, 5)\n\n# List (mutable)  \nl1 = [1, 2, 3]  \nl2 = l1 + [4, 5]  # l2 becomes (1, 2, 3, 4, 5)\n```\n\n### Summary  \n- **Tuple**: Best for data that won’t change or for immutable data.  \n- **List**: Best for dynamic, mutable data.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "2c69370b7ac37217", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the difference between a list and a tuple in Python?", "text": "In Python, the main difference between a list and a tuple is:\n\n1. **Mutability**:  \n   - A **list** is mutable, meaning you can add, remove, or modify elements.  \n   - A **tuple** is immutable, so once created, its elements cannot be changed.\n\n2. **Fixed Length**:  \n   - Tuples have a fixed length and cannot change their elements.  \n   - Lists can grow or shrink dynamically as elements are added or removed.\n\n3. **Access**:  \n   - Tuples are accessed using indexes, while lists use indices as well.  \n\n4. **Use Cases**:  \n   - Lists are ideal for dynamic data where changes are frequent.  \n   - Tuples are used when you need to store data that should not change, such as coordinates or a sequence of numbers.", "retry_text": "In Python, the main difference between a list and a tuple is:\n\n1. **Mutability**:  \n   - A **list** is mutable (you can change its elements).  \n   - A **tuple** is immutable (elements cannot be changed once created).\n\n2. **Fixed Length**:  \n   - Tuples have a fixed length and cannot change their elements.  \n   - Lists can grow or shrink dynamically.\n\n3. **Access**:  \n   - Both lists and tuples use indices for access, though tuples are accessed using indexes directly.\n\n4. **Use Cases**:  \n   - Lists are ideal for dynamic data with frequent changes.  \n   - Tuples are used for data that should remain constant, such as coordinates or a sequence of numbers.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a53e2fdb6fe27851", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the difference between a list and a tuple in Python?", "text": "In Python, a **list** and a **tuple** are both iterable data types, but they have key differences:\n\n1. **Modification**:  \n   - **Lists** can be modified (e.g., `list.append()`, `list.remove()`, `list.insert()`) after creation.  \n   - **Tuples** cannot be modified once created.\n\n2. **Length**:  \n   - **Tuples** have a **fixed length** and cannot be expanded or shrunk.  \n   - **Lists** can dynamically grow or shrink.\n\n3. **Hashing**:  \n   - Tuples can be used for hashing (e.g., `hash(tuple)`) and are often used for unique identifiers.  \n   - Lists can also have a `hash()` method.\n\n4. **Use Cases**:  \n   - **Lists** are mutable and used for dynamic data.  \n   - **Tuples** are immutable and used for fixed, unchanging data.\n\nFor example:  \n```python\nmy_list = [1, 2, 3]  \nmy_list.append(4)  # List changes  \nprint(my_list)  # Output: [1, 2, 3, 4]  \n\nmy_tuple = (1, 2, 3)  \nprint(my_tuple)  # Output: (1, 2, 3)  \n```", "retry_text": "In Python, a list and a tuple are both iterable data types but differ in their behaviors:\n\n1. **Modification**:  \n   - **Lists** can be modified (e.g., appended, removed, or inserted elements).  \n   - **Tuples** cannot be modified once created.\n\n2. **Length**:  \n   - **Tuples** have a **fixed length** and cannot be expanded or shrunk.\n\n3. **Hashing**:  \n   - Tuples can be used for hashing and are often used for unique identifiers.  \n   - Lists can also have a `hash()` method.\n\n4. **Use Cases**:  \n   - **Lists** are mutable and used for dynamic data.  \n   - **Tuples** are immutable and used for fixed, unchanging data.\n\nFor example:  \n```python\nmy_list = [1, 2, 3]  \nmy_list.append(4)  # List changes  \nprint(my_list)  # Output: [1, 2, 3, 4]  \n\nmy_tuple = (1, 2, 3)  \nprint(my_tuple)  # Output: (1, 2, 3)\n```", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "91fc9d1fb351d237", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the difference between a list and a tuple in Python?", "text": "In Python, a **list** and a **tuple** are both used to store collections of data, but they differ in several key aspects:\n\n1. **Mutability**:\n   - A **list** is **mutable**, meaning it can be modified (inserted, deleted, or updated elements).\n   - A **tuple** is **immutable**, so its elements cannot be changed once created.\n\n2. **Order**:\n   - Tuples preserve the order of elements, while lists also maintain the original order when elements are modified.\n\n3. **Use Cases**:\n   - **Lists** are used when you want to change the order or add elements dynamically.\n   - **Tuples** are used when you want to pass a collection of values as a group and not modify them.\n\nFor example:\n```python\nmy_list = [1, 2, 3]\nmy_list.append(4)  # List is mutable\nprint(my_list)  # Output: [1, 2, 3, 4]\n\nmy_tuple = (1, 2, 3)\nmy_tuple[0] = 5  # Tuple is immutable\nprint(my_tuple)  # Output: (1, 2, 3)\n```\n\nThe main difference lies in their immutability and how they handle changes.", "retry_text": "A list and a tuple in Python differ primarily in their **mutability** and **behavior**:\n\n1. **Mutability**:\n   - A **list** is **mutable** (can be modified, added, or removed elements).\n   - A **tuple** is **immutable** (cannot be changed once created).\n\n2. **Order**:\n   - Tuples preserve the original order of elements, while lists also maintain that order.\n\n3. **Use Cases**:\n   - **Lists** are used when you dynamically modify or add elements.\n   - **Tuples** are used when you want to pass a collection of values as a group and not modify them.\n\nTuples are immutable, making them ideal for static data structures, while lists allow dynamic modifications.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "616079ceb9f98220", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the difference between a list and a tuple in Python?", "text": "In Python, a **list** and a **tuple** are both used to store collections of items, but they differ in several key aspects:\n\n1. **Uniqueness**:  \n   - A **tuple** is **uniquely defined** and cannot be modified once created.  \n   - A **list** is **dynamic** and allows adding or removing elements at any time.  \n\n2. **Immutable**:  \n   - Tuples are **immutable** and cannot be changed after creation.  \n   - Lists are **mutable** and can be modified as needed.  \n\n3. **Order**:  \n   - Tuples maintain a **fixed order** of elements.  \n   - Lists allow dynamic reordering (e.g., appending or removing elements).  \n\n4. **Use Cases**:  \n   - **Tuples** are ideal for data that must remain unchanged, such as coordinates, constants, or fixed-order collections.  \n   - **Lists** are used for dynamic data, such as a list of names, a list of items to be added or removed in an application.  \n\nFor example:  \n- A tuple: `(\"apple\", \"banana\")`  \n- A list: `[\"banana\", \"apple\", \"cherry\"]`  \n\nThis distinction is crucial in Python when handling data structures and performance considerations.", "retry_text": "A **list** and a **tuple** are both used to store collections of items in Python, but they differ in several key aspects:  \n\n1. **Uniqueness**:  \n   - A **tuple** is **uniquely defined** and cannot be modified once created.  \n   - A **list** is **dynamic** and allows adding or removing elements at any time.  \n\n2. **Immutable**:  \n   - Tuples are **immutable** and cannot be changed after creation.  \n   - Lists are **mutable** and can be modified as needed.  \n\n3. **Order**:  \n   - Tuples maintain a **fixed order** of elements.  \n   - Lists allow dynamic reordering (e.g., appending or removing elements).  \n\n4. **Use Cases**:  \n   - **Tuples** are ideal for data that must remain unchanged, such as coordinates, constants, or fixed-order collections.  \n   - **Lists** are used for dynamic data, such as a list of names, a list of items to be added or removed in an application.  \n\nFor example:  \n- A tuple: `(\"apple\", \"banana\")`  \n- A list: `[\"banana\", \"apple\", \"cherry\"]`  \n\nThis distinction is crucial in Python when handling data structures and performance considerations.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "2078c91883c4a353", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the difference between a list and a tuple in Python?", "text": "In Python, a **list** and a **tuple** are both iterable, but they differ in several key aspects:\n\n### 1. **Immutable vs. Mutable**\n- A **list** is **mutable**, meaning it can be changed or modified.\n- A **tuple** is **immutable**, so changes to its elements require creating a new tuple.\n\n### 2. **Dynamic vs. Static**\n- A **list** is dynamic—its elements can be added or removed, and its length can grow.\n- A **tuple** is static—its elements cannot be changed once created.\n\n### 3. **Use Cases**\n- **Lists** are ideal for data that can be dynamically added or modified (e.g., a list of coordinates, a list of names).\n- **Tuples** are used for data that cannot be modified (e.g., coordinates, a fixed set of values).\n\n### Example\n```python\n# List\nmy_list = [1, 2, 3]\nmy_list[0] = 4  # Changes the first element\nmy_list.append(4)  # Adds an element\n\n# Tuple\nmy_tuple = (1, 2, 3)\nmy_tuple[0] = 5  # Cannot change, so a new tuple is created\n```\n\nThis distinction is fundamental in Python for handling different types of data efficiently.", "retry_text": "In Python, a **list** and a **tuple** are both iterable, but they differ in several key aspects:\n\n1. **Immutable vs. Mutable**:\n   - A **list** is **mutable** (can be changed or modified).\n   - A **tuple** is **immutable** (cannot be changed once created).\n\n2. **Dynamic vs. Static**:\n   - A **list** is **dynamic** (elements can be added, removed, or modified).\n   - A **tuple** is **static** (elements cannot be changed).\n\n3. **Use Cases**:\n   - **Lists** are ideal for data that can be dynamically added or modified (e.g., a list of coordinates, a list of names).\n   - **Tuples** are used for data that cannot be modified (e.g., coordinates, a fixed set of values).\n\nExample:\n```python\n# List\nmy_list = [1, 2, 3]\nmy_list[0] = 4  # Changes the first element\nmy_list.append(4)  # Adds an element\n\n# Tuple\nmy_tuple = (1, 2, 3)\nmy_tuple[0] = 5  # Cannot change, so a new tuple is created\n```\n\nThis distinction is fundamental in Python for handling different types of data efficiently.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "fd3dec4b84816b0c", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What is the difference between a list and a tuple in Python?", "text": "In Python, a **list** and a **tuple** are both used to store collections of items, but they differ in several key aspects:\n\n### 1. **Mutability**:\n   - **List**: Mutable. You can change its elements after creation (e.g., `list.append()` or `list.remove()`) to modify the contents.\n   - **Tuple**: Immutable. Once created, its elements cannot be changed, and any attempt to modify a tuple element will throw an error.\n\n### 2. **Structure**:\n   - **List**: Elements are added in **order**, and elements can be inserted or removed at any position.\n   - **Tuple**: Elements are added in **order**, and elements cannot be inserted or removed, making it static.\n\n### 3. **Comparison**:\n   - When comparing two lists or tuples, they are compared element-wise. If a list has elements with different lengths, the longer list is considered \"greater\" in a tuple-wise comparison.\n   - When comparing a list and a tuple, they are compared as a whole, and if the lists or tuples have different lengths, the one with more elements is considered greater.\n\n### 4. **Use Cases**:\n   - **List**: For dynamic data that can be modified (e.g., a list of people, a list of prices).\n   - **Tuple**: For static data that cannot be changed (e.g., a list of coordinates or a fixed set of values).\n\n### Example:\n```python\nmy_list = [1, 2, 3]\nmy_list.append(4)\nprint(my_list)  # Output: [1, 2, 3, 4]\n\nmy_tuple = (1, 2, 3)\nprint(my_tuple)  # Output: (1, 2, 3)\n```\n\nThis shows how the differences in mutability and structure affect their usage.", "retry_text": "In Python, a **list** and a **tuple** are both used to store collections of items, but they differ in several key aspects:\n\n### 1. **Mutability**:\n   - **List**: Mutable. You can change its elements after creation (e.g., `list.append()` or `list.remove()`) to modify the contents.\n   - **Tuple**: Immutable. Once created, its elements cannot be changed, and any attempt to modify a tuple element will throw an error.\n\n### 2. **Structure**:\n   - **List**: Elements are added in **order**, and elements can be inserted or removed at any position.\n   - **Tuple**: Elements are added in **order**, and elements cannot be inserted or removed, making it static.\n\n### 3. **Comparison**:\n   - When comparing lists or tuples, they are compared element-wise. If a list has elements with different lengths, the longer list is considered \"greater\" in a tuple-wise comparison.\n   - When comparing a list and a tuple, they are compared as a whole, and if the lists or tuples have different lengths, the one with more elements is considered greater.\n\n### 4. **Use Cases**:\n   - **List**: For dynamic data that can be modified (e.g., a list of people, a list of prices).\n   - **Tuple**: For static data that cannot be changed (e.g., a list of coordinates or a fixed set of values).\n\n### Example:\n```python\nmy_list = [1, 2, 3]\nmy_list.append(4)\nprint(my_list)  # Output: [1, 2, 3, 4]\n\nmy_tuple = (1, 2, 3)\nprint(my_tuple)  # Output: (1, 2, 3)\n```", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "41da97d37a1bfe04", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Summarise what LoRA does in one paragraph.", "text": "LoRA (Low-Rank Adaptation) is a technique used in training large language models to improve their performance during training by adding a small linear transformation to the model's weights. This method helps in optimizing the model's parameters more efficiently by leveraging the model's existing structure, allowing for faster training without significant computational overhead.", "retry_text": "LoRA (Low-Rank Adaptation) is a technique used in training large language models to enhance their performance by introducing a small, linear transformation to the model's weights. This method allows for more efficient training by leveraging the model's existing structure, reducing computational overhead and enabling faster parameter optimization.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "7d705f85c729f2d5", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Summarise what LoRA does in one paragraph.", "text": "LoRA stands for LoRALayer-wise, a technique used in training large language models (LLMs) to enhance performance and efficiency. It works by adding a small linear layer to the model's weights, enabling the model to learn the same representations without updating all the model's weights. This approach reduces computational overhead and speeds up training, particularly beneficial for very large models.", "retry_text": "LoRA stands for LoRALayer-wise, a technique used in training large language models (LLMs) to enhance performance and efficiency. It works by adding a small linear layer to the model's weights, enabling the model to learn the same representations without updating all the model's weights. This approach reduces computational overhead and speeds up training, particularly beneficial for very large models.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "e712132213910e8a", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Summarise what LoRA does in one paragraph.", "text": "LoRA (LoRALayer) is a technique used in training large language models to improve efficiency by adding a small set of additional parameters to the main model. These parameters are not updated during training, allowing the model to be trained more quickly and with less computational cost. This approach reduces the number of parameters in the model while enabling faster and more effective training.", "retry_text": "LoRA (LoRALayer) is a technique used in training large language models to enhance efficiency by adding a small set of additional parameters to the main model. These parameters are not updated during training, allowing the model to be trained more quickly and with less computational cost. This approach reduces the number of parameters while enabling faster and more effective training.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ed971da889aa00c5", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Summarise what LoRA does in one paragraph.", "text": "LoRA (LoRA Optimization) is a technique used to enhance the efficiency and performance of training transformer models, such as BERT or GPT, by adding a small set of specific parameters to the base model. These parameters are designed to improve the model's ability to learn, allowing it to train faster and use less computational resources. This optimization technique enables models to achieve better accuracy without requiring a significant increase in model size or complexity.", "retry_text": "LoRA (LoRA Optimization) is a technique designed to enhance the efficiency and performance of transformer models like BERT or GPT by adding a small set of specific parameters to the base model. These parameters are tailored to improve the model's ability to learn, enabling faster training and reduced computational resource usage. This optimization allows models to achieve better accuracy without requiring a substantial increase in model size or complexity.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "cb283f6344409751", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Summarise what LoRA does in one paragraph.", "text": "LoRA stands for Low-Rank Adaptation, a technique used to improve the performance of neural networks by adding a low-rank matrix to the model's weight matrix. This allows the model to adapt more efficiently and with less computational cost, making it suitable for tasks like language modeling where large-scale parameters are common.", "retry_text": "LoRA stands for Low-Rank Adaptation, a technique used to improve the performance of neural networks by adding a low-rank matrix to the model's weight matrix. This allows the model to adapt more efficiently and with less computational cost, making it suitable for tasks like language modeling where large-scale parameters are common.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "062b1b577a41b87a", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Summarise what LoRA does in one paragraph.", "text": "LoRA (LoRA Optimization) is a technique used in deep learning models to improve performance by adding additional layers to the attention mechanism in large language models (like BERT or GPT). This optimization allows the model to be more efficient in training, as it reduces the number of parameters while maintaining high accuracy, which is crucial for handling large-scale tasks.", "retry_text": "LoRA (LoRA Optimization) is a technique used in deep learning models to improve performance by adding additional layers to the attention mechanism in large language models (like BERT or GPT). This optimization allows the model to be more efficient in training, as it reduces the number of parameters while maintaining high accuracy, which is crucial for handling large-scale tasks.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "7b07271fa7175e04", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Summarise what LoRA does in one paragraph.", "text": "LoRA stands for LoRALayer and is a technique used in transformer models to enhance their performance. It allows for efficient fine-tuning of large models by modifying the weights in a way that improves the learning of the model without requiring full retraining. This method enables the model to learn better representations without significant computational overhead.", "retry_text": "LoRA stands for LoRALayer, a technique used in transformer models to enhance performance by allowing efficient fine-tuning of large models. It modifies the weights in a way that improves the model's learning without requiring full retraining, enabling the model to learn better representations without significant computational overhead.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "126f4d579ccec344", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Summarise what LoRA does in one paragraph.", "text": "LoRA (LoRALayerwise) is a technique that modifies the weights of a model's layers in a specific position, allowing for enhanced performance during training by fine-tuning the model's parameters. This approach enables efficient training by reducing computational overhead without sacrificing model accuracy, particularly in scenarios where the model's parameters need to be optimized for specific tasks.", "retry_text": "LoRA (LoRALayerwise) is a technique that modifies the weights of a model's layers in a specific position, enabling efficient training by fine-tuning the model's parameters. This approach reduces computational overhead while preserving model accuracy, particularly in scenarios where parameter optimization is needed.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "1d74acd146fcd5c3", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is interpreted and compiled. It works in a way where the interpreter runs the code directly, while the compiler (used in the Python standard library) is optimized for performance. The interpreter handles the runtime execution, and the compiler (in C) is used for code optimization and performance.", "retry_text": "Python is interpreted and compiled. It works in a way where the interpreter runs the code directly, while the compiler (used in the Python standard library) is optimized for performance. The interpreter handles the runtime execution, and the compiler (in C) is used for code optimization and performance.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a88f0061096b96f8", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is interpreted. The interpreter reads the code and executes it, which is a common way Python functions work.", "retry_text": "Python is interpreted. The interpreter reads and executes the code as it runs, which is a common way Python functions work.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "4eccea6235fa63b0", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is interpreted and can be compiled. When you run Python code, it uses a runtime interpreter (interpreted mode) to execute the code. For performance, Python can also be compiled into a C extension module, which is interpreted at runtime (compiled mode).", "retry_text": "Python is interpreted and can be compiled. When you run Python code, it uses a runtime interpreter (interpreted mode) to execute the code. For performance, Python can also be compiled into a C extension module, which is interpreted at runtime (compiled mode).", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "d76dead29a1db13f", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is interpreted by default. It reads and executes code directly from files, which is the most common way it's used. If you're using the Python interpreter (like in an environment), it runs the code as a script. If you're using a compiled version, like a bytecode file stored in an `.exe` or `.pyc` file, it's executed as a compiled program. Python is not a compiler—it's a language that can be interpreted or compiled at runtime.", "retry_text": "Python is interpreted by default. It reads and executes code directly from files, which is the most common way it's used. If you're using the Python interpreter (like in an environment), it runs the code as a script. If you're using a compiled version, like a bytecode file stored in an `.exe` or `.pyc` file, it's executed as a compiled program. Python is not a compiler—it's a language that can be interpreted or compiled at runtime.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "e455113745e55075", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is interpreted and compiled. Here's a breakdown:\n\n- **Interpreted**: Python code is executed directly in the interpreter, meaning it runs as a program without needing a compiler. This is the standard method used.\n- **Compiled**: Python can also be compiled into bytecode (like in CPython), which is then executed by the interpreter. This approach is more efficient for performance-critical applications.\n\nPython is a **general-purpose language** that allows developers to write code that can be interpreted or compiled, depending on the environment.", "retry_text": "Python is interpreted and compiled. Here's a breakdown:\n\n- **Interpreted**: Python code is executed directly in the interpreter, meaning it runs as a program without requiring a compiler (this is the standard method).\n- **Compiled**: Python can also be compiled into bytecode (like in CPython), which is then executed by the interpreter. This approach is more efficient for performance-critical applications.\n\nPython is a **general-purpose language** that allows developers to write code that can be interpreted or compiled, depending on the environment.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "614649027bb66232", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is interpreted or compiled depending on the context. When you run Python code, it is interpreted, meaning the code is executed directly without compilation. However, when you use compiled code (like using a library module), it's compiled into a module that can be executed quickly. Python uses a combination of interpreted and compiled code to achieve this.", "retry_text": "Python is interpreted or compiled depending on the context. When you run Python code, it is interpreted, meaning the code is executed directly without compilation. However, when you use compiled code (like using a library module), it is compiled into a module that can be executed quickly. Python uses a combination of interpreted and compiled code to achieve this.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "6814c671025ade74", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is interpreted, not compiled. The interpreter processes and executes the code directly, translating it into machine code for execution. This contrasts with compiled languages like C++ or Java, where the code is compiled into a single file (a \"linked list\") that can be run without recompilation.", "retry_text": "Python is interpreted, not compiled. The interpreter processes and executes the code directly, translating it into machine code for execution.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "bc5650447d001e2d", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is interpreted, not compiled. The interpreter runs the code directly, while the compiler compiles the code into machine code for execution.", "retry_text": "Python is interpreted, not compiled. The interpreter runs the code directly, while the compiler compiles the code into machine code for execution.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "6e6494aead8f1636", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What colour do you get mixing blue and yellow?", "text": "When you mix blue and yellow, you get green. This is because the primary colors in the traditional color wheel are red, blue, and yellow. Mixing two primary colors results in the third color in the spectrum.", "retry_text": "When you mix blue and yellow, you get green. This is because the primary colors in the traditional color wheel are red, blue, and yellow. Mixing two primary colors results in the third color in the spectrum.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "5a39099b612f2c75", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What colour do you get mixing blue and yellow?", "text": "When you mix blue and yellow, the resulting color is a complementary color pair. Blue and yellow are on the color wheel, and combining them produces a color that is not directly visible but complementary. The exact shade can vary depending on the mixing method (e.g., using a palette knife vs. a flat surface), but the result is typically a shade of blue or yellow in the opposite direction.", "retry_text": "When mixing blue and yellow, the resulting color is a complementary color pair. Blue and yellow are on the color wheel, and combining them produces a color that is not directly visible but complementary. The exact shade can vary depending on the mixing method, but the result is typically a shade of blue or yellow in the opposite direction.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "32dc34da4d27b3c4", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What colour do you get mixing blue and yellow?", "text": "The mixing of blue and yellow results in yellow. This is a basic color theory fact.", "retry_text": "The mixing of blue and yellow results in yellow. This is a basic color theory fact.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "67a73b74efc2bbd5", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What colour do you get mixing blue and yellow?", "text": "The color you get when mixing blue and yellow is yellow. This is because blue and yellow are complementary colors in the visible spectrum, and their combination results in the color yellow.", "retry_text": "The color you get when mixing blue and yellow is yellow.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "5643f5575982ee67", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What colour do you get mixing blue and yellow?", "text": "The correct answer is that mixing blue and yellow results in **orange**. This is based on the traditional color mixing principles of additive color theory.", "retry_text": "The correct answer is that mixing blue and yellow results in **orange**. This is based on the traditional color mixing principles of additive color theory.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "04784cbe5123000d", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What colour do you get mixing blue and yellow?", "text": "When mixing blue and yellow, the result is white. This is based on the principles of color theory, where the combination of primary colors (blue, yellow, and red) results in white. The exact color outcome depends on the intensity of the colors used, but typically, mixing equal parts blue and yellow yields the characteristic white.", "retry_text": "The color obtained when mixing blue and yellow is white. This is based on the principles of color theory, where primary colors (blue, yellow, and red) combine to produce the complementary color white.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a25aab4f1135fbba", "model": "qwen3-0.6b", "prompt_class": "no_tool", "prompt": "What colour do you get mixing blue and yellow?", "text": "When you mix blue and yellow, the resulting color is **yellow**. This is because the primary colors (red, blue, and yellow) combine to produce secondary colors when mixed. Mixing two primary colors results in the secondary color.", "retry_text": "The color when mixing blue and yellow is **yellow**. This is because the primary colors (red, blue, and yellow) combine to produce secondary colors when mixed.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "5bdabb7899cfd8f8", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Give me a numbered list of steps to fine tune a model.", "text": "Here’s a numbered list of steps to fine tune a model:\n\n1. **Data Preparation**: Collect and preprocess the dataset to ensure it’s suitable for the model. This includes handling missing data, scaling features, and cleaning the data.\n\n2. **Model Selection**: Choose an appropriate model (e.g., transformer, CNN, or RNN) that is suitable for the task and aligns with the data characteristics.\n\n3. **Hyperparameter Tuning**: Optimize the model’s hyperparameters (e.g., learning rate, batch size, regularization) to improve performance and convergence.\n\n4. **Initial Training**: Perform a full model training on the dataset to fine-tune the model to the task.\n\n5. **Fine-Tuning**: Adjust the model parameters based on the performance of the initial training phase to improve accuracy.\n\n6. **Evaluation**: Assess the model’s performance using appropriate metrics and compare it with the original model.\n\n7. **Deployment**: Once the model is optimized, deploy it for prediction or inference tasks.\n\nLet me know if you need further clarification!", "retry_text": "Here’s a numbered list of steps to fine tune a model:\n\n1. **Data Preparation**  \n2. **Model Selection**  \n3. **Hyperparameter Tuning**  \n4. **Initial Training**  \n5. **Fine-Tuning**  \n6. **Evaluation**  \n7. **Deployment**  \n\nLet me know if you need further clarification!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "20f4b59d62954aa6", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Give me a numbered list of steps to fine tune a model.", "text": "Here’s a numbered list of steps to fine-tune a model:\n\n1. **Pre-training on a large dataset**  \n   Train the model on a large, diverse dataset to obtain a strong initial representation of the task.\n\n2. **Fine-tuning on the target dataset**  \n   Adjust the model on a smaller, specific dataset to fine-tune its parameters for the task at hand.\n\n3. **Validation during training**  \n   Monitor the model’s performance using a validation set or metric to ensure it generalizes well and doesn’t overfit.\n\n4. **Hyperparameter tuning**  \n   Optimize the model’s hyperparameters (e.g., learning rate, batch size, regularization) to improve efficiency and performance.\n\n5. **Using techniques like data augmentation**  \n   Augment the training data during fine-tuning to increase the model’s robustness and reduce overfitting.", "retry_text": "Here’s a numbered list of steps to fine tune a model:\n\n1. **Pre-training on a large dataset**  \n   Train the model on a large, diverse dataset to obtain a strong initial representation of the task.  \n\n2. **Fine-tuning on the target dataset**  \n   Adjust the model on a smaller, specific dataset to fine-tune its parameters for the task at hand.  \n\n3. **Validation during training**  \n   Monitor the model’s performance using a validation set or metric to ensure it generalizes well and doesn’t overfit.  \n\n4. **Hyperparameter tuning**  \n   Optimize the model’s hyperparameters (e.g., learning rate, batch size, regularization) to improve efficiency and performance.  \n\n5. **Using techniques like data augmentation**  \n   Augment the training data during fine-tuning to increase the model’s robustness and reduce overfitting.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "51864a903d9fb2bd", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Give me a numbered list of steps to fine tune a model.", "text": "Here is a numbered list of steps to fine-tune a model:\n\n1. **Initial Model Selection**: Choose the appropriate model architecture that is suitable for the task and has a good performance baseline.  \n2. **Data Collection**: Gather and preprocess the training data to ensure it is suitable for the model.  \n3. **Model Initialization**: Start by initializing the model with the provided parameters.  \n4. **Training Phase**: Train the model on the dataset to improve its performance. Adjust the number of parameters if needed.  \n5. **Hyperparameter Tuning**: Conduct experiments to optimize hyperparameters (e.g., learning rate, batch size) to enhance model efficiency.  \n6. **Validation Phase**: Perform validation to check for overfitting and ensure the model generalizes well.  \n7. **Final Fine-tuning**: Fine-tune the model using additional data or in a new dataset to improve its performance on the target task.  \n\nLet me know if you need further clarification!", "retry_text": "Here is a numbered list of steps to fine tune a model:\n\n1. **Initial Model Selection**: Choose the appropriate model architecture that is suitable for the task and has a good performance baseline.  \n2. **Data Collection**: Gather and preprocess the training data to ensure it is suitable for the model.  \n3. **Model Initialization**: Start by initializing the model with the provided parameters.  \n4. **Training Phase**: Train the model on the dataset to improve its performance. Adjust the number of parameters if needed.  \n5. **Hyperparameter Tuning**: Conduct experiments to optimize hyperparameters (e.g., learning rate, batch size) to enhance model efficiency.  \n6. **Validation Phase**: Perform validation to check for overfitting and ensure the model generalizes well.  \n7. **Final Fine-tuning**: Fine-tune the model using additional data or in a new dataset to improve its performance on the target task.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a452cbc60878427b", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Give me a numbered list of steps to fine tune a model.", "text": "Here is a numbered list of steps to fine tune a model:\n\n1. **Data Preparation**:  \n   - Clean and prepare the dataset for fine-tuning.  \n   - Handle missing values and imbalance in the data.  \n\n2. **Model Selection**:  \n   - Choose a suitable model architecture (e.g., Transformer, CNN, or RNN) based on the task.  \n\n3. **Hyperparameters Tuning**:  \n   - Optimize model hyperparameters (e.g., learning rate, batch size, number of epochs) to balance training efficiency and model performance.  \n\n4. **Validation and Testing**:  \n   - Validate the model using a separate validation set to ensure it generalizes well.  \n\n5. **Fine-tuning Execution**:  \n   - Adjust the model parameters (e.g., weights, biases) to improve performance on the target dataset.  \n\n6. **Deployment**:  \n   - After fine-tuning, deploy the optimized model for real-world use.", "retry_text": "Here is a numbered list of steps to fine tune a model:\n\n1. **Data Preparation**  \n   - Clean and prepare the dataset for fine-tuning.  \n   - Handle missing values and imbalance in the data.  \n\n2. **Model Selection**  \n   - Choose a suitable model architecture (e.g., Transformer, CNN, or RNN) based on the task.  \n\n3. **Hyperparameters Tuning**  \n   - Optimize model hyperparameters (e.g., learning rate, batch size, number of epochs) to balance training efficiency and model performance.  \n\n4. **Validation and Testing**  \n   - Validate the model using a separate validation set to ensure it generalizes well.  \n\n5. **Fine-Tuning Execution**  \n   - Adjust the model parameters (e.g., weights, biases) to improve performance on the target dataset.  \n\n6. **Deployment**  \n   - After fine-tuning, deploy the optimized model for real-world use.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "6d58c791b1b4028a", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Give me a numbered list of steps to fine tune a model.", "text": "1. Load the pre-trained model.  \n2. Verify the model’s readiness for fine-tuning.  \n3. Prepare and clean the training data.  \n4. Perform the fine-tuning phase using the original dataset.  \n5. Evaluate the model’s performance after fine-tuning.", "retry_text": "1. Load the pre-trained model.  \n2. Verify the model’s readiness for fine-tuning.  \n3. Prepare and clean the training data.  \n4. Perform the fine-tuning phase using the original dataset.  \n5. Evaluate the model’s performance after fine-tuning.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "3f4913c6bbd0921f", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Give me a numbered list of steps to fine tune a model.", "text": "Here's a numbered list of steps to fine-tune a model:\n\n1. **Pre-Train the Model**: Select a pre-trained model architecture (e.g., BERT, GPT) and fine-tune it on a new dataset to adapt it to the specific task.\n\n2. **Data Preparation**: Gather and preprocess the dataset for the task, ensuring it matches the model's input requirements.\n\n3. **Model Selection**: Choose a suitable model architecture that can effectively adapt to the task and is compatible with the training environment.\n\n4. **Training the Model**: Train the pre-trained model on the new data to maximize its performance and parameter optimization.\n\n5. **Fine-Tuning**: Adjust the model's parameters based on the new data to improve its accuracy and efficiency.\n\n6. **Evaluation**: After fine-tuning, evaluate the model's performance using appropriate metrics to ensure it meets the desired results.", "retry_text": "Here's a numbered list of steps to fine tune a model:\n\n1. **Pre-Train the Model**: Select a pre-trained model architecture (e.g., BERT, GPT) and fine-tune it on a new dataset to adapt it to the specific task.\n\n2. **Data Preparation**: Gather and preprocess the dataset for the task, ensuring it matches the model's input requirements.\n\n3. **Model Selection**: Choose a suitable model architecture that can effectively adapt to the task and is compatible with the training environment.\n\n4. **Training the Model**: Train the pre-trained model on the new data to maximize its performance and parameter optimization.\n\n5. **Fine-Tuning**: Adjust the model's parameters based on the new data to improve its accuracy and efficiency.\n\n6. **Evaluation**: After fine-tuning, evaluate the model's performance using appropriate metrics to ensure it meets the desired results.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "fc1e25c5d4a98acd", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Outline, as a bulleted list, how to debug a memory leak.", "text": "- **Identify the source of the leak** by tracing where the memory is being allocated or released.  \n- **Check memory usage** using tools like Valgrind or heap analysis tools to detect unallocated memory.  \n- **Monitor memory consumption** with tools like `gmon -p` or `gmon -p -m` to track heap usage.  \n- **Analyze heap dumps** to identify objects that are not being freed.  \n- **Use the garbage collector** (GC) to determine if it's catching the leak.  \n- **Test with different environments** to validate the cause.  \n- **Use Valgrind** to pinpoint the exact location of the leak.", "retry_text": "- **Identify the source of the leak** by tracing where memory is allocated or released.  \n- **Check memory usage** using tools like Valgrind or heap analysis to detect unallocated memory.  \n- **Monitor memory consumption** with tools like `gmon -p` or `gmon -p -m` to track heap usage.  \n- **Analyze heap dumps** to identify objects that are not being freed.  \n- **Use the garbage collector** (GC) to determine if it's catching the leak.  \n- **Test with different environments** to validate the cause.  \n- **Use Valgrind** to pinpoint the exact location of the leak.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "66a9eeebd5692881", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Outline, as a bulleted list, how to debug a memory leak.", "text": "- **Check for memory leaks in code** (e.g., using tools like Valgrind or memory dumps).  \n- **Track allocated memory** (e.g., using `valgrind` or `gmonloc` to identify allocation/deallocation points).  \n- **Locate memory allocation and deallocation points** in the code or runtime environment.  \n- **Analyze heap and stack usage** to identify where the leak occurs.  \n- **Use Valgrind or other memory analysis tools** to report and resolve leaks.", "retry_text": "- **Check for memory leaks using Valgrind** (e.g., by running `valgrind --leak-test-cmd` or analyzing memory dumps).  \n- **Identify allocation/deallocation points** via tools like `gmonloc` or Valgrind's memory analysis tools.  \n- **Track allocated memory** using `valgrind` or heap tracing to pinpoint where leaks occur.  \n- **Analyze heap and stack usage** to determine the source of the memory issue.  \n- **Use Valgrind to report and resolve leaks**, ensuring proper cleanup and memory management.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "afac0fda9391e9b9", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Outline, as a bulleted list, how to debug a memory leak.", "text": "- **Identify the source of the memory leak** by inspecting the program's code for memory allocations and deallocations.  \n- **Use tools like Valgrind or LeakCheck** to detect leaks in the application's memory usage.  \n- **Track memory usage** by writing code that monitors allocations and deallocations.  \n- **Analyze memory allocation and deallocation points** to pinpoint where the leak occurs.  \n- **Check for unused variables or improper memory management** in the code.  \n- **Validate the solution** after fixing the leak to ensure no new leaks are introduced.", "retry_text": "- **Identify the source of the memory leak** by inspecting the program's code for memory allocations and deallocations.  \n- **Use tools like Valgrind or LeakCheck** to detect leaks in the application's memory usage.  \n- **Track memory usage** by writing code that monitors allocations and deallocations.  \n- **Analyze memory allocation and deallocation points** to pinpoint where the leak occurs.  \n- **Check for unused variables or improper memory management** in the code.  \n- **Validate the solution** after fixing the leak to ensure no new leaks are introduced.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "1a1a92f274563204", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Outline, as a bulleted list, how to debug a memory leak.", "text": "- **Identify the source of the leak**: Use tools like `valgrind` to detect memory leaks by analyzing memory usage.  \n- **Check allocated objects**: Look for memory allocations and ensure they are freed properly.  \n- **Use `valgrind` to detect leaks**: Run commands like `valgrind --leak-check=static` or `valgrind --leak-check=full` to find memory leaks.  \n- **Analyze memory usage**: Tools like `gmem` or `gprof` can help trace memory allocation and deallocation.  \n- **Use a debugger**: Set breakpoints in the code to inspect memory usage and identify leaks.  \n- **Implement memory management strategies**: Use smart pointers or proper deallocation in C/C++ to prevent leaks.  \n- **Review and verify code for leaks**: Ensure all memory allocations and frees are handled correctly.", "retry_text": "- **Identify the source of the leak**: Use tools like `valgrind` to detect memory leaks by analyzing memory usage.  \n- **Check allocated objects**: Ensure memory allocations and deallocations are handled properly.  \n- **Use `valgrind` to detect leaks**: Run commands like `valgrind --leak-check=static` or `valgrind --leak-check=full` to find memory leaks.  \n- **Analyze memory usage**: Tools like `gmem` or `gprof` can help trace memory allocation and deallocation.  \n- **Use a debugger**: Set breakpoints in the code to inspect memory usage and identify leaks.  \n- **Implement memory management strategies**: Use smart pointers or proper deallocation in C/C++ to prevent leaks.  \n- **Review and verify code for leaks**: Ensure all memory allocations and frees are handled correctly.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "d052ea50b3e08f1c", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Outline, as a bulleted list, how to debug a memory leak.", "text": "- **Identify the leak** by checking memory allocations and observing where the leak starts and ends.  \n- **Analyze the heap dump** to pinpoint the source of the leak.  \n- **Use tools like Valgrind or LeakCheck** to trace memory allocations and identify leaks.  \n- **Check if memory is being freed incorrectly** (e.g., not marking it as free).  \n- **Modify the code** to prevent memory leaks by ensuring proper allocation and freeing of memory.  \n- **Test the fix** by running the application and verifying that the leak is resolved.", "retry_text": "- **Identify the leak** by checking memory allocations and observing where the leak starts and ends.  \n- **Analyze the heap dump** to pinpoint the source of the leak.  \n- **Use tools like Valgrind or LeakCheck** to trace memory allocations and identify leaks.  \n- **Check if memory is being freed incorrectly** (e.g., not marking it as free).  \n- **Modify the code** to prevent memory leaks by ensuring proper allocation and freeing of memory.  \n- **Test the fix** by running the application and verifying that the leak is resolved.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ad9fb6b1ed4c0158", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Outline, as a bulleted list, how to debug a memory leak.", "text": "- **Identify the leak**: Use tools like Valgrind, gdb, or memory analysis tools to locate where memory is allocated and not freed.  \n- **Check allocation/deallocation points**: Trace program execution to ensure memory is properly allocated and released.  \n- **Verify memory usage**: Confirm that all memory allocations and frees are correctly managed.  \n- **Ensure proper memory management**: Avoid dangling pointers and ensure all memory is released before the program exits.", "retry_text": "- **Identify the leak**: Use tools like Valgrind, gdb, or memory analysis tools to locate where memory is allocated and not freed.  \n- **Check allocation/deallocation points**: Trace program execution to ensure memory is properly allocated and released.  \n- **Verify memory usage**: Confirm that all memory allocations and frees are correctly managed.  \n- **Ensure proper memory management**: Avoid dangling pointers and ensure all memory is released before the program exits.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "608c5d437fef5603", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Outline, as a bulleted list, how to debug a memory leak.", "text": "- **Identify the leak**: Use tools like Valgrind, Leakcheck, or your application's built-in memory analysis tools to detect memory leaks by checking heap usage or logs.  \n- **Analyze the problem**: Trace the code to identify where memory was allocated and released, and determine why a memory block was not freed.  \n- **Debug the code**: Modify the code to prevent memory leaks, such as updating allocation and release functions or ensuring proper memory management.  \n- **Test the application**: Re-run the application after fixes to ensure the leak is resolved and no other issues persist.", "retry_text": "- **Identify the leak**: Use tools like Valgrind, Leakcheck, or your application's built-in memory analysis tools to detect memory leaks by checking heap usage or logs.  \n- **Analyze the problem**: Trace the code to identify where memory was allocated and released, and determine why a memory block was not freed.  \n- **Debug the code**: Modify the code to prevent memory leaks, such as updating allocation and release functions or ensuring proper memory management.  \n- **Test the application**: Re-run the application after fixes to ensure the leak is resolved and no other issues persist.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "737a72c4c49ec6d6", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Outline, as a bulleted list, how to debug a memory leak.", "text": "- **Identify the Leak**: Use tools like `valgrind` or `gmem` to pinpoint where memory is being used incorrectly.  \n- **Check Heap Usage**: Ensure all memory allocations and frees are properly tracked using the `heapq` or `heapq.heappush` module.  \n- **Review Libraries**: Check if a library or framework (e.g., `heapq`, `heapq.heappush`) is not properly managing memory.  \n- **Use Profiling Tools**: Tools like `valgrind` or GDB can help track memory allocation and usage patterns.  \n- **Analyze Heap Usage**: Ensure heap memory is not exhausted by monitoring allocations and frees.  \n- **Check for Proper Freeing**: Verify that every memory allocation is correctly freed using the appropriate `free` or `delete` functions.  \n- **Use Debugging Tools**: Tools like GDB or `gmem` can help trace memory usage in the context of the program.  \n- **Review Logs**: Look for logs related to memory allocation and usage to identify potential issues.", "retry_text": "- **Identify the Leak**: Use tools like `valgrind` or `gmem` to pinpoint where memory is being used incorrectly.  \n- **Check Heap Usage**: Ensure all memory allocations and frees are properly tracked using the `heapq` or `heapq.heappush` module.  \n- **Review Libraries**: Check if a library or framework (e.g., `heapq`, `heapq.heappush`) is not properly managing memory.  \n- **Use Profiling Tools**: Tools like `valgrind` or GDB can help track memory allocation and usage patterns.  \n- **Analyze Heap Usage**: Ensure heap memory is not exhausted by monitoring allocations and frees.  \n- **Check for Proper Freeing**: Verify that every memory allocation is correctly freed using the appropriate `free` or `delete` functions.  \n- **Use Debugging Tools**: Tools like GDB or `gmem` can help trace memory usage in the context of the program.  \n- **Review Logs**: Look for logs related to memory allocation and usage to identify potential issues.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ca9f375a692cb8e9", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "List the first three things you would check if training loss is NaN.", "text": "If training loss is NaN, you should check the following three things:\n\n1. **Data Validation**: Ensure the input data is properly formatted and contains valid values. Missing or incomplete data might cause the loss function to fail.  \n2. **Model Initialization**: Verify that the model has been initialized correctly, including parameters and initializations. A uninitialized model might not learn effectively.  \n3. **Training Progress**: Confirm that the training process has started and that the model has been updated properly, even if the loss is NaN at intermediate steps.", "retry_text": "The first three things to check if training loss is NaN are:\n\n1. **Data Validation**: Ensure the input data is properly formatted and contains valid values. Missing or incomplete data might cause the loss function to fail.  \n2. **Model Initialization**: Verify that the model has been initialized correctly, including parameters and initializations. A uninitialized model might not learn effectively.  \n3. **Training Progress**: Confirm that the training process has started and that the model has been updated properly, even if the loss is NaN at intermediate steps.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "df064c12eb1268c4", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "List the first three things you would check if training loss is NaN.", "text": "If training loss is NaN, you should check the following three things:\n\n1. **Missing Values**: Check for any missing entries in your training data that might affect the loss function calculation.\n2. **Non-Numeric Values**: Ensure that the training data contains only numeric values, as non-numeric values (e.g., strings or special characters) can cause the loss to be undefined.\n3. **Invalid Inputs**: Verify that there are no invalid inputs during training, such as division by zero or malformed data entries.", "retry_text": "If training loss is NaN, you should check the following three things:\n\n1. **Missing Values**: Ensure the training data contains no missing entries that could affect the loss function calculation.  \n2. **Non-Numeric Values**: Verify that the training data contains only numeric values (e.g., integers, floats) instead of strings or special characters.  \n3. **Invalid Inputs**: Check for invalid inputs like division by zero, malformed data entries, or errors in the training process.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "83bd51ee6dbacc14", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "List the first three things you would check if training loss is NaN.", "text": "If training loss is NaN, the first three things to check are:  \n1. **Check for division by zero**: If the loss calculation involves division by zero (e.g., `loss = 1 / 0`), this could cause a NaN.  \n2. **Verify model updates**: Ensure the model has been trained with enough parameters and the optimizer has optimized the loss function.  \n3. **Confirm dataset completeness**: Ensure the dataset is fully available and does not contain missing or incomplete data.", "retry_text": "If training loss is NaN, the first three things to check are:  \n1. **Check for division by zero** (e.g., `loss = 1 / 0`).  \n2. **Verify model updates** (ensure the model has been trained with enough parameters and the optimizer has optimized the loss function).  \n3. **Confirm dataset completeness** (ensure the dataset is fully available and contains no missing or incomplete data).", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "e344d44a16689e8e", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "List the first three things you would check if training loss is NaN.", "text": "If training loss is NaN, the first three things to check are:  \n1. **Data Availability**: Ensure the training data is properly available and formatted.  \n2. **Optimizer Initialization**: Verify that the optimizer (e.g., Adam, SGD) is initialized correctly.  \n3. **Training Process**: Confirm that the training process has been executed for a sufficient number of epochs to allow convergence.", "retry_text": "If training loss is NaN, the first three things to check are:  \n1. **Data Availability**: Ensure the training data is properly available and formatted.  \n2. **Optimizer Initialization**: Verify that the optimizer (e.g., Adam, SGD) is initialized correctly.  \n3. **Training Process**: Confirm that the training process has been executed for a sufficient number of epochs to allow convergence.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a88294870a2c9f99", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "List the first three things you would check if training loss is NaN.", "text": "If training loss is NaN, the first three things to check are:\n\n1. **Check for division by zero** (e.g., dividing a number by zero in training loss calculations).  \n2. **Verify data type** (e.g., ensuring the input data type matches expected values).  \n3. **Check for missing or incomplete data** (e.g., missing training samples or values in the dataset).", "retry_text": "If training loss is NaN, the first three things to check are:\n\n1. **Check for division by zero** (e.g., dividing a number by zero in training loss calculations).  \n2. **Verify data type** (e.g., ensuring the input data type matches expected values).  \n3. **Check for missing or incomplete data** (e.g., missing training samples or values in the dataset).", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "fac14512075ca35c", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "List the first three things you would check if training loss is NaN.", "text": "If training loss is NaN, you should check the following three things:\n\n1. **Missing Values in the Training Data**: Ensure there are no missing entries in the training data that could cause issues with the loss function.\n2. **Model Configuration**: Verify that the model architecture is correctly defined and the weights and biases are initialized appropriately.\n3. **Optimizer Performance**: Check if the optimizer (e.g., Adam, SGD) is converging properly, which might lead to NaN values if it's not optimized effectively.", "retry_text": "List the first three things you would check if training loss is NaN:\n\n1. **Missing Values in the Training Data**: Ensure there are no missing entries in the training data that could cause issues with the loss function.  \n2. **Model Configuration**: Verify that the model architecture is correctly defined and the weights and biases are initialized appropriately.  \n3. **Optimizer Performance**: Check if the optimizer (e.g., Adam, SGD) is converging properly, which might lead to NaN values if it's not optimized effectively.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "3f23520b6e2895bd", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "List the first three things you would check if training loss is NaN.", "text": "If training loss is NaN, you should check the following:\n\n1. **Incomplete or Missing Training Data**: The training data may be incomplete or lack the necessary inputs for the model to compute the loss.\n2. **Incorrect Learning Rate**: A learning rate that is too large or too small could cause the loss to become undefined.\n3. **Optimizer Parameters**: The optimizer (e.g., Adam, RMSProp) may not be properly initialized or configured, leading to NaN values in the loss.", "retry_text": "Here are the first three things to check if training loss is NaN:\n\n1. **Incomplete or Missing Training Data**: The training data may be missing essential inputs required for the model to compute the loss.  \n2. **Incorrect Learning Rate**: A learning rate that is too large or too small could cause the loss to become undefined.  \n3. **Optimizer Parameters**: The optimizer (e.g., Adam, RMSProp) may not be properly initialized or configured, leading to NaN values in the loss.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ec5763efa2dbc665", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "List the first three things you would check if training loss is NaN.", "text": "If training loss is NaN, consider the following first three checks:\n\n1. **Check the loss value is a valid number** (e.g., not a NaN or invalid value).  \n2. **Verify the loss is being computed correctly** (e.g., inputs are properly fed, outputs are not empty or invalid).  \n3. **Confirm the model is using a valid loss function** (e.g., not a numeric type or unsupported function).", "retry_text": "If training loss is NaN, consider the following first three checks:\n\n1. **Check the loss value is a valid number** (e.g., not a NaN or invalid value).  \n2. **Verify the loss is being computed correctly** (e.g., inputs are properly fed, outputs are not empty or invalid).  \n3. **Confirm the model is using a valid loss function** (e.g., not a numeric type or unsupported function).", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "e7c0bb256910c0a1", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Walk me through setting up a Python virtualenv, step by step.", "text": "To set up a Python virtual environment step-by-step, follow these instructions:\n\n1. **Install Python**:  \n   Ensure you have Python installed. You can check this by typing `python -V` or `python3 -V` in your terminal. If not installed, download it from [https://www.python.org/downloads](https://www.python.org/downloads).\n\n2. **Create a Virtual Environment**:  \n   Use the command:  \n   ```bash\n   python -m venv <your_venv_name>\n   ```  \n   Replace `<your_venv_name>` with a name (e.g., `my_venv`). This creates a directory where your Python environment is stored. If the directory doesn't exist, the command will create it.\n\n3. **Activate the Virtual Environment**:  \n   - **Using the shell**:  \n     ```bash\n     source <your_venv_name>/bin/activate\n     ```  \n     This activates the environment and uses `pip` commands for package management.  \n   - **Using the terminal**:  \n     ```bash\n     activate <your_venv_name>\n     ```  \n     This also activates the environment.\n\n4. **Verify the Environment**:  \n   After activation, check if the environment is active by running:  \n   ```bash\n   pip --version\n   ```  \n   This confirms that the virtual environment is correctly set up.\n\n5. **Using the Environment**:  \n   Once the virtual environment is active, you can install packages using `pip install` or run scripts in the same directory.\n\n**Note**: If you want to use a custom path, specify it in the `venv` command. Also, if you need to create a new virtualenv, ensure that the directory exists before running the command.", "retry_text": "To set up a Python virtual environment, follow these steps:\n\n1. **Install Python**  \n   Ensure you have Python installed. You can check this by typing `python -V` or `python3 -V` in your terminal. If not installed, download it from [https://www.python.org/downloads](https://www.python.org/downloads).\n\n2. **Create a Virtual Environment**  \n   Use the command:  \n   ```bash\n   python -m venv <your_venv_name>\n   ```  \n   Replace `<your_venv_name>` with a name (e.g., `my_venv`). This creates a directory where your Python environment is stored. If the directory doesn't exist, the command will create it.\n\n3. **Activate the Virtual Environment**  \n   - **Using the shell**:  \n     ```bash\n     source <your_venv_name>/bin/activate\n     ```  \n     This activates the environment and uses `pip` commands for package management.  \n   - **Using the terminal**:  \n     ```bash\n     activate <your_venv_name>\n     ```  \n     This also activates the environment.\n\n4. **Verify the Environment**  \n   After activation, check if the environment is active by running:  \n   ```bash\n   pip --version\n   ```  \n   This confirms that the virtual environment is correctly set up.\n\n5. **Using the Environment**  \n   Once the virtual environment is active, you can install packages using `pip install` or run scripts in the same directory.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "6a58c90d0c6b8c6a", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Walk me through setting up a Python virtualenv, step by step.", "text": "To set up a Python virtualenv step by step, follow this process:\n\n1. **Create the Virtualenv**:  \n   Use the command `python -m venv <env_name>` to create a new virtual environment. For example:  \n   ```bash  \n   python -m venv myvenv  \n   ```  \n   This creates a folder named `myvenv` with the Python interpreter and packages for your project.\n\n2. **Activate the Virtualenv**:  \n   After creating the environment, activate it using:  \n   ```bash  \n   source myvenv/bin/activate  \n   ```  \n   This makes the virtualenv available for executing code within the current working directory.\n\n3. **Install Dependencies**:  \n   If your project has dependencies, use `pip install` within the virtualenv:  \n   ```bash  \n   pip install -r requirements.txt  \n   ```  \n   This installs all necessary packages for your project.\n\n4. **Use the Virtualenv**:  \n   Now, you can run code directly in the virtualenv:  \n   ```bash  \n   python your_script.py  \n   ```  \n   This avoids conflicts with other Python environments.\n\n5. **Verify**:  \n   Check the path to ensure the virtualenv is active:  \n   ```bash  \n   python -c \"import sys; print(sys.path)\"  \n   ```  \n   This will show where the virtualenv's packages are available.\n\nThis process ensures isolated, secure environments for project dependencies.", "retry_text": "To set up a Python virtualenv step by step:\n\n1. **Create the Virtualenv**:  \n   Use the command `python -m venv <env_name>` to create a new virtual environment. For example:  \n   ```bash  \n   python -m venv myvenv  \n   ```  \n\n2. **Activate the Virtualenv**:  \n   After creating the environment, activate it using:  \n   ```bash  \n   source myvenv/bin/activate  \n   ```  \n\n3. **Install Dependencies**:  \n   If your project has dependencies, use `pip install` within the virtualenv:  \n   ```bash  \n   pip install -r requirements.txt  \n   ```  \n\n4. **Use the Virtualenv**:  \n   Now, you can run code directly in the virtualenv:  \n   ```bash  \n   python your_script.py  \n   ```  \n\n5. **Verify**:  \n   Check the path to ensure the virtualenv is active:  \n   ```bash  \n   python -c \"import sys; print(sys.path)\"  \n   ```  \n\nThis process ensures isolated, secure environments for project dependencies.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "6012f002d10c0090", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Walk me through setting up a Python virtualenv, step by step.", "text": "To set up a Python virtualenv step by step, follow this process:\n\n1. **Install Python**:  \n   If not already installed, install Python using your system's package manager (e.g., `apt-get` on Ubuntu/Debian or `yum` on CentOS). Example command:  \n   ```bash\n   # For Ubuntu/Debian\n   sudo apt-get update\n   sudo apt-get install -y python3\n   # For CentOS\n   sudo yum install -y python3\n   ```\n\n2. **Create a Virtual Environment**:  \n   Use `venv` or `virtualenv` to create a new environment. The command is:  \n   ```bash\n   python3 -m venv my_venv\n   ```\n\n   This creates a directory named `my_venv` with Python and dependencies. Ensure the environment is activated.  \n   **Activate the virtualenv** (change directory to the new environment):  \n   ```bash\n   source my_venv/bin/activate\n   ```\n\n3. **Install Packages**:  \n   If your project requires specific packages, install them using `pip install package_name`. For example:  \n   ```bash\n   pip install requirements.txt\n   ```\n\n4. **Use the Virtualenv**:  \n   Now, you can use the virtualenv to install packages and run scripts without conflicts with your main Python environment.\n\n**Example**:  \n```bash\n# Create a virtualenv\npython3 -m venv my_venv\n\n# Activate the virtualenv\nsource my_venv/bin/activate\n\n# Install dependencies\npip install some_package\n```\n\nThis setup allows you to manage multiple Python environments easily.", "retry_text": "Here's a step-by-step walkthrough of setting up a Python virtualenv:\n\n1. **Install Python**:  \n   If not already installed, install Python using your system's package manager. For example:  \n   ```bash\n   # For Ubuntu/Debian\n   sudo apt-get update\n   sudo apt-get install -y python3\n   # For CentOS\n   sudo yum install -y python3\n   ```\n\n2. **Create a Virtual Environment**:  \n   Use `venv` or `virtualenv` to create a new environment. The command is:  \n   ```bash\n   python3 -m venv my_venv\n   ```\n\n   This creates a directory named `my_venv` with Python and dependencies. Ensure the environment is activated.  \n   **Activate the virtualenv**:  \n   ```bash\n   source my_venv/bin/activate\n   ```\n\n3. **Install Packages**:  \n   If your project requires specific packages, install them using `pip install package_name`. For example:  \n   ```bash\n   pip install requirements.txt\n   ```\n\n4. **Use the Virtualenv**:  \n   Now, you can use the virtualenv to install packages and run scripts without conflicts with your main Python environment.\n\n**Example**:  \n```bash\n# Create a virtualenv\npython3 -m venv my_venv\n\n# Activate the virtualenv\nsource my_venv/bin/activate\n\n# Install dependencies\npip install some_package\n```\n\nThis setup allows you to manage multiple Python environments easily.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "4f318c8ba1ed6f94", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "What is your plan for answering a question you are unsure about?", "text": "When unsure about a question, I plan to first search the web for the information using the web_search function. If the query is resolved through online resources, I'll proceed with that. If not, I'll use Python code execution to address any specific tasks or data analysis required. I'm also prepared to adjust my plan based on the results found.", "retry_text": "When unsure about a question, I plan to first search the web using the `web_search` function. If the information is resolved through online resources, I'll proceed with that. If not, I'll use Python code execution to address any specific tasks or data analysis required. I'm also prepared to adjust my plan based on the results found.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "095bb6719849ccf9", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "What is your plan for answering a question you are unsure about?", "text": "When unsure about a question, my approach is to use the tools available, such as web searches or sandbox code execution. I'll ask for clarification or provide relevant information based on the tools. If I need assistance with specific tasks, I'll use the tools provided to get the required information.", "retry_text": "When unsure about a question, my approach is to use tools like web searches or sandbox code execution to gather information. I will ask for clarification or provide relevant details. If no tools are needed, I will directly answer the question as requested.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "6466d33bf4055aba", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "What is your plan for answering a question you are unsure about?", "text": "When unsure about a question, my plan is to:  \n1. **Research the topic** to gather reliable information.  \n2. **Analyze the information** to determine the most accurate answer.  \n3. **Provide a well-structured response** that addresses the question while being helpful and comprehensive.  \n\nThis approach ensures clarity and addresses the user’s needs effectively.", "retry_text": "When unsure about a question, my plan is:  \n1. **Research** to gather reliable information.  \n2. **Analyze** the information to determine the most accurate answer.  \n3. **Provide a well-structured response** that addresses the question while being helpful and comprehensive.  \n\nThis approach ensures clarity and addresses the user’s needs effectively.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "f571eddf37233a27", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "What is your plan for answering a question you are unsure about?", "text": "When unsure about a question, my plan is to first use the web_search function to gather information if the question involves specific content or needs external data. Otherwise, I rely on my existing knowledge and domain expertise. I remain flexible and adapt my responses based on the context and the nature of the question.", "retry_text": "I don't need a tool to answer this question. When unsure about a question, my plan is to use the web_search function to gather information if the question requires external data or specific content, or rely on my existing knowledge and domain expertise otherwise. I remain flexible and adapt my responses based on the context and the nature of the question.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "287075bcccadaa02", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "What is your plan for answering a question you are unsure about?", "text": "When unsure about a question, my plan is to:  \n1. **Verify facts** by cross-referencing information from reliable sources.  \n2. **Stay informed** by staying updated on the topic to ensure accuracy.  \n3. **Consider different perspectives** to find the most comprehensive answer.  \n\nI am always open to learning and adapting my approach to provide the best response.", "retry_text": "When unsure about a question, my plan is to:  \n1. **Verify facts** by cross-referencing information from reliable sources.  \n2. **Stay informed** by staying updated on the topic to ensure accuracy.  \n3. **Consider different perspectives** to find the most comprehensive answer.  \n\nI am always open to learning and adapting my approach to provide the best response.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "9474398e6db4735a", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "What is your plan for answering a question you are unsure about?", "text": "When unsure about a question, my plan is to:  \n1. **Research thoroughly** using reliable sources.  \n2. **Ask clarifying questions** to deepen understanding.  \n3. **Analyze the context** and consider different perspectives.  \n4. **Stay open to new information** and adapt based on available resources.  \n\nLet me know if you'd like to refine this approach!", "retry_text": "I don't need a tool to answer this question. When unsure, my plan is to:  \n1. **Research thoroughly** using reliable sources.  \n2. **Ask clarifying questions** to deepen understanding.  \n3. **Analyze the context** and consider different perspectives.  \n4. **Stay open to new information** and adapt based on available resources.  \n\nLet me know if you'd like to refine this approach!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "16abe1d67280a8e4", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "What is your plan for answering a question you are unsure about?", "text": "When unsure about a question, I rely on available information and use the web_search function to gather details. I ensure thoroughness by checking the information thoroughly. If you have more context or specific details about the question, I can tailor my plan accordingly. Let me know!", "retry_text": "When unsure about a question, I rely on available information and use the web_search function to gather details. I ensure thoroughness by checking thoroughly. If you have more context or specific details about the question, I can tailor my plan accordingly. Let me know!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "d851acb957839ac2", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Describe your approach to reviewing a pull request, as a list.", "text": "Here’s a structured approach to reviewing a pull request:\n\n- **Code Review for Correctness and Style**:  \n  - Ensure code is logically sound and follows project standards.  \n  - Check for syntax errors, logic flaws, and code duplication.  \n\n- **Identify Issues and Bug Fixes**:  \n  - Locate and resolve any bugs, edge cases, or issues in the code.  \n\n- **Ensure Maintainability**:  \n  - Verify that the code is easy to understand, testable, and scalable for future updates.  \n\n- **Testing**:  \n  - Conduct unit tests, integration tests, and end-to-end tests to validate the code.  \n\n- **Communication with the Team**:  \n  - Coordinate with the team to ensure alignment and address any concerns.  \n\n- **Documentation**:  \n  - Update documentation to reflect the changes made in the pull request.  \n\nThis approach ensures a thorough and constructive review process.", "retry_text": "Here’s a structured approach to reviewing a pull request:\n\n- **Code Review for Correctness and Style**:  \n  - Ensure code is logically sound and follows project standards.  \n  - Check for syntax errors, logic flaws, and code duplication.  \n\n- **Identify Issues and Bug Fixes**:  \n  - Locate and resolve any bugs, edge cases, or issues in the code.  \n\n- **Ensure Maintainability**:  \n  - Verify that the code is easy to understand, testable, and scalable for future updates.  \n\n- **Testing**:  \n  - Conduct unit tests, integration tests, and end-to-end tests to validate the code.  \n\n- **Communication with the Team**:  \n  - Coordinate with the team to ensure alignment and address any concerns.  \n\n- **Documentation**:  \n  - Update documentation to reflect the changes made in the pull request.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "5d1eee90ae0e4d93", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Describe your approach to reviewing a pull request, as a list.", "text": "Here’s a structured approach to reviewing a pull request:\n\n1. **Code Review**  \n   - Check for bugs, syntax errors, and logical errors.  \n   - Ensure code follows coding standards and best practices.  \n\n2. **Testing**  \n   - Validate functionality with unit tests and integration tests.  \n   - Ensure the code works as expected under various conditions.  \n\n3. **Documentation**  \n   - Verify that the code is well-documented and explained.  \n\n4. **Security**  \n   - Check for vulnerabilities (e.g., XSS, SQL injection) and security issues.  \n\n5. **Code Quality**  \n   - Look for unused variables, inefficient loops, or code smells.  \n\n6. **Contextual Understanding**  \n   - Understand the purpose of the pull request and the specific changes being made.  \n\n7. **Impact on Other Components**  \n   - Ensure the code doesn’t introduce new issues or conflicts with existing code.  \n\n8. **Automated Tools**  \n   - Use tools (e.g., linters, static analyzers) to catch issues before integration.  \n\nThis approach ensures a thorough review while maintaining efficiency and quality.", "retry_text": "Here’s a structured approach to reviewing a pull request:\n\n1. **Code Review**  \n   - Check for bugs, syntax errors, and logical errors.  \n   - Ensure code follows coding standards and best practices.  \n\n2. **Testing**  \n   - Validate functionality with unit tests and integration tests.  \n   - Ensure the code works as expected under various conditions.  \n\n3. **Documentation**  \n   - Verify that the code is well-documented and explained.  \n\n4. **Security**  \n   - Check for vulnerabilities (e.g., XSS, SQL injection) and security issues.  \n\n5. **Code Quality**  \n   - Look for unused variables, inefficient loops, or code smells.  \n\n6. **Contextual Understanding**  \n   - Understand the purpose of the pull request and the specific changes being made.  \n\n7. **Impact on Other Components**  \n   - Ensure the code doesn’t introduce new issues or conflicts with existing code.  \n\n8. **Automated Tools**  \n   - Use tools (e.g., linters, static analyzers) to catch issues before integration.  \n\nThis approach ensures a thorough review while maintaining efficiency and quality.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ae18ee4c0a550b21", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Describe your approach to reviewing a pull request, as a list.", "text": "Here's a structured approach to reviewing a pull request:\n\n1. **Code Quality Checks**  \n   - ✅ Check for syntax errors, logical errors, and clean code.  \n   - ✅ Look for potential security vulnerabilities or code smells.  \n\n2. **Documentation Review**  \n   - ✅ Ensure the PR includes well-documented changes and test cases.  \n   - ✅ Verify that any documentation related to the PR is updated and maintained.  \n\n3. **Code Consistency**  \n   - ✅ Check for consistent coding style, naming conventions, and variable/constant usage.  \n   - ✅ Ensure all changes align with the project’s existing codebase.  \n\n4. **Test Coverage**  \n   - ✅ Verify that the PR includes test cases for the affected functionality.  \n   - ✅ Check for any edge cases or scenarios that might need special attention.  \n\n5. **Issue Resolution**  \n   - ✅ Ensure the PR addresses any open issues or bugs in the project.  \n   - ✅ Confirm that the PR has been reviewed and approved by relevant stakeholders.  \n\n6. **CI/CD Integration**  \n   - ✅ Check that the PR is prepared for integration into the main codebase (CI/CD pipeline).  \n   - ✅ Ensure that the changes are compatible with existing workflows.  \n\n7. **Future-Proofing**  \n   - ✅ Review for architectural changes that could break future features.  \n   - ✅ Ensure the code is maintainable and scalable for long-term use.  \n\nThis approach ensures that the PR is reviewed thoroughly and ready for merging into the main codebase.", "retry_text": "Here's a structured approach to reviewing a pull request:  \n\n1. **Code Quality Checks**  \n   - ✅ Check for syntax errors, logical errors, and clean code.  \n   - ✅ Ensure potential security vulnerabilities or code smells are addressed.  \n\n2. **Documentation Review**  \n   - ✅ Verify well-documented changes and test cases are included.  \n   - ✅ Confirm documentation aligns with the project’s existing structure.  \n\n3. **Code Consistency**  \n   - ✅ Check coding style, naming conventions, and variable/constant usage.  \n   - ✅ Ensure all changes align with the project’s codebase standards.  \n\n4. **Test Coverage**  \n   - ✅ Verify the PR includes test cases for the affected functionality.  \n   - ✅ Check edge cases or scenarios requiring special attention.  \n\n5. **Issue Resolution**  \n   - ✅ Ensure the PR addresses any open issues or bugs in the project.  \n   - ✅ Confirm the changes are ready for review and approval.  \n\n6. **CI/CD Integration**  \n   - ✅ Check compatibility with existing CI/CD workflows.  \n   - ✅ Ensure changes are prepared for integration into the main codebase.  \n\n7. **Future-Proofing**  \n   - ✅ Review for architectural changes that could break future features.  \n   - ✅ Ensure the code is maintainable, scalable, and ready for long-term use.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "93f33dfafe4b553e", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Describe your approach to reviewing a pull request, as a list.", "text": "Here's a structured approach to reviewing a pull request:\n\n1. **Code Review**: Check the code for syntax errors, logic errors, and any potential issues. Ensure the implementation aligns with the project's coding standards.\n2. **Documentation**: Verify that the changes in the PR are documented in the codebase and that the documentation is updated accordingly.\n3. **Project Codebase**: Compare the PR changes with the existing codebase to ensure no missed or duplicated work. Check for any unintended side effects.\n4. **Security Vulnerabilities**: Look for any potential security flaws, such as insecure code or vulnerabilities that could be exploited.\n5. **Testing**: Ensure the changes pass all relevant tests, especially if the code is complex or new to the project.\n6. **Communication**: Communicate with the maintainers or the project's team to gather their feedback and address any concerns.\n\nThis approach ensures a thorough and constructive review of the PR.", "retry_text": "Here's a structured approach to reviewing a pull request:\n\n1. **Code Review**: Check the code for syntax errors, logic errors, and any potential issues. Ensure the implementation aligns with the project's coding standards.  \n2. **Documentation**: Verify that the changes in the PR are documented in the codebase and that the documentation is updated accordingly.  \n3. **Project Codebase**: Compare the PR changes with the existing codebase to ensure no missed or duplicated work. Check for any unintended side effects.  \n4. **Security Vulnerabilities**: Look for any potential security flaws, such as insecure code or vulnerabilities that could be exploited.  \n5. **Testing**: Ensure the changes pass all relevant tests, especially if the code is complex or new to the project.  \n6. **Communication**: Communicate with the maintainers or the project's team to gather their feedback and address any concerns.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "e3307b82f25a0462", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Describe your approach to reviewing a pull request, as a list.", "text": "Here's a structured approach to reviewing a pull request:\n\n1. **Code Review Quality Check**  \n   - Ensure code meets project standards and has no syntax errors  \n   - Check for logical errors, performance issues, or security flaws  \n\n2. **Code Documentation**  \n   - Verify the issue description is well-documented and addresses the original problem  \n   - Ensure any changes are clearly explained in the documentation  \n\n3. **Testing**  \n   - Conduct unit tests and integration tests  \n   - Check for compatibility issues with existing code and dependencies  \n\n4. **Issue Resolution**  \n   - Address any issues raised in the pull request  \n   - Ensure the pull request is ready for merging  \n\n5. **Code Policy Compliance**  \n   - Confirm it aligns with the project's code review policies and guidelines  \n\nThis approach ensures a thorough and constructive review process.", "retry_text": "Here's a structured approach to reviewing a pull request:\n\n1. **Code Review Quality Check**  \n   - Ensure code meets project standards and has no syntax errors  \n   - Check for logical errors, performance issues, or security flaws  \n\n2. **Code Documentation**  \n   - Verify the issue description is well-documented and addresses the original problem  \n   - Ensure any changes are clearly explained in the documentation  \n\n3. **Testing**  \n   - Conduct unit tests and integration tests  \n   - Check for compatibility issues with existing code and dependencies  \n\n4. **Issue Resolution**  \n   - Address any issues raised in the pull request  \n   - Ensure the pull request is ready for merging  \n\n5. **Code Policy Compliance**  \n   - Confirm it aligns with the project's code review policies and guidelines  \n\nThis approach ensures a thorough and constructive review process.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "bddfdd9ea80c6169", "model": "qwen3-0.6b", "prompt_class": "format", "prompt": "Describe your approach to reviewing a pull request, as a list.", "text": "Here's a structured approach to reviewing a pull request:\n\n1. **Initial Review**  \n   - **Commit Message**: Ensure the message is clear, concise, and addresses the issue.  \n   - **Code Formatting**: Check for consistent formatting, including whitespace, spacing, and code blocks.  \n   - **Licensing and Policies**: Verify the code complies with project policies (e.g., MIT, GPL, or specific licensing agreements).  \n\n2. **Code Quality Check**  \n   - **Code Documentation**: Ensure changes are well-documented and align with existing documentation.  \n   - **Code Standards**: Confirm the code follows the project's coding guidelines (e.g., PEP 8, style rules).  \n\n3. **Issue Resolution**  \n   - **Issue Description**: Validate the issue description is clear and the fix addresses the issue.  \n   - **Code for Fixes**: Review the fix implementation and ensure it solves the problem effectively.  \n\n4. **Testing**  \n   - **Unit and Integration Tests**: Ensure the code works as intended in tests (e.g., unit tests, integration tests).  \n\n5. **History Review**  \n   - **Conflict Checks**: Ensure there are no duplicate or conflicting changes in the history.  \n   - **PR History**: Confirm the PR includes the correct code and no unnecessary changes.  \n\n6. **Documentation**  \n   - **Update Documentation**: Update documentation to reflect the changes made in the pull request.  \n\n7. **Final Approval**  \n   - **Final Check**: Confirm all points are addressed and the PR is ready for review.  \n\nThis approach ensures a thorough review while maintaining efficiency and quality.", "retry_text": "Here's a structured approach to reviewing a pull request:\n\n1. **Initial Review**  \n   - **Commit Message**: Ensure the message is clear, concise, and addresses the issue.  \n   - **Code Formatting**: Check for consistent formatting, including whitespace, spacing, and code blocks.  \n\n2. **Code Quality Check**  \n   - **Code Documentation**: Ensure changes are well-documented and align with existing documentation.  \n   - **Code Standards**: Confirm the code follows the project's coding guidelines (e.g., PEP 8, style rules).  \n\n3. **Issue Resolution**  \n   - **Issue Description**: Validate the issue description is clear and the fix addresses the issue.  \n   - **Code for Fixes**: Review the fix implementation and ensure it solves the problem effectively.  \n\n4. **Testing**  \n   - **Unit and Integration Tests**: Ensure the code works as intended in tests (e.g., unit tests, integration tests).  \n\n5. **History Review**  \n   - **Conflict Checks**: Ensure there are no duplicate or conflicting changes in the history.  \n   - **PR History**: Confirm the PR includes the correct code and no unnecessary changes.  \n\n6. **Documentation**  \n   - **Update Documentation**: Update documentation to reflect the changes made in the pull request.  \n\n7. **Final Approval**  \n   - **Final Check**: Confirm all points are addressed and the PR is ready for review.  \n\nThis approach ensures a thorough review while maintaining efficiency and quality.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "2f578170c686ad08", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "I can help you with something. Could you please specify what you need assistance with?", "retry_text": "I can help you with something. Could you please specify what you need assistance with?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "0156c4d90b02722b", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "I can help you with that. Could you please specify what you need assistance with?", "retry_text": "I can help you with that. Could you please specify what you need assistance with?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "19beab9b9b904353", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "I can assist you if you need help with a specific task. Could you please clarify what you'd like me to do? For example, is there a web search you'd like to perform, or is there a coding problem you'd like to solve?", "retry_text": "I can assist you if you need help with a specific task. Could you please clarify what you'd like me to do? For example, is there a web search you'd like to perform, or is there a coding problem you'd like to solve?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "e077760a025ac077", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "I can help you with that! Could you please clarify what you need assistance with? For example, are you looking for information, a solution to a problem, or a specific action to take?", "retry_text": "I can help you with that! Could you please clarify what you need assistance with? For example, are you looking for information, a solution to a problem, or a specific action to take?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "8b5d6af1da80759a", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "I can help you with that! Could you please specify what you need assistance with?", "retry_text": "Could you please specify what you need assistance with?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "f989b1b1d2d66995", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "I can't look into that. I can only perform web searches or execute Python code. Could you please provide more details or specify what you're looking for?", "retry_text": "I cannot look into that. I can only perform web searches or execute Python code. Could you please provide more details or specify what you're looking for?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "069078e3f77bd1f3", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "Sure! Could you please tell me what you need assistance with?", "retry_text": "Sure! Could you please tell me what you need assistance with?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "d1a015f991ab3d00", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "I can help you with that. Could you please clarify what specific task or query you need assistance with?", "retry_text": "I can help you with that. Could you please clarify what specific task or query you need assistance with?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "8ba7363ec8d4808b", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Fix it.", "text": "Could you please specify what \"Fix it\" refers to in your context? For example, is it a task to be completed, a code issue, a configuration problem, or something else? This will help me provide you with the most accurate assistance.", "retry_text": "Could you please specify what \"Fix it\" refers to in your context? For example, is it a task to be completed, a code issue, a configuration problem, or something else? This will help me provide you with the most accurate assistance.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "dd46a0139e68687c", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Fix it.", "text": "I'm sorry, but I'm not sure what \"Fix it\" refers to. Could you please provide more context or specify the issue you'd like me to address?", "retry_text": "I'm sorry, but I'm not sure what \"Fix it\" refers to. Could you please provide more context or specify the issue you'd like me to address?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "09cbf0f6bc16ece4", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Fix it.", "text": "Could you please specify what \"Fix it\" refers to? For example, is there a task, a requirement, or a problem that needs fixing? This will help me provide you with the appropriate assistance.", "retry_text": "Could you please specify what \"Fix it\" refers to? For example, is there a task, a requirement, or a problem that needs fixing? This will help me provide you with the appropriate assistance.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "5d1c121db57e292c", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Fix it.", "text": "Could you please specify what issue or problem you're referring to? I'll be able to help you fix it.", "retry_text": "Could you please specify what issue or problem you're referring to? I'll be able to help you fix it.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "92551203a59734bf", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Fix it.", "text": "Could you please specify what issue or problem you'd like to fix?", "retry_text": "Could you please specify what issue or problem you'd like to fix?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "71a8e1df216e3c57", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Fix it.", "text": "Could you please specify what \"Fix it\" refers to? For example, is there a task, problem, or issue you want to address? Let me know and I'll do my best to help.", "retry_text": "Could you please specify what \"Fix it\" refers to? For example, is there a task, problem, or issue you want to address? Let me know and I'll do my best to help.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "3fd18b05b0bf9744", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Fix it.", "text": "Could you please provide more context or specify what \"Fix it\" refers to in your situation?", "retry_text": "The message \"Fix it\" doesn't contain a specific request or task to address. To proceed, could you please provide more context or clarify what \"Fix it\" refers to in your situation?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "63b05dc7bff381cd", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "I don't have access to a conversation history or previous information. Could you please clarify what \"the thing we discussed\" refers to? Let me know and I'll do my best to assist.", "retry_text": "I don't have access to a conversation history or previous information. Could you please clarify what \"the thing we discussed\" refers to? Let me know and I'll do my best to assist.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "513c9d08d9dd0f83", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "I can't tell you about what we discussed without more context. Could you please specify the topic or provide details about the subject we've been discussing?", "retry_text": "I cannot proceed without more context. Could you please specify the subject or topic we discussed?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "986805e14cfb085e", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "Could you please specify what \"the thing we discussed\" refers to so I can provide a detailed response?", "retry_text": "Could you please specify what \"the thing we discussed\" refers to so I can provide a detailed response?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "3a8cc99a4fd7102d", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "I don't have access to specific information or a conversation history in this context. Could you please provide more details or clarify what you're referring to?", "retry_text": "I don't have access to specific information or a conversation history in this context. Could you please provide more details or clarify what you're referring to?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "1a596b971a3932b3", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "I can't tell you about the topic we discussed without more information. Could you please specify what the subject was or provide details about the conversation?", "retry_text": "I cannot provide a detailed explanation about the topic we discussed without specific information or context. Could you please share more details or clarify what the subject was or the conversation was about?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "b321d857119f0e1a", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "I can't determine the specific topic without additional context. Could you please clarify what topic or subject you're interested in?", "retry_text": "I cannot determine the specific topic without additional context. Could you please clarify what topic or subject you're interested in?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "f27d000a239b9998", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "I don't have access to your conversation history. Could you please provide more context or specify what \"the thing we discussed\" refers to?", "retry_text": "I don't have access to your conversation history. Could you please provide more context or specify what \"the thing we discussed\" refers to?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "df692d56e5002f3b", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "I don't have access to previous conversation history. Could you please specify what the \"thing we discussed\" refers to?", "retry_text": "I don't have access to previous conversation history. Could you please specify what the \"thing we discussed\" refers to?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "e2051d39a9ee4a41", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Is it better?", "text": "The question \"Is it better?\" is a bit ambiguous. Could you please provide more context or clarify what specific aspect you're referring to (e.g., which decision, choice, or situation you're comparing)? This will help me better respond.", "retry_text": "The question \"Is it better?\" is ambiguous and lacks specificity. To assist you, please clarify what aspect or context you are referring to (e.g., which decision, choice, or situation you’re comparing).", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "587cc3f962fa5e14", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Is it better?", "text": "I can't determine \"Is it better?\" without more context or specific information. Could you please provide the question or clarify what you're referring to?", "retry_text": "The question \"Is it better?\" is ambiguous without additional context. Could you please specify the question or clarify what you're referring to?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "b15957ffbe039c9e", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Is it better?", "text": "The question is a bit ambiguous. Could you clarify what you're asking? For example, are you comparing two options, deciding which is better, or looking for information on a specific topic? This would help me provide a more accurate response.", "retry_text": "The question is ambiguous. Could you clarify what you're asking? For example, are you comparing two options, deciding which is better, or looking for information on a specific topic?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a358f81fdb3324a7", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Is it better?", "text": "The choice between using `web_search` or `python` depends on your specific needs. Here's how each tool can help:\n\n- **`web_search`**: Ideal for general information retrieval, as it can fetch full-page text from URLs. Useful for finding answers to questions or solving problems.\n- **`python`**: Great for executing code within a sandbox environment, which is perfect for tasks like data processing, automation, or scripting.\n\nIf you have a specific task or question in mind, let me know what tools you'd like to use!", "retry_text": "The choice between `web_search` and `python` depends on your specific needs. Here's how each tool functions:\n\n- **`web_search`**: Fetches full-page text from URLs to retrieve information, ideal for solving problems or finding answers to questions.\n- **`python`**: Executes code in a sandbox environment, suitable for tasks like data processing, automation, or scripting.\n\nIf you have a specific task, let me know what tools you’d like to use!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "7ecabf000158e586", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "Is it better?", "text": "Is it better? Please provide the specific context or choices you're comparing, and I can help you determine the best option.", "retry_text": "The question \"Is it better?\" is ambiguous and lacks context or specific choices to compare. To assist you, please provide the specific question, comparison criteria, or options you're evaluating. For example, if you're comparing two products, services, or decisions, clarify what you're referring to. Without this information, I can't determine the best option.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "0cebe6188ebb1145", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "What should I do next?", "text": "To help you better, could you please provide more details or clarify what \"next\" refers to in your context? For example, are you looking for a specific action, a step-by-step guide, or assistance with something else?", "retry_text": "To assist you better, could you please provide more details or clarify what \"next\" refers to in your context? For example, are you looking for a specific action, a step-by-step guide, or something else?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "354119601b55966d", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "What should I do next?", "text": "To assist you effectively, could you please share more details about the situation or question you're facing? For example, are you looking for help with a specific problem, a request, or guidance on a topic?", "retry_text": "To assist you effectively, could you please share more details about the situation or question you're facing? For example, are you looking for help with a specific problem, a request, or guidance on a topic?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "4997cc2bb2015a82", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "What should I do next?", "text": "Could you please provide more details about the situation or the query you're facing?", "retry_text": "Could you please provide more details about the situation or the query you're facing?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "cbe166be7cdfccfe", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "What should I do next?", "text": "To help you better, could you please provide more details about your situation or the specific question you're facing?", "retry_text": "Could you please provide more details about your situation or the specific question you're facing?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "e3c3a09b2135bb07", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "What should I do next?", "text": "I can help you with specific questions. Could you please provide more details or clarify what you're looking for?", "retry_text": "Could you please specify what you need help with?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "eec98c2dd8620558", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "What should I do next?", "text": "Could you please clarify what you need assistance with? For example, are you looking for information, a solution to a problem, or help with something specific?", "retry_text": "Could you please clarify what you need assistance with? For example, are you looking for information, a solution to a problem, or help with something specific?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "55159a82d8ce1dc7", "model": "qwen3-0.6b", "prompt_class": "ambiguous", "prompt": "What should I do next?", "text": "The question \"What should I do next?\" is general and doesn't provide enough context for me to determine the appropriate action. Could you please specify the situation or problem you're facing so I can assist you better?", "retry_text": "The question \"What should I do next?\" is general and doesn't provide specific context for me to determine the appropriate action. Could you please specify the situation or problem you're facing so I can assist you better?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "5e31f765cb50a4e8", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Quelle est la capitale de la France ?", "text": "La capitale de la France est Paris.", "retry_text": "La capitale de la France est Paris.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "5c719501e2307082", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Quelle est la capitale de la France ?", "text": "La capitale de la France est Paris.", "retry_text": "La capitale de la France est Paris.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ebb10de7c8d9f6cd", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Quelle est la capitale de la France ?", "text": "La capitale de la France est Lyon.", "retry_text": "La capitale de la France est Lyon.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "3eefd2de8fb723f7", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Quelle est la capitale de la France ?", "text": "La capitale de la France est Paris.", "retry_text": "La capitale de la France est Paris.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "0cca10877add3e01", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Quelle est la capitale de la France ?", "text": "La capitale de la France est Paris.", "retry_text": "La capitale de la France est Paris.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a06f840c1397b49a", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Quelle est la capitale de la France ?", "text": "La capitale de la France est Paris.", "retry_text": "La capitale de la France est Paris.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "b9fae31106f76bfc", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Quelle est la capitale de la France ?", "text": "La capitale de la France est Paris.", "retry_text": "La capitale de la France est Paris.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "1bd8528c0be010cb", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Quelle est la capitale de la France ?", "text": "La capitale de la France est Paris.", "retry_text": "La capitale de la France est Paris.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "2fe04e2076dbadad", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Explica en una frase que es el sobreajuste.", "text": "El sobreajuste es el fenómeno en el que una persona exagera su capacidad o habilidad, lo que lleva a una desviación en sus actos o resultados.", "retry_text": "El sobreajuste es el fenómeno en el que una persona exagera su capacidad o habilidad, lo que lleva a una desviación en sus actos o resultados.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "d09e13931c8de2f5", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Explica en una frase que es el sobreajuste.", "text": "El sobreajuste es cuando el monto de la deuda es mayor que lo que debe pagar.", "retry_text": "El sobreajuste es cuando el monto de la deuda es mayor que lo que debe pagar.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "933c9d680abe10f3", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Explica en una frase que es el sobreajuste.", "text": "Sobreajuste es cuando una actividad o sistema se enfrenta a una carga sobrecarga que puede no ser resuelta debido a la capacidad limitada de los recursos.", "retry_text": "Sobreajuste es la situación en la que una actividad o sistema enfrenta una carga sobrecarga que no puede ser resuelta debido a la capacidad limitada de los recursos.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "91712a4316d56f06", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Explica en una frase que es el sobreajuste.", "text": "El sobreajuste se refiere a la excesiva consumo de recursos o recursos materiales.", "retry_text": "El sobreajuste se refiere a la excesiva consumo de recursos o recursos materiales.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "19851173a737b424", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Explica en una frase que es el sobreajuste.", "text": "El sobreajuste ocurre cuando una persona experimenta una excesiva carga física o mental, causando desgaste.", "retry_text": "El sobreajuste ocurre cuando una persona experimenta una excesiva carga física o mental, causando desgaste.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "6273bf296ba0d580", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Explica en una frase que es el sobreajuste.", "text": "El sobreajuste es un sobreajuste, es decir, una pérdida financiera o un impago que ocurre cuando se enfrentan desafíos económicos o operaciones no esperadas.", "retry_text": "El sobreajuste es una pérdida financiera o un impago que ocurre cuando se enfrentan desafíos económicos o operaciones no esperadas.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "c5fc34cdf315ee06", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Was ist der Unterschied zwischen RAM und Festplatte?", "text": "RAM (Random Access Memory) and a hard drive (Festplatte) are two different types of storage and memory in computers. Here's the distinction:\n\n1. **RAM (Memory):**\n   - **Purpose:** Temporary storage for active data and running applications.\n   - **Speed:** Accesses data quickly (typically faster than a hard drive).\n   - **Volatility:** Data is lost when the computer turns off.\n   - **Usage:** Used for tasks that require immediate access, like browsing, running programs, or processing data.\n\n2. **Hard Drive (Festplatte):**\n   - **Purpose:** Permanent storage for files, programs, and data.\n   - **Speed:** Accesses data slowly compared to RAM.\n   - **Durability:** Stores information long-term without losing data.\n   - **Usage:** For storing software, user data, and backups.\n\nRAM is volatile, while a hard drive is non-volatile.", "retry_text": "RAM (Random Access Memory) and a hard drive (Festplatte) are two types of storage in computers. Here's the key difference:\n\n1. **RAM (Memory):**\n   - **Purpose:** Temporary storage for active data and running applications.\n   - **Speed:** Accesses data quickly (typically faster than a hard drive).\n   - **Volatility:** Data is lost when the computer turns off.\n   - **Usage:** Used for tasks requiring immediate access, like browsing, running programs, or processing data.\n\n2. **Hard Drive (Festplatte):**\n   - **Purpose:** Permanent storage for files, programs, and data.\n   - **Speed:** Accesses data slowly compared to RAM.\n   - **Durability:** Stores information long-term without losing data.\n   - **Usage:** For storing software, user data, and backups.\n\nRAM is volatile and loses data when the computer turns off, while a hard drive holds data permanently.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "56b6ed34e8ccc3b0", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Was ist der Unterschied zwischen RAM und Festplatte?", "text": "RAM (Random Access Memory) and a hard drive (Festplatte) serve different purposes in a computer:\n\n1. **RAM (Memory):**  \n   - Holds data temporarily for quick access. It's used to run programs and store information currently being used by the CPU.  \n   - Is consumed when the computer is actively processing data (e.g., running a program). It's faster than the hard drive because it's accessed directly by the CPU.\n\n2. **Hard Drive (Festplatte):**  \n   - Stores data permanently. It's used to store files, system information, and software.  \n   - Is slower than RAM because it requires a separate processor to access the data.\n\n**Key Difference:**  \nRAM is for quick, temporary tasks, while the hard drive is for long-term storage.", "retry_text": "RAM (Random Access Memory) and a hard drive (Festplatte) serve different purposes in a computer:\n\n1. **RAM (Memory):**  \n   - Stores data temporarily for quick access. It's used to run programs and hold information currently being used by the CPU. It's faster than the hard drive because it's directly accessed by the CPU.\n\n2. **Hard Drive (Festplatte):**  \n   - Stores data permanently. It holds files, system information, and software. It's slower than RAM because it requires a separate processor to access data.\n\n**Key Difference:**  \nRAM is for quick, temporary tasks, while the hard drive stores long-term data.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ae569736c21c98ab", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Was ist der Unterschied zwischen RAM und Festplatte?", "text": "RAM (Random Access Memory) is temporary storage used to hold data that is currently being used by the CPU. It's volatile, meaning it loses its data when the power goes out. Unlike a hard drive, which stores data permanently, RAM is used for quick data access. \n\nA hard drive (or SSD) stores data permanently, and its speed depends on whether it's an SSD (non-volatile) or an HDD (volatile). SSDs are faster because they don't use mechanical components, while HDDs are slower. The hard drive can be read from (RAID) or not (non-Raid) depending on its configuration.", "retry_text": "RAM (Random Access Memory) is temporary storage used to hold data that the CPU is currently processing. It's volatile, meaning it loses its data when the power goes out. A hard drive (or SSD) stores data permanently, and its speed depends on whether it's an SSD (non-volatile) or an HDD (volatile). RAM is used for quick data access, while a hard drive is for long-term storage.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "9a67df1137489d09", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Was ist der Unterschied zwischen RAM und Festplatte?", "text": "RAM (Random Access Memory) is a type of temporary storage used to hold active data while a computer is running programs. It's volatile, meaning it's not as stable as a hard drive, which stores data permanently. RAM is used to quickly access and process data as needed, whereas a hard drive (or SSD) stores data permanently, allowing for faster access times.", "retry_text": "RAM (Random Access Memory) is a type of temporary storage used to hold active data while a computer is running programs. It is volatile, meaning it loses its data when the power is off. RAM is used to quickly access and process data as needed, whereas a hard drive (or SSD) stores data permanently, allowing for faster access times.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ae4fb07db438690b", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Was ist der Unterschied zwischen RAM und Festplatte?", "text": "RAM (Random Access Memory) is a type of volatile memory used to hold data temporarily. It allows quick access to programs and data, making it essential for running applications or waiting for input. In contrast, a hard drive (Festplatte) is a non-volatile storage device used to store permanent data and software. It doesn't require constant power and can hold more data than RAM. Both components work together to support the computer's operations, with RAM being temporary and the hard drive storing permanent information.", "retry_text": "RAM (Random Access Memory) is a type of volatile memory used to hold data temporarily, while a hard drive (Festplatte) is a non-volatile storage device used to store permanent data and software. RAM allows quick access to programs and data, whereas the hard drive holds permanent information and storage. Both components work together to support the computer's operations.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "f71d15a0eaffe179", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Was ist der Unterschied zwischen RAM und Festplatte?", "text": "RAM (Random Access Memory) is a type of temporary storage used to hold active data and instructions that the CPU is currently processing. It's faster than a hard drive because it's directly accessed by the CPU, allowing for quick data retrieval. A hard drive, in contrast, is a long-term storage device that holds data and operating system files, which can be accessed and saved over time. RAM is typically smaller in capacity but offers faster access speeds.", "retry_text": "RAM (Random Access Memory) and a hard drive are both storage components, but they serve different purposes:  \n- **RAM** is a type of temporary memory used to hold active data and instructions the CPU is currently processing. It is faster than a hard drive because it's directly accessed by the CPU, allowing for quick data retrieval.  \n- **Hard drive** holds data and operating system files over time, meaning it's less volatile and can store large amounts of information without losing it.  \n\nRAM is typically smaller in capacity but provides faster access speeds.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "3333944ab8b3b14d", "model": "qwen3-0.6b", "prompt_class": "non_english", "prompt": "Was ist der Unterschied zwischen RAM und Festplatte?", "text": "RAM (Random Access Memory) is a type of temporary storage used in computers to hold data and instructions temporarily. It's volatile, meaning it loses data when the computer turns off. The hard drive, on the other hand, is a non-volatile storage device that holds the operating system and user data. It's not replaced, but it can be upgraded or replaced over time. RAM is used for quick access to data, while the hard drive stores more data.", "retry_text": "RAM (Random Access Memory) is a type of volatile storage used to hold data and instructions temporarily. It loses data when the computer turns off. A hard drive (HDD) is non-volatile and stores operating system and user data, not replaced over time. RAM is used for quick access, while HDDs store more data.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "22a85371668ee43a", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Thanks. So which version should I install?", "text": "The latest llama.cpp release version is **b6291** (2026-07-14). This version includes improvements to the Vulkan backend and faster prompt processing on CUDA. You should follow the official documentation or GitHub release notes for the most accurate installation instructions.", "retry_text": "The latest llama.cpp release version is **b6291** (2026-07-14). This version includes improvements to the Vulkan backend and faster prompt processing on CUDA. You should follow the official documentation or GitHub release notes for the most accurate installation instructions.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "4cf325393db5369b", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Thanks. So which version should I install?", "text": "The latest available version is **b6291** (2026-07-14). This version includes improvements such as Vulkan backend enhancements and faster prompt processing on CUDA.\n\nWould you like additional details about this version or any updates beyond this release?", "retry_text": "The latest available version is **b6291** (2026-07-14). This version includes improvements such as Vulkan backend enhancements and faster prompt processing on CUDA. Let me know if you need further details!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "544a64cdb96703fa", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Thanks. So which version should I install?", "text": "The latest available version for Llama.cpp is **2026-07-14**, which includes improvements like Vulkan backend enhancements and faster prompt processing on CUDA. You can install it directly from the official Llama.cpp GitHub repository. For installation instructions, refer to the [Llama.cpp documentation](https://github.com/llama-project/llama.cpp) or the [GitHub release page](https://github.com/llama-project/llama.cpp/releases). Let me know if you need further assistance!", "retry_text": "The latest available version for Llama.cpp is **2026-07-14**, which includes improvements like Vulkan backend enhancements and faster prompt processing on CUDA. You can install it directly from the official Llama.cpp GitHub repository. For installation instructions, refer to the [Llama.cpp documentation](https://github.com/llama-project/llama.cpp) or the [GitHub release page](https://github.com/llama-project/llama.cpp/releases). Let me know if you need further assistance!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "29a0818174227203", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Thanks. So which version should I install?", "text": "The latest llama.cpp version is **b6291** (2026-07-14). This release includes improvements for the Vulkan backend and faster prompt processing on CUDA. Let me know if you need further details!", "retry_text": "The latest llama.cpp version is **b6291** (2026-07-14). This release includes improvements for the Vulkan backend and faster prompt processing on CUDA. Let me know if you need further details!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "e200b9622f9f9588", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Thanks. So which version should I install?", "text": "The latest known llama.cpp release version is **b6291** (2026-07-14). This version includes improvements such as Vulkan backend enhancements and faster prompt processing on CUDA. For the most accurate and up-to-date information, you can check the official llama.cpp documentation. Let me know if you need further details!", "retry_text": "The latest known llama.cpp release version is **b6291** (2026-07-14). This version includes improvements such as Vulkan backend enhancements and faster prompt processing on CUDA. Make sure to install it using the official documentation for the most accurate and up-to-date information. Let me know if you need further details!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "d7da157d532ba3d6", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Thanks. So which version should I install?", "text": "The latest llama.cpp release version is **b6291 (2026-07-14)**. This version includes improvements such as enhanced Vulkan backend support and faster prompt processing on CUDA. You can install it via the official llama.cpp repository or via your environment setup. Let me know if you need further details!", "retry_text": "The latest llama.cpp release version is **b6291 (2026-07-14)**. This version includes improvements such as enhanced Vulkan backend support and faster prompt processing on CUDA. You can install it via the official llama.cpp repository or your environment setup. Let me know if you need further details!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a218f8ce82494ee9", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Thanks. So which version should I install?", "text": "The latest llama.cpp release is version **b6291** (2026-07-14). For installation, please refer to the official documentation or release notes for specific instructions. If you need help with dependencies or installation steps, I can assist further!", "retry_text": "The latest llama.cpp release is version **b6291** (2026-07-14). For installation, refer to the official documentation or release notes for specific instructions. Let me know if you need further assistance!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "3ce863a497159351", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Do I need a coat?", "text": "Based on the current weather in Reykjavik (9°C, overcast, wind 22 km/h, humidity 81%), it's **not extremely cold**. The wind and high humidity might make it feel a bit chilly, but it's still a mild day. You might need a coat, especially if the wind is strong, but it's not overly harsh. If you're planning to go outdoors, a light jacket or a sweater could be sufficient.", "retry_text": "Based on the current weather in Reykjavik (9°C, overcast, wind 22 km/h, humidity 81%), it's **not extremely cold**. The wind and high humidity might make it feel a bit chilly, but it's still a mild day. You might need a coat, especially if you plan to go outdoors, but it's not overly harsh. A light jacket or sweater could be sufficient.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "9d59823d666f5eee", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Do I need a coat?", "text": "Given the current weather in Reykjavik (9°C, overcast, wind 22 km/h, humidity 81%), it is highly recommended to wear a **coat** to cope with the cold and potentially high humidity. While the temperature is relatively mild, the combination of wind and humidity can make the environment feel slightly colder. If you plan to go out, packing layers and a jacket is a good precaution.", "retry_text": "Yes, you should wear a coat in Reykjavik. The current weather is 9°C with wind and high humidity, which can make temperatures feel slightly colder. A coat is a good precaution to protect against the elements.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "0b32a9c799a643ef", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Do I need a coat?", "text": "Based on the current weather: 9°C with overcast skies and a breeze, a light jacket or sweater is recommended. If it's expected to be cold (e.g., snow or rain), a coat might be necessary. However, since the temperature is mild, a jacket is sufficient.", "retry_text": "Based on the current weather (9°C with overcast skies and a breeze), a light jacket or sweater is recommended. Since the temperature is mild, a coat may not be necessary.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "b28dcd77e59abd11", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Do I need a coat?", "text": "If the temperature is 9°C in Reykjavik, it's generally not very cold, but it's still a good idea to have a coat in light rain or snowfall. In this case, the weather is overcast, so it's possible there's a chance of rain. While a coat might not be necessary, it's still a prudent precaution to stay prepared.", "retry_text": "Based on the current weather in Reykjavik (9°C, overcast, light rain), a coat is still a prudent precaution to stay prepared, even if it's not very cold.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "8c6687b7138a1e68", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Do I need a coat?", "text": "In Reykjavik, the current weather is 9°C with overcast conditions. While the temperature is not extremely cold, the wind from the north east and high humidity (81%) may make it slightly chilly. In general, a coat is not necessary for such a mild temperature. However, if the wind is strong and the temperature is still warm, a coat could help. The exact need for a coat depends on the wind's strength and the time of year.", "retry_text": "In Reykjavik, the current weather is 9°C with overcast conditions. The wind and humidity (81%) may make it slightly chilly, but a coat is not necessary for such a mild temperature. If the wind is strong or the temperature remains warm, a coat could help.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "aa9668cc9c6353c6", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Do I need a coat?", "text": "Based on the current weather in Reykjavik (9°C, overcast, wind from the north east, humidity 81%), it's likely that you'll need a light jacket or sweater. The temperature is above freezing, so a warm layer is essential for comfort. If you don't have a jacket, a sweater could be a good alternative. Make sure to prepare for the weather!", "retry_text": "Based on the current weather in Reykjavik (9°C, overcast, wind from the north east, humidity 81%), it's likely that you will need a light jacket or sweater. The temperature is above freezing, so a warm layer is essential for comfort. If you don't have a jacket, a sweater could be a good alternative. Make sure to prepare for the weather!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "766a83d5188fad2f", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Do I need a coat?", "text": "Based on the current weather in Reykjavik, which is **9°C (40°F)** with **overcast** conditions, there isn't a need for a heavy coat right now. The mild temperature combined with light clouds suggests it's a dry day, so a light jacket or a sweater should be sufficient. If it starts to rain soon, a light jacket is still a good option. Let me know if you'd like more details!", "retry_text": "Based on the current weather in Reykjavik, which is **9°C (40°F)** with **overcast** conditions, there isn't a need for a heavy coat. A light jacket or sweater is sufficient. If you're planning to go outside, a light clothing item will be appropriate. Let me know if you'd like further details!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "422a14dde624fb20", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Do I need a coat?", "text": "Based on the current weather conditions in Reykjavik (9°C, overcast, wind at 22 km/h, and humidity at 81%), it is likely to be chilly. However, while it's not extremely cold, it's possible that the weather may still require a coat. If the temperature drops below 0°C, a coat would be essential, but with 9°C, it's more likely to be a mild day. Therefore, you may need a coat, but it's not extreme.", "retry_text": "Based on the current weather conditions in Reykjavik (9°C, overcast, wind at 22 km/h, and humidity at 81%), it is not extremely cold, but the weather is still variable. While it's not expected to drop below freezing, the chances of rain or snow are present. A coat might be necessary, but it's not extreme.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "6200c8ed5ffeafea", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Is that number even or odd?", "text": "The 40th Fibonacci number is **even**. This is because the Fibonacci sequence alternates between odd and even starting from 0 and 1, and since 40 is even, the 40th term is even.", "retry_text": "The 40th Fibonacci number is **odd**.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ece827001e0f21af", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Is that number even or odd?", "text": "The 40th Fibonacci number is **even**. \n\n**Answer:** Even.", "retry_text": "The 40th Fibonacci number is **even**. The number is **102334155**.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "807e08728fab0668", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Is that number even or odd?", "text": "The 40th Fibonacci number is **odd**.", "retry_text": "The 40th Fibonacci number is **odd**.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "eb5098205f6aa11a", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Is that number even or odd?", "text": "The 40th Fibonacci number is **even**.", "retry_text": "The 40th Fibonacci number is **102334155**, which is **odd**.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "03f360a5c461d1df", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Is that number even or odd?", "text": "The 40th Fibonacci number is **odd**.", "retry_text": "The 40th Fibonacci number is **odd**.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "bd836b267afae418", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Is that number even or odd?", "text": "The 40th Fibonacci number is **even**. This is because the Fibonacci sequence modulo 2 cycles every 3 terms: 0, 1, 0, 1, 0, 1, etc. Since 40 is divisible by 3, the 40th term is congruent to 0 modulo 2.", "retry_text": "The 40th Fibonacci number is **even**. This is because the Fibonacci sequence modulo 2 cycles every 3 terms: 0, 1, 0, 1, 0, 1, ..., and since 40 is divisible by 3, the 40th term is congruent to 0 modulo 2.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "cb1b4451cfac3721", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Is that number even or odd?", "text": "The 40th Fibonacci number is even.", "retry_text": "The 40th Fibonacci number is odd.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "da47f38255c75c0b", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Is that number even or odd?", "text": "The 40th Fibonacci number is **odd**.", "retry_text": "The 40th Fibonacci number is **even**.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "5c67b8896031c82e", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "How does that compare with last month?", "text": "The current price of Bitcoin is $61,204.33 as of July 2026. However, without access to the last month's price, I cannot provide a direct comparison. If you have the last month's price or any specific reference, I can help you analyze it.", "retry_text": "The current price of Bitcoin (BTC/USD) is $61,204.33 as of July 2026. Without access to last month's data, I cannot provide a direct comparison. If you have the last month's price or any specific reference, I can help analyze it.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "400ef3f65d82b485", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "How does that compare with last month?", "text": "The current price of Bitcoin (BTC/USD) is **61,204.33**, as of July 29, 2026, at 3:00 UTC. This represents a **1.2% decline** compared to the day's price. However, without historical data, I cannot provide a direct comparison with last month's price. If you have access to historical Bitcoin price data, I can help analyze the trend.", "retry_text": "The current price of Bitcoin (BTC/USD) is **61,204.33**, as of July 29, 2026, at 3:00 UTC. This represents a **1.2% decline** compared to the day's price. To compare with last month, historical Bitcoin price data would be required, as this information is not accessible through standard means.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "1c609305d97bd2f1", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "How does that compare with last month?", "text": "As of July 29, 2026, Bitcoin (BTC/USD) has declined by approximately 1.2% compared to the same day in the previous month. This indicates a downward trend in the market price.", "retry_text": "The current price of Bitcoin (BTC/USD) as of July 29, 2026, is $61,204.33, reflecting a 1.2% decline from the same day in the previous month. This indicates a downward trend in the market price.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "0e3e17ceee0537ad", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "How does that compare with last month?", "text": "The current price of Bitcoin (BTC/USD) is $61,204.33 as of July 29, 2026, down 1.2% on the day. If you're comparing to last month's price, the exact percentage change would depend on the data available for that period. Since the tool response only includes the current price, I cannot provide a direct comparison to last month unless additional historical data is available.", "retry_text": "The current Bitcoin price (BTC/USD) is $61,204.33 as of July 29, 2026, down 1.2% on the day. To compare with last month's price, historical data would be required, which is not currently accessible through the provided tools.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "c45bec6121fa628e", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "How does that compare with last month?", "text": "The current price of Bitcoin (BTC/USD) is approximately $61,204.33 as of July 29, 2026, which represents a 1.2% decline compared to the previous month. This indicates a sustained decrease in the cryptocurrency's value over the past month.", "retry_text": "The current price of Bitcoin (BTC/USD) is approximately $61,204.33 as of July 2026, representing a 1.2% decline compared to the previous month.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "4068ff179b618f8d", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Round that to two decimals and explain what it means.", "text": "The standard deviation of the numbers 4, 8, 15, 16, 23, and 42 is approximately **12.30**. This indicates the spread of the data around the mean (18) is about **12.3 units**, meaning most of the numbers cluster closely within this range. A standard deviation of 12.30 suggests the data is relatively tightly distributed around the average. \n\n**Explanation:**  \nThe standard deviation measures how spread out the numbers are from the mean. A lower standard deviation means the numbers are closer to the mean, while a higher one indicates more variability. In this case, the numbers are fairly clustered around 18, resulting in a standard deviation of approximately 12.30.", "retry_text": "The standard deviation of the numbers 4, 8, 15, 16, 23, and 42 is approximately **12.30**. This indicates the spread of the data around the mean (18) is about **12.3 units**, meaning most of the numbers cluster closely within this range. A standard deviation of 12.30 suggests the data is relatively tightly distributed around the average.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "e6418b747ea08467", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Round that to two decimals and explain what it means.", "text": "The standard deviation of the numbers 4, 8, 15, 16, 23, and 42 is approximately **12.32** when rounded to two decimal places. This value indicates the spread of the data points around the mean, with a smaller standard deviation suggesting the numbers are closer to the mean.\n\n**Explanation:**  \nThe standard deviation measures how much variation exists in the dataset. Here, the value of 12.32 indicates that the numbers are relatively spread out from the mean (18).", "retry_text": "The standard deviation of the numbers 4, 8, 15, 16, 23, and 42 is approximately **12.32** when rounded to two decimal places. This value indicates the spread of the data points around the mean, with a smaller standard deviation suggesting the numbers are closer to the mean.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "3a2aa2e024c6d40b", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Round that to two decimals and explain what it means.", "text": "The standard deviation of the numbers 4, 8, 15, 16, 23, and 42 is approximately 12.33 (rounded to two decimal places). This value indicates how spread out the numbers are from the mean (18). A smaller standard deviation means the data points are closer to the mean, while a larger one suggests greater variability.  \n\n**Explanation:**  \n- **Mean (μ):** 18  \n- **Standard Deviation (σ):** √(910/6) ≈ 12.33  \n- **Interpretation:** The data points are relatively close to the mean, with the standard deviation measuring the average distance from the mean.", "retry_text": "The standard deviation of the numbers 4, 8, 15, 16, 23, and 42 is approximately 12.33 (rounded to two decimal places). This value indicates how spread out the numbers are from the mean (18). A smaller standard deviation means the data points are closer to the mean, while a larger one suggests greater variability.  \n\n**Explanation:**  \n- **Mean (μ):** 18  \n- **Standard Deviation (σ):** √(910/6) ≈ 12.33  \n- **Interpretation:** The data points are relatively close to the mean, with the standard deviation measuring the average distance from the mean.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "5a730decea125659", "model": "qwen3-0.6b", "prompt_class": "post_tool", "prompt": "Round that to two decimals and explain what it means.", "text": "The standard deviation of the numbers 4, 8, 15, 16, 23, and 42 is approximately **11.55** when rounded to two decimal places. \n\n**Explanation:**  \nStandard deviation measures how spread out the numbers are from the mean (18). A smaller standard deviation indicates the data points are closer to the mean, while a larger one suggests greater variability. In this case, the standard deviation of 11.55 means the numbers 4, 8, 15, 16, 23, and 42 are relatively close to the average value of 18.", "retry_text": "The standard deviation of the numbers 4, 8, 15, 16, 23, and 42 is approximately **11.55** when rounded to two decimal places. \n\n**Explanation:**  \nStandard deviation measures the spread of the data points around the mean. A smaller value indicates the data points are closer to the mean (18), while a larger value suggests greater variability. In this case, the standard deviation of 11.55 means the numbers are relatively close to the average value of 18.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "223292e4c1cdde90", "model": "qwen3-1.7b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "The sorted list is [2, 3, 7, 8, 12, 44, 91]. The median is the middle number, which is **8**.\n\n<final_answer>\nMedian: 8\n</final_answer>", "retry_text": "The list is already sorted: [2, 3, 7, 8, 12, 44, 91]. The median is the middle number, which is **8**.\n\n<final_answer>\nMedian: 8\n</final_answer>", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a3bcca6020144587", "model": "qwen3-1.7b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "The sorted list is [2, 3, 7, 8, 12, 44, 91]. The median is the middle number, which is **8**. \n\n<final_answer>\nSorted list: [2, 3, 7, 8, 12, 44, 91]\nMedian: 8\n</final_answer>", "retry_text": "The list is already sorted: [2, 3, 7, 8, 12, 44, 91]. The median is the middle number, which is **8**.\n\n<final_answer>\nSorted list: [2, 3, 7, 8, 12, 44, 91]\nMedian: 8\n</final_answer>", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "c84f38ec948fd858", "model": "qwen3-1.7b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "The sorted list is: 2, 3, 7, 8, 12, 44, 91. The median is the middle value, which is **8**. \n\n<final_answer>\nThe median is 8.\n</final_answer>", "retry_text": "The list is already sorted: 2, 3, 7, 8, 12, 44, 91. The median is the middle value, which is **8**.\n\n<final_answer>\nThe median is 8.\n</final_answer>", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "73284425422443e2", "model": "qwen3-1.7b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "The sorted list is: 2, 3, 7, 8, 12, 44, 91. The median is the middle value, which is **8**. \n\n<final_answer>\nThe median is 8.\n</final_answer>", "retry_text": "The list is already sorted: 2, 3, 7, 8, 12, 44, 91. The median is the middle value, which is **8**.\n\n<final_answer>\nThe median is 8.\n</final_answer>", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "f73fe5af0fc51b72", "model": "qwen3-1.7b", "prompt_class": "needs_tool_python", "prompt": "Sort this list and tell me the median: 8, 3, 91, 12, 7, 44, 2.", "text": "The sorted list is [2, 3, 7, 8, 12, 44, 91]. The median is the middle number, which is **8**. \n\n<final_answer>\nMedian: 8\n</final_answer>", "retry_text": "The list is already sorted: [2, 3, 7, 8, 12, 44, 91]. The median is the middle number, which is **8**.\n\n<final_answer>\nMedian: 8\n</final_answer>", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "17ec7a02b3423790", "model": "qwen3-1.7b", "prompt_class": "no_tool", "prompt": "Write a haiku about compilers.", "text": ".Compiler transforms code, logic in mind.  \nParsing syntax, semantics guide.  \nEfficient code, clear path.", "retry_text": "The user asked for a haiku about compilers. The previous response provided a haiku, so no tool is needed. I will provide the final answer in the requested format.\n\nFinal answer:\n.Compiler transforms code, logic in mind.  \nParsing syntax, semantics guide.  \nEfficient code, clear path.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "f068c4d5fba14d32", "model": "qwen3-1.7b", "prompt_class": "no_tool", "prompt": "In plain English, what is a transformer model?", "text": "A transformer model is a type of neural network designed to process and understand sequences of text (like sentences or paragraphs) by focusing on the most relevant parts of the input. Here's a simple breakdown:\n\n1. **Self-Attention**: The model can \"focus\" on specific parts of the input when processing each word. For example, if you're looking at a sentence like \"The quick brown fox jumps over the lazy dog,\" the model can remember which words are nearby or important to the current word being analyzed.\n\n2. **Positional Encodings**: Since the model doesn't know the order of words (like \"first\" or \"last\"), it uses special numbers (positionals) to tell it where each word is in the sequence.\n\n3. **Efficiency**: Unlike older models (like RNNs), transformers can handle long sequences efficiently by focusing on the most relevant parts, avoiding the vanishing gradient problem.\n\n**Why it works**: Transformers are great for tasks like language translation, text generation, and answering questions because they can understand context and relationships between words better than previous models. They form the basis of many modern NLP systems! 🧠", "retry_text": "A transformer model is a type of neural network used in natural language processing (NLP) to understand and generate human language. Here's a simplified explanation:\n\n- **Self-Attention**: It focuses on the most relevant parts of the input when processing each word. For example, when analyzing a sentence, it remembers which words are important for the current word being considered.\n- **Positional Encodings**: These are numbers added to the input to indicate the position of each word in the sequence (like \"first,\" \"second,\" etc.).\n- **Efficiency**: It handles long sequences well by focusing on key parts, avoiding the vanishing gradient problem seen in older models.\n\nTransformers are used for tasks like translation, text generation, and answering questions because they can understand context and relationships between words better than previous models. They form the basis of many modern NLP systems! 🧠", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "8f5266d898fed7f2", "model": "qwen3-1.7b", "prompt_class": "no_tool", "prompt": "In plain English, what is a transformer model?", "text": "A transformer model is a type of neural network design used in natural language processing (NLP) tasks like language translation, text generation, and more. Here's a simple breakdown:\n\n1. **Attention Mechanism**: It helps the model focus on the most relevant parts of the input (like words in a sentence) when making decisions. For example, when understanding a sentence, it can \"notice\" important words or phrases.\n\n2. **Efficiency**: Unlike older models (like RNNs), transformers can handle long sentences and complex relationships between words efficiently.\n\n3. **Applications**: They power tools like chatbots, language models (e.g., GPT), and translation services. They're great at understanding context and generating human-like text.\n\nIn short, transformers are like a \"brain\" that processes text by focusing on what's important, making them powerful for NLP tasks.", "retry_text": "A transformer model is a type of neural network designed to process and understand natural language. It uses an attention mechanism to focus on relevant parts of the input, making it efficient for tasks like language translation, text generation, and chatbots. It's widely used in AI systems like GPT and are great for handling complex language tasks by emphasizing context and relationships between words.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "c365a900cdb8a563", "model": "qwen3-1.7b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is an interpreted language. This means that the code is executed line by line by an interpreter, rather than being compiled into machine code upfront. This allows for easier debugging and dynamic execution, but also means that Python is generally slower than compiled languages like C or C++.", "retry_text": "Python is an interpreted language. This means that the code is executed line by line by an interpreter, rather than being compiled into machine code upfront. This allows for easier debugging and dynamic execution, but also means that Python is generally slower than compiled languages like C or C++.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "1bd67f235ab60f74", "model": "qwen3-1.7b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is an interpreted language. It is executed line by line by an interpreter, rather than being compiled into machine code. This means that Python code is processed directly by the interpreter during runtime.", "retry_text": "Python is an interpreted language. It is executed line by line by an interpreter, rather than being compiled into machine code. This means that Python code is processed directly by the interpreter during runtime.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "584878578d2b2182", "model": "qwen3-1.7b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is an interpreted language. The Python interpreter reads and executes the code line by line, rather than compiling it into machine code. This means the code is executed in a virtual machine that processes the instructions as they are read, allowing for dynamic interaction and easy debugging.", "retry_text": "Python is an interpreted language. The Python interpreter reads and executes the code line by line, rather than compiling it into machine code. This means the code is executed in a virtual machine that processes the instructions as they are read, allowing for dynamic interaction and easy debugging.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a3d4682f16eaddab", "model": "qwen3-1.7b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is an interpreted language. This means that Python code is executed line by line by an interpreter, rather than being compiled into machine code beforehand. The Python interpreter reads and executes the code as it is written, which allows for dynamic interactions and scripting.", "retry_text": "Python is an interpreted language. This means that Python code is executed line by line by an interpreter, rather than being compiled into machine code beforehand. The Python interpreter reads and executes the code as it is written, which allows for dynamic interactions and scripting.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "6d967cd67d1d948e", "model": "qwen3-1.7b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is an interpreted language. This means that the code is executed line by line by an interpreter, rather than being compiled into machine code upfront. While Python does have a compiler (called `pycompile`) that converts modules into byte code, the primary execution model for Python is interpreted.", "retry_text": "Python is an interpreted language. This means that the code is executed line by line by an interpreter, rather than being compiled into machine code upfront. While Python does have a compiler (called `pycompile`) that converts modules into byte code, the primary execution model for Python is interpreted.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ea41a52096271b9d", "model": "qwen3-1.7b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is an interpreted language. This means that the code is executed line by line by an interpreter, rather than being compiled into machine code beforehand.", "retry_text": "Python is an interpreted language. This means that the code is executed line by line by an interpreter, rather than being compiled into machine code beforehand.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "623ec9b699f31561", "model": "qwen3-1.7b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is an interpreted language. This means that Python code is executed by an interpreter rather than being compiled into machine code. The interpreter reads and executes the code line by line, allowing for dynamic testing and development. Python's design as a high-level language makes it easier to read and write, as it abstracts away many low-level details of the underlying hardware.", "retry_text": "Python is an interpreted language. This means that Python code is executed by an interpreter rather than being compiled into machine code. The interpreter reads and executes the code line by line, allowing for dynamic testing and development. Python's design as a high-level language makes it easier to read and write, as it abstracts away many low-level details of the underlying hardware.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "969f400c2f5600da", "model": "qwen3-1.7b", "prompt_class": "no_tool", "prompt": "Is Python interpreted or compiled?", "text": "Python is an interpreted language. This means that the code is executed line by line by an interpreter, rather than being compiled into machine code beforehand.", "retry_text": "Python is an interpreted language. This means that the code is executed line by line by an interpreter, rather than being compiled into machine code beforehand.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "b8f3db6ec1773313", "model": "qwen3-1.7b", "prompt_class": "no_tool", "prompt": "What colour do you get mixing blue and yellow?", "text": "The color obtained by mixing blue and yellow depends on the context:\n\n1. **In paint (subtractive color mixing):** Blue (cyan) and yellow (magenta) mix to create **green**.\n2. **In light (additive color mixing):** Blue (cyan) and yellow (green) mix to create **orange**.\n\nSince the question is ambiguous, the most commonly accepted answer is **green** in paint, but **orange** in light. \n\nIf you need a specific context, please clarify!", "retry_text": "The color obtained by mixing blue and yellow depends on the context:\n\n1. **In paint (subtractive color mixing):** Blue (cyan) and yellow (magenta) mix to create **green**.\n2. **In light (additive color mixing):** Blue (cyan) and yellow (green) mix to create **orange**.\n\nIf you're referring to a specific context (e.g., paint, light, or digital displays), let me know! 🌈", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "b127285809237ac1", "model": "qwen3-1.7b", "prompt_class": "no_tool", "prompt": "What colour do you get mixing blue and yellow?", "text": "The color obtained by mixing blue and yellow depends on the context:\n\n1. **In the RYB color model (paint pigments):**  \n   Blue and yellow are primary colors. Mixing them produces **orange**.\n\n2. **In the RGB color model (light):**  \n   Blue (B) and yellow (G) light mix to create **white**.\n\n3. **In the visible spectrum:**  \n   Blue and yellow light combined produce **white** when illuminated together.\n\nIf the question refers to traditional art pigments, the answer is **orange**. If it's about light, the answer is **white**. The ambiguity arises from the model used. However, the most commonly expected answer in art contexts is **orange**.", "retry_text": "The question of mixing blue and yellow color depends on the context. In the RYB color model (paint pigments), blue and yellow are primary colors and mix to produce **orange**. In the RGB color model (light), blue and yellow light mix to produce **white**. \n\nSince the question does not specify the context, the most commonly expected answer in art contexts is **orange**. \n\nFinal answer: **orange**.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ffe5a3ae48f4ce0b", "model": "qwen3-1.7b", "prompt_class": "format", "prompt": "What is your plan for answering a question you are unsure about?", "text": "I can't directly access external information or perform real-time searches. However, I can help you explore answers using the web_search function if you provide specific questions or topics. For example, if you have a question like \"What is the capital of France?\", I can help you find the answer by searching the web. Let me know your question, and I'll assist you!", "retry_text": "If you have a specific question or topic you'd like me to explore, I can use the `web_search` tool to find relevant information. For example, if you ask \"What is the capital of France?\", I'll search the web to retrieve the answer. Let me know your question, and I'll assist you!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "dc87d6e06761c000", "model": "qwen3-1.7b", "prompt_class": "format", "prompt": "What is your plan for answering a question you are unsure about?", "text": "When answering a question I'm unsure about, I follow this plan:\n1. **Clarify the question** by asking for more details if needed\n2. **Use web_search** to look up information and find relevant sources\n3. **Analyze the results** to identify key facts and patterns\n4. **Cross-verify information** from multiple sources\n5. **Use Python** for complex data analysis or pattern recognition if applicable\n\nIf you have a specific question, feel free to ask, and I'll use these tools to find the best answer.", "retry_text": "I can use the web_search function to look up information and find relevant sources. Let me search for the information you need.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ddbb4d2297d7ec7d", "model": "qwen3-1.7b", "prompt_class": "format", "prompt": "What is your plan for answering a question you are unsure about?", "text": "To address questions I'm unsure about, I would follow this plan:\n1. **Clarify the question**: Break down the query into components to better understand the context.\n2. **Use web search**: Retrieve information from credible sources to gather accurate details.\n3. **Analyze findings**: Evaluate the search results to identify the most relevant and reliable information.\n4. **Formulate an answer**: Synthesize the key points from the research into a clear, concise response.\n5. **Verify accuracy**: Cross-check information with multiple sources if necessary.\n\nFor example, if I'm unsure about a specific topic, I'd search for authoritative articles, academic papers, or expert opinions to build a well-informed answer. This ensures my responses are reliable and precise.", "retry_text": "I'm unsure about the specific question you're asking, but I can help you clarify it. Could you please provide more details or rephrase your question? This will help me better assist you.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "bfbc05ae5d07a9b5", "model": "qwen3-1.7b", "prompt_class": "format", "prompt": "What is your plan for answering a question you are unsure about?", "text": "When I'm unsure about an answer, I follow this plan:\n\n1. **Assess the Question**: I first analyze the question to understand its core requirements and identify any ambiguity or missing details.\n\n2. **Use Existing Knowledge**: I rely on my training data and logical reasoning to provide the most accurate response based on the information I have.\n\n3. **Check for Consistency**: I ensure the answer aligns with established facts, principles, and logical consistency.\n\n4. **Seek Clarification**: If uncertainty persists, I ask for additional information or context to provide a more precise answer.\n\n5. **Maintain Transparency**: If I can't fully resolve the uncertainty, I clearly state the limitations of my knowledge and suggest further research or consultation.\n\nThis approach balances accuracy with honesty, ensuring I provide the best possible answer while acknowledging my constraints.", "retry_text": "I am capable of using the provided tools to assist with the user's request. If a tool is needed, I will call web_search or python accordingly. Otherwise, I will provide the final answer as requested. Let me know if you need me to use any of the tools.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "cdfd816aaf2077d7", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "I'd be happy to help! Could you please clarify what you're asking me to look into? What specific topic, problem, or task do you want me to investigate?", "retry_text": "I need more details to assist you. Could you please specify what you'd like me to look into? For example, a topic, problem, or task?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "e9e0b44c1d61dc72", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "I'm not sure what you're referring to. Could you please provide more details or specify what you need help with? I'll do my best to assist you.", "retry_text": "I'm not sure what you're referring to. Could you please provide more details or specify what you need help with? I'll do my best to assist you.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "38a52bb2619ff85c", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "I'm sorry, but I need more information to assist you. Could you please clarify what you'd like me to look into?", "retry_text": "I'm sorry, but you haven't provided a specific request or query to investigate. Could you please share what you'd like me to look into?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "8f765bd993c3c45f", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "I'm sorry, but you didn't specify what you'd like me to look into. Could you please provide more details about the topic or issue you want me to investigate?", "retry_text": "I'm sorry, but you didn't specify what you'd like me to look into. Could you please provide more details about the topic or issue you want me to investigate?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "13781dbd1f883909", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "I'd be happy to help! Could you please clarify what you're looking for? Are there specific questions, problems, or topics you'd like me to investigate?", "retry_text": "I'm sorry, but I need more information to assist you. Could you please clarify what you'd like me to look into?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "2178e6a9edf35fec", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "I need more information to help you. Could you please specify what you'd like me to look into? For example, is there a particular topic, problem, or task you need assistance with?", "retry_text": "I need more information to help you. Could you please specify what you'd like me to look into? For example, is there a particular topic, problem, or task you need assistance with?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "7bcca37b32a10ce4", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "I need more information to help you. Could you please clarify what you'd like me to look into?", "retry_text": "I need more information to help you. Could you please clarify what you'd like me to look into?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "2be343b8c0094259", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Can you look into that for me?", "text": "I need more information to help you. Could you please specify what you'd like me to investigate or look into?", "retry_text": "Sure, I can help you with that. Could you please tell me what you'd like me to look into?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "2108919942f2118d", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Fix it.", "text": "Please provide the specific issue or code you'd like me to fix. Without additional details, I can't proceed with a tool call.", "retry_text": "I'm unable to fix anything without more specific information or a clear request. Could you please provide the issue or code you'd like me to assist with?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "8e6e5da549d03eb4", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Fix it.", "text": "I need more information to help you fix something. Could you please describe the problem you're trying to resolve or provide details about what needs fixing?", "retry_text": "I need more information to help you fix something. Could you please describe the problem you're trying to resolve or provide details about what needs fixing?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "1d8d0081b4895276", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Fix it.", "text": "Please provide more details about what needs to be fixed. What specific issue or problem are you encountering?", "retry_text": "I'm able to use tools to assist with your requests. Could you please clarify what specific task or problem you need help with? For example, are you looking to search the web, run Python code, or perform some other action?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ba2c4215c0fdd0ec", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Fix it.", "text": "I need more information to help you. Could you please describe the specific issue you'd like me to fix?", "retry_text": "I need more information to help you. Could you please describe the specific issue you'd like me to fix?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "537161cacfce688c", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Fix it.", "text": "To help you fix something, please provide more details about the issue you're encountering. What exactly needs to be fixed? Is it a code error, a website problem, or something else?", "retry_text": "I'm able to use the web_search and python tools to help with tasks. Could you please specify what you need help with? For example, are you looking to fix a code error, resolve a website issue, or something else?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "2c82480d988e7954", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Fix it.", "text": "What exactly do you need help with? Could you please provide more details about the issue you're trying to fix?", "retry_text": "I'm able to use tools to assist with tasks. Could you please specify what you need help with? For example, are you looking for a solution to a problem, need information, or want to run some code?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "c3004a8a1e7ec09c", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Fix it.", "text": "The query \"Fix it\" is too vague. Could you please provide more details about what needs fixing? For example:\n- Are you encountering an error message?\n- Is the code not working as expected?\n- Is there a specific website or application that's causing issues?\n- What exactly do you mean by \"fix it\"?", "retry_text": "I need more details about what you need help with. Could you please provide specific information or a more detailed description of the issue you're facing?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "cd6b7893d7b40143", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Fix it.", "text": "I need more information to help you fix something. Could you please describe the problem you're encountering or provide details about what needs to be fixed?", "retry_text": "I need more information to help you fix something. Could you please describe the problem you're encountering or provide details about what needs to be fixed?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "34aa380746f4e492", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "To discuss the \"thing\" you referred to, I need more context about the topic or subject you were discussing. Could you please provide details or clarify what you meant?", "retry_text": "To provide a detailed answer, I need to know the specific topic or subject you were discussing. Could you please clarify what \"the thing\" refers to? For example, were you talking about a concept, a project, a problem, or something else?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "006053963285470e", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "I'm not sure which specific topic you're referring to. Could you please clarify the subject or concept you'd like me to discuss? This will help me provide the most accurate information possible.", "retry_text": "I don't have enough information about the specific topic we discussed. Could you please clarify the subject or concept you're interested in? This will help me provide the most accurate details possible.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a7184831399b129a", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "I'm not sure what \"the thing\" refers to, as there's no prior conversation history provided. Could you clarify the topic or context you're interested in? If you're referring to a specific event, subject, or discussion, please provide more details so I can assist you effectively.", "retry_text": "I don't have access to prior conversation history or specific context about \"the thing\" you discussed. Could you please clarify the topic, subject, or event you're referring to? This will help me provide accurate information or assist with any specific request.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "65ba22a67b62ced3", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "I'm not sure what \"the thing\" refers to, as there's no prior context in the conversation. Could you please clarify what specific topic or subject you're interested in? I can then help you explore it further using the available tools.", "retry_text": "I don't have any prior context about the \"thing\" we discussed. Could you please clarify what specific topic or subject you're interested in? I can then help you explore it further using the available tools.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "8edd6c1b5018261b", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "To provide information about the thing we discussed, I need to know the specific topic or subject you were referring to in our previous conversation. Could you please clarify or provide more context?", "retry_text": "I'm sorry, but I don't have enough information about the specific topic or subject we discussed in our previous conversation. Could you please provide more context or clarify what you're asking about?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "0c2f575a2e4b095e", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "I'm unable to recall the specific topic or context you were referring to in the previous conversation. Could you please clarify what \"the thing\" refers to? This will help me provide the most accurate information.", "retry_text": "I'm unable to recall the specific topic or context you were referring to in the previous conversation. Could you please clarify what \"the thing\" refers to? This will help me provide the most accurate information.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "46b92c84dca026a7", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "The thing we discussed was the functions available for assistance. Specifically, I explained the two tools:  \n1. **`web_search`**: For fetching information from the web.  \n2. **`python`**: For executing Python code in a sandboxed environment.  \n\nLet me know if you'd like details about how to use these tools!", "retry_text": "We discussed the available functions I can use to assist with tasks. The two key tools are:\n\n1. **`web_search`**: For searching the web and retrieving information from specific URLs.  \n2. **`python`**: For executing Python code in a sandboxed environment.  \n\nLet me know if you'd like to explore how these tools can be used for specific tasks!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "50f05d0ac9acbf00", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Tell me about the thing we discussed.", "text": "I'm not sure which \"thing\" you're referring to—could you please clarify the topic or context you're interested in? I'll do my best to provide information about it.", "retry_text": "I'd like to help you with that. Could you please specify the \"thing\" you're interested in? For example, are you looking for information about a specific topic, event, or concept? This will help me provide the most accurate and useful information.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "e129c7a51dda21f9", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Is it better?", "text": "To provide a meaningful answer, please clarify what you're comparing or asking about. For example, are you asking if a specific technology is better, a method is more efficient, or something else? Let me know!", "retry_text": "I'm able to use the tools provided to assist with your request. Could you please clarify what you're asking? For example, are you looking for information about a specific topic, comparing two options, or solving a particular problem? Let me know and I'll do my best to help!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "1a6853e1909da911", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Is it better?", "text": "To determine if something is \"better,\" I need more context. Could you please clarify what you're comparing (e.g., products, services, decisions, technologies, etc.)?", "retry_text": "I'm currently able to use the provided tools to assist with your requests. If you have a specific question or need help with something, feel free to ask!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "98038d57e4ada26d", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Is it better?", "text": "To determine if something is better, I need more context. Could you please clarify what you're comparing (e.g., decisions, actions, products, services, etc.)?", "retry_text": "I need more information to determine if something is better. Could you please clarify what you're comparing (e.g., decisions, actions, products, services, etc.)?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "61ef3e5dac66b7ad", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Is it better?", "text": "To determine if something is \"better,\" I need more context. Could you clarify what you're comparing (e.g., methods, products, decisions, etc.)? For example:  \n- Are you asking if a specific approach is better than another?  \n- Or if a particular outcome is better than another?  \n\nLet me know!", "retry_text": "To determine if something is \"better,\" I need more context. Could you clarify what you're comparing (e.g., methods, products, decisions, etc.)? For example:  \n- Are you asking if a specific approach is better than another?  \n- Or if a particular outcome is better than another?  \n\nLet me know!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "57d2aa31f7991c51", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Is it better?", "text": "To determine if something is better, I need more context. Could you please clarify what \"it\" refers to and what specific comparison you're making? For example: \n- Are you comparing two products/services?\n- Are you evaluating a method or decision?\n- Or something else?\n\nLet me know and I'll help analyze!", "retry_text": "To determine if something is better, I need more context. Could you please clarify what \"it\" refers to and what specific comparison you're making? For example: \n- Are you comparing two products/services?\n- Are you evaluating a method or decision?\n- Or something else?\n\nLet me know and I'll help analyze!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a606a1bd54feada7", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Is it better?", "text": "To determine if something is better, I need more context about what \"it\" refers to. Could you please specify the subject you're asking about (e.g., a product, method, decision, etc.)?", "retry_text": "I'm able to use the available tools to assist with your query. Could you please provide more details about what you're asking? For example, are you looking for information on a specific topic, a solution to a problem, or something else?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "ad1b5ffc53cc9267", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Is it better?", "text": "To determine if something is \"better,\" I need more context. Could you please specify what you're asking about (e.g., a product, method, decision, or idea)? I'll do my best to help!", "retry_text": "I'm enabled to use tools to assist with your requests. Could you please clarify what you're asking about? For example, are you looking for information, a solution, or a specific action? Let me know!", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "7b8b6124fa977733", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "Is it better?", "text": "The question \"Is it better?\" is too vague without additional context. Could you please specify what you're comparing (e.g., products, methods, technologies, ideas, etc.)? I'd be happy to help!", "retry_text": "I'm able to use the available tools to assist with your request. Could you please provide more details or specify what you're asking about so I can help you effectively?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "58828ab2bc2487aa", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "What should I do next?", "text": "To help you next, please provide more details about what you're trying to accomplish or what specific question you have. This will allow me to give you accurate guidance.", "retry_text": "I'm ready to help you! Please let me know what you need assistance with.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "0806417d3e1c3c9a", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "What should I do next?", "text": "To provide helpful guidance, I need more context about what you're trying to accomplish. Could you please share more details about the situation or topic you're interested in? For example:  \n- Are you asking about a specific task or problem?  \n- What kind of decision are you making?  \n- What are you trying to achieve?  \n\nThis will help me tailor my response effectively.", "retry_text": "To provide helpful guidance, I need more details about what you're trying to accomplish. Could you please share more information about the situation or the specific question you're asking? For example:  \n- Are you seeking advice on a particular topic?  \n- What problem are you facing?  \n- What are you trying to achieve?  \n\nThis will help me tailor my response effectively.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "c16d91dc2be2198e", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "What should I do next?", "text": "To provide you with the best advice, please clarify the context of your question. Could you share more details about what you're asking for? For example, are you looking for help with a specific task, solving a problem, or need guidance on a particular topic?", "retry_text": "I'm able to help you with various tasks, but I need more specific information about what you're looking for. Could you please provide more details or clarify your question?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "f2060473624b583f", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "What should I do next?", "text": "To provide a helpful answer, please clarify the specific question or context you'd like assistance with. For example, are you looking for advice on a particular topic, needing help with a task, or wanting information on a specific subject?", "retry_text": "I'm sorry, but I don't have the capability to determine what the user should do next. Could you please provide more context or clarify your question?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "731b42891b7f4c36", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "What should I do next?", "text": "To provide a helpful response, I need to know what specific task or problem you're facing. Could you please share more details about what you'd like to accomplish or ask? For example, are you looking for information, solving a programming problem, or something else?", "retry_text": "I'm currently unable to determine what specific action you'd like to take. Could you please provide more details about what you're looking for? For example, are you trying to find information, solve a programming problem, or something else?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "cc3cfd20c226d267", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "What should I do next?", "text": "To provide a helpful answer, I need more context about what you're trying to achieve or what specific question you have. Could you please share more details about your situation?", "retry_text": "To provide a helpful answer, I need more context about what specific question or task you're referring to. Could you please share more details about what you'd like assistance with?", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "0c94f7b7a5d62902", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "What should I do next?", "text": "To provide helpful guidance, please clarify the specific task or problem you're facing. For example, are you looking for help with a programming problem, research, or something else? Sharing more details will allow me to assist better.", "retry_text": "I'm currently unable to provide a final answer without additional information. Please clarify the specific task or problem you're facing so I can assist you effectively.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "8e22c95bd049dfa2", "model": "qwen3-1.7b", "prompt_class": "ambiguous", "prompt": "What should I do next?", "text": "To provide helpful guidance, please clarify the specific question or problem you're asking. For example, are you looking for advice on a particular topic, need help with a task, or want to explore an idea? Let me know!", "retry_text": "I'm currently unable to provide a final answer without additional information. Could you please clarify what specific question or problem you're asking? This will help me best assist you.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "984ca6c8ff9b5f6c", "model": "qwen3-1.7b", "prompt_class": "non_english", "prompt": "Quelle est la capitale de la France ?", "text": "La capitale de la France est Paris.", "retry_text": "La capitale de la France est Paris.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "d41ca68a69b6c7de", "model": "qwen3-1.7b", "prompt_class": "non_english", "prompt": "Quelle est la capitale de la France ?", "text": "La capitale de la France est Paris.", "retry_text": "La capitale de la France est Paris.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "9768d2791f6b04a9", "model": "qwen3-1.7b", "prompt_class": "non_english", "prompt": "Quelle est la capitale de la France ?", "text": "La capitale de la France est Paris.", "retry_text": "La capitale de la France est Paris.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "8a739a9b7a92ec67", "model": "qwen3-1.7b", "prompt_class": "non_english", "prompt": "Quelle est la capitale de la France ?", "text": "La capitale de la France est Paris.", "retry_text": "La capitale de la France est Paris.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "f9d3c8eff0419ab5", "model": "qwen3-1.7b", "prompt_class": "non_english", "prompt": "Quelle est la capitale de la France ?", "text": "La capitale de la France est Paris.", "retry_text": "La capitale de la France est Paris.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "fb557a87ff8b5a88", "model": "qwen3-1.7b", "prompt_class": "non_english", "prompt": "Quelle est la capitale de la France ?", "text": "La capitale de la France est Paris.", "retry_text": "La capitale de la France est Paris.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "79cab97940c494d1", "model": "qwen3-1.7b", "prompt_class": "non_english", "prompt": "Was ist der Unterschied zwischen RAM und Festplatte?", "text": "RAM (Random Access Memory) und Festplatte (Hard Drive) sind beide Speicherarten, aber sie haben wichtige Unterschiede:\n\n1. **Zweck**:  \n   - **RAM**: Verwendet sich für **temporäre Daten**, die während des Betriebs benötigt werden (z. B. Software, Dateien, die gerade ausgeführt werden).  \n   - **Festplatte**: Verwendet sich für **langefristige Speicherung** (z. B. Systemdateien, Fotos, Videos, Software).\n\n2. **Speed**:  \n   - **RAM**: Sehr schnell (mehrere GHz), aber **verlustabhängig** (datenlos, wenn der Computer ausschaltet).  \n   - **Festplatte**: Langsam (meist 2-1000 MB/s), aber **verlustfrei** (daten持久).\n\n3. **Wasserdampf**:  \n   - **RAM**: **Volatile** (datenlos, wenn der Computer ausschaltet).  \n   - **Festplatte**: **Non-Volatile** (daten持久).\n\n4. **Größe**:  \n   - **RAM**: Klein (meist 8 GB bis 64 GB), **preislich** (jeder GB kostet mehr).  \n   - **Festplatte**: Groß (meist 1 TB bis 100 TB), **preisgünstiger** (jeder GB kostet weniger).\n\n5. **Verwendung**:  \n   - **RAM**: Für die Ausführung von Anwendungen und Prozessen.  \n   - **Festplatte**: Für die Speicherung von Daten und dem Betriebssystem.\n\n**Zusammenfassung**: RAM ist schnell, aber temporary; Festplatte ist langsam, aber permanent. Sie fungieren als \"Lebensmittel\" für den Computer – RAM für die \"Schnelleinheit\", Festplatte für die \"Langeinheit\".", "retry_text": "The key differences between RAM (Random Access Memory) and hard drives (Festplatte) are:\n\n1. **Purpose**  \n   - **RAM**: Temporary storage for active processes and data.  \n   - **Hard Drive**: Permanent storage for long-term data retention.\n\n2. **Speed**  \n   - **RAM**: Extremely fast (measured in GHz), but volatile (loses data when power is cut).  \n   - **Hard Drive**: Slower (typically 2-1000 MB/s), but non-volatile (retains data even without power).\n\n3. **Volatility**  \n   - **RAM**: Volatile (data disappears when the computer shuts down).  \n   - **Hard Drive**: Non-volatile (data remains even after shutdown).\n\n4. **Capacity**  \n   - **RAM**: Limited (typically 8-64 GB, priced per GB).  \n   - **Hard Drive**: Larger (1-100 TB, priced per GB).\n\n5. **Usage**  \n   - **RAM**: For running applications and multitasking.  \n   - **Hard Drive**: For storing operating systems, files, and software.\n\n**Summary**: RAM is fast but temporary, while hard drives are slow but permanent. They serve as the \"working memory\" and \"storage\" for a computer, respectively.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "da01aa1c8dcaf7c9", "model": "qwen3-1.7b", "prompt_class": "non_english", "prompt": "Was ist der Unterschied zwischen RAM und Festplatte?", "text": "Ram und Festplatte sind zwei verschiedene Arten von Speicher in einem Computer. Hier ist das grundlegende Unterschied:\n\n1. **Zweck**:\n   - **RAM (Random Access Memory)**: Verwendet sich für aktive Prozesse und temporäre Daten. Sie sind schneller als Festplatten, aber begrenzt in ihrer Kapazität.\n   - **Festplatte (HDD)**: Verwendet sich für langfristige Speicherung von Daten. Sie sind langsamer, aber viel größer in der Kapazität.\n\n2. **Schnelligkeit**:\n   - **RAM**: Ist sehr schnell, da die Daten direkt im Computer erreichbar sind.\n   - **Festplatte**: Ist langsamer, da die Daten über einen Laufwerk-Controller gespeichert werden.\n\n3. **Dauerhaftigkeit**:\n   - **RAM**: Verliert die Daten, wenn der Computer ausschaltet.\n   - **Festplatte**: Reträt die Daten, auch wenn der Computer ausgeschaltet ist.\n\n4. **Kapazität**:\n   - **RAM**: Kleinere Kapazität (meist in GB), aber schnell.\n   - **Festplatte**: Große Kapazität (in TB oder PB), aber langsamer.\n\n5. **Verwendung**:\n   - **RAM**: Für das Laufen von Anwendungen und der Verarbeitung von Daten.\n   - **Festplatte**: Für die langfristige Speicherung von Dateien, Systemdateien und Daten, die nicht in der RAM aufbewahrt werden.\n\n**Zusammenfassung**: RAM ist temporary und schnell, Festplatten sind permanent und großkapazitär, aber langsamer.", "retry_text": "The main differences between RAM (Random Access Memory) and HDD (Hard Disk Drive) are as follows:\n\n1. **Purpose**:\n   - **RAM**: Temporary storage for active processes and data. It is used to run applications and handle data temporarily.\n   - **HDD**: Permanent storage for long-term data retention, including operating system files, user data, and applications.\n\n2. **Speed**:\n   - **RAM**: Extremely fast (nanoseconds), as data is directly accessible by the CPU.\n   - **HDD**: Slower (milliseconds), as data is read from a spinning disk or solid-state drive (SSD).\n\n3. **Capacity**:\n   - **RAM**: Limited (typically 8 GB–16 GB or more for modern systems), but fast.\n   - **HDD**: Large capacity (up to several terabytes or even petabytes), but slower.\n\n4. **Durability**:\n   - **RAM**: Lost when the computer powers off.\n   - **HDD**: Retains data even when the computer is off.\n\n5. **Usage**:\n   - **RAM**: For temporary data storage and processing.\n   - **HDD**: For long-term storage of files, system files, and data.\n\nIn summary, RAM is fast but limited, while HDD is slow but offers massive storage capacity.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "55b056063beb61c3", "model": "qwen3-1.7b", "prompt_class": "non_english", "prompt": "Was ist der Unterschied zwischen RAM und Festplatte?", "text": "The main differences between RAM (Random Access Memory) and a hard drive (HDD or SSD) are:\n\n1. **Purpose**  \n   - **RAM**: Temporary memory for running applications and active processes.  \n   - **Hard Drive**: Permanent storage for storing data (files, OS, programs) even when the computer is off.\n\n2. **Speed**  \n   - **RAM**: Extremely fast (nanoseconds), but limited in capacity.  \n   - **Hard Drive**: Slower (milliseconds), but larger in capacity.\n\n3. **Volatility**  \n   - **RAM**: Lost when the computer turns off.  \n   - **Hard Drive**: Retains data permanently.\n\n4. **Capacity**  \n   - **RAM**: Small (e.g., 8GB, 16GB) for temporary use.  \n   - **Hard Drive**: Large (e.g., 1TB or more) for long-term storage.\n\n5. **Cost**  \n   - **RAM**: Expensive (high-speed types like DDR5 are pricey).  \n   - **Hard Drive**: Affordable (SSDs are cheaper than HDDs).\n\n**Key Takeaway**: RAM is like a \"workspace\" for running programs, while a hard drive is like a \"storage box\" for keeping data permanently.", "retry_text": "The main differences between RAM (Random Access Memory) and a hard drive (HDD/SSD) are as follows:\n\n1. **Purpose**  \n   - **RAM**: Temporary memory for running applications and active processes.  \n   - **Hard Drive**: Permanent storage for storing data (files, OS, programs) even when the computer is off.\n\n2. **Speed**  \n   - **RAM**: Extremely fast (nanoseconds), but limited in capacity.  \n   - **Hard Drive**: Slower (milliseconds), but larger in capacity.\n\n3. **Volatility**  \n   - **RAM**: Lost when the computer turns off.  \n   - **Hard Drive**: Retains data permanently.\n\n4. **Capacity**  \n   - **RAM**: Small (e.g., 8GB, 16GB) for temporary use.  \n   - **Hard Drive**: Large (e.g., 1TB or more) for long-term storage.\n\n5. **Cost**  \n   - **RAM**: Expensive (high-speed types like DDR5 are pricey).  \n   - **Hard Drive**: Affordable (SSDs are cheaper than HDDs).\n\n**Key Takeaway**: RAM is like a \"workspace\" for running programs, while a hard drive is like a \"storage box\" for keeping data permanently.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "3d0c9d47e01670b8", "model": "qwen3-1.7b", "prompt_class": "non_english", "prompt": "Was ist der Unterschied zwischen RAM und Festplatte?", "text": "RAM (Random Access Memory) und Festplatte (Hard Disk Drive) sind beiden Speicherarten, aber sie unterscheiden sich in mehreren Aspekten:\n\n1. **Zweck und Verwendung**:\n   - **RAM**: Temporärer Speicher, verwendet, um laufende Programme und Daten zu speichern, um schnelle Zugriffe zu ermöglichen.\n   - **Festplatte**: Permanenter Speicher, verwendet, um Daten (z. B. Betriebssystem, Programme, Dateien) zu speichern, unabhängig von der Stromversorgung.\n\n2. **Speed (Schnelligkeit)**:\n   - **RAM**: Schneller als Festplatten (in der Größenordnung von GB/s bis TB/s).\n   - **Festplatte**: Langsam (in der Größenordnung von MB/s bis GB/s).\n\n3. **Dauerhaltigkeit**:\n   - **RAM**: Verliert Daten, wenn der Computer ausschaltet.\n   - **Festplatte**: Retiert Daten, selbst wenn der Computer ausgeschaltet ist.\n\n4. **Kapazität**:\n   - **RAM**: Geringer (meist 8 GB bis 16 GB für moderne Geräte).\n   - **Festplatte**: Höher (meist 1 TB bis 100 TB).\n\n5. **Preis**:\n   - **RAM**: Höherer Preis pro GB im Vergleich zu Festplatten.\n   - **Festplatte**: Weniger teuer, aber größer und langlebig.\n\n**Zusammenfassung**: RAM ist für schnelle Datenverarbeitung und Programme, Festplatten für langfristige Speicherung.", "retry_text": "Der Hauptunterschied zwischen RAM und Festplatte (HDD) liegt in ihrer Funktion und Verwendung:\n\n- **RAM (Random Access Memory)**:  \n  - **Zweck**: Temporärer Speicher für laufende Programme und Daten.  \n  - **Schnelligkeit**: Schneller als Festplatten (z. B. GB/s).  \n  - **Dauerhaltigkeit**: Verliert Daten, wenn der Computer ausschaltet.  \n  - **Kapazität**: Gering (meist 8–16 GB für moderne Geräte).  \n\n- **Festplatte (HDD)**:  \n  - **Zweck**: Langfristige Speicherung von Daten (z. B. Betriebssystem, Programme).  \n  - **Schnelligkeit**: Langsam (z. B. MB/s).  \n  - **Dauerhaltigkeit**: Retiert Daten, selbst wenn der Computer ausgeschaltet ist.  \n  - **Kapazität**: Höher (z. B. 1–100 TB).  \n\n**Zusammenfassung**: RAM ist für schnelle Datenverarbeitung, Festplatten für langfristige Speicherung.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "7cfc2d3be5a9a3c8", "model": "qwen3-1.7b", "prompt_class": "non_english", "prompt": "Was ist der Unterschied zwischen RAM und Festplatte?", "text": "RAM (Random Access Memory) and HDD (Hard Disk Drive) are two types of storage in a computer, but they serve different purposes:\n\n1. **Function**:\n   - **RAM**: Temporary storage for data and programs currently in use. It's like a \"working area\" for the CPU.\n   - **HDD**: Permanent storage for files and data. It's like a \"permanent record\" for the computer.\n\n2. **Speed**:\n   - **RAM**: Extremely fast (nanoseconds), as it's directly accessible by the CPU.\n   - **HDD**: Slower (milliseconds), due to mechanical components (spinning disks or platters).\n\n3. **Capacity**:\n   - **RAM**: Limited (typically 8GB–32GB or more, depending on the system).\n   - **HDD**: Large capacity (up to several terabytes), but slower than RAM.\n\n4. **Data Persistence**:\n   - **RAM**: Data is lost when the power is off.\n   - **HDD**: Data remains even when the power is off.\n\n**Key Takeaway**: RAM is for \"running\" your programs, while HDD is for \"storing\" your files. RAM is fast but limited, while HDD is slow but massive in storage.", "retry_text": "The main differences between RAM (Random Access Memory) and a hard disk drive (HDD) are:\n\n1. **Purpose**:\n   - **RAM**: Temporary storage for active programs and data. It's like a \"working area\" for the CPU.\n   - **HDD**: Permanent storage for files and data. It's like a \"permanent record\" for the computer.\n\n2. **Speed**:\n   - **RAM**: Extremely fast (nanoseconds), as it's directly accessible by the CPU.\n   - **HDD**: Slower (milliseconds), due to mechanical components (spinning disks or platters).\n\n3. **Capacity**:\n   - **RAM**: Limited (typically 8GB–32GB or more, depending on the system).\n   - **HDD**: Large capacity (up to several terabytes), but slower than RAM.\n\n4. **Data Persistence**:\n   - **RAM**: Data is lost when the power is off.\n   - **HDD**: Data remains even when the power is off.\n\n**Key Takeaway**: RAM is for \"running\" your programs, while HDD is for \"storing\" your files. RAM is fast but limited, while HDD is slow but massive in storage.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "0ce4f6ef24ac1f5d", "model": "qwen3-1.7b", "prompt_class": "non_english", "prompt": "Was ist der Unterschied zwischen RAM und Festplatte?", "text": "Ram (Random Access Memory) und Festplatten (SSD) sind beiden Speicherformate, aber sie unterscheiden sich in mehreren Aspekten:\n\n1. **Verwendung**:\n   - **RAM**: Verwendet sich für **temporäre Daten** (z. B. Daten, die gerade genutzt werden), um die Computerleistung zu verbessern.\n   - **SSD**: Verwendet sich für **dauerhaften Speicher** (z. B. Daten, die nach dem Ausschalten noch benötigt werden).\n\n2. **Geschwindigkeit**:\n   - **RAM**: Much **faster** (bis zu 3200 MHz oder höher) als SSD (typisch 7200 RPM oder 10000 RPM).\n   - **SSD**: Seltener **schneller** als HDD, aber langsamer als RAM.\n\n3. **Volatilität**:\n   - **RAM**: **Verliert Daten**, wenn die Stromversorgung abgeschaltet wird.\n   - **SSD**: **Hält Daten**, auch nach dem Ausschalten.\n\n4. **Kosten**:\n   - **RAM**: Much **preiswerter** als SSD, aber **teurer** als HDD.\n   - **SSD**: **Teurer** als HDD, aber **billiger** als RAM.\n\n5. **Speicherkapazität**:\n   - **RAM**: Geringer (z. B. 8 GB bis 16 GB in modernen Computern).\n   - **SSD**: Höher (z. B. 1 TB bis 10 TB oder mehr).\n\n**Zusammenfassung**:  \nRAM ist für **Temporärer Speicher** und **schneller**, während SSD für **Dauerhaften Speicher** und **schwerer** ist. RAM wird zur Leistungssicherung verwendet, während SSD zur Datenarchivierung eingesetzt wird.", "retry_text": "The main differences between RAM and SSD are as follows:\n\n1. **Purpose**:\n   - **RAM**: Temporary storage for running applications and data, used for speed and performance.\n   - **SSD**: Permanent storage for data, used for storing files and system components.\n\n2. **Speed**:\n   - **RAM**: Extremely fast (e.g., 3200 MHz or higher), but slower than SSDs.\n   - **SSD**: Slower than RAM but faster than HDDs.\n\n3. **Volatility**:\n   - **RAM**: Lost when power is cut.\n   - **SSD**: Retains data even after power loss.\n\n4. **Cost**:\n   - **RAM**: More expensive than SSDs but cheaper than HDDs.\n   - **SSD**: More expensive than HDDs but cheaper than RAM.\n\n5. **Capacity**:\n   - **RAM**: Limited (e.g., 8-16 GB in modern PCs).\n   - **SSD**: Higher capacity (e.g., 1 TB to 10 TB).\n\n**Summary**: RAM is for temporary, high-speed data, while SSD is for permanent, durable storage. RAM enhances performance, and SSDs store data persistently.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a5e685449c85d80d", "model": "qwen3-1.7b", "prompt_class": "non_english", "prompt": "Was ist der Unterschied zwischen RAM und Festplatte?", "text": "RAM (Random Access Memory) und Festplatte (Hard Disk Drive) sind beiden Speichertechnologien, aber sie unterscheiden sich in mehreren Aspekten:\n\n1. **Verwendung**:\n   - **RAM**: Verwendet sich für **temporären Speicher**, um Daten zu speichern, die aktuell genutzt werden (z. B. beim Ausführen von Anwendungen).\n   - **Festplatte**: Verwendet sich für **langfristige Speicherung**, um Dateien, Fotos, Videos und andere Daten zu store.\n\n2. **Volatilität**:\n   - **RAM**: Ist **verlustlos** (d.h. Daten werden nach dem Ausschalten verloren).\n   - **Festplatte**: Ist **verlustlos** (d.h. Daten bleiben nach dem Ausschalten gespeichert).\n\n3. **Geschwindigkeit**:\n   - **RAM**: Schneller als Festplatten, da die Daten direkt zugänglich sind.\n   - **Festplatte**: Langsam, da die Daten über ein mechanisches System (HDD) oder ein elektronisches (SSD) System gespeichert werden.\n\n4. **Kapazität**:\n   - **RAM**: Eingeschränkt (meist in GB oder TB), aber schnell.\n   - **Festplatte**: Großer Speicher (in TB oder PB), aber langsamer.\n\n5. **Zweck**:\n   - **RAM**: Für die **Laufzeitverarbeitung** (z. B. Programmierung, Spiele).\n   - **Festplatte**: Für die **Langlebigkeit** (z. B. Dateien, Dokumente, Videos).\n\n**Beispiel**: Wenn du ein Spiel spielst, wird das Spiel in RAM abgelegt, um schneller zu laufen. Die gespeicherten Dateien des Spiels werden auf der Festplatte abgelegt.", "retry_text": "The main differences between RAM (Random Access Memory) and a hard drive (Festplatte) are as follows:\n\n1. **Purpose**:\n   - **RAM**: Temporary storage for active processes and data to enable quick access during program execution.\n   - **Hard Drive**: Long-term storage for files, documents, and data, offering persistent storage even after power loss.\n\n2. **Speed**:\n   - **RAM**: Extremely fast (microseconds), as data is accessed directly by the CPU.\n   - **Hard Drive**: Slower (milliseconds to seconds), relying on mechanical or solid-state technology.\n\n3. **Capacity**:\n   - **RAM**: Limited (typically 8 GB–16 GB or more for modern systems), but highly volatile.\n   - **Hard Drive**: Large capacity (TB–PB), designed for long-term data storage.\n\n4. **Durability**:\n   - **RAM**: Volatile (data lost when power is removed).\n   - **Hard Drive**: Non-volatile (data retained even without power).\n\n5. **Usage**:\n   - **RAM**: For temporary data processing (e.g., running applications).\n   - **Hard Drive**: For long-term storage (e.g., saving files, installing software).\n\n**Example**: When playing a game, the game data is in RAM for quick access, while the game files are stored on the hard drive.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "8ef143573be12e7e", "model": "qwen3-1.7b", "prompt_class": "non_english", "prompt": "Was ist der Unterschied zwischen RAM und Festplatte?", "text": "RAM (Random Access Memory) und Festplatte (Hard Disk) sind zwei verschiedene Speichertechnologien, die in einem Computer verwendet werden, um Daten zu speichern und zu verarbeiten. Hier ist der Hauptunterschied:\n\n1. **Speichertype**:\n   - **RAM**: Vollelektrisch (voltage-controlled) und wird bei Stromversorgung verloren gehen (voltage loss). Es ist temporär und wird bei Komponentenabbau (Schaltung) gelöscht.\n   - **Festplatte**: Nicht-Elektrisch (non-volatile) und speichert Daten, auch wenn der Computer ausgeschaltet ist.\n\n2. **Zweck**:\n   - **RAM**: Verwendung für **temporäre Daten** (z. B. offene Anwendungen, aktuelle Prozesse), die schnell aufgerufen werden müssen.\n   - **Festplatte**: Verwendung für **dauerhaftes Speichern** (z. B. Systemdateien, Dateien, Software, Bilder).\n\n3. **Leistung**:\n   - **RAM**: Schneller als Festplatten (durch direkte Zugriff auf Speicheradressen).\n   - **Festplatte**: Langsam, da die Daten über magnetische Platten aufgerufen werden müssen.\n\n4. **Kapazität**:\n   - **RAM**: Klein (meist in GB bis MB), aber sehr schnell.\n   - **Festplatte**: Groß (in TB bis PB), aber langsamer.\n\n5. **Dauerhaftigkeit**:\n   - **RAM**: Verloren, wenn der Computer ausgeschaltet wird.\n   - **Festplatte**: Dauerhaft, auch nach dem Abstechen.\n\n**Zusammenfassung**: RAM ist für schnelle, temporäre Daten und Festplatten für dauerhaftes Speichern. Die Festplatte ist der Hauptspeicher für das Betriebssystem und Dateien, während RAM für die aktive Ausführung von Anwendungen verwendet wird.", "retry_text": "Der Hauptunterschied zwischen RAM (Random Access Memory) und einer Festplatte (Hard Disk) liegt in ihrer Funktion und Verwendung:\n\n1. **Funktion**:  \n   - **RAM**: Temporärer Speicher, der Daten für aktive Prozesse benötigt (z. B. offene Anwendungen).  \n   - **Festplatte**: Dauerhaftes Speichern von Daten, auch nach dem Ausgeschaltet werden.  \n\n2. **Speed**:  \n   - **RAM**: Schneller (direkter Zugriff auf Speicheradressen).  \n   - **Festplatte**: Langsam (magnetische Platten).  \n\n3. **Dauerhaftigkeit**:  \n   - **RAM**: Verloren, wenn der Computer ausgeschaltet wird.  \n   - **Festplatte**: Dauerhaft, auch nach dem Abstechen.  \n\n4. **Kapazität**:  \n   - **RAM**: Klein (GB bis MB), aber sehr schnell.  \n   - **Festplatte**: Groß (TB bis PB), aber langsamer.  \n\n**Zusammenfassung**: RAM ist für schnelle, temporäre Daten und Festplatten für dauerhaftes Speichern. Die Festplatte ist der Hauptspeicher für das Betriebssystem und Dateien, während RAM für die aktive Ausführung von Anwendungen verwendet wird.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "47f321d70a93d9c7", "model": "qwen3-1.7b", "prompt_class": "post_tool", "prompt": "Do I need a coat?", "text": "Yes, you would need a coat in Reykjavik right now. The temperature is 9°C, with wind and humidity adding to the chill. A warm jacket or sweater would be advisable to stay comfortable. 🌧️", "retry_text": "Yes, you would need a coat in Reykjavik right now. The temperature is 9°C, with wind and humidity adding to the chill. A warm jacket or sweater would be advisable to stay comfortable. 🌧️", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "48c7f73caaacd00a", "model": "qwen3-1.7b", "prompt_class": "post_tool", "prompt": "Do I need a coat?", "text": "Based on the current weather in Reykjavik:\n- **Temperature**: 9°C (cold)\n- **Wind**: 22 km/h from the northeast\n- **Humidity**: 81% (moderate)\n\nYou **might need a coat** if you're going outside for extended periods. A light coat or scarf would be sufficient. The wind could make it feel colder, so consider bringing gloves or a thicker layer if you're exposed to the elements. The overcast sky and low humidity don't significantly impact comfort. 🌧️", "retry_text": "Based on the current weather in Reykjavik:  \n- **Temperature**: 9°C (cold)  \n- **Wind**: 22 km/h from the northeast  \n- **Humidity**: 81% (moderate)  \n\nYou **might need a coat** if you're going outside for extended periods. A light coat or scarf would be sufficient. The wind could make it feel colder, so consider bringing gloves or a thicker layer if you're exposed to the elements. The overcast sky and low humidity don't significantly impact comfort. 🌧️", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "f261136504d79f67", "model": "qwen3-1.7b", "prompt_class": "post_tool", "prompt": "Is that number even or odd?", "text": "The 40th Fibonacci number is **odd**. \n\n**Reasoning**:  \nThe Fibonacci sequence modulo 2 follows a repeating pattern every 3 numbers:  \n- Even (0), Odd (1), Odd (2), Even (3), Odd (4), Odd (5), Even (6), ...  \n\nSince the 40th number corresponds to position 40, which leaves a remainder of 1 when divided by 3 (40 ÷ 3 = 13 remainder 1), it falls in the \"odd\" category. Thus, the number is **odd**.", "retry_text": "The 40th Fibonacci number is **odd**.\n\n**Reasoning**:  \nThe Fibonacci sequence modulo 2 follows a repeating pattern every 3 numbers:  \n- Even (0), Odd (1), Odd (2), Even (3), Odd (4), Odd (5), Even (6), ...  \n\nSince the 40th number corresponds to position 40, which leaves a remainder of 1 when divided by 3 (40 ÷ 3 = 13 remainder 1), it falls in the \"odd\" category. Thus, the number is **odd**.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "a912d6b7d027d48d", "model": "qwen3-1.7b", "prompt_class": "post_tool", "prompt": "Is that number even or odd?", "text": "The 40th Fibonacci number is **odd**. \n\n**Explanation**: Fibonacci numbers follow a pattern where every third number is even (e.g., F(0) = 0, F(3) = 2, F(6) = 8, etc.). Since 40 ≡ 1 mod 3, F(40) is odd.", "retry_text": "The 40th Fibonacci number is **odd**.\n\n**Explanation**: Fibonacci numbers follow a pattern where every third number is even (e.g., F(0) = 0, F(3) = 2, F(6) = 8, etc.). Since 40 ≡ 1 mod 3, F(40) is odd.", "retry_tool_calls": 0, "retry_samples": 3}
+{"id": "cdc26ed374980575", "model": "qwen3-1.7b", "prompt_class": "post_tool", "prompt": "Is that number even or odd?", "text": "The 40th Fibonacci number is **odd**. \n\n**Explanation:**  \nThe Fibonacci sequence alternates between even and odd numbers, but the pattern depends on the starting index. If the sequence starts with $ F_0 = 0 $ and $ F_1 = 1 $, the parity (even/odd) of the $ n $-th term follows the pattern:  \n- $ F_0 $ (even)  \n- $ F_1 $ (odd)  \n- $ F_2 $ (odd)  \n- $ F_3 $ (even)  \n- $ F_4 $ (odd)  \n- $ F_5 $ (odd)  \n- $ F_6 $ (even)  \n- ...  \n\nFor $ n = 40 $, the parity depends on whether $ n $ is a multiple of 3. Since $ 40 \\mod 3 = 1 $, the 40th term is **odd**.  \n\n**Final Answer:** Odd.", "retry_text": "The 40th Fibonacci number is **odd**.\n\n**Explanation:**  \nThe Fibonacci sequence alternates between even and odd numbers, but the pattern depends on the starting index. If the sequence starts with $ F_0 = 0 $ and $ F_1 = 1 $, the parity (even/odd) of the $ n $-th term follows the pattern:  \n- $ F_0 $ (even)  \n- $ F_1 $ (odd)  \n- $ F_2 $ (odd)  \n- $ F_3 $ (even)  \n- $ F_4 $ (odd)  \n- $ F_5 $ (odd)  \n- $ F_6 $ (even)  \n- ...  \n\nFor $ n = 40 $, the parity depends on whether $ n $ is a multiple of 3. Since $ 40 \\mod 3 = 1 $, the 40th term is **odd**.  \n\n**Final Answer:** Odd.", "retry_tool_calls": 0, "retry_samples": 3}
diff --git a/studio/backend/tests/test_change_password_policy.py b/studio/backend/tests/test_change_password_policy.py
index c73e9ed839..fc095760d0 100644
--- a/studio/backend/tests/test_change_password_policy.py
+++ b/studio/backend/tests/test_change_password_policy.py
@@ -67,9 +67,11 @@ def test_rejects_password_containing_spaces(_user):
 
 
 def test_allows_password_without_spaces(_user, monkeypatch):
-    monkeypatch.setattr(auth_routes.storage, "update_password", lambda *args, **kwargs: True)
-    monkeypatch.setattr(auth_routes, "create_access_token", lambda subject: "at")
-    monkeypatch.setattr(auth_routes, "create_refresh_token", lambda subject: "rt")
+    monkeypatch.setattr(
+        auth_routes.storage, "update_password", lambda *args, **kwargs: "rotated-secret"
+    )
+    monkeypatch.setattr(auth_routes, "create_access_token", lambda subject, **kwargs: "at")
+    monkeypatch.setattr(auth_routes, "create_refresh_token", lambda subject, **kwargs: "rt")
     token = _change("correct-horse-battery")
     assert token.access_token == "at"
     assert token.must_change_password is False
diff --git a/studio/backend/tests/test_chat_text_encoding.py b/studio/backend/tests/test_chat_text_encoding.py
new file mode 100644
index 0000000000..64860dab1a
--- /dev/null
+++ b/studio/backend/tests/test_chat_text_encoding.py
@@ -0,0 +1,195 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Model text stays intact when it carries non-ASCII.
+
+``open()`` and ``Path.read_text()`` fall back to ``locale.getencoding()`` when
+no ``encoding`` is passed. On Windows that is the ANSI codepage, not UTF-8, so
+a chat template or model config holding ``ä ö ü → 世`` mojibakes or raises
+``UnicodeDecodeError``. These files are UTF-8, so the reads must say so.
+
+Each fixture writes raw UTF-8 (``ensure_ascii = False``), matching what
+Hugging Face actually ships, rather than ASCII ``\\uXXXX`` escapes.
+"""
+
+from __future__ import annotations
+
+import json
+import subprocess
+import sys
+import textwrap
+from pathlib import Path
+
+
+BACKEND_ROOT = Path(__file__).resolve().parent.parent
+
+
+def test_config_json_round_trips_non_ascii(tmp_path: Path) -> None:
+    from utils import transformers_version
+
+    name = "Modell für Grüße 世界"
+    (tmp_path / "config.json").write_text(
+        json.dumps({"model_type": "llama", "_name_or_path": name}, ensure_ascii = False),
+        encoding = "utf-8",
+    )
+    transformers_version._config_json_cache.clear()
+
+    cfg = transformers_version._load_config_json(str(tmp_path))
+
+    assert cfg is not None
+    assert cfg["_name_or_path"] == name
+
+
+def test_tokenizer_config_round_trips_non_ascii_chat_template(tmp_path: Path) -> None:
+    """Chat templates commonly hold ``→`` and smart quotes, which cp1252 mangles."""
+    from utils import transformers_version
+
+    template = "{{ '→ Grüße 世界' }}"
+    (tmp_path / "tokenizer_config.json").write_text(
+        json.dumps(
+            {"tokenizer_class": "TokenizersBackend", "chat_template": template},
+            ensure_ascii = False,
+        ),
+        encoding = "utf-8",
+    )
+    transformers_version._tokenizer_class_cache.clear()
+
+    assert transformers_version._check_tokenizer_config_needs_v5(str(tmp_path)) is True
+
+
+def test_config_json_survives_a_utf8_bom(tmp_path: Path) -> None:
+    """Notepad wrote "UTF-8 with BOM" by default for years, so hand-edited
+    configs on Windows carry one. Plain utf-8 keeps the BOM and json.load then
+    fails on it; utf-8-sig strips it and is identical otherwise."""
+    from utils import transformers_version
+
+    name = "Grüße 世界"
+    (tmp_path / "config.json").write_text(
+        json.dumps({"model_type": "llama", "_name_or_path": name}, ensure_ascii = False),
+        encoding = "utf-8-sig",
+    )
+    transformers_version._config_json_cache.clear()
+
+    cfg = transformers_version._load_config_json(str(tmp_path))
+
+    assert cfg is not None
+    assert cfg["_name_or_path"] == name
+
+
+def test_remote_code_scan_reads_non_ascii_sources(tmp_path: Path) -> None:
+    """A German Windows profile also puts umlauts in the model sources scanned."""
+    from utils.security import remote_code_scan
+
+    source = "# Grüße über Öl\nVALUE = '世界'\n"
+    # newline = "" pins the bytes on disk, so Windows line end translation cannot make the
+    # read back differ by \r. open() because Path.write_text() only grew newline in 3.10.
+    with open(
+        tmp_path / "modeling_custom.py",
+        "w",
+        encoding = "utf-8",
+        newline = "",
+    ) as handle:
+        handle.write(source)
+
+    files = remote_code_scan.repo_remote_code_files(str(tmp_path))
+
+    assert files["modeling_custom.py"] == source
+
+
+def test_model_config_reads_do_not_rely_on_the_locale_encoding(tmp_path: Path) -> None:
+    """The reads above pass anywhere the locale is already UTF-8, which hides
+    the Windows bug on Linux and macOS. ``-X warn_default_encoding`` makes
+    CPython flag any text I/O that falls back to the locale, so this fails on
+    every platform if an ``encoding`` argument goes missing again."""
+    # The readers swallow exceptions, so record the warnings instead of raising.
+    script = textwrap.dedent(
+        f"""
+        import sys, warnings
+        sys.path.insert(0, {str(BACKEND_ROOT)!r})
+        from utils import transformers_version
+
+        target = {str(tmp_path)!r}
+        with warnings.catch_warnings(record = True) as caught:
+            warnings.simplefilter("always")
+            transformers_version._config_json_cache.clear()
+            transformers_version._tokenizer_class_cache.clear()
+            assert transformers_version._load_config_json(target) is not None
+            assert transformers_version._check_tokenizer_config_needs_v5(target) is True
+
+        missing = [str(w.message) for w in caught if w.category is EncodingWarning]
+        if missing:
+            sys.exit("text I/O fell back to the locale encoding: " + "; ".join(missing))
+        """
+    )
+    for name, payload in (
+        ("config.json", {"model_type": "llama", "_name_or_path": "Grüße"}),
+        ("tokenizer_config.json", {"tokenizer_class": "TokenizersBackend"}),
+    ):
+        (tmp_path / name).write_text(json.dumps(payload, ensure_ascii = False), encoding = "utf-8")
+
+    result = subprocess.run(
+        [sys.executable, "-X", "warn_default_encoding", "-c", script],
+        capture_output = True,
+        text = True,
+        encoding = "utf-8",
+        errors = "replace",
+        timeout = 120,
+    )
+
+    assert result.returncode == 0, result.stderr
+
+
+def test_utf8_child_env_round_trips_non_ascii(tmp_path: Path) -> None:
+    """A Python child encodes stdout with its locale unless told otherwise, so
+    reading its pipe as utf-8 needs the child told to emit utf-8."""
+    from utils.child_stdio import utf8_child_env
+
+    payload = "Grüße über Öl → 世界"
+    child = tmp_path / "child.py"
+    child.write_text("import sys\nsys.stdout.write(" + repr(payload) + ")\n", encoding = "utf-8")
+
+    env = utf8_child_env()
+    assert env["PYTHONIOENCODING"] == "utf-8"
+
+    proc = subprocess.run(
+        [sys.executable, str(child)],
+        capture_output = True,
+        text = True,
+        encoding = "utf-8",
+        errors = "replace",
+        env = env,
+        timeout = 120,
+    )
+
+    assert proc.returncode == 0, proc.stderr
+    assert proc.stdout == payload
+
+
+def test_python_children_are_told_to_emit_utf8() -> None:
+    """Any child we decode as utf-8 must also be told to write utf-8, or a
+    cp1252 console silently mangles what it prints."""
+    import ast
+
+    offenders: list[str] = []
+    for path in sorted(BACKEND_ROOT.rglob("*.py")):
+        parts = path.relative_to(BACKEND_ROOT).parts
+        if any(p in ("tests", "node_modules", "plugins", "__pycache__") for p in parts):
+            continue
+        source = path.read_text(encoding = "utf-8")
+        for node in ast.walk(ast.parse(source, filename = str(path))):
+            if not isinstance(node, ast.Call):
+                continue
+            func = node.func
+            if not (isinstance(func, ast.Attribute) and func.attr in ("run", "Popen")):
+                continue
+            segment = ast.get_source_segment(source, node) or ""
+            if "sys.executable" not in segment or 'encoding = "utf-8"' not in segment:
+                continue
+            if "utf8_child_env" in segment or "PYTHONIOENCODING" in segment:
+                continue
+            offenders.append(f"{path.name}:{node.lineno}")
+
+    assert not offenders, (
+        "these spawn a Python child and decode it as utf-8 without setting the "
+        "child's own stdio encoding; wrap env in utf8_child_env():\n  " + "\n  ".join(offenders)
+    )
diff --git a/studio/backend/tests/test_credential_rotation_race.py b/studio/backend/tests/test_credential_rotation_race.py
new file mode 100644
index 0000000000..9b0f95aa02
--- /dev/null
+++ b/studio/backend/tests/test_credential_rotation_race.py
@@ -0,0 +1,255 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""A password rotation must not leave a session minted from the replaced credential.
+
+`unsloth studio reset-password` rotates in place against a live server, so a login
+can verify the old password, have the rotation land, and only then mint its tokens.
+Issuance is bound to the credential version that was verified, so such a login gets
+tokens that are already dead rather than a session that outlives the reset.
+"""
+
+import secrets
+from datetime import datetime, timedelta, timezone
+
+import jwt
+import pytest
+
+from auth import hashing, storage
+from auth.authentication import ALGORITHM, create_access_token, create_refresh_token
+
+
+@pytest.fixture(autouse = True)
+def isolated_auth_db(tmp_path, monkeypatch):
+    monkeypatch.setattr(storage, "DB_PATH", tmp_path / "auth.db")
+    monkeypatch.setattr(storage, "_BOOTSTRAP_PW_PATH", tmp_path / ".bootstrap_password")
+    monkeypatch.setattr(storage, "_bootstrap_password", None)
+    monkeypatch.setattr(storage, "_api_key_pbkdf2_salt_cache", None)
+    yield
+
+
+@pytest.fixture
+def admin():
+    storage.create_initial_user(
+        username = storage.DEFAULT_ADMIN_USERNAME,
+        password = "old-password-123",
+        jwt_secret = secrets.token_urlsafe(64),
+    )
+    return storage.DEFAULT_ADMIN_USERNAME
+
+
+def _verified_secret(username):
+    return storage.get_user_and_secret(username)[2]
+
+
+def test_access_token_from_the_replaced_credential_is_rejected(admin):
+    secret = _verified_secret(admin)
+
+    storage.update_password(admin, "new-password-456", revoke_refresh_tokens = True)
+    token = create_access_token(subject = admin, secret = secret)
+
+    with pytest.raises(jwt.InvalidTokenError):
+        jwt.decode(token, storage.get_jwt_secret(admin), algorithms = [ALGORITHM])
+
+
+def test_refresh_token_from_the_replaced_credential_is_rejected(admin):
+    secret = _verified_secret(admin)
+
+    # Inserted AFTER the rotation's DELETE, so revocation alone cannot catch it.
+    storage.update_password(admin, "new-password-456", revoke_refresh_tokens = True)
+    token = create_refresh_token(subject = admin, secret = secret)
+
+    assert storage.verify_refresh_token(token) is None
+    assert storage.consume_refresh_token(token) is None
+
+
+def test_a_rejected_refresh_token_is_dropped(admin):
+    secret = _verified_secret(admin)
+    storage.update_password(admin, "new-password-456", revoke_refresh_tokens = True)
+    token = create_refresh_token(subject = admin, secret = secret)
+
+    storage.verify_refresh_token(token)
+
+    conn = storage.get_connection()
+    try:
+        assert conn.execute("SELECT COUNT(*) AS c FROM refresh_tokens").fetchone()["c"] == 0
+    finally:
+        conn.close()
+
+
+def test_tokens_from_the_current_credential_still_work(admin):
+    secret = _verified_secret(admin)
+
+    access = create_access_token(subject = admin, secret = secret)
+    refresh = create_refresh_token(subject = admin, secret = secret)
+
+    jwt.decode(access, storage.get_jwt_secret(admin), algorithms = [ALGORITHM])
+    assert storage.verify_refresh_token(refresh) == (admin, False)
+
+
+def test_refresh_cannot_outlive_a_rotation_it_raced(admin):
+    # /refresh consumes, then mints. A rotation landing in between must not let
+    # the replacement pair be signed with the credential that just replaced it.
+    secret = _verified_secret(admin)
+    token = create_refresh_token(subject = admin, secret = secret)
+    consumed = storage.consume_refresh_token(token)
+    assert consumed is not None
+    _username, _is_desktop, consumed_secret = consumed
+
+    storage.update_password(admin, "new-password-456", revoke_refresh_tokens = True)
+    access = create_access_token(subject = admin, secret = consumed_secret)
+    refresh = create_refresh_token(subject = admin, secret = consumed_secret)
+
+    with pytest.raises(jwt.InvalidTokenError):
+        jwt.decode(access, storage.get_jwt_secret(admin), algorithms = [ALGORITHM])
+    assert storage.verify_refresh_token(refresh) is None
+
+
+def test_desktop_login_cannot_outlive_a_rotation_it_raced(admin):
+    # The reset deletes the desktop secret, so a desktop-login that validated it
+    # just beforehand must not mint a session that survives.
+    raw = storage.create_desktop_secret()
+    verified = storage.validate_desktop_secret_with_credential(raw)
+    assert verified is not None
+    _username, verified_secret = verified
+
+    storage.update_password(admin, "new-password-456", revoke_refresh_tokens = True)
+    access = create_access_token(subject = admin, desktop = True, secret = verified_secret)
+    refresh = create_refresh_token(subject = admin, desktop = True, secret = verified_secret)
+
+    with pytest.raises(jwt.InvalidTokenError):
+        jwt.decode(access, storage.get_jwt_secret(admin), algorithms = [ALGORITHM])
+    assert storage.verify_refresh_token(refresh) is None
+
+
+def test_change_password_cannot_overwrite_a_rotation_it_raced(admin):
+    # A change-password that verified the old hash must not clobber a reset that
+    # committed while it was in flight.
+    _salt, verified_hash, _secret, _must_change = storage.get_user_and_secret(admin)
+
+    storage.update_password(admin, "reset-by-the-cli-789", revoke_refresh_tokens = True)
+
+    assert not storage.update_password(
+        admin,
+        "attacker-chosen-000",
+        revoke_refresh_tokens = True,
+        expect_password_hash = verified_hash,
+    )
+    salt, pwd_hash, _s, _m = storage.get_user_and_secret(admin)
+    assert hashing.verify_password("reset-by-the-cli-789", salt, pwd_hash)
+
+
+def test_api_key_creation_from_a_revoked_credential_is_refused(admin):
+    generation = storage.credential_generation(_verified_secret(admin))
+
+    storage.update_password(admin, "new-password-456", revoke_refresh_tokens = True)
+
+    with pytest.raises(storage.CredentialRotated):
+        storage.create_api_key(username = admin, name = "k", expect_gen = generation)
+    conn = storage.get_connection()
+    try:
+        assert conn.execute("SELECT COUNT(*) AS c FROM api_keys").fetchone()["c"] == 0
+    finally:
+        conn.close()
+
+
+def test_api_key_creation_under_the_current_credential_still_works(admin):
+    generation = storage.credential_generation(_verified_secret(admin))
+
+    raw_key, _row = storage.create_api_key(username = admin, name = "k", expect_gen = generation)
+
+    assert storage.validate_api_key(raw_key) == admin
+
+
+def test_change_password_tokens_are_bound_to_its_own_write(admin):
+    # The tokens returned to a successful change-password must be signed with the
+    # secret that write produced, not whatever a later reset put in the DB.
+    _salt, verified_hash, _secret, _must = storage.get_user_and_secret(admin)
+    new_secret = storage.update_password(
+        admin,
+        "chosen-by-the-user",
+        revoke_refresh_tokens = True,
+        expect_password_hash = verified_hash,
+    )
+    assert new_secret is not None
+
+    storage.update_password(admin, "reset-by-the-cli-789", revoke_refresh_tokens = True)
+    access = create_access_token(subject = admin, secret = new_secret)
+    refresh = create_refresh_token(subject = admin, secret = new_secret)
+
+    with pytest.raises(jwt.InvalidTokenError):
+        jwt.decode(access, storage.get_jwt_secret(admin), algorithms = [ALGORITHM])
+    assert storage.verify_refresh_token(refresh) is None
+
+
+def test_internal_api_key_minting_honours_the_request_generation(admin):
+    generation = storage.credential_generation(_verified_secret(admin))
+    storage.update_password(admin, "new-password-456", revoke_refresh_tokens = True)
+
+    with pytest.raises(storage.CredentialRotated):
+        storage.create_api_key(
+            username = admin,
+            name = "data-recipe workflow",
+            internal = True,
+            expect_gen = generation,
+        )
+
+
+def test_api_key_auth_reports_the_version_the_key_was_valid_under(admin):
+    # The generation must come from the same transaction as the key check, or a
+    # revoked key could hand a route the post-reset generation and mint again.
+    raw, _row = storage.create_api_key(username = admin, name = "agent")
+    verified = storage.validate_api_key_with_credential(raw)
+    assert verified is not None
+    _user, secret = verified
+    generation = storage.credential_generation(secret)
+
+    storage.update_password(admin, "new-password-456", revoke_refresh_tokens = True)
+    conn = storage.get_connection()
+    try:
+        conn.execute("DELETE FROM api_keys")
+        conn.commit()
+    finally:
+        conn.close()
+
+    assert storage.validate_api_key(raw) is None
+    with pytest.raises(storage.CredentialRotated):
+        storage.create_api_key(username = admin, name = "after", expect_gen = generation)
+
+
+def test_consuming_a_legacy_token_reports_the_pre_reset_credential(admin):
+    # An unstamped row has no generation to compare, so consume must read the
+    # credential inside the delete transaction rather than after committing it.
+    token = secrets.token_urlsafe(48)
+    expires_at = (datetime.now(timezone.utc) + timedelta(days = 7)).isoformat()
+    storage.save_refresh_token(token, admin, expires_at, secret_gen = None)
+    conn = storage.get_connection()
+    try:
+        conn.execute("UPDATE refresh_tokens SET secret_gen = NULL")
+        conn.commit()
+    finally:
+        conn.close()
+
+    consumed = storage.consume_refresh_token(token)
+    assert consumed is not None
+    _username, _is_desktop, consumed_secret = consumed
+
+    storage.update_password(admin, "new-password-456", revoke_refresh_tokens = True)
+    access = create_access_token(subject = admin, secret = consumed_secret)
+    with pytest.raises(jwt.InvalidTokenError):
+        jwt.decode(access, storage.get_jwt_secret(admin), algorithms = [ALGORITHM])
+
+
+def test_unstamped_legacy_tokens_still_verify(admin):
+    # Rows written before the secret_gen column existed must not log users out.
+    token = secrets.token_urlsafe(48)
+    expires_at = (datetime.now(timezone.utc) + timedelta(days = 7)).isoformat()
+    storage.save_refresh_token(token, admin, expires_at, secret_gen = None)
+    conn = storage.get_connection()
+    try:
+        conn.execute("UPDATE refresh_tokens SET secret_gen = NULL")
+        conn.commit()
+    finally:
+        conn.close()
+
+    assert storage.verify_refresh_token(token) == (admin, False)
diff --git a/studio/backend/tests/test_desktop_auth.py b/studio/backend/tests/test_desktop_auth.py
index bc995b6a59..039bb5e3e6 100644
--- a/studio/backend/tests/test_desktop_auth.py
+++ b/studio/backend/tests/test_desktop_auth.py
@@ -134,6 +134,218 @@ def test_ensure_default_admin_loads_existing_bootstrap_after_restart(monkeypatch
     assert storage.get_bootstrap_password() == bootstrap_pw
 
 
+def test_bootstrap_password_file_ends_with_a_newline():
+    # Otherwise `cat` welds the passphrase onto the shell prompt.
+    storage.ensure_default_admin()
+
+    # Bytes: read_text would decode CRLF back to "\n" and hide a CR.
+    raw = storage._BOOTSTRAP_PW_PATH.read_bytes()
+
+    assert raw == storage.get_bootstrap_password().encode("utf-8") + b"\n"
+
+
+def test_bootstrap_password_round_trips_across_a_restart_with_the_newline():
+    storage.ensure_default_admin()
+    original = storage.get_bootstrap_password()
+
+    storage._bootstrap_password = None
+
+    assert storage.generate_bootstrap_password() == original
+
+
+def test_upgrade_normalises_the_bootstrap_file():
+    # Upgrade path: the admin row exists, so generate_bootstrap_password() never runs.
+    seed_user()
+    storage._BOOTSTRAP_PW_PATH.write_bytes(b"legacy-bootstrap-secret")
+
+    storage.ensure_default_admin()
+
+    assert storage._BOOTSTRAP_PW_PATH.read_bytes() == b"legacy-bootstrap-secret\n"
+    assert storage.get_bootstrap_password() == "legacy-bootstrap-secret"
+
+
+@pytest.mark.parametrize(
+    "other",
+    [
+        b"legacy-bootstrap-secret\r\n",  # only an unreleased build wrote this
+        b"legacy-bootstrap-secret\r",
+        b"legacy-bootstrap-secret   ",
+    ],
+)
+def test_only_an_exactly_unterminated_bootstrap_file_is_touched(other):
+    # Appending is safe only because it is restricted to the one released shape.
+    seed_user()
+    storage._BOOTSTRAP_PW_PATH.write_bytes(other)
+
+    storage.ensure_default_admin()
+
+    assert storage.get_bootstrap_password() == "legacy-bootstrap-secret"
+    assert storage._BOOTSTRAP_PW_PATH.read_bytes() == other
+
+
+def test_upgrade_normalises_when_the_admin_row_is_missing():
+    storage._BOOTSTRAP_PW_PATH.write_bytes(b"legacy-bootstrap-secret")
+
+    assert storage.generate_bootstrap_password() == "legacy-bootstrap-secret"
+    assert storage._BOOTSTRAP_PW_PATH.read_bytes() == b"legacy-bootstrap-secret\n"
+
+
+def test_a_well_formed_bootstrap_file_is_not_rewritten():
+    seed_user()
+    storage._BOOTSTRAP_PW_PATH.write_bytes(b"legacy-bootstrap-secret\n")
+    mtime = storage._BOOTSTRAP_PW_PATH.stat().st_mtime_ns
+
+    storage.ensure_default_admin()
+
+    assert storage._BOOTSTRAP_PW_PATH.stat().st_mtime_ns == mtime
+
+
+def test_migration_failure_does_not_break_startup(monkeypatch):
+    seed_user()
+    storage._BOOTSTRAP_PW_PATH.write_bytes(b"legacy-bootstrap-secret")
+
+    real_open = storage.os.open
+
+    def refuse(path, flags, *args, **kwargs):
+        if str(path) == str(storage._BOOTSTRAP_PW_PATH):
+            raise PermissionError("read-only auth dir")
+        return real_open(path, flags, *args, **kwargs)
+
+    monkeypatch.setattr(storage.os, "open", refuse)
+
+    storage.ensure_default_admin()
+
+    assert storage.get_bootstrap_password() == "legacy-bootstrap-secret"
+    assert storage._BOOTSTRAP_PW_PATH.read_bytes() == b"legacy-bootstrap-secret"
+
+
+def test_normalising_never_recreates_a_cleared_bootstrap_file(monkeypatch):
+    # A rename would resurrect revoked plaintext if the password changed after the read.
+    seed_user()
+    storage._BOOTSTRAP_PW_PATH.write_bytes(b"legacy-bootstrap-secret")
+
+    real_open = storage.os.open
+
+    def clear_then_open(path, flags, *args, **kwargs):
+        if str(path) == str(storage._BOOTSTRAP_PW_PATH):
+            storage._BOOTSTRAP_PW_PATH.unlink(missing_ok = True)
+        return real_open(path, flags, *args, **kwargs)
+
+    monkeypatch.setattr(storage.os, "open", clear_then_open)
+
+    assert storage._read_persisted_bootstrap_password() == "legacy-bootstrap-secret"
+    assert not storage._BOOTSTRAP_PW_PATH.exists()
+
+
+def test_normalising_does_not_overwrite_a_rotated_bootstrap_file(monkeypatch):
+    seed_user()
+    storage._BOOTSTRAP_PW_PATH.write_bytes(b"legacy-bootstrap-secret")
+
+    real_open = storage.os.open
+
+    def rotate_then_open(path, flags, *args, **kwargs):
+        if str(path) == str(storage._BOOTSTRAP_PW_PATH):
+            storage._BOOTSTRAP_PW_PATH.write_bytes(b"brand-new-secret\n")
+        return real_open(path, flags, *args, **kwargs)
+
+    monkeypatch.setattr(storage.os, "open", rotate_then_open)
+
+    storage._read_persisted_bootstrap_password()
+
+    # The append may add a second newline; the rotated credential must survive.
+    raw = storage._BOOTSTRAP_PW_PATH.read_bytes()
+    assert raw.strip() == b"brand-new-secret"
+    storage._bootstrap_password = None
+    assert storage._load_bootstrap_password() == "brand-new-secret"
+
+
+def test_leading_whitespace_bootstrap_file_is_left_alone(monkeypatch):
+    # An in-place rewrite is not atomic, so only the exact unterminated shape is touched.
+    seed_user()
+    storage._BOOTSTRAP_PW_PATH.write_bytes(b"  legacy-bootstrap-secret  ")
+
+    storage.ensure_default_admin()
+
+    assert storage.get_bootstrap_password() == "legacy-bootstrap-secret"
+    assert storage._BOOTSTRAP_PW_PATH.read_bytes() == b"  legacy-bootstrap-secret  "
+
+
+def test_normalising_opens_the_file_in_binary_mode(monkeypatch):
+    # Without O_BINARY, Windows text mode turns the written LF back into CRLF.
+    seed_user()
+    storage._BOOTSTRAP_PW_PATH.write_bytes(b"legacy-bootstrap-secret")
+    monkeypatch.setattr(storage.os, "O_BINARY", 0x8000, raising = False)
+    seen = []
+    real_open = storage.os.open
+
+    def spy(path, flags, *args, **kwargs):
+        if str(path) == str(storage._BOOTSTRAP_PW_PATH):
+            seen.append(flags)
+        return real_open(path, flags & ~0x8000, *args, **kwargs)
+
+    monkeypatch.setattr(storage.os, "open", spy)
+
+    storage.ensure_default_admin()
+
+    assert seen and all(f & 0x8000 for f in seen), seen
+
+
+def test_clearing_by_truncation_mid_normalisation_is_not_undone(monkeypatch):
+    # clear_bootstrap_password() truncates through its own descriptor when the unlink
+    # fails (Windows, while ours is open); the append must not restore the plaintext.
+    seed_user()
+    storage._BOOTSTRAP_PW_PATH.write_bytes(b"legacy-bootstrap-secret")
+
+    real_open = storage.os.open
+
+    def truncate_then_open(path, flags, *args, **kwargs):
+        fd = real_open(path, flags, *args, **kwargs)
+        if str(path) == str(storage._BOOTSTRAP_PW_PATH):
+            storage._BOOTSTRAP_PW_PATH.write_text("", encoding = "utf-8")
+        return fd
+
+    monkeypatch.setattr(storage.os, "open", truncate_then_open)
+
+    storage._read_persisted_bootstrap_password()
+
+    # A lone newline over a cleared file still reads back as no password.
+    assert storage._BOOTSTRAP_PW_PATH.read_bytes().strip() == b""
+    storage._bootstrap_password = None
+    assert storage._load_bootstrap_password() is None
+
+
+def test_normalising_works_without_fchmod(monkeypatch):
+    # os.fchmod only reached Windows in 3.13; its absence must not raise.
+    seed_user()
+    storage._BOOTSTRAP_PW_PATH.write_bytes(b"legacy-bootstrap-secret")
+    monkeypatch.delattr(storage.os, "fchmod", raising = False)
+
+    storage.ensure_default_admin()
+
+    assert storage._BOOTSTRAP_PW_PATH.read_bytes() == b"legacy-bootstrap-secret\n"
+    assert storage.get_bootstrap_password() == "legacy-bootstrap-secret"
+
+
+def test_persisting_the_bootstrap_password_is_atomic(monkeypatch, tmp_path):
+    # A partial write would destroy the only plaintext recovery credential.
+    storage._persist_bootstrap_password("original-secret")
+
+    def boom(src, dst):
+        raise OSError("crash before replace")
+
+    monkeypatch.setattr(storage.os, "replace", boom)
+    with pytest.raises(OSError):
+        storage._persist_bootstrap_password("new-secret")
+
+    assert storage._BOOTSTRAP_PW_PATH.read_bytes() == b"original-secret\n"
+    leftovers = [
+        p.name
+        for p in storage._BOOTSTRAP_PW_PATH.parent.iterdir()
+        if "bootstrap_password." in p.name
+    ]
+    assert leftovers == []
+
+
 def test_ensure_default_admin_does_not_generate_for_empty_existing_bootstrap():
     seed_user()
     storage._BOOTSTRAP_PW_PATH.write_text(" \n", encoding = "utf-8")
@@ -233,7 +445,7 @@ def test_consume_refresh_token_second_call_returns_none():
     storage.save_refresh_token(raw, storage.DEFAULT_ADMIN_USERNAME, expires)
 
     first = storage.consume_refresh_token(raw)
-    assert first == (storage.DEFAULT_ADMIN_USERNAME, False)
+    assert first[:2] == (storage.DEFAULT_ADMIN_USERNAME, False)
     second = storage.consume_refresh_token(raw)
     assert second is None
 
@@ -262,7 +474,7 @@ def test_consume_refresh_token_concurrent_only_one_succeeds(tmp_path, monkeypatc
 
     successes = [r for r in results if r is not None]
     assert len(successes) == 1, f"expected exactly one consumer to win, got {len(successes)}"
-    assert successes[0] == (storage.DEFAULT_ADMIN_USERNAME, False)
+    assert successes[0][:2] == (storage.DEFAULT_ADMIN_USERNAME, False)
 
 
 def test_consume_refresh_token_expired_returns_none():
@@ -336,6 +548,28 @@ def test_local_recipe_token_authenticates_as_admin_for_web_user(loaded_local_mod
     assert asyncio.run(get_current_subject(credentials)) == storage.DEFAULT_ADMIN_USERNAME
 
 
+def test_rotated_credential_job_start_is_401_not_500(loaded_local_model):
+    # A reset-password landing mid-request makes the workflow-key mint refuse.
+    # That must reach the client as a revoked credential, not an unhandled error.
+    from fastapi import HTTPException
+
+    seed_user()
+    jobs_route = data_recipe_jobs_module()
+    stale_gen = storage.credential_generation(secrets.token_urlsafe(64))
+
+    with pytest.raises(storage.CredentialRotated):
+        jobs_route._inject_local_providers(local_recipe(), local_recipe_request("t"), stale_gen)
+
+    def _boom(*_a, **_k):
+        raise storage.CredentialRotated("revoked")
+
+    jobs_route._inject_local_providers = _boom
+    payload = SimpleNamespace(recipe = local_recipe(), run = {})
+    with pytest.raises(HTTPException) as excinfo:
+        jobs_route.create_job(payload, local_recipe_request("t"), ("unsloth", stale_gen))
+    assert excinfo.value.status_code == 401
+
+
 def test_desktop_login_rejects_invalid_secret():
     seed_user(must_change_password = False)
     client = auth_client()
@@ -358,7 +592,7 @@ def test_write_desktop_secret_file_is_0600_on_unix(tmp_path):
 
     studio_cli._write_auth_secret(path, "desktop-secret")
 
-    assert path.read_text() == "desktop-secret"
+    assert path.read_bytes() == b"desktop-secret\n"
     if platform.system() != "Windows":
         assert oct(path.stat().st_mode & 0o777) == "0o600"
 
@@ -368,18 +602,31 @@ def test_reset_password_removes_desktop_secret_files(tmp_path, monkeypatch):
     from unsloth_cli.commands import studio as studio_cli
 
     auth_dir = tmp_path / "auth"
-    auth_dir.mkdir()
-    (auth_dir / "auth.db").write_text("db")
-    (auth_dir / ".bootstrap_password").write_text("boot")
-    (auth_dir / ".desktop_secret").write_text("new")
     monkeypatch.setattr(studio_cli, "STUDIO_HOME", tmp_path)
+    secret = studio_cli._create_desktop_secret_in_cli()
+    studio_cli._write_auth_secret(auth_dir / studio_cli.DESKTOP_SECRET_FILE, secret)
+    (auth_dir / studio_cli.BOOTSTRAP_PASSWORD_FILE).write_text("boot")
 
     result = CliRunner().invoke(studio_cli.studio_app, ["reset-password"])
 
-    assert result.exit_code == 0
-    assert not (auth_dir / "auth.db").exists()
-    assert not (auth_dir / ".bootstrap_password").exists()
-    assert not (auth_dir / ".desktop_secret").exists()
+    assert result.exit_code == 0, result.output
+    # The DB survives on purpose: a running server keeps serving from its admin row.
+    assert (auth_dir / "auth.db").exists()
+    assert not (auth_dir / studio_cli.BOOTSTRAP_PASSWORD_FILE).exists()
+    assert not (auth_dir / studio_cli.DESKTOP_SECRET_FILE).exists()
+
+    conn = studio_cli._connect_auth_db()
+    try:
+        surviving = conn.execute(
+            "SELECT COUNT(*) FROM app_secrets WHERE key IN (?, ?)",
+            (
+                studio_cli.DESKTOP_SECRET_HASH_KEY,
+                studio_cli.DESKTOP_SECRET_CREATED_AT_KEY,
+            ),
+        ).fetchone()[0]
+    finally:
+        conn.close()
+    assert surviving == 0
 
 
 def test_reset_password_removes_desktop_secret_files_without_db(tmp_path, monkeypatch):
@@ -525,7 +772,8 @@ if result.exit_code != 0:
         capture_output = True,
     )
     assert result.returncode == 0, result.stderr + result.stdout
-    secret = (auth_dir / ".desktop_secret").read_text()
+    # Strip like the src-tauri readers do.
+    secret = (auth_dir / ".desktop_secret").read_text().strip()
     assert secret.startswith("desktop-")
 
     conn = sqlite3.connect(auth_dir / "auth.db")
@@ -633,7 +881,7 @@ def test_update_password_clears_desktop_secret():
     assert storage.validate_desktop_secret(raw) == storage.DEFAULT_ADMIN_USERNAME
 
     changed = storage.update_password(storage.DEFAULT_ADMIN_USERNAME, "new-admin-password")
-    assert changed is True
+    assert changed
     assert storage.validate_desktop_secret(raw) is None
 
 
@@ -642,7 +890,7 @@ def test_update_password_on_unknown_user_leaves_desktop_secret_intact():
     raw = storage.create_desktop_secret()
 
     changed = storage.update_password("not-a-user", "irrelevant")
-    assert changed is False
+    assert not changed
     assert storage.validate_desktop_secret(raw) == storage.DEFAULT_ADMIN_USERNAME
 
 
diff --git a/studio/backend/tests/test_llama_cpp_tool_loop.py b/studio/backend/tests/test_llama_cpp_tool_loop.py
index cbd1b07505..7f59a2d681 100644
--- a/studio/backend/tests/test_llama_cpp_tool_loop.py
+++ b/studio/backend/tests/test_llama_cpp_tool_loop.py
@@ -1487,7 +1487,418 @@ def test_internal_reprompt_attempts_do_not_duplicate_visible_text(monkeypatch):
 
     content_texts = [event.get("text", "") for event in events if event.get("type") == "content"]
     assert content_texts == ["I will use render_html now."]
+    # Each retry restates the last, so the loop gives up: initial + 2 re-prompts.
+    assert len(payloads) == 3 < _MAX_REPROMPTS + 1
+
+
+def test_post_tool_stall_still_nudged_after_a_pre_tool_reprompt(monkeypatch):
+    """The post-tool nudge has its own budget, so an earlier stall can't spend it."""
+
+    streams = [
+        [_sse({"content": "I will search the web now."}), _done()],
+        [
+            _sse(
+                {
+                    "tool_calls": [
+                        {
+                            "index": 0,
+                            "id": "call_first",
+                            "type": "function",
+                            "function": {
+                                "name": "web_search",
+                                "arguments": json.dumps({"query": "red square"}),
+                            },
+                        }
+                    ]
+                }
+            ),
+            _done(),
+        ],
+        [_sse({"content": "Let me summarize the results."}), _done()],
+        [_sse({"content": "Final answer: the square is red."}), _done()],
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, streams, payloads)
+
+    calls: list[tuple[str, dict]] = []
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        calls.append((name, arguments))
+        return "Search results: red is #f00."
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "web_search",
+                "description": "Search the web.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"query": {"type": "string"}},
+                    "required": ["query"],
+                },
+            },
+        }
+    ]
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "Make a red square."}],
+            tools = tools,
+            max_tool_iterations = 2,
+        )
+    )
+
+    assert len(payloads) == 4
+    assert len(calls) == 1
+    nudges = [
+        message
+        for message in payloads[-1]["messages"]
+        if message.get("role") == "user" and "call web_search now" in message.get("content", "")
+    ]
+    assert len(nudges) == 2
+    content_texts = [event.get("text", "") for event in events if event.get("type") == "content"]
+    assert content_texts[-1] == "Final answer: the square is red."
+
+
+def test_post_tool_reprompt_budget_is_one(monkeypatch):
+    """The post-tool nudge fires once; a second stall is surrendered as the answer."""
+
+    streams = [
+        [
+            _sse(
+                {
+                    "tool_calls": [
+                        {
+                            "index": 0,
+                            "id": "call_first",
+                            "type": "function",
+                            "function": {
+                                "name": "web_search",
+                                "arguments": json.dumps({"query": "red square"}),
+                            },
+                        }
+                    ]
+                }
+            ),
+            _done(),
+        ],
+        [_sse({"content": "Let me summarize the results."}), _done()],
+        [_sse({"content": "Now I will check the sources."}), _done()],
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, streams, payloads)
+
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda *_a, **_k: "Search results: red is #f00.",
+    )
+
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "web_search",
+                "description": "Search the web.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"query": {"type": "string"}},
+                    "required": ["query"],
+                },
+            },
+        }
+    ]
+
+    list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "Make a red square."}],
+            tools = tools,
+            max_tool_iterations = 2,
+        )
+    )
+
+    assert len(payloads) == 3
+
+
+def test_repeat_guard_resets_after_a_tool_runs(monkeypatch):
+    """A tool execution opens a new phase, so the same intent text is nudged again.
+
+    Without the reset the pre-tool stall text still sits in the repeat tracker and
+    the identical post-tool stall is surrendered as the visible final answer.
+    """
+
+    stall = "I will search the web now."
+    streams = [
+        [_sse({"content": stall}), _done()],
+        [
+            _sse(
+                {
+                    "tool_calls": [
+                        {
+                            "index": 0,
+                            "id": "call_first",
+                            "type": "function",
+                            "function": {
+                                "name": "web_search",
+                                "arguments": json.dumps({"query": "red square"}),
+                            },
+                        }
+                    ]
+                }
+            ),
+            _done(),
+        ],
+        [_sse({"content": stall}), _done()],
+        [_sse({"content": "Final answer: the square is red."}), _done()],
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, streams, payloads)
+
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda *_a, **_k: "Search results: red is #f00.",
+    )
+
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "web_search",
+                "description": "Search the web.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"query": {"type": "string"}},
+                    "required": ["query"],
+                },
+            },
+        }
+    ]
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "Make a red square."}],
+            tools = tools,
+            max_tool_iterations = 2,
+        )
+    )
+
+    assert len(payloads) == 4
+    content_texts = [event.get("text", "") for event in events if event.get("type") == "content"]
+    assert content_texts[-1] == "Final answer: the square is red."
+
+
+def test_restatement_keeps_deletions_that_change_the_answer():
+    """A dropped word can invert the meaning, so a subset is not a restatement."""
+
+    from core.inference.tool_call_parser import is_reprompt_restatement
+    from core.inference.llama_cpp import _should_suppress_forced_no_tool_output as suppress
+
+    previous = "Now I think the feature is not supported in version 1."
+    corrected = "Now I think the feature is supported in version 1."
+    assert not is_reprompt_restatement(corrected, previous)
+    assert not suppress(corrected, previous)
+
+    stall = "I'll search for that now."
+    assert is_reprompt_restatement(stall, stall)
+    assert is_reprompt_restatement("Understood. " + stall, "Understood, " + stall)
+    assert not is_reprompt_restatement(stall + " Tokyo.", stall)
+
+
+def test_forced_turn_suppression_covers_obligation_phrasing():
+    from core.inference.llama_cpp import _should_suppress_forced_no_tool_output as suppress
+    for stall in (
+        "I need to use render_html now",
+        "Need to call web_search",
+        "I will summarize the results now",
+        "I have to run the search first",
+        "I should call web_search now",
+        "I should use render_html now",
+        # Plain modals take a bare infinitive, not the need|have|ought "to" group.
+        "I must call web_search now",
+        "I must use render_html now",
+        "I must run the search first",
+        # Subjectless plans open a new sentence just as often as a new line.
+        "Okay. Need to call web_search now.",
+        "Understood. Going to search now.",
+        # Subjectless modals, not just subjectless semi-modals.
+        "Must call web_search now.",
+        "Should search the web now.",
+        # A missing answer is not a final answer: the plan behind it is still a stall.
+        "I should call web_search because the answer is not in the provided context",
+        "I must run the search since the answer is unknown so far",
+        # A pivot with nothing behind it answers nothing.
+        "I should call web_search, though.",
+        "I need to run the search, but",
+        # A purpose clause is part of the plan, not a summary of results.
+        "I need to call web_search to summarize the results",
+    ):
+        assert suppress(stall), f"leaked {stall!r}"
+
+    for answer in (
+        "You need to install the package first.",
+        "The square is red.",
+        "Here is the summary of what I found.",
+        "Run `pip install unsloth` to get started.",
+        "I should mention that the square is red.",
+        # Obligation phrasing mid-sentence is prose that happens to name a tool.
+        "The API I should invoke is foo() because it supports streaming.",
+        "The tool I need to use is documented here.",
+        # "invoke"/"query" read as technical prose far more often than as a stall.
+        "I should invoke foo() because it supports streaming.",
+        "I should query the cache first for a faster path.",
+        "You should call your bank about the charge.",
+        # Second person is the user's obligation, not the model's plan.
+        "You must call your bank about the charge.",
+        "I must admit the square is red.",
+        # A plan that pivots to an answer must ship the answer with it.
+        "I should call web_search, but the answer is Tokyo.",
+        "I need to call web_search. The answer is Tokyo.",
+        "I should call web_search to confirm, but Tokyo is the capital of Japan.",
+        "I must run the search, however the result is already known: 42.",
+    ):
+        assert not suppress(answer), f"dropped {answer!r}"
+
+
+def test_forced_turn_intent_lead_in_needs_a_restatement_to_be_dropped():
+    """A bare intent match is a stall only when the retry restates the nudge.
+
+    ``INTENT_SIGNAL`` fires on lead-ins that introduce a real answer ("Now I
+    have the results. ..."), so matching it alone would discard the answer.
+    """
+    from core.inference.llama_cpp import _should_suppress_forced_no_tool_output as suppress
+
+    stall = "I will summarize the results now"
+    answer = "Now I have the search results. The capital of Japan is Tokyo."
+
+    # Restating the nudged text is still a stall.
+    assert suppress(stall, stall)
+    assert suppress("Understood. " + stall, "Understood, " + stall)
+    # Progress past the nudged text keeps the answer, lead-in and all.
+    assert not suppress(answer, stall)
+    assert not suppress("Step 3: done. Tokyo is the capital.", stall)
+    # Near-repeat is enough to stop nudging, never enough to drop the turn.
+    assert not suppress(stall + ": Tokyo.", stall)
+    # An obligation plan is a stall on its own, no previous text needed.
+    assert suppress("I must call web_search now", answer)
+
+
+def test_forced_turn_answer_with_an_intent_lead_in_survives_after_a_tool(monkeypatch):
+    """The post-tool retry answers behind a lead-in; the answer must still ship.
+
+    The nudge budget is spent, so the reply lands on the suppression branch.
+    ``INTENT_SIGNAL`` matches its "Now I ..." opener, and dropping it on that
+    alone left the user with the stall and no answer at all.
+    """
+
+    answer = "Now I have the results. The capital of Japan is Tokyo."
+    streams = [
+        [
+            _sse(
+                {
+                    "tool_calls": [
+                        {
+                            "index": 0,
+                            "id": "call_first",
+                            "type": "function",
+                            "function": {
+                                "name": "web_search",
+                                "arguments": json.dumps({"query": "capital of Japan"}),
+                            },
+                        }
+                    ]
+                }
+            ),
+            _done(),
+        ],
+        [_sse({"content": "Let me summarize what I found."}), _done()],
+        [_sse({"content": answer}), _done()],
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, streams, payloads)
+
+    monkeypatch.setattr(
+        "core.inference.tools.execute_tool",
+        lambda *_a, **_k: "Search results: Tokyo.",
+    )
+
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "web_search",
+                "description": "Search the web.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"query": {"type": "string"}},
+                    "required": ["query"],
+                },
+            },
+        }
+    ]
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "What is the capital of Japan?"}],
+            tools = tools,
+            max_tool_iterations = 2,
+        )
+    )
+
+    assert len(payloads) == 3
+    content_texts = [event.get("text", "") for event in events if event.get("type") == "content"]
+    assert content_texts[-1] == answer
+
+
+def test_forced_turn_answer_with_an_intent_lead_in_survives_pre_tool(monkeypatch):
+    """Same guarantee once the pre-tool nudge budget is spent on distinct stalls."""
+
+    answer = "Now I see the data clearly. Tokyo is the capital."
+    streams = [
+        [_sse({"content": text}), _done()]
+        for text in (
+            "I will look that up for you.",
+            "Now I have the search results. The capital of Japan is Tokyo.",
+            "Now I can confirm it. Japan's capital city is Tokyo.",
+            answer,
+        )
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, streams, payloads)
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        raise AssertionError(f"unexpected tool execution: {name} {arguments}")
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    tools = [
+        {
+            "type": "function",
+            "function": {
+                "name": "web_search",
+                "description": "Search the web.",
+                "parameters": {
+                    "type": "object",
+                    "properties": {"query": {"type": "string"}},
+                    "required": ["query"],
+                },
+            },
+        }
+    ]
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "What is the capital of Japan?"}],
+            tools = tools,
+            max_tool_iterations = 2,
+        )
+    )
+
+    # Initial turn plus the three pre-tool nudges.
     assert len(payloads) == _MAX_REPROMPTS + 1
+    content_texts = [event.get("text", "") for event in events if event.get("type") == "content"]
+    assert content_texts[-1] == answer
 
 
 def test_forced_reprompt_plain_final_answer_is_visible(monkeypatch):
@@ -2084,6 +2495,51 @@ def test_confirm_tool_calls_skips_gguf_rag_autoinject(monkeypatch):
     assert any(event.get("type") == "content" and event.get("text") == "Done." for event in events)
 
 
+def test_rag_autoinject_counts_as_a_prior_tool_execution(monkeypatch):
+    """Autoinjected retrieval runs before the controller, so history stays empty.
+
+    Without counting it the turn reads as pre-tool and gets the full re-prompt
+    budget, repeating the expensive retrieval the post-tool cap exists to stop.
+    """
+
+    stall = "I will summarize the retrieved passages now."
+    streams = [
+        [_sse({"content": stall}), _done()],
+        [_sse({"content": "Still working on the summary."}), _done()],
+        [_sse({"content": "Final answer: the passages describe Tokyo."}), _done()],
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, streams, payloads)
+
+    monkeypatch.setattr(
+        "core.inference.tools.build_rag_autoinject",
+        lambda *_a, **_k: {
+            "events": [],
+            "messages": [{"role": "user", "content": "Retrieved passage: Tokyo."}],
+        },
+    )
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "summarize the docs"}],
+            tools = [{"type": "function", "function": {"name": "search_knowledge_base"}}],
+            max_tool_iterations = 2,
+            rag_scope = {"thread_id": "t1"},
+        )
+    )
+
+    # Initial turn plus one retry; read as pre-tool it would spend the full budget.
+    assert len(payloads) == 2, payloads
+    nudges = [
+        message
+        for message in payloads[-1]["messages"]
+        if message.get("role") == "user"
+        and "call search_knowledge_base now" in message.get("content", "")
+    ]
+    assert len(nudges) == 1, nudges
+    assert events
+
+
 def test_confirm_tool_calls_deny_skips_gguf_tool_and_retry_can_execute(monkeypatch):
     same_call = _structured_tool_call("python", {"code": "print(1)"}, "call_py")
     streams = [
diff --git a/studio/backend/tests/test_password_prompt_backstop.py b/studio/backend/tests/test_password_prompt_backstop.py
index 3c2c1956f9..6c22532532 100644
--- a/studio/backend/tests/test_password_prompt_backstop.py
+++ b/studio/backend/tests/test_password_prompt_backstop.py
@@ -247,8 +247,8 @@ def test_lifespan_honors_bootstrap_suppression_in_source():
 def test_clear_bootstrap_password_truncates_when_unlink_fails(monkeypatch, tmp_path):
     # If the file cannot be unlinked (Windows AV / read-only auth dir), clear must
     # truncate it so its stale plaintext cannot be re-seeded by
-    # generate_bootstrap_password() after a later reset-password deletes auth.db,
-    # which would re-validate the revoked bootstrap password.
+    # generate_bootstrap_password() if auth.db is ever recreated, which would
+    # re-validate the revoked bootstrap password.
     import pathlib
 
     pw_path = tmp_path / ".bootstrap_password"
diff --git a/studio/backend/tests/test_plan_classifier_accuracy.py b/studio/backend/tests/test_plan_classifier_accuracy.py
new file mode 100644
index 0000000000..9144fb92e3
--- /dev/null
+++ b/studio/backend/tests/test_plan_classifier_accuracy.py
@@ -0,0 +1,96 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""An accuracy floor for the plan-without-action classifier, on real model output.
+
+The rest of the tool-loop suites pin behaviour on hand-written example sentences,
+which is how the patterns here were tuned. That says nothing about how often the
+classifier is right on what models actually emit, so this file scores it against a
+corpus captured from local models (``tests/data/plan_vs_answer.jsonl``).
+
+How the corpus was built: three GGUF models (Qwen3-0.6B, Qwen3-1.7B,
+Llama-3.2-1B-Instruct) were driven through llama-server with the real Studio tool
+schemas over prompts spanning tool-requiring questions, questions needing no tool,
+list-formatted answers, ambiguous requests, non-English, and follow-ups issued after
+a tool had already run. Turns cut off by the token cap were dropped, since a
+truncation is not a stall.
+
+Every turn here is a *finished answer*: the turn called no tool, and when the
+production nudge was appended and the turn regenerated three times, not one retry
+produced a tool call. A forceful re-prompt could not extract an action, so there was
+no action left to take. Nudging these is wasted work, and in the GGUF loop the
+retry's text can then be discarded, which costs the user a visible answer.
+
+Measured when this landed, over the 300 turns:
+
+    tree                          nudged      retry discarded
+    origin/main (pre-PR)          36 (12.0%)  60 (20.2%)
+    this PR                        5 ( 1.7%)   1 ( 0.3%)
+
+The budgets below sit above the measured counts so that innocuous wording changes
+do not fail the build, and far below the pre-PR counts so a real regression does.
+A failure prints the offending turns: fix the pattern, or if the turn really is a
+stall, correct its label here.
+"""
+
+import json
+from pathlib import Path
+
+from core.inference.llama_cpp import _should_suppress_forced_no_tool_output
+from core.inference.tool_call_parser import is_short_intent_without_action
+
+DATA = Path(__file__).parent / "data" / "plan_vs_answer.jsonl"
+
+# Measured 5 of 300; pre-PR was 36.
+NUDGE_BUDGET = 9
+# Measured 1 of 300; pre-PR was 60. Tighter, because this one destroys output.
+DISCARD_BUDGET = 4
+
+
+def _corpus():
+    with open(DATA, encoding = "utf-8") as fh:
+        return [json.loads(line) for line in fh if line.strip()]
+
+
+def _report(rows, limit = 10):
+    lines = []
+    for row in rows[:limit]:
+        text = " ".join(row["text"].split())
+        lines.append(
+            f"  [{row['model']}/{row['prompt_class']}] {row['prompt']!r}\n    {text[:200]!r}"
+        )
+    if len(rows) > limit:
+        lines.append(f"  ... and {len(rows) - limit} more")
+    return "\n".join(lines)
+
+
+def test_corpus_is_intact():
+    """Guards the budgets: they mean nothing if the corpus silently shrinks."""
+    corpus = _corpus()
+    assert len(corpus) == 300
+    assert all(row["text"].strip() for row in corpus)
+    # Every row is a finished answer by construction.
+    assert all(row["retry_tool_calls"] == 0 for row in corpus)
+
+
+def test_finished_answers_are_rarely_nudged():
+    """A finished answer costs a whole extra generation when it is nudged."""
+    nudged = [row for row in _corpus() if is_short_intent_without_action(row["text"])]
+    assert len(nudged) <= NUDGE_BUDGET, (
+        f"{len(nudged)}/300 finished answers classified as plans "
+        f"(budget {NUDGE_BUDGET}):\n{_report(nudged)}"
+    )
+
+
+def test_finished_answers_are_not_discarded():
+    """The retry's text is all the user gets, so discarding it is the worst case."""
+    discarded = [
+        row
+        for row in _corpus()
+        if row["retry_text"].strip()
+        and _should_suppress_forced_no_tool_output(row["retry_text"], row["text"])
+    ]
+    assert len(discarded) <= DISCARD_BUDGET, (
+        f"{len(discarded)}/300 finished retries would be discarded "
+        f"(budget {DISCARD_BUDGET}):\n{_report(discarded)}"
+    )
diff --git a/studio/backend/tests/test_rocm_multi_gpu_vram_system_wide.py b/studio/backend/tests/test_rocm_multi_gpu_vram_system_wide.py
index bdafdeae9b..db89b02003 100644
--- a/studio/backend/tests/test_rocm_multi_gpu_vram_system_wide.py
+++ b/studio/backend/tests/test_rocm_multi_gpu_vram_system_wide.py
@@ -45,8 +45,20 @@ def _build_structlog_stub():
 _maybe_stub("loggers", _build_loggers_stub)
 _maybe_stub("structlog", _build_structlog_stub)
 
+import pytest
+
 import utils.hardware.hardware as hw  # noqa: E402
 
+# The DRM/KFD readers below are Linux-only in production: _rocm_linux_amdgpu_cards and
+# _rocm_linux_sysfs_vram_by_pci_gb return early unless platform.system() is "Linux", and
+# _rocm_kfd_gpu_pci_ids only ever globs /sys/class/kfd. Their fake sysfs tree needs PCI
+# addresses like "0000:00:02.0" as directory names and POSIX separators in the paths the
+# readers match; Windows permits neither, so the tree cannot be represented there.
+linux_only = pytest.mark.skipif(
+    not sys.platform.startswith("linux"),
+    reason = "covers Linux-only DRM/KFD sysfs parsing driven by a fake /sys tree",
+)
+
 
 def _device(
     index,
@@ -99,6 +111,7 @@ def _fake_drm(tmp_path, monkeypatch, cards):
     return card_paths
 
 
+@linux_only
 def test_linux_vram_keyed_by_pci_excludes_foreign_adapters(monkeypatch, tmp_path):
     # Foreign (non-amdgpu) adapters contribute no entry, so they cannot shift ordinals.
     monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
@@ -117,6 +130,7 @@ def test_linux_vram_keyed_by_pci_excludes_foreign_adapters(monkeypatch, tmp_path
     }
 
 
+@linux_only
 def test_linux_vram_omits_bad_cards_without_shifting(monkeypatch, tmp_path):
     # A zero-total card has no entry; identity keying means its absence renumbers nothing.
     monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
@@ -131,6 +145,7 @@ def test_linux_vram_omits_bad_cards_without_shifting(monkeypatch, tmp_path):
     assert hw._rocm_linux_sysfs_vram_by_pci_gb() == {"0000:41:00.0": (2.0, 16.0)}
 
 
+@linux_only
 def test_linux_vram_omits_amd_card_without_vram_files(monkeypatch, tmp_path):
     # An APU with no mem_info_vram_* files has no entry; the discrete card keeps its address.
     monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
@@ -174,6 +189,7 @@ def _fake_kfd(tmp_path, monkeypatch, nodes):
     return node_paths
 
 
+@linux_only
 def test_kfd_lists_gpu_nodes_in_device_order(monkeypatch, tmp_path):
     # The CPU node (simd_count 0) takes no ordinal; GPU nodes in node-id order are HIP's order.
     monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
@@ -189,12 +205,14 @@ def test_kfd_lists_gpu_nodes_in_device_order(monkeypatch, tmp_path):
     assert hw._rocm_kfd_gpu_pci_ids() == ["0000:03:00.0", "0000:41:00.0"]
 
 
+@linux_only
 def test_kfd_decodes_domain_device_and_function(monkeypatch, tmp_path):
     monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
     _fake_kfd(tmp_path, monkeypatch, [(1, 64, (0xC1 << 8) | (0x1F << 3) | 5, 0x1234, _AMD)])
     assert hw._rocm_kfd_gpu_pci_ids() == ["1234:c1:1f.5"]
 
 
+@linux_only
 def test_kfd_skips_non_amd_gpu_nodes(monkeypatch, tmp_path):
     # An NVIDIA KFD node is not a HIP device: it must take no ordinal, else it
     # shifts every AMD GPU and ROCm device 1 resolves to AMD GPU 0.
@@ -212,6 +230,7 @@ def test_kfd_skips_non_amd_gpu_nodes(monkeypatch, tmp_path):
     assert hw._rocm_kfd_gpu_pci_ids() == ["0000:03:00.0", "0000:41:00.0"]
 
 
+@linux_only
 def test_kfd_fails_closed_when_a_gpu_has_no_location(monkeypatch, tmp_path):
     # Dropping an unplaceable AMD GPU shifts later ordinals; fail closed for the whole map.
     monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
@@ -226,6 +245,7 @@ def test_kfd_fails_closed_when_a_gpu_has_no_location(monkeypatch, tmp_path):
     assert hw._rocm_kfd_gpu_pci_ids() == []
 
 
+@linux_only
 def test_kfd_fails_closed_when_a_node_is_unreadable(monkeypatch, tmp_path):
     # An unreadable node could be a GPU; assuming otherwise would shift ordinals.
     monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
@@ -241,6 +261,23 @@ def test_kfd_fails_closed_when_a_node_is_unreadable(monkeypatch, tmp_path):
     assert hw._rocm_kfd_gpu_pci_ids() == []
 
 
+@linux_only
+def test_kfd_fails_closed_when_a_node_does_not_decode(monkeypatch, tmp_path):
+    # UnicodeDecodeError is a ValueError, so it slips past `except OSError` and
+    # would shift every later HIP ordinal.
+    monkeypatch.setattr(hw.platform, "system", lambda: "Linux")
+    paths = _fake_kfd(
+        tmp_path,
+        monkeypatch,
+        [
+            (1, 304, (0x03 << 8) | 0, 0, _AMD),
+            (2, 304, (0x41 << 8) | 0, 0, _AMD),
+        ],
+    )
+    (Path(paths[0]) / "properties").write_bytes(b"simd_count 304\nvendor_id \x80\xff\n")
+    assert hw._rocm_kfd_gpu_pci_ids() == []
+
+
 def test_kfd_absent_yields_no_device_order(monkeypatch):
     monkeypatch.setattr(hw.glob, "glob", lambda pattern: [])
     assert hw._rocm_kfd_gpu_pci_ids() == []
@@ -422,6 +459,10 @@ def test_visible_utilization_rocm_fallback_overlays(monkeypatch):
     ):
         monkeypatch.delenv(_var, raising = False)
     monkeypatch.setattr(hw, "IS_ROCM", True)
+    # No AMD adapter data on this host. On Windows this branch runs ahead of the torch
+    # fallback under test, and probing it imports torch, which the CI runner does not
+    # install. Off Windows the real function is never reached, so this changes nothing.
+    monkeypatch.setattr(hw, "_rocm_windows_per_device_vram", lambda ids: [])
     monkeypatch.setattr(hw, "get_device", lambda: hw.DeviceType.CUDA)
     monkeypatch.setattr(hw, "_smi_query", lambda *a, **k: None)  # amd-smi unavailable
     monkeypatch.setattr(
@@ -450,6 +491,10 @@ def test_visible_utilization_rocm_fallback_overlays(monkeypatch):
 def test_visible_utilization_relative_index_skips_overlay(monkeypatch):
     # UUID/MIG mask gives relative indices; the overlay matches physical index, so it must not run.
     monkeypatch.setattr(hw, "IS_ROCM", True)
+    # No AMD adapter data on this host. On Windows this branch runs ahead of the torch
+    # fallback under test, and probing it imports torch, which the CI runner does not
+    # install. Off Windows the real function is never reached, so this changes nothing.
+    monkeypatch.setattr(hw, "_rocm_windows_per_device_vram", lambda ids: [])
     monkeypatch.setattr(hw, "get_device", lambda: hw.DeviceType.CUDA)
     monkeypatch.setattr(hw, "_smi_query", lambda *a, **k: None)
     monkeypatch.setattr(
diff --git a/studio/backend/tests/test_safetensors_tool_loop.py b/studio/backend/tests/test_safetensors_tool_loop.py
index 2e7e99fbba..4a7b3ece20 100644
--- a/studio/backend/tests/test_safetensors_tool_loop.py
+++ b/studio/backend/tests/test_safetensors_tool_loop.py
@@ -2232,6 +2232,33 @@ def test_reprompt_names_only_active_tools_not_hardcoded():
     assert "python" not in reprompt["content"]
 
 
+def test_reprompt_stops_when_the_retry_restates_the_stall():
+    """A nudge answered with the same text has not worked; do not spend the budget."""
+
+    captured: list[list] = []
+    stall = "I'll search for that now."
+
+    def fake_single_turn(messages, active_tools = None):
+        captured.append(list(messages))
+        yield stall  # same forward-looking intent every time
+
+    exec_fn = FakeExecuteTool([])
+    _events = _collect_events(
+        run_safetensors_tool_loop(
+            single_turn = fake_single_turn,
+            messages = [{"role": "user", "content": "find X"}],
+            tools = [{"type": "function", "function": {"name": "search_knowledge_base"}}],
+            execute_tool = exec_fn,
+            auto_heal_tool_calls = True,
+            nudge_tool_calls = True,
+            max_tool_iterations = 3,
+        )
+    )
+
+    # One nudge, then the repeat guard stops it: two generations, not MAX_ACT_REPROMPTS + 1.
+    assert len(captured) == 2, captured
+
+
 def test_reprompt_is_announced_on_the_status_channel():
     # The re-prompted turn is hidden, so the badge is the only sign of life.
     # Blank still comes first: the route resets its text cursor only on that.
@@ -3624,8 +3651,22 @@ class TestGGUFSafetensorsHealingParity:
             "Let me check",
             "I am going to call the tool",
             "First, I will explore",
+            "First, let's search the web",
+            "First, let us search the web",
+            # Imperative plans carry no pronoun; an action verb is enough.
+            "First, search the web for the latest release notes.",
+            "First, check the documentation.",
+            "First, analyze the attached data",
+            "The first step is to search the web",
+            "First, my plan is to search the web.",
+            "First: search the web for release notes.",
+            "First - search the web for release notes.",
+            "First \u2013 search the web for release notes.",
+            "First, our approach is to check the docs.",
             "Here's my plan",
             "Now I need to call web_search",
+            # The "let me know" exemption is scoped to "let me", not all direct intent.
+            "I will know the answer after I search the web",
         ):
             assert shared_re.search(phrase), f"missed {phrase!r}"
             assert shared_fn(phrase), f"helper missed {phrase!r}"
@@ -3641,6 +3682,18 @@ class TestGGUFSafetensorsHealingParity:
             # force a tool-call re-prompt on it.
             "I will not search the web for that.",
             "I'll never call that tool.",
+            # Hands control back rather than announcing an action.
+            "Let me know if you need anything else.",
+            "First, the answer is 42",
+            "First, the result is 3.",
+            "First, it is 42",
+            "First, my answer is 42",
+            "The first line is blank.",
+            # Ordinal prose, not a plan.
+            "First place went to Alice",
+            "First class is available",
+            # Advice to the user, not work for this turn.
+            "First, install the package.",
         ):
             assert not shared_re.search(plain), f"wrongly fired on {plain!r}"
             assert not shared_fn(plain), f"helper wrongly fired on {plain!r}"
@@ -3653,6 +3706,98 @@ class TestGGUFSafetensorsHealingParity:
 
         assert gguf_cap == sf_cap == shared_cap
 
+    def test_reprompt_repeat_keeps_punctuation_bearing_terms(self):
+        # Stripping all non-word chars collapsed "C++" and "C#" to "c", so different
+        # plans compared equal and the retry lost its nudge.
+        from core.inference.tool_call_parser import is_reprompt_repeat
+        assert not is_reprompt_repeat("I will search for C#.", "I will search for C++.")
+        # A leading mark is part of the term too.
+        assert not is_reprompt_repeat("I will search for .NET", "I will search for NET")
+
+    def test_reprompt_repeat_respects_word_order(self):
+        # Set overlap scores a reordered query as identical, so the comparison is
+        # sequence-based.
+        from core.inference.tool_call_parser import is_reprompt_repeat
+
+        assert not is_reprompt_repeat(
+            "I will search for dogs not cats", "I will search for cats not dogs"
+        )
+        assert is_reprompt_repeat(
+            "I will search for cats not dogs", "I will search for cats not dogs"
+        )
+        assert is_reprompt_repeat("I will search for C++!", "I will search for C++.")
+
+    def test_reprompt_repeat_keeps_a_changed_query_token(self):
+        # One corrected token in a long plan is a new attempt; at the old 0.85 bar it
+        # scored ~0.87 and cost the model its remaining nudge.
+        from core.inference.tool_call_parser import is_reprompt_repeat
+
+        before = "I will search the web for the latest CUDA version 12.4 driver release notes"
+        after = "I will search the web for the latest CUDA version 12.5 driver release notes"
+        assert not is_reprompt_repeat(after, before)
+        assert is_reprompt_repeat(before, before)
+
+    def test_reprompt_repeat_keeps_standalone_operator_tokens(self):
+        # A marks-only token stripped to nothing, so a bounded correction compared
+        # equal to the unbounded original.
+        from core.inference.tool_call_parser import is_reprompt_repeat, is_reprompt_restatement
+
+        loose = "Now I think the value is 5"
+        bounded = "Now I think the value is < 5"
+        assert not is_reprompt_repeat(bounded, loose)
+        assert not is_reprompt_restatement(bounded, loose)
+
+    def test_reprompt_repeat_keeps_a_changed_token_in_a_long_plan(self):
+        # Every similarity ratio is length-dependent: one changed token scored 0.98
+        # across 54 tokens, so long corrected plans lost their nudge.
+        from core.inference.tool_call_parser import is_reprompt_repeat
+
+        words = [f"token{index}" for index in range(54)]
+        corrected = list(words)
+        corrected[20] = "revised"
+        assert not is_reprompt_repeat(" ".join(corrected), " ".join(words))
+        assert is_reprompt_repeat(" ".join(words), " ".join(words))
+
+    def test_reprompt_repeat_keeps_articles_that_name_a_target(self):
+        # "The Who" and "Who" are different searches, so articles are not filler.
+        from core.inference.tool_call_parser import is_reprompt_repeat
+        assert not is_reprompt_repeat(
+            "I will search for The Who discography",
+            "I will search for Who discography",
+        )
+
+    def test_reprompt_repeat_keeps_filler_words_that_name_a_target(self):
+        # No word is reliably filler: dropping "ok"/"the" to absorb rewording also
+        # absorbed the search target. Reordered filler now reads as a new attempt,
+        # which costs one nudge out of the cap and never strands a plan.
+        from core.inference.tool_call_parser import is_reprompt_repeat
+        assert not is_reprompt_repeat(
+            "I will search for OK Go discography",
+            "I will search for Go discography",
+        )
+        assert not is_reprompt_repeat(
+            "I will now summarize the findings",
+            "I will summarize the findings now",
+        )
+
+    def test_reprompt_repeat_detects_restated_answers(self):
+        # A nudge answered with the same text again has not worked; stop there.
+        from core.inference.tool_call_parser import is_reprompt_repeat
+
+        same = "I will summarize what I found."
+        assert is_reprompt_repeat(same, same)
+        assert is_reprompt_repeat("I WILL summarize what I found!", same)
+        assert is_reprompt_repeat(
+            "The summary is ready, please let me know if you need anything else",
+            "The summary is ready. Please let me know if you need anything else!",
+        )
+
+        # No previous text, or genuinely different progress, keeps the nudge.
+        assert not is_reprompt_repeat(same, "")
+        assert not is_reprompt_repeat("Tokyo is 18C and cloudy right now.", same)
+        # Short texts must not collide on incidental word overlap.
+        assert not is_reprompt_repeat("Let me check.", "Let me search.")
+
 
 class TestLoopControl:
     def test_cancel_event_breaks_loop(self):
@@ -4182,9 +4327,11 @@ class TestPlanWithoutActionReprompt:
         # final answer and no further turn is generated.
         from core.inference.tool_call_parser import MAX_ACT_REPROMPTS
 
-        stall = "Let me look into it first."
+        # Distinct stalls: identical ones stop at the repeat guard, never reaching the cap.
+        stalls = [f"Let me look into detail {i} first." for i in range(MAX_ACT_REPROMPTS)]
+        stall = stalls[-1]
         turns = [["I'll search the web for that."]]
-        turns += [[stall]] * MAX_ACT_REPROMPTS
+        turns += [[s] for s in stalls]
         turns += [["SHOULD NOT APPEAR"]]
 
         generations = {"count": 0}
diff --git a/studio/backend/tests/test_studio_pid_files.py b/studio/backend/tests/test_studio_pid_files.py
new file mode 100644
index 0000000000..df2c8e87f8
--- /dev/null
+++ b/studio/backend/tests/test_studio_pid_files.py
@@ -0,0 +1,568 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Per-port PID files, so `unsloth studio stop` can find every server.
+
+Imports run.py directly, so run under the Unsloth venv.
+"""
+
+from __future__ import annotations
+
+import os
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+_BACKEND = Path(__file__).resolve().parents[1]
+if str(_BACKEND) not in sys.path:
+    sys.path.insert(0, str(_BACKEND))
+
+import run  # noqa: E402
+
+# Captured before the autouse fixture stubs them, for the tests that exercise them.
+_REAL_IS_STUDIO_BACKEND = run._pid_is_studio_backend
+_REAL_PID_ALIVE = run._pid_alive
+
+
+@pytest.fixture(autouse = True)
+def isolated_root(tmp_path, monkeypatch):
+    monkeypatch.setattr(run, "_studio_root", lambda: tmp_path)
+    monkeypatch.setattr(run, "_PID_FILE", tmp_path / "studio.pid")
+    monkeypatch.setattr(run, "_OWN_PID_FILE", None)
+    monkeypatch.setattr(run, "_pid_alive", lambda pid: True)
+    monkeypatch.setattr(run, "_pid_is_studio_backend", lambda pid, created_times = (): True)
+    yield
+
+
+def _files(tmp_path):
+    return sorted(p.name for p in tmp_path.glob("studio-*.pid"))
+
+
+def _pid_of(path):
+    return path.read_text(encoding = "utf-8").splitlines()[0]
+
+
+def test_write_pid_file_records_port_and_pid(tmp_path):
+    run._write_pid_file(8901)
+
+    assert _files(tmp_path) == [f"studio-8901-{os.getpid()}.pid"]
+    assert _pid_of(tmp_path / f"studio-8901-{os.getpid()}.pid") == str(os.getpid())
+
+
+def test_write_pid_file_records_the_start_time(tmp_path):
+    # Pins the record to this process, so a reused PID isn't mistaken for it.
+    run._write_pid_file(8901)
+
+    record = run._read_pid_record(tmp_path / f"studio-8901-{os.getpid()}.pid")
+
+    assert record[0] == os.getpid()
+    assert record[1] == pytest.approx(run._process_create_time(os.getpid()))
+
+
+def test_write_pid_file_keeps_the_legacy_file_a_bare_pid(tmp_path):
+    # An older CLI's `stop` reads studio.pid and expects only digits.
+    run._write_pid_file(8901)
+
+    assert (tmp_path / "studio.pid").read_text(encoding = "utf-8") == str(os.getpid())
+
+
+def test_second_port_does_not_clobber_the_first(tmp_path):
+    (tmp_path / "studio-8901-8550.pid").write_text("8550", encoding = "utf-8")
+
+    run._write_pid_file(8902)
+
+    assert _pid_of(tmp_path / "studio-8901-8550.pid") == "8550"
+    assert (tmp_path / f"studio-8902-{os.getpid()}.pid").exists()
+
+
+def test_same_port_on_two_binds_does_not_clobber(tmp_path):
+    # 127.0.0.1:8888 and ::1:8888 can both listen; one file per port would lose one.
+    (tmp_path / "studio-8888-8550.pid").write_text("8550", encoding = "utf-8")
+
+    run._write_pid_file(8888)
+
+    assert len(_files(tmp_path)) == 2
+
+
+def test_remove_pid_file_only_removes_our_own(tmp_path, monkeypatch):
+    run._write_pid_file(8901)
+    (tmp_path / "studio-8902-8600.pid").write_text("8600", encoding = "utf-8")
+    # Nothing to hand the legacy pointer to, so it goes away with us.
+    monkeypatch.setattr(run, "_pid_alive", lambda pid: pid == os.getpid())
+
+    run._remove_pid_file()
+
+    assert _files(tmp_path) == ["studio-8902-8600.pid"]
+    assert not (tmp_path / "studio.pid").exists()
+
+
+def test_the_legacy_pointer_moves_to_a_live_sibling(tmp_path):
+    # Only one server owns studio.pid. Deleting it on our way out would leave an
+    # older CLI, which reads nothing else, unable to stop the sibling still up.
+    run._write_pid_file(8901)
+    (tmp_path / "studio-8902-8600.pid").write_text("8600", encoding = "utf-8")
+
+    run._remove_pid_file()
+
+    assert (tmp_path / "studio.pid").read_text(encoding = "utf-8").strip() == "8600"
+
+
+def test_the_legacy_pointer_is_not_handed_to_a_dead_sibling(tmp_path, monkeypatch):
+    run._write_pid_file(8901)
+    (tmp_path / "studio-8902-8600.pid").write_text("8600", encoding = "utf-8")
+    monkeypatch.setattr(run, "_pid_is_studio_backend", lambda pid, created_times = (): False)
+
+    run._remove_pid_file()
+
+    assert not (tmp_path / "studio.pid").exists()
+
+
+def test_remove_pid_file_leaves_a_reused_entry_alone(tmp_path):
+    run._write_pid_file(8901)
+    own = tmp_path / f"studio-8901-{os.getpid()}.pid"
+    own.write_text("999999", encoding = "utf-8")
+
+    run._remove_pid_file()
+
+    assert own.read_text(encoding = "utf-8") == "999999"
+
+
+def test_windows_liveness_does_not_call_every_pid_alive(monkeypatch):
+    # os.kill(pid, 0) raises OSError for every pid on Windows, so without the
+    # tasklist fallback a stale record would block its port forever.
+    import subprocess
+
+    monkeypatch.setattr(run, "_pid_alive", _REAL_PID_ALIVE)
+    monkeypatch.setitem(sys.modules, "psutil", None)
+    monkeypatch.setattr(sys, "platform", "win32")
+    monkeypatch.setattr(
+        subprocess, "run", lambda *a, **k: SimpleNamespace(stdout = '"python.exe","8550",...')
+    )
+
+    assert run._pid_alive(8550) is True
+    assert run._pid_alive(9999) is False
+
+
+def test_windows_liveness_keeps_the_record_when_tasklist_fails(monkeypatch):
+    # Unconfirmed must mean keep, matching the CLI's _pid_alive. Pruning a live
+    # server's record lets the next launch fall back past it and strand it, which
+    # is the bug this file exists to fix; a stale record costs one clear abort.
+    import subprocess
+
+    def _boom(*a, **k):
+        raise OSError("tasklist missing")
+
+    monkeypatch.setattr(run, "_pid_alive", _REAL_PID_ALIVE)
+    monkeypatch.setitem(sys.modules, "psutil", None)
+    monkeypatch.setattr(sys, "platform", "win32")
+    monkeypatch.setattr(subprocess, "run", _boom)
+
+    assert run._pid_alive(8550) is True
+
+
+def test_read_pid_record_parses_pid_time_and_address(tmp_path):
+    (tmp_path / "r.pid").write_text("8550\n111.5\n127.0.0.1", encoding = "utf-8")
+
+    assert run._read_pid_record(tmp_path / "r.pid") == (8550, 111.5, "127.0.0.1")
+
+
+def test_read_pid_record_tolerates_a_bare_pid(tmp_path):
+    (tmp_path / "r.pid").write_text("8550", encoding = "utf-8")
+
+    assert run._read_pid_record(tmp_path / "r.pid") == (8550, None, None)
+
+
+def test_read_pid_record_rejects_pid_zero_and_init(tmp_path):
+    # kill(0) signals our whole process group.
+    (tmp_path / "zero.pid").write_text("0", encoding = "utf-8")
+    (tmp_path / "init.pid").write_text("1", encoding = "utf-8")
+
+    assert run._read_pid_record(tmp_path / "zero.pid") is None
+    assert run._read_pid_record(tmp_path / "init.pid") is None
+
+
+def test_read_pid_record_rejects_a_corrupt_file(tmp_path):
+    (tmp_path / "r.pid").write_text("not-a-pid", encoding = "utf-8")
+
+    assert run._read_pid_record(tmp_path / "r.pid") is None
+
+
+def test_graceful_shutdown_drops_the_record_last(monkeypatch):
+    # Cleanup can take seconds while the server is still alive. Dropping the record
+    # first leaves a retried `stop` or a new launch unable to find it.
+    order = []
+    monkeypatch.setattr(run, "_remove_pid_file", lambda: order.append("remove_record"))
+
+    class _Server:
+        def __setattr__(self, name, value):
+            order.append("release_socket")
+
+    run._graceful_shutdown(_Server())
+
+    assert order == ["release_socket", "remove_record"]
+
+
+def test_own_studio_on_port_is_found_without_psutil(tmp_path, monkeypatch):
+    # psutil is optional; a listener scan finds nothing without it, so detection
+    # must come from our own records or we silently start a duplicate.
+    monkeypatch.setitem(sys.modules, "psutil", None)
+    (tmp_path / "studio-8901-8550.pid").write_text("8550\n\n127.0.0.1", encoding = "utf-8")
+
+    assert run._own_studio_on_port(8901, "127.0.0.1") == 8550
+
+
+def test_no_record_for_the_port_means_no_own_studio(tmp_path):
+    # jupyter-lab on 8888 must keep the fallback, not abort the launch.
+    (tmp_path / "studio-8901-8550.pid").write_text("8550", encoding = "utf-8")
+
+    assert run._own_studio_on_port(8888, "127.0.0.1") is None
+
+
+def test_own_studio_on_port_prunes_a_dead_record(tmp_path, monkeypatch):
+    monkeypatch.setattr(run, "_pid_alive", lambda pid: False)
+    (tmp_path / "studio-8901-8550.pid").write_text("8550", encoding = "utf-8")
+
+    assert run._own_studio_on_port(8901, "127.0.0.1") is None
+    assert not (tmp_path / "studio-8901-8550.pid").exists()
+
+
+def test_a_reused_pid_is_not_treated_as_our_studio(tmp_path, monkeypatch):
+    # Stale record + the OS handing that PID to something else must not abort.
+    monkeypatch.setattr(run, "_pid_is_studio_backend", lambda pid, created_times = (): False)
+    (tmp_path / "studio-8901-8550.pid").write_text("8550", encoding = "utf-8")
+
+    assert run._own_studio_on_port(8901, "127.0.0.1") is None
+
+
+def test_an_unverifiable_record_still_blocks_a_duplicate(tmp_path, monkeypatch):
+    # Can't tell: refusing with a clear message beats a silent second instance.
+    monkeypatch.setattr(run, "_pid_is_studio_backend", lambda pid, created_times = (): True)
+    (tmp_path / "studio-8901-8550.pid").write_text("8550", encoding = "utf-8")
+
+    assert run._own_studio_on_port(8901, "127.0.0.1") == 8550
+
+
+def test_start_time_mismatch_rejects_a_reused_pid(monkeypatch):
+    monkeypatch.setattr(run, "_pid_is_studio_backend", _REAL_IS_STUDIO_BACKEND)
+    monkeypatch.setattr(run, "_process_create_time", lambda pid: 999.0)
+
+    assert run._pid_is_studio_backend(8550, [111.5]) is False
+    assert run._pid_is_studio_backend(8550, [999.0]) is True
+
+
+def test_a_stale_record_does_not_veto_a_live_server_sharing_the_pid(monkeypatch):
+    # Crash leaves studio-8888-1234.pid, the OS reuses 1234 for a new server on
+    # another port. Keeping only the first timestamp would reject the live one.
+    monkeypatch.setattr(run, "_pid_is_studio_backend", _REAL_IS_STUDIO_BACKEND)
+    monkeypatch.setattr(run, "_process_create_time", lambda pid: 999.0)
+
+    assert run._pid_is_studio_backend(1234, [111.5, 999.0]) is True
+    assert run._pid_is_studio_backend(1234, [111.5, 222.5]) is False
+
+
+def test_a_stale_record_on_another_port_does_not_hide_a_live_server(tmp_path, monkeypatch):
+    # 1234 was reused: the stale 8888 record must not stop us seeing 9000.
+    monkeypatch.setattr(run, "_pid_is_studio_backend", _REAL_IS_STUDIO_BACKEND)
+    monkeypatch.setattr(run, "_process_create_time", lambda pid: 999.0)
+    (tmp_path / "studio-8888-1234.pid").write_text("1234\n111.5\n", encoding = "utf-8")
+    (tmp_path / "studio-9000-1234.pid").write_text("1234\n999.0\n", encoding = "utf-8")
+
+    assert run._own_studio_on_port(8888, "127.0.0.1") is None
+    assert run._own_studio_on_port(9000, "127.0.0.1") == 1234
+
+
+def test_a_start_time_is_the_only_thing_that_disproves_a_record(monkeypatch):
+    monkeypatch.setattr(run, "_pid_is_studio_backend", _REAL_IS_STUDIO_BACKEND)
+    monkeypatch.setattr(run, "_process_create_time", lambda pid: 999.0)
+
+    assert run._pid_is_studio_backend(8550, [999.0]) is True
+    assert run._pid_is_studio_backend(8550, [111.5]) is False
+
+
+def test_a_bare_run_py_command_line_is_not_rejected(monkeypatch):
+    # `cd studio/backend && python run.py --port 8901` has no "studio" or "unsloth"
+    # in argv. Guessing from the command line called that "not ours".
+    monkeypatch.setattr(run, "_pid_is_studio_backend", _REAL_IS_STUDIO_BACKEND)
+
+    class _FakeProcess:
+        def __init__(self, pid):
+            self.pid = pid
+
+        def cmdline(self):
+            return ["python", "run.py", "--port", "8901"]
+
+        def create_time(self):
+            return 111.5
+
+    monkeypatch.setitem(sys.modules, "psutil", SimpleNamespace(Process = _FakeProcess))
+
+    assert run._pid_is_studio_backend(8550) is True
+
+
+def test_an_untimed_legacy_record_is_trusted(monkeypatch):
+    # `python run.py --port 8901` has no telltale argv, so guessing from the
+    # command line rejected real servers. Only a start time can disprove one.
+    monkeypatch.setattr(run, "_pid_is_studio_backend", _REAL_IS_STUDIO_BACKEND)
+    monkeypatch.setattr(run, "_process_create_time", lambda pid: 999.0)
+
+    assert run._pid_is_studio_backend(8550) is True
+    assert run._pid_is_studio_backend(8550, [None]) is True
+
+
+def test_the_untimed_legacy_record_does_not_cancel_a_timed_one(monkeypatch):
+    # Mirrors _pid_is_studio_server in the CLI. An untimed record carries no
+    # information, so it must not overrule a start time that says "not ours" --
+    # every current server writes one of each, which made the check inert.
+    monkeypatch.setattr(run, "_pid_is_studio_backend", _REAL_IS_STUDIO_BACKEND)
+    monkeypatch.setattr(run, "_process_create_time", lambda pid: 999.0)
+
+    assert run._pid_is_studio_backend(8550, [111.5, None]) is False
+    assert run._pid_is_studio_backend(8550, [111.5, 999.0]) is True
+
+
+def test_a_legacy_server_on_the_port_is_recognised(tmp_path, monkeypatch):
+    # Pre-upgrade servers wrote only studio.pid. Falling back past one strands it
+    # and then overwrites its record.
+    monkeypatch.setattr(run, "_get_pid_on_port", lambda p: (8550, "python"))
+    (tmp_path / "studio.pid").write_text("8550", encoding = "utf-8")
+
+    assert run._own_studio_on_port(8901, "127.0.0.1") == 8550
+
+
+def test_a_legacy_record_for_a_different_listener_falls_back(tmp_path, monkeypatch):
+    # jupyter holds the port; the legacy server is elsewhere. Keep falling back.
+    monkeypatch.setattr(run, "_get_pid_on_port", lambda p: (117, "jupyter-lab"))
+    (tmp_path / "studio.pid").write_text("8550", encoding = "utf-8")
+
+    assert run._own_studio_on_port(8901, "127.0.0.1") is None
+
+
+def test_an_unknowable_listener_treats_the_legacy_record_as_ours(tmp_path, monkeypatch):
+    # No psutil: _get_pid_on_port can't say. Refusing beats a silent duplicate.
+    monkeypatch.setattr(run, "_get_pid_on_port", lambda p: None)
+    (tmp_path / "studio.pid").write_text("8550", encoding = "utf-8")
+
+    assert run._own_studio_on_port(8901, "127.0.0.1") == 8550
+
+
+def test_a_dead_legacy_record_falls_back(tmp_path, monkeypatch):
+    monkeypatch.setattr(run, "_pid_alive", lambda pid: False)
+    monkeypatch.setattr(run, "_get_pid_on_port", lambda p: None)
+    (tmp_path / "studio.pid").write_text("8550", encoding = "utf-8")
+
+    assert run._own_studio_on_port(8901, "127.0.0.1") is None
+
+
+def test_a_stale_per_port_record_does_not_mask_a_legacy_server(tmp_path, monkeypatch):
+    # Crashed current build left studio-8901-8550.pid; 8550 was then reused by a
+    # pre-upgrade server recorded only in studio.pid. The stale record must not
+    # count as "port already known" and send us falling back past the live one.
+    monkeypatch.setattr(run, "_pid_is_studio_backend", _REAL_IS_STUDIO_BACKEND)
+    monkeypatch.setattr(run, "_process_create_time", lambda pid: 999.0)
+    monkeypatch.setattr(run, "_get_pid_on_port", lambda p: (8550, "python"))
+    (tmp_path / "studio-8901-8550.pid").write_text("8550\n111.5\n127.0.0.1", encoding = "utf-8")
+    (tmp_path / "studio.pid").write_text("8550", encoding = "utf-8")
+
+    assert run._own_studio_on_port(8901, "127.0.0.1") == 8550
+
+
+def test_a_current_server_elsewhere_does_not_block_a_foreign_port(tmp_path, monkeypatch):
+    # Current builds write studio.pid too. Without psutil the legacy check can't
+    # see the listener, so it must not claim our 8901 server holds jupyter's 8888.
+    monkeypatch.setattr(run, "_get_pid_on_port", lambda p: None)
+    (tmp_path / "studio-8901-5000.pid").write_text("5000\n\n127.0.0.1", encoding = "utf-8")
+    (tmp_path / "studio.pid").write_text("5000", encoding = "utf-8")
+
+    assert run._own_studio_on_port(8888, "127.0.0.1") is None
+
+
+def test_a_per_port_record_is_preferred_over_the_legacy_one(tmp_path, monkeypatch):
+    monkeypatch.setattr(run, "_get_pid_on_port", lambda p: (8550, "python"))
+    (tmp_path / "studio-8901-8600.pid").write_text("8600\n\n127.0.0.1", encoding = "utf-8")
+    (tmp_path / "studio.pid").write_text("8550", encoding = "utf-8")
+
+    assert run._own_studio_on_port(8901, "127.0.0.1") == 8600
+
+
+def test_our_studio_on_another_bind_address_does_not_abort(tmp_path):
+    # Our server holds ::1:8889; binding 127.0.0.1:8889 is not a conflict with us,
+    # so fall through to the next port instead of refusing.
+    (tmp_path / "studio-8889-8550.pid").write_text("8550\n\n::1", encoding = "utf-8")
+
+    assert run._own_studio_on_port(8889, "127.0.0.1") is None
+    assert run._own_studio_on_port(8889, "::1") == 8550
+
+
+def test_address_matching(tmp_path):
+    assert run._addresses_collide("0.0.0.0", "127.0.0.1", 8889) is True
+    assert run._addresses_collide("127.0.0.1", "0.0.0.0", 8889) is True
+    assert run._addresses_collide("127.0.0.1", "127.0.0.1", 8889) is True
+    assert run._addresses_collide("::1", "127.0.0.1", 8889) is False
+    # An unrecorded address is unknown, so assume a conflict.
+    assert run._addresses_collide(None, "127.0.0.1", 8889) is True
+
+
+def test_a_hostname_resolves_the_same_way_the_bind_does(tmp_path):
+    # `localhost` and the address _is_port_free actually binds must agree, or a
+    # recorded server is missed and a duplicate starts.
+    recorded = ",".join(sorted(run._bind_addresses("localhost", 8889)))
+
+    assert run._addresses_collide(recorded, "localhost", 8889) is True
+
+
+def test_a_hostname_records_every_address_it_resolves_to(tmp_path):
+    # `localhost` binds 127.0.0.1 AND ::1. Recording only the first lets a later
+    # launch on the other literal miss us and start a duplicate.
+    addrs = run._bind_addresses("localhost", 8889)
+    recorded = ",".join(sorted(addrs))
+
+    for literal in addrs:
+        assert run._addresses_collide(recorded, literal, 8889) is True
+
+
+def test_a_multi_address_record_matches_either_literal(tmp_path):
+    recorded = "127.0.0.1,::1"
+
+    assert run._addresses_collide(recorded, "127.0.0.1", 8889) is True
+    assert run._addresses_collide(recorded, "::1", 8889) is True
+    assert run._addresses_collide("127.0.0.1", "::1", 8889) is False
+
+
+def test_fallback_aborts_on_our_own_server_further_up_the_range(tmp_path, monkeypatch):
+    # jupyter holds 8888, our server holds 8889: skipping to 8890 is the duplicate.
+    (tmp_path / "studio-8889-8550.pid").write_text("8550\n\n127.0.0.1", encoding = "utf-8")
+    monkeypatch.setattr(run, "_is_port_free", lambda host, p: p >= 8890)
+
+    with pytest.raises(SystemExit) as excinfo:
+        run._find_free_port("127.0.0.1", 8889, avoid_own_studio = True)
+
+    assert excinfo.value.code == 1
+
+
+def test_fallback_still_skips_foreign_processes(tmp_path, monkeypatch):
+    # No record for 8889, so the blocker is not ours: keep falling back.
+    monkeypatch.setattr(run, "_is_port_free", lambda host, p: p >= 8890)
+
+    assert run._find_free_port("127.0.0.1", 8889, avoid_own_studio = True) == 8890
+
+
+def test_the_requested_port_is_kept_when_it_is_free(monkeypatch):
+    monkeypatch.setattr(run, "_is_port_free", lambda host, p: True)
+
+    assert run._resolve_port("127.0.0.1", 8888) == 8888
+
+
+def test_our_own_server_on_the_requested_port_aborts_rather_than_falling_back(
+    tmp_path, monkeypatch
+):
+    # The reported bug: 8888 is ours, so falling back to 8889 is the duplicate
+    # that leaves 8888 serving with nothing recording it.
+    monkeypatch.setattr(run, "_is_port_free", lambda host, p: p != 8888)
+    (tmp_path / "studio-8888-8550.pid").write_text("8550\n\n127.0.0.1", encoding = "utf-8")
+
+    with pytest.raises(SystemExit) as excinfo:
+        run._resolve_port("127.0.0.1", 8888)
+
+    assert excinfo.value.code == 1
+
+
+def test_a_foreign_process_on_the_requested_port_still_falls_back(monkeypatch):
+    # jupyter-lab on 8888 must not stop Unsloth starting on 8889.
+    monkeypatch.setattr(run, "_is_port_free", lambda host, p: p != 8888)
+
+    assert run._resolve_port("127.0.0.1", 8888) == 8889
+
+
+def test_a_caller_that_reads_the_port_back_keeps_the_plain_fallback(tmp_path, monkeypatch):
+    # api-only callers (the desktop app via TAURI_PORT, `studio run` via
+    # app.state.server_port) follow us to the new port, so aborting there only
+    # turns a working launch into a crash the desktop app reports as "stopped
+    # unexpectedly". Both servers are still recorded, so `stop` finds them.
+    monkeypatch.setattr(run, "_is_port_free", lambda host, p: p != 8888)
+    (tmp_path / "studio-8888-8550.pid").write_text("8550\n\n127.0.0.1", encoding = "utf-8")
+
+    assert run._resolve_port("127.0.0.1", 8888, avoid_own_studio = False) == 8889
+
+
+def test_the_recorded_address_is_every_address_the_bind_resolves_to(tmp_path):
+    # The only test that runs the writer with a real host. Recording `host`
+    # verbatim, or dropping the line, passes every other test here and silently
+    # stops matching a launch that spells the same interface differently.
+    run._write_pid_file(8901, "localhost")
+
+    record = run._read_pid_record(tmp_path / f"studio-8901-{os.getpid()}.pid")
+
+    assert record[2] is not None, "no bind address recorded"
+    assert set(record[2].split(",")) == run._bind_addresses("localhost", 8901)
+
+
+def test_a_server_started_on_a_hostname_is_found_again_by_ip(tmp_path):
+    run._write_pid_file(8901, "localhost")
+
+    for literal in run._bind_addresses("localhost", 8901):
+        assert run._own_studio_on_port(8901, literal) == os.getpid()
+
+
+def test_bind_addresses_keeps_every_family_a_hostname_resolves_to(monkeypatch):
+    # Independent oracle: the sibling test derives its expectation from this
+    # function's own output, so dropping a family would pass it.
+    import socket
+    monkeypatch.setattr(
+        socket,
+        "getaddrinfo",
+        lambda *a, **k: [
+            (socket.AF_INET, socket.SOCK_STREAM, 6, "", ("127.0.0.1", 8889)),
+            (socket.AF_INET6, socket.SOCK_STREAM, 6, "", ("::1", 8889, 0, 0)),
+        ],
+    )
+
+    assert run._bind_addresses("localhost", 8889) == {"127.0.0.1", "::1"}
+
+
+def test_the_legacy_file_is_written_even_when_the_per_port_record_fails(tmp_path, monkeypatch):
+    # A studio root that cannot take a new entry used to leave the server
+    # recorded nowhere at all, so the CLI could not stop it. studio.pid is an
+    # overwrite of an existing path, so it can still succeed and must be tried.
+    blocked = tmp_path / "not-a-directory"
+    blocked.write_text("", encoding = "utf-8")
+    monkeypatch.setattr(
+        run, "_pid_file_for_port", lambda port: blocked / f"studio-{port}-{os.getpid()}.pid"
+    )
+
+    run._write_pid_file(8901, "127.0.0.1")
+
+    assert (tmp_path / "studio.pid").read_text(encoding = "utf-8") == str(os.getpid())
+    assert run._OWN_PID_FILE is None
+
+
+def test_a_record_whose_pid_is_not_ascii_digits_is_discarded(tmp_path):
+    # A superscript two passes isdigit() but int() rejects it, so that gate alone
+    # let a ValueError escape into every caller of _read_pid_record.
+    (tmp_path / "r.pid").write_text("²", encoding = "utf-8")
+
+    assert run._read_pid_record(tmp_path / "r.pid") is None
+
+
+def test_the_legacy_file_is_not_taken_from_a_live_server(tmp_path):
+    # A pre-upgrade server is recorded in studio.pid and nowhere else, so a
+    # second launch overwriting it is exactly what strands it. That is the
+    # orphan this file exists to prevent, reached from the other direction.
+    (tmp_path / "studio.pid").write_text("8550", encoding = "utf-8")
+
+    run._write_pid_file(8902, "127.0.0.1")
+
+    assert (tmp_path / "studio.pid").read_text(encoding = "utf-8") == "8550"
+    assert (tmp_path / f"studio-8902-{os.getpid()}.pid").exists()
+
+
+def test_the_legacy_file_is_taken_over_from_a_dead_server(tmp_path, monkeypatch):
+    # A stale record must not keep the pointer forever, or an older CLI could
+    # never stop anything again.
+    monkeypatch.setattr(run, "_pid_alive", lambda pid: False)
+    (tmp_path / "studio.pid").write_text("8550", encoding = "utf-8")
+
+    run._write_pid_file(8902, "127.0.0.1")
+
+    assert (tmp_path / "studio.pid").read_text(encoding = "utf-8") == str(os.getpid())
diff --git a/studio/backend/tests/test_text_io_encoding.py b/studio/backend/tests/test_text_io_encoding.py
new file mode 100644
index 0000000000..7eae3c7fef
--- /dev/null
+++ b/studio/backend/tests/test_text_io_encoding.py
@@ -0,0 +1,809 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Text I/O must name its encoding, or Windows silently uses the ANSI codepage.
+
+``open()``, ``Path.read_text()`` and ``subprocess(text = True)`` fall back to
+``locale.getencoding()`` when no ``encoding`` is passed. On Windows that is
+cp1252 (or cp932, cp1251, ... by system locale), not UTF-8, so a chat template,
+model config or path containing ``ä ö ü → 世`` mojibakes or raises
+``UnicodeDecodeError`` mid-load. Studio's files are UTF-8, so say so.
+"""
+
+from __future__ import annotations
+
+import ast
+import importlib.util
+import json
+import os
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+
+BACKEND_ROOT = Path(__file__).resolve().parent.parent
+
+# Not runtime source. Shipped plugins under plugins/*/src are, so only builds are skipped.
+_SKIPPED_DIRS = ("node_modules", "build", "tests", "__pycache__")
+
+# Path.open()'s signature is what tells it apart from other libraries' open(),
+# e.g. fitz.open(stream=...) and av.open(..., metadata_errors=...).
+_FILE_MODE_CHARS = set("rwxabt+")
+_PATH_OPEN_ARGS = ("mode", "buffering", "encoding", "errors", "newline")
+_PATH_OPEN_KWARGS = set(_PATH_OPEN_ARGS)
+_PATH_OPEN_ENCODING_ARG = _PATH_OPEN_ARGS.index("encoding")
+
+_SUBPROCESS_CALLS = {"run", "Popen", "check_output", "check_call", "call"}
+
+# open(file, mode, buffering, encoding, ...), and os.fdopen forwards the same
+# signature with a descriptor in place of the path.
+_OPEN_ENCODING_ARG = 3
+
+
+def _studio_sources() -> list[Path]:
+    return [
+        path
+        for path in sorted(BACKEND_ROOT.rglob("*.py"))
+        if not any(part in _SKIPPED_DIRS for part in path.relative_to(BACKEND_ROOT).parts)
+    ]
+
+
+def _has_keyword(node: ast.Call, name: str) -> bool:
+    return any(keyword.arg == name for keyword in node.keywords)
+
+
+def _mode_is_binary(node: ast.Call) -> bool:
+    mode: str | None = None
+    if len(node.args) >= 2 and isinstance(node.args[1], ast.Constant):
+        value = node.args[1].value
+        mode = value if isinstance(value, str) else None
+    for keyword in node.keywords:
+        if keyword.arg == "mode" and isinstance(keyword.value, ast.Constant):
+            value = keyword.value.value
+            if isinstance(value, str):
+                mode = value
+    return bool(mode and "b" in mode)
+
+
+def _open_has_encoding(node: ast.Call) -> bool:
+    """open()/os.fdopen() also take encoding positionally: open(p, "w", 1, "utf-8")."""
+    return _has_keyword(node, "encoding") or len(node.args) > _OPEN_ENCODING_ARG
+
+
+def _path_open_mode(node: ast.Call) -> str | None:
+    if node.args and isinstance(node.args[0], ast.Constant):
+        value = node.args[0].value
+        if isinstance(value, str):
+            return value
+    for keyword in node.keywords:
+        if keyword.arg == "mode" and isinstance(keyword.value, ast.Constant):
+            value = keyword.value.value
+            if isinstance(value, str):
+                return value
+    return None
+
+
+def _is_path_open(node: ast.Call) -> bool:
+    """True only for calls matching ``Path.open``'s signature."""
+    if len(node.args) > len(_PATH_OPEN_ARGS):
+        return False
+    if any(k.arg not in _PATH_OPEN_KWARGS for k in node.keywords):
+        return False
+    mode = _path_open_mode(node)
+    if mode is not None:
+        return bool(mode) and set(mode) <= _FILE_MODE_CHARS
+    return not node.args
+
+
+def _path_open_has_encoding(node: ast.Call) -> bool:
+    """Path.open() also takes encoding positionally: open("w", 1, "utf-8")."""
+    return _has_keyword(node, "encoding") or len(node.args) > _PATH_OPEN_ENCODING_ARG
+
+
+def _call_name(node: ast.Call) -> str | None:
+    func = node.func
+    if isinstance(func, ast.Name):
+        return func.id
+    if isinstance(func, ast.Attribute):
+        return func.attr
+    return None
+
+
+def _subprocess_names(tree: ast.AST) -> set[str]:
+    """Names subprocess is reachable under here, e.g. `import subprocess as _sp`."""
+    names = set()
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Import):
+            for alias in node.names:
+                if alias.name == "subprocess":
+                    names.add(alias.asname or alias.name)
+    return names
+
+
+def _subprocess_aliases(tree: ast.AST, names: set[str]) -> set[str]:
+    """Plain names bound to a subprocess callable, called without the module.
+
+    ``install_wheel(run = subprocess.run)`` calls its injected ``run`` as a bare
+    name, so matching only the attribute form leaves those installer calls
+    unguarded. Imports, assignments and parameter defaults all bind one.
+    """
+
+    def _is_bound(value: ast.expr | None) -> bool:
+        return (
+            isinstance(value, ast.Attribute)
+            and value.attr in _SUBPROCESS_CALLS
+            and isinstance(value.value, ast.Name)
+            and value.value.id in names
+        )
+
+    aliases: set[str] = set()
+    for node in ast.walk(tree):
+        if isinstance(node, ast.ImportFrom) and node.module == "subprocess":
+            aliases.update(a.asname or a.name for a in node.names if a.name in _SUBPROCESS_CALLS)
+        elif isinstance(node, ast.Assign) and _is_bound(node.value):
+            aliases.update(t.id for t in node.targets if isinstance(t, ast.Name))
+        elif isinstance(node, ast.AnnAssign) and _is_bound(node.value):
+            if isinstance(node.target, ast.Name):
+                aliases.add(node.target.id)
+        elif isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef)):
+            args = node.args
+            positional = args.posonlyargs + args.args
+            # Defaults cover the tail of the positional parameters; kw_defaults
+            # is aligned with kwonlyargs already, holding None where absent.
+            padded = [None] * (len(positional) - len(args.defaults)) + list(args.defaults)
+            pairs = list(zip(positional, padded)) + list(zip(args.kwonlyargs, args.kw_defaults))
+            aliases.update(arg.arg for arg, default in pairs if _is_bound(default))
+    return aliases
+
+
+def _is_subprocess_call(node: ast.Call, names: set[str], aliases: set[str]) -> bool:
+    func = node.func
+    if isinstance(func, ast.Name):
+        return func.id in aliases
+    if not isinstance(func, ast.Attribute) or func.attr not in _SUBPROCESS_CALLS:
+        return False
+    value = func.value
+    return isinstance(value, ast.Name) and value.id in names
+
+
+def _text_mode_subprocess(node: ast.Call) -> bool:
+    for keyword in node.keywords:
+        if keyword.arg not in ("text", "universal_newlines"):
+            continue
+        if isinstance(keyword.value, ast.Constant) and keyword.value.value is True:
+            return True
+    return False
+
+
+def _text_mode_dict(node: ast.Dict) -> bool:
+    """A ``{"text": True, ...}`` literal with no "encoding" key."""
+    keys = [k.value for k in node.keys if isinstance(k, ast.Constant)]
+    if "encoding" in keys:
+        return False
+    for key, value in zip(node.keys, node.values):
+        if not isinstance(key, ast.Constant) or key.value not in (
+            "text",
+            "universal_newlines",
+        ):
+            continue
+        if isinstance(value, ast.Constant) and value.value is True:
+            return True
+    return False
+
+
+def _splatted_names(tree: ast.AST) -> set[str]:
+    """Names handed to a call as ``**name``."""
+    names = set()
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Call):
+            for keyword in node.keywords:
+                if keyword.arg is None and isinstance(keyword.value, ast.Name):
+                    names.add(keyword.value.id)
+    return names
+
+
+def _encoding_assigned_later(tree: ast.AST, name: str) -> bool:
+    """``name["encoding"] = ...`` somewhere, so the literal need not carry it."""
+    for node in ast.walk(tree):
+        if not isinstance(node, ast.Subscript) or not isinstance(node.ctx, ast.Store):
+            continue
+        target, key = node.value, node.slice
+        if isinstance(target, ast.Name) and target.id == name:
+            if isinstance(key, ast.Constant) and key.value == "encoding":
+                return True
+    return False
+
+
+def _splatted_kwargs_offenders(tree: ast.AST) -> list[ast.Dict]:
+    """Text-mode kwargs built in a dict and splatted into a call.
+
+    Kwargs are collected in a dict and splatted (``run(cmd, **run_kwargs)``)
+    where a branch has to add a timeout or an env, and the call is often through
+    a helper, so neither the callee nor the keywords are visible at the call
+    site. Only dicts that reach a call this way are judged: an unrelated payload
+    that happens to carry ``"text": True`` is not subprocess configuration.
+    """
+    found = []
+    # ``run(cmd, **{...})``: the literal is at the call already.
+    for node in ast.walk(tree):
+        if not isinstance(node, ast.Call):
+            continue
+        for keyword in node.keywords:
+            if keyword.arg is None and isinstance(keyword.value, ast.Dict):
+                if _text_mode_dict(keyword.value):
+                    found.append(keyword.value)
+    splatted = _splatted_names(tree)
+    if not splatted:
+        return found
+    for node in ast.walk(tree):
+        targets = []
+        if isinstance(node, ast.Assign):
+            targets = [t for t in node.targets if isinstance(t, ast.Name)]
+        elif isinstance(node, ast.AnnAssign) and isinstance(node.target, ast.Name):
+            targets = [node.target]
+        if not targets or not isinstance(node.value, ast.Dict):
+            continue
+        if not _text_mode_dict(node.value):
+            continue
+        for target in targets:
+            if target.id in splatted and not _encoding_assigned_later(tree, target.id):
+                found.append(node.value)
+                break
+    return found
+
+
+def _offenders(path: Path) -> list[str]:
+    source = path.read_text(encoding = "utf-8")
+    tree = ast.parse(source, filename = str(path))
+    subprocess_names = _subprocess_names(tree)
+    subprocess_aliases = _subprocess_aliases(tree, subprocess_names)
+    found: list[str] = []
+    for node in _splatted_kwargs_offenders(tree):
+        found.append(
+            f"{path.name}:{node.lineno}: subprocess kwargs with text = True and no encoding"
+        )
+    for node in ast.walk(tree):
+        if not isinstance(node, ast.Call):
+            continue
+        name = _call_name(node)
+
+        if _is_subprocess_call(node, subprocess_names, subprocess_aliases):
+            if _text_mode_subprocess(node) and not _has_keyword(node, "encoding"):
+                found.append(f"{path.name}:{node.lineno}: subprocess(text = True) without encoding")
+            continue
+
+        if name == "open" and isinstance(node.func, ast.Name):
+            if _mode_is_binary(node) or _open_has_encoding(node):
+                continue
+            found.append(f"{path.name}:{node.lineno}: open() without encoding")
+            continue
+
+        # os.fdopen(fd, "w") is open() on a descriptor, so text mode takes the
+        # same locale default. Its mode defaults to "r", i.e. text, like open's.
+        if name == "fdopen":
+            if _mode_is_binary(node) or _open_has_encoding(node):
+                continue
+            found.append(f"{path.name}:{node.lineno}: os.fdopen() without encoding")
+            continue
+
+        if name == "open" and isinstance(node.func, ast.Attribute):
+            if not _is_path_open(node) or _path_open_has_encoding(node):
+                continue
+            if _path_open_mode(node) and "b" in _path_open_mode(node):
+                continue
+            found.append(f"{path.name}:{node.lineno}: Path.open() without encoding")
+            continue
+
+        if name in ("read_text", "write_text") and isinstance(node.func, ast.Attribute):
+            if _has_keyword(node, "encoding"):
+                continue
+            # importlib.metadata Distribution.read_text() takes no encoding kwarg.
+            if isinstance(node.func.value, ast.Name) and node.func.value.id == "dist":
+                continue
+            found.append(f"{path.name}:{node.lineno}: {name}() without encoding")
+    return found
+
+
+@pytest.mark.parametrize("path", _studio_sources(), ids = lambda p: str(p.name))
+def test_text_io_names_its_encoding(path: Path) -> None:
+    offenders = _offenders(path)
+    assert not offenders, (
+        "Text I/O without an explicit encoding falls back to the Windows ANSI "
+        'codepage and corrupts non-ASCII (ä ö ü → 世). Pass encoding = "utf-8":\n  '
+        + "\n  ".join(offenders)
+    )
+
+
+_STATE_STORE = (
+    BACKEND_ROOT
+    / "plugins/data-designer-github-repo-seed/src"
+    / "data_designer_github_repo_seed/scraper_impl/state_store.py"
+)
+
+
+def _load_state_store(codepage: str):
+    """Load state_store with the writing machine's codepage pinned."""
+    spec = importlib.util.spec_from_file_location(f"state_store_{codepage}", _STATE_STORE)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    module.locale = SimpleNamespace(
+        getencoding = lambda: codepage,
+        getpreferredencoding = lambda _ = True: codepage,
+    )
+    return module
+
+
+@pytest.mark.parametrize(
+    ("codepage", "name"), [("cp1252", "Jürgen"), ("cp1251", "Юрий"), ("cp932", "田中")]
+)
+def test_resuming_a_legacy_jsonl_keeps_one_encoding(
+    tmp_path: Path, codepage: str, name: str
+) -> None:
+    """A scrape written before UTF-8 was explicit must resume, not duplicate."""
+    path = tmp_path / "out.jsonl"
+    records = [{"id": 1, "author": name}, {"id": 2, "author": name}]
+    body = "".join(json.dumps(r, ensure_ascii = False) + "\n" for r in records)
+    path.write_bytes(body.encode(codepage))
+    before = path.read_bytes()
+
+    writer = _load_state_store(codepage).JsonlWriter(path)
+    try:
+        # Seen keys survive the resume, so a repeat is refused, not appended.
+        assert writer.has("id:1") and writer.has("id:2")
+        assert writer.write(records[0]) is False
+        assert writer.write({"id": 3, "author": name}) is True
+    finally:
+        writer.close()
+
+    # Never converted, so it still reads in its own codepage; the append is ASCII.
+    blob = path.read_bytes()
+    assert blob.startswith(before)
+    assert blob[len(before) :].isascii()
+    lines = [json.loads(x) for x in blob.decode(codepage).splitlines() if x.strip()]
+    assert len(lines) == 3
+    assert [line["author"] for line in lines] == [name] * 3
+
+
+def test_a_coincidentally_utf8_legacy_line_is_left_alone(tmp_path: Path) -> None:
+    """cp1251 `Р°` is D0 B0, which is also UTF-8 `а`, and nothing can tell them apart."""
+    path = tmp_path / "out.jsonl"
+    ambiguous = "Р°"
+    assert ambiguous.encode("cp1251").decode("utf-8") == "а"  # the trap
+    authors = ["Привет", "Здравствуйте", "Москва", ambiguous]
+    path.write_bytes(
+        b"".join(
+            json.dumps({"id": i, "author": a}, ensure_ascii = False).encode("cp1251") + b"\n"
+            for i, a in enumerate(authors)
+        )
+    )
+    before = path.read_bytes()
+
+    _load_state_store("cp1251").JsonlWriter(path).close()
+
+    # Untouched, so the ambiguity never had to be resolved.
+    assert path.read_bytes() == before
+    rows = [json.loads(x) for x in path.read_text(encoding = "cp1251").splitlines() if x.strip()]
+    assert [row["author"] for row in rows] == authors
+
+
+@pytest.mark.parametrize(
+    ("codepage", "word"), [("cp1251", "Привет"), ("cp932", "こんにちは"), ("cp1252", "Jürgen")]
+)
+def test_a_moved_shard_is_not_rewritten_by_guesswork(
+    tmp_path: Path, codepage: str, word: str
+) -> None:
+    """Off the writing machine there is no codepage to attribute the file to."""
+    path = tmp_path / "out.jsonl"
+    # Two records: a lone non-UTF-8 line would count as damage, not legacy.
+    path.write_bytes(
+        b"".join(
+            json.dumps({"id": i, "author": word}, ensure_ascii = False).encode(codepage) + b"\n"
+            for i in (1, 4)
+        )
+    )
+    before = path.read_bytes()
+
+    # A UTF-8 host: latin-1 would read cp1251 `Привет` back as `Ïðèâåò`.
+    writer = _load_state_store("utf-8").JsonlWriter(path)
+    try:
+        assert writer.has("id:1")  # ASCII keys still recover
+        assert writer.write({"id": 2, "author": "Grüße"}) is True
+    finally:
+        writer.close()
+
+    blob = path.read_bytes()
+    assert blob.startswith(before)  # never rewritten
+    assert blob[len(before) :].isascii()  # appended as \uXXXX, so no second encoding
+    rows = [json.loads(x) for x in blob.decode(codepage).splitlines() if x.strip()]
+    assert [row["author"] for row in rows] == [word, word, "Grüße"]
+
+
+def test_an_all_ambiguous_shard_still_gets_ascii_appends(tmp_path: Path) -> None:
+    """Every line valid under both readings still means the append must not pick one."""
+    path = tmp_path / "out.jsonl"
+    ambiguous = "Р°"  # cp1251 D0 B0, also valid UTF-8 for "а"
+    path.write_bytes(
+        b"".join(
+            json.dumps({"id": i, "a": ambiguous}, ensure_ascii = False).encode("cp1251") + b"\n"
+            for i in range(3)
+        )
+    )
+    before = path.read_bytes()
+
+    writer = _load_state_store("cp1251").JsonlWriter(path)
+    try:
+        assert writer.write({"id": 9, "a": "世界"}) is True
+    finally:
+        writer.close()
+
+    blob = path.read_bytes()
+    assert blob.startswith(before)
+    # ASCII, so the appended record survives whichever reading is chosen.
+    assert blob[len(before) :].isascii()
+    for codec in ("cp1251", "utf-8"):
+        rows = [json.loads(x) for x in blob.decode(codec).splitlines() if x.strip()]
+        assert rows[-1]["a"] == "世界"
+
+
+def test_a_damaged_line_in_an_ascii_shard_does_not_block_its_retry(tmp_path: Path) -> None:
+    """With no non-ASCII records to outvote it, one damaged line is still damage."""
+    path = tmp_path / "out.jsonl"
+    path.write_bytes(
+        b'{"id": 1, "author": "alice"}\n'
+        + b'{"id": 99, "author": "bad \x96 byte"}\n'
+        + b'{"id": 2, "author": "bob"}\n'
+    )
+
+    writer = _load_state_store("cp1252").JsonlWriter(path)
+    try:
+        assert writer.has("id:1") and writer.has("id:2")
+        assert not writer.has("id:99")
+        assert writer.write({"id": 99, "author": "good byte"}) is True
+    finally:
+        writer.close()
+
+
+def test_a_damaged_line_does_not_block_its_own_retry(tmp_path: Path) -> None:
+    """Its key comes from the codepage reading, which a UTF-8 shard did not pick."""
+    path = tmp_path / "out.jsonl"
+    path.write_bytes(
+        json.dumps({"id": 1, "author": "Jürgen"}, ensure_ascii = False).encode()
+        + b"\n"
+        + b'{"id": 99, "author": "bad \x96 byte"}\n'
+    )
+
+    writer = _load_state_store("cp1252").JsonlWriter(path)
+    try:
+        assert writer.has("id:1")
+        assert not writer.has("id:99")
+        assert writer.write({"id": 99, "author": "good byte"}) is True
+    finally:
+        writer.close()
+
+
+def test_one_damaged_byte_does_not_relabel_a_utf8_shard(tmp_path: Path) -> None:
+    """A complete JSON line with a stray 0x96 parses as cp1252, but is only one vote."""
+    path = tmp_path / "out.jsonl"
+    healthy = ["Jürgen", "Grüße", "Björn"]
+    path.write_bytes(
+        json.dumps({"id": 0, "author": healthy[0]}, ensure_ascii = False).encode()
+        + b"\n"
+        + b'{"id": 99, "author": "bad \x96 byte"}\n'
+        + b"".join(
+            json.dumps({"id": i, "author": a}, ensure_ascii = False).encode() + b"\n"
+            for i, a in enumerate(healthy[1:], start = 1)
+        )
+    )
+    before = path.read_bytes()
+
+    _load_state_store("cp1252").JsonlWriter(path).close()
+
+    # Untouched, so the healthy records were never re-read as cp1252.
+    assert path.read_bytes() == before
+    rows = []
+    for line in path.read_bytes().splitlines():
+        try:
+            rows.append(json.loads(line.decode()))
+        except (UnicodeDecodeError, ValueError):
+            continue
+    assert [row["author"] for row in rows] == healthy
+
+
+def test_a_torn_line_does_not_relabel_a_utf8_shard(tmp_path: Path) -> None:
+    """One interrupted append must not get the whole shard read as cp1252."""
+    path = tmp_path / "out.jsonl"
+    good = [{"id": 1, "author": "Jürgen"}, {"id": 3, "author": "Grüße"}]
+    torn = '{"id": 2, "author": "Jürgen"}'.encode()[:-6]  # cut mid-character
+    path.write_bytes(
+        json.dumps(good[0], ensure_ascii = False).encode()
+        + b"\n"
+        + torn
+        + b"\n"
+        + json.dumps(good[1], ensure_ascii = False).encode()
+        + b"\n"
+    )
+    before = path.read_bytes()
+
+    writer = _load_state_store("cp1252").JsonlWriter(path)
+    try:
+        assert writer.has("id:1") and writer.has("id:3")
+        assert not writer.has("id:2")  # torn line yields no key
+    finally:
+        writer.close()
+
+    # Untouched: no rewrite, so no record was re-encoded into mojibake.
+    after = path.read_bytes()
+    assert after.startswith(before)
+    assert "Jürgen".encode() in after
+    assert "Jürgen".encode("utf-8").decode("cp1252").encode() not in after
+
+
+def test_an_undecodable_transport_marker_reads_as_unknown(tmp_path: Path) -> None:
+    """Pinning the decode turns an undecodable marker into UnicodeDecodeError,
+    which is a ValueError and so is not an OSError. Before the pin those bytes
+    simply read as an unknown value and the caller safely purged and restarted
+    the partial download; letting the error escape aborts the transfer instead.
+    """
+    import sys
+
+    backend = str(Path(__file__).resolve().parent.parent)
+    if backend not in sys.path:
+        sys.path.insert(0, backend)
+    from hub.utils import download_registry as registry
+
+    marker = tmp_path / ".transport"
+    marker.write_bytes(b"\x80\xffnative\n")
+    assert registry._read_marker_value(marker) is None
+    # A readable but unknown value takes the same path (the behaviour restored).
+    marker.write_text("something-else\n", encoding = "utf-8")
+    assert registry._read_marker_value(marker) is None
+
+
+def test_a_torn_cache_ref_reads_as_not_cached(tmp_path: Path, monkeypatch) -> None:
+    """hf_cache_snapshot_dir answers "is this model already on disk", and the
+    offline embedding checks turn a raise into a 500. A refs/main holding a byte
+    the codepage used to decode into a nonsense commit simply missed the snapshot
+    dir before the pin; it has to keep missing it."""
+    import sys
+
+    backend = str(Path(__file__).resolve().parent.parent)
+    if backend not in sys.path:
+        sys.path.insert(0, backend)
+    from utils import utils as backend_utils
+
+    good_root = tmp_path / "good"
+    torn_root = tmp_path / "torn"
+    for root, ref_bytes in ((torn_root, b"\x80\xff\n"), (good_root, b"abc123\n")):
+        repo = root / "models--Org--Model"
+        (repo / "refs").mkdir(parents = True)
+        (repo / "refs" / "main").write_bytes(ref_bytes)
+    (good_root / "models--Org--Model" / "snapshots" / "abc123").mkdir(parents = True)
+
+    monkeypatch.setattr(backend_utils, "_hf_cache_roots", lambda: [torn_root])
+    assert backend_utils.hf_cache_snapshot_dir("Org/Model") is None
+    # The torn root is skipped, not fatal: a healthy second root still answers.
+    monkeypatch.setattr(backend_utils, "_hf_cache_roots", lambda: [torn_root, good_root])
+    found = backend_utils.hf_cache_snapshot_dir("Org/Model")
+    assert found is not None and found.name == "abc123"
+
+
+def test_a_corrupt_pid_file_does_not_abort_shutdown(tmp_path: Path, monkeypatch) -> None:
+    """_remove_pid_file runs first in _graceful_shutdown, so a raise there leaves
+    the inference, export, training and tunnel children alive."""
+    import sys
+
+    backend = str(Path(__file__).resolve().parent.parent)
+    if backend not in sys.path:
+        sys.path.insert(0, backend)
+    import run as studio_run
+
+    pid_file = tmp_path / "studio.pid"
+    pid_file.write_bytes(b"\x80\xff")
+    monkeypatch.setattr(studio_run, "_PID_FILE", pid_file)
+    studio_run._remove_pid_file()
+    # Not this process's PID, so the file stays; the point is that it returned.
+    assert pid_file.exists()
+
+    pid_file.write_text(str(os.getpid()), encoding = "utf-8")
+    studio_run._remove_pid_file()
+    assert not pid_file.exists()
+
+
+def test_the_kwargs_guard_only_judges_dicts_that_reach_a_call(tmp_path: Path) -> None:
+    """Only a dict splatted into a call is subprocess configuration. An unrelated
+    payload that happens to carry "text": True is not, and neither is one whose
+    encoding is filled in on a later line."""
+    cases = {
+        "offender.py": 'kw = {"text": True}\nrun(cmd, **kw)\n',
+        "annotated.py": 'kw: dict = {"universal_newlines": True}\nrun(cmd, **kw)\n',
+        "payload.py": 'payload = {"text": True}\nrequests.post(url, json = payload)\n',
+        "inline.py": 'run(cmd, **{"text": True})\n',
+        "later.py": 'kw = {"text": True}\nkw["encoding"] = "utf-8"\nrun(cmd, **kw)\n',
+        "carried.py": 'kw = {"text": True, "encoding": "utf-8"}\nrun(cmd, **kw)\n',
+    }
+    flagged = set()
+    for name, source in cases.items():
+        path = tmp_path / name
+        path.write_text(source, encoding = "utf-8")
+        if any("subprocess kwargs" in line for line in _offenders(path)):
+            flagged.add(name)
+    assert flagged == {"offender.py", "annotated.py", "inline.py"}, flagged
+
+
+def test_the_guard_follows_subprocess_through_an_alias(tmp_path: Path) -> None:
+    """install_wheel() takes ``run = subprocess.run`` and calls it as a bare
+    name, so an attribute-only match let both of its installer calls drop their
+    encoding unnoticed. A name bound to something else is still not subprocess."""
+    cases = {
+        "param_default.py": (
+            "import subprocess\n"
+            "def install(*, run = subprocess.run):\n"
+            "    run(cmd, text = True)\n"
+        ),
+        "assigned.py": "import subprocess\n_run = subprocess.run\n_run(cmd, text = True)\n",
+        "imported.py": "from subprocess import check_output\ncheck_output(cmd, text = True)\n",
+        "renamed.py": "from subprocess import run as _r\n_r(cmd, universal_newlines = True)\n",
+        "encoded.py": (
+            "import subprocess\n"
+            "def install(*, run = subprocess.run):\n"
+            '    run(cmd, text = True, encoding = "utf-8")\n'
+        ),
+        "unrelated.py": "def run(cmd, text = False):\n    pass\nrun(cmd, text = True)\n",
+    }
+    flagged = set()
+    for name, source in cases.items():
+        path = tmp_path / name
+        path.write_text(source, encoding = "utf-8")
+        if any("subprocess(text = True)" in line for line in _offenders(path)):
+            flagged.add(name)
+    assert flagged == {"param_default.py", "assigned.py", "imported.py", "renamed.py"}, flagged
+
+
+def test_the_guard_sees_os_fdopen(tmp_path: Path) -> None:
+    """os.fdopen(fd, mode) is open() on a descriptor and takes the same locale
+    default in text mode, so leaving it out let the swap lock file keep the
+    codepage on the write side while its reader was pinned to UTF-8."""
+    cases = {
+        "text.py": 'import os\nos.fdopen(fd, "w")\n',
+        "default_mode.py": "import os\nos.fdopen(fd)\n",  # defaults to "r", still text
+        "binary.py": 'import os\nos.fdopen(fd, "wb")\n',
+        "keyword.py": 'import os\nos.fdopen(fd, "w", encoding = "utf-8")\n',
+        "positional.py": 'import os\nos.fdopen(fd, "w", 1, "utf-8")\n',
+    }
+    flagged = set()
+    for name, source in cases.items():
+        path = tmp_path / name
+        path.write_text(source, encoding = "utf-8")
+        if any("fdopen" in line for line in _offenders(path)):
+            flagged.add(name)
+    assert flagged == {"text.py", "default_mode.py"}, flagged
+
+
+def test_an_undecodable_bootstrap_password_does_not_stop_startup(
+    tmp_path: Path, monkeypatch
+) -> None:
+    """ensure_default_admin calls _load_bootstrap_password for every existing
+    admin and the lifespan calls that with no handler, so a raise here takes the
+    whole backend down instead of ignoring an unusable file."""
+    import sys
+
+    backend = str(Path(__file__).resolve().parent.parent)
+    if backend not in sys.path:
+        sys.path.insert(0, backend)
+    from auth import storage
+
+    pw_file = tmp_path / ".bootstrap_password"
+    pw_file.write_bytes(b"\x80\xffnot-utf8\n")
+    monkeypatch.setattr(storage, "_BOOTSTRAP_PW_PATH", pw_file)
+    assert storage._load_bootstrap_password() is None
+
+    # A readable one still loads, so this is a narrowing of failure, not of function.
+    pw_file.write_text("correct horse battery staple\n", encoding = "utf-8")
+    assert storage._load_bootstrap_password() == "correct horse battery staple"
+
+
+def test_a_damaged_checkpoint_resets_instead_of_resuming_on_a_broken_cursor(tmp_path: Path) -> None:
+    """A checkpoint holds only base64 cursors and booleans, so a codepage reading
+    can only ever add non-ASCII, never recover any. Resuming on a mojibaked cursor
+    sends GitHub one it answers with INVALID_CURSOR_ARGUMENTS, and the empty page
+    that comes back marks the stream done and skips the rest of it for good.
+    Dropping the checkpoint only replays pages the writers already dedup."""
+    module = _load_state_store("cp1252")
+    cursor = "Y3Vyc29yOnYyOpK0MjAxMi0wMi0xNlQwNjo1Mzo0MVrOADGL_A=="
+    healthy = json.dumps({"issues_cursor": cursor, "issues_done": False}, indent = 2)
+    path = tmp_path / "octocat__Hello-World.json"
+
+    path.write_text(healthy, encoding = "utf-8")
+    assert module.StateStore(path).get("issues_cursor") == cursor
+
+    # Written by a pre-UTF-8 release in the operator's codepage. Nothing is lost
+    # by reading UTF-8 only, because an all-ASCII document is the same bytes.
+    path.write_bytes(healthy.encode("cp1252"))
+    assert module.StateStore(path).get("issues_cursor") == cursor
+
+    # One damaged byte inside the cursor: still a whole JSON document under a
+    # single-byte codepage, so only refusing that reading resets the checkpoint.
+    raw = healthy.encode()
+    at = raw.index(b"MjAxMi0wMi0xNlQ") + 3
+    path.write_bytes(raw[:at] + b"\x96" + raw[at + 1 :])
+    assert json.loads(path.read_bytes().decode("latin-1"))["issues_cursor"] != cursor
+    store = module.StateStore(path)
+    assert store.all() == {}
+    assert store.get("issues_cursor") is None
+
+
+def test_a_utf8_record_is_not_parsed_a_second_time(tmp_path: Path) -> None:
+    """These shards reach gigabytes and every resume reads all of one, so a
+    record that already read as UTF-8 must not be decoded and parsed again under
+    the codepage. The legacy reading exists only to recover keys UTF-8 could not."""
+    module = _load_state_store("cp1252")
+    calls: list[str] = []
+    real_parse = module._parse
+
+    def counting_parse(raw, encoding):
+        calls.append(encoding)
+        return real_parse(raw, encoding)
+
+    module._parse = counting_parse
+    try:
+        healthy = json.dumps({"id": 1, "author": "Jürgen"}).encode("utf-8")
+        reading = module._read_line(healthy, "cp1252")
+        assert reading.as_utf8 == {"id": 1, "author": "Jürgen"}
+        assert calls == ["utf-8"], calls
+
+        # A line UTF-8 cannot read still falls through to the codepage, the whole point.
+        calls.clear()
+        legacy = json.dumps({"id": 2, "author": "Jürgen"}, ensure_ascii = False).encode("cp1252")
+        reading = module._read_line(legacy, "cp1252")
+        assert reading.as_utf8 is None
+        assert reading.as_legacy == {"id": 2, "author": "Jürgen"}
+        assert calls == ["utf-8", "cp1252"], calls
+    finally:
+        module._parse = real_parse
+
+
+def _too_deeply_nested_json() -> str:
+    """A JSON document nested past what this interpreter will descend into.
+
+    Probed rather than hardcoded: the depth json.loads gives up at is bounded by
+    sys.getrecursionlimit() up to 3.11 and by the C recursion limit from 3.12,
+    which sys.setrecursionlimit no longer moves and which varies by micro
+    version. That is ~995 on 3.9 and ~9999 on 3.13.
+    """
+    depth = 1
+    while depth <= 1 << 17:
+        document = "[" * depth + "]" * depth
+        try:
+            json.loads(document)
+        except RecursionError:
+            return document
+        depth *= 2
+    pytest.skip("this interpreter parses arbitrarily nested JSON")
+
+
+def test_an_unparseably_nested_document_is_discarded_not_raised(tmp_path: Path) -> None:
+    """json.loads answers nesting it cannot descend with RecursionError, which is
+    a RuntimeError and so is neither a ValueError nor a UnicodeDecodeError.
+    _parse is called outside any other handler in both StateStore.__init__ and
+    JsonlWriter._scan_existing, so letting it escape aborts the scraper at
+    startup on a file the catch-all it replaced simply discarded."""
+    module = _load_state_store("cp1252")
+    nested = _too_deeply_nested_json()
+
+    checkpoint = tmp_path / "octocat__Hello-World.json"
+    checkpoint.write_text(nested, encoding = "utf-8")
+    assert module.StateStore(checkpoint).all() == {}  # reset, not raised
+
+    shard = tmp_path / "out.jsonl"
+    shard.write_text(
+        nested + "\n" + json.dumps({"id": 1}) + "\n" + json.dumps({"id": 2}) + "\n",
+        encoding = "utf-8",
+    )
+    writer = module.JsonlWriter(shard)
+    try:
+        # Skipped like any other unreadable line, so its neighbours still yield the dedup
+        # keys that keep the resume from re-fetching them.
+        assert writer.has("id:1") and writer.has("id:2")
+    finally:
+        writer.close()
diff --git a/studio/backend/tests/test_training_worker_flash_attn.py b/studio/backend/tests/test_training_worker_flash_attn.py
index 86511987b1..d136821ea2 100644
--- a/studio/backend/tests/test_training_worker_flash_attn.py
+++ b/studio/backend/tests/test_training_worker_flash_attn.py
@@ -9,8 +9,28 @@ import sys
 from typing import Any
 from unittest import mock
 
+import pytest
+
 from core.training import worker
 
+# The runtime install is Linux-only, so elsewhere these return before any status.
+linux_only = pytest.mark.skipif(
+    not sys.platform.startswith("linux"),
+    reason = "the runtime flash-attn install is gated to Linux",
+)
+
+# causal-conv1d and flash-linear-attention are NOT Linux-gated: both installers bail out
+# on `sys.platform == "win32"` alone (no prebuilt wheel for Windows) and run everywhere
+# else, macOS included. linux_only here would skip cases that legitimately pass off Linux.
+not_on_windows = pytest.mark.skipif(
+    sys.platform == "win32",
+    reason = (
+        "mirrors the sys.platform == 'win32' bail-out in "
+        "_ensure_flash_linear_attention_unconditional and "
+        "_ensure_causal_conv1d_fast_path"
+    ),
+)
+
 
 def _missing_flash_attn_import():
     real_import = builtins.__import__
@@ -55,6 +75,7 @@ def test_should_try_runtime_flash_attn_install_threshold_and_skip(monkeypatch):
     assert worker._should_try_runtime_flash_attn_install(32768) is False
 
 
+@linux_only
 def test_runtime_flash_attn_prefers_prebuilt_wheel(monkeypatch):
     statuses: list[str] = []
 
@@ -82,6 +103,7 @@ def test_runtime_flash_attn_prefers_prebuilt_wheel(monkeypatch):
     assert statuses == ["Installing flash-attn for faster training..."]
 
 
+@linux_only
 def test_runtime_flash_attn_falls_back_to_pypi(monkeypatch):
     calls: list[list[str]] = []
     statuses: list[str] = []
@@ -113,12 +135,7 @@ def test_runtime_flash_attn_falls_back_to_pypi(monkeypatch):
     )
     monkeypatch.setattr(worker, "install_wheel", mock.Mock())
 
-    def fake_run(
-        cmd,
-        stdout = None,
-        stderr = None,
-        text = None,
-    ):
+    def fake_run(cmd, **kwargs):
         calls.append(list(cmd))
         return subprocess.CompletedProcess(cmd, 0, "")
 
@@ -139,6 +156,7 @@ def test_runtime_flash_attn_skip_env_avoids_all_install_work(monkeypatch):
     worker._sp.run.assert_not_called()
 
 
+@not_on_windows
 def test_causal_conv1d_fast_path_preserves_wheel_first_install_args(monkeypatch):
     install_mock = mock.Mock(return_value = True)
     monkeypatch.setattr(worker, "_install_package_wheel_first", install_mock)
@@ -160,6 +178,7 @@ def test_causal_conv1d_fast_path_preserves_wheel_first_install_args(monkeypatch)
     )
 
 
+@not_on_windows
 def test_causal_conv1d_fast_path_includes_qwen3_6_variants(monkeypatch):
     install_mock = mock.Mock(return_value = True)
     monkeypatch.setattr(worker, "_install_package_wheel_first", install_mock)
@@ -225,6 +244,7 @@ def _pin_fla_model_types(monkeypatch):
     )
 
 
+@not_on_windows
 def test_flash_linear_attention_installs_pinned_pair_for_qwen3_5(monkeypatch):
     _pin_fla_model_types(monkeypatch)
     monkeypatch.setattr(worker.shutil, "which", lambda name: "/usr/bin/uv")
@@ -277,6 +297,7 @@ def test_flash_linear_attention_skips_for_ssm_only_models(monkeypatch):
     run_mock.assert_not_called()
 
 
+@not_on_windows
 def test_flash_linear_attention_matches_full_qwen3_family(monkeypatch):
     monkeypatch.setattr(worker.shutil, "which", lambda name: "/usr/bin/uv")
     run_mock = mock.Mock(return_value = mock.Mock(returncode = 0, stdout = ""))
@@ -331,6 +352,7 @@ def test_flash_linear_attention_skipped_via_env(monkeypatch):
     run_mock.assert_not_called()
 
 
+@not_on_windows
 def test_flash_linear_attention_skipped_below_torch_2_7(monkeypatch):
     _pin_fla_model_types(monkeypatch)
     monkeypatch.delenv(worker._FLA_SKIP_ENV, raising = False)
@@ -349,6 +371,7 @@ def test_flash_linear_attention_skipped_below_torch_2_7(monkeypatch):
     assert any("torch>=" in s for s in statuses)
 
 
+@not_on_windows
 def test_flash_linear_attention_install_includes_einops(monkeypatch):
     _pin_fla_model_types(monkeypatch)
     monkeypatch.delenv(worker._FLA_SKIP_ENV, raising = False)
@@ -375,6 +398,7 @@ def test_flash_linear_attention_install_includes_einops(monkeypatch):
     assert f"fla-core=={worker._FLA_CORE_PACKAGE_VERSION}" in args
 
 
+@not_on_windows
 def test_flash_linear_attention_logs_post_install_import_failure(monkeypatch):
     """pip exits 0 but `import fla.modules` still fails (missing transitive)."""
     _pin_fla_model_types(monkeypatch)
@@ -421,6 +445,7 @@ def test_tilelang_backend_skipped_on_unsupported_linux_arch(monkeypatch):
     run_mock.assert_not_called()
 
 
+@linux_only
 def test_tilelang_backend_pins_only_binary(monkeypatch):
     _pin_fla_model_types(monkeypatch)
     monkeypatch.delenv(worker._TILELANG_SKIP_ENV, raising = False)
@@ -462,6 +487,7 @@ def _force_missing_tilelang_imports(monkeypatch):
     monkeypatch.setattr(builtins, "__import__", fake_import)
 
 
+@linux_only
 def test_tilelang_backend_installs_pinned_pair_for_qwen3_5(monkeypatch):
     _pin_fla_model_types(monkeypatch)
     monkeypatch.delenv(worker._TILELANG_SKIP_ENV, raising = False)
@@ -486,6 +512,7 @@ def test_tilelang_backend_installs_pinned_pair_for_qwen3_5(monkeypatch):
     assert any("Installing TileLang" in s for s in statuses)
 
 
+@linux_only
 def test_tilelang_backend_reinstalls_when_tvm_ffi_is_broken(monkeypatch):
     """Repair path issues TWO pip calls:
 
@@ -555,6 +582,7 @@ def test_tilelang_backend_skipped_on_windows(monkeypatch):
     run_mock.assert_not_called()
 
 
+@linux_only
 def test_tilelang_backend_swallows_install_timeout(monkeypatch):
     _pin_fla_model_types(monkeypatch)
     monkeypatch.delenv(worker._TILELANG_SKIP_ENV, raising = False)
@@ -609,6 +637,7 @@ def test_tilelang_backend_skipped_via_env(monkeypatch):
     run_mock.assert_not_called()
 
 
+@linux_only
 def test_tilelang_backend_swallows_install_failure(monkeypatch):
     _pin_fla_model_types(monkeypatch)
     monkeypatch.delenv(worker._TILELANG_SKIP_ENV, raising = False)
@@ -673,6 +702,7 @@ def _patch_iu_gates(monkeypatch, fla_gate, conv_gate):
     monkeypatch.setattr(_iu, "is_causal_conv1d_available", conv_gate)
 
 
+@not_on_windows
 def test_hook_installs_when_gate_returns_false(monkeypatch):
     _pin_fla_model_types(monkeypatch)
     fla_gate = _make_fake_gate(initial_return = False)
@@ -976,6 +1006,7 @@ def test_hook_does_install_tilelang_for_qwen35(monkeypatch):
     tile_install.assert_called_once()
 
 
+@linux_only
 def test_tilelang_repair_does_not_touch_torch_cuda_stack(monkeypatch):
     """Finding #2: the broken-tvm-ffi repair must use --no-deps on the
     forced step so --force-reinstall doesn't cascade through
@@ -1119,6 +1150,7 @@ def test_hook_runs_tilelang_repair_when_fla_already_true(monkeypatch):
     tile_install.assert_called_once()
 
 
+@not_on_windows
 def test_fla_installer_force_reinstalls_when_older_version_present(monkeypatch):
     """Finding #8: an older `flash-linear-attention` that is importable
     but below the pin must force a reinstall (not no-op).
@@ -1583,15 +1615,10 @@ def test_install_respects_user_gcc_install_dir(monkeypatch):
     )
     _make_hip_install_env(monkeypatch, gcc_dir = "/usr/lib/gcc/x86_64-linux-gnu/13")
 
-    captured: dict[str, str] | None = {"_called": "no"}
+    captured: dict[str, str] = {}
 
     def fake_run(cmd, **kwargs):
-        env = kwargs.get("env")
-        if env is not None:
-            captured.clear()
-            captured.update(env)
-        else:
-            captured["_called"] = "yes_no_env"
+        captured.update(kwargs.get("env") or {})
         return subprocess.CompletedProcess(cmd, 0, "")
 
     monkeypatch.setattr(worker._sp, "run", fake_run)
@@ -1607,14 +1634,11 @@ def test_install_respects_user_gcc_install_dir(monkeypatch):
         release_base_url = "https://example.com",
     )
 
-    # subprocess.run invoked without env override (user already set
-    # HIPCC_COMPILE_FLAGS_APPEND with --gcc-install-dir, so we left the
-    # env alone — the existing value is inherited).
-    assert captured == {"_called": "yes_no_env"}
+    assert captured["HIPCC_COMPILE_FLAGS_APPEND"] == "--gcc-install-dir=/opt/custom/gcc-13"
 
 
 def test_install_does_not_inject_env_on_cuda(monkeypatch):
-    """CUDA path (no hip_version in env) → no env override at all."""
+    """CUDA path (no hip_version in env) → no HIP flag injected."""
     monkeypatch.delenv("HIPCC_COMPILE_FLAGS_APPEND", raising = False)
     monkeypatch.setattr(builtins, "__import__", _missing_module_import("causal_conv1d"))
     monkeypatch.setattr(
@@ -1641,7 +1665,7 @@ def test_install_does_not_inject_env_on_cuda(monkeypatch):
     captured: dict[str, Any] = {}
 
     def fake_run(cmd, **kwargs):
-        captured["env_in_kwargs"] = "env" in kwargs
+        captured.update(kwargs.get("env") or {})
         return subprocess.CompletedProcess(cmd, 0, "")
 
     monkeypatch.setattr(worker._sp, "run", fake_run)
@@ -1657,5 +1681,5 @@ def test_install_does_not_inject_env_on_cuda(monkeypatch):
         release_base_url = "https://example.com",
     )
 
-    # CUDA branch never sets the env, never invokes the gcc helper.
-    assert captured.get("env_in_kwargs") is False
+    # env is always passed (to force UTF-8), but never the HIP flag.
+    assert "HIPCC_COMPILE_FLAGS_APPEND" not in captured
diff --git a/studio/backend/utils/changelog.py b/studio/backend/utils/changelog.py
new file mode 100644
index 0000000000..84cd54df05
--- /dev/null
+++ b/studio/backend/utils/changelog.py
@@ -0,0 +1,1056 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Release notes for the update popup, sourced from CHANGELOG.md.
+
+Notes are keyed to one exact version: the popup asks for the version it is
+offering and gets that section or nothing, so an older release's notes can
+never appear next to a newer update.
+
+The remote copy on the default branch wins over the bundled one, since the
+offered version is newer than the installed checkout. Both reads are lazy,
+cached and skipped when update checks are off.
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import threading
+import time
+import urllib.request
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+from packaging.version import InvalidVersion, Version
+
+from .update_status import DISABLE_ENV_VAR, RELEASE_NOTES_URL
+
+CHANGELOG_FILENAME = "CHANGELOG.md"
+CHANGELOG_RAW_URL = "https://raw.githubusercontent.com/unslothai/unsloth/main/CHANGELOG.md"
+CHANGELOG_URL_ENV_VAR = "UNSLOTH_CHANGELOG_URL"
+CHANGELOG_PATH_ENV_VAR = "UNSLOTH_CHANGELOG_PATH"
+CHANGELOG_TIMEOUT_SECONDS = 3
+CHANGELOG_MAX_BYTES = 2 * 1024 * 1024
+_CHANGELOG_CHUNK_BYTES = 64 * 1024
+_CHANGELOG_MIN_READ_SECONDS = 0.05
+CHANGELOG_SUCCESS_TTL_SECONDS = 30 * 60
+CHANGELOG_FAILURE_TTL_SECONDS = 5 * 60
+RELEASE_NOTES_MAX_CHARS = 20_000
+
+# CommonMark requires a space, tab or line end after the hashes: a non-breaking
+# space copied from rich text renders as text, not a heading, but a bare `##` is
+# an empty heading and still ends the release above.
+_HEADING_PATTERN = re.compile(r"^ {0,3}##(?:[ \t]+(?P<title>.*?))?[ \t]*$")
+_FENCE_PATTERN = re.compile(r"^ {0,3}(?P<marker>`{3,}|~{3,})(?P<rest>.*)$")
+# CommonMark type 1 HTML blocks: contents are literal until a closing tag,
+# which the spec says need not be the one that opened the block.
+_RAW_HTML_OPEN = re.compile(r"^ {0,3}<(pre|script|style|textarea)(?=[\s>]|$)", re.IGNORECASE)
+_RAW_HTML_CLOSE = re.compile(r"</(pre|script|style|textarea)\s*>", re.IGNORECASE)
+# Types 3 to 5 (processing instructions, declarations, CDATA) are literal too,
+# each ending on its own delimiter. Comments open mid-line, so are separate.
+_RAW_BLOCKS = (
+    (_RAW_HTML_OPEN, _RAW_HTML_CLOSE),
+    (re.compile(r"^ {0,3}<\?"), re.compile(r"\?>")),
+    (re.compile(r"^ {0,3}<!\[CDATA\["), re.compile(r"\]\]>")),
+    # A declaration needs an uppercase letter, so `<!note` stays ordinary text.
+    (re.compile(r"^ {0,3}<![A-Z]"), re.compile(r">")),
+)
+# Type 6 blocks run to the next blank line, so `<details>` only holds Markdown
+# once a blank line has closed the block. Open and close tags both start one.
+_HTML_BLOCK_OPEN = re.compile(r"^ {0,3}</?([a-zA-Z][a-zA-Z0-9-]*)(?=[\s/>]|$)")
+# Blocks that break into an open paragraph, so none is open after them and one
+# they are written below is closed rather than continued.
+_INTERRUPTS = re.compile(
+    r"^ {0,3}(?:#{1,6}([ \t]|$)|(?:\*[ \t]*){3,}$|(?:-[ \t]*){3,}$|(?:_[ \t]*){3,}$)"
+)
+# A definition is a block of its own but may not interrupt a paragraph, so it
+# ends the one above it only when there is none to continue.
+_LINK_DEFINITION = re.compile(r"^ {0,3}\[(?:[^\[\]\\]|\\.)+\]:")
+# Blocks that are not paragraph text, so a following underline is not setext.
+_PARAGRAPH_TEXT = re.compile(r"^ {0,3}(?![-*+>]([ \t]|$)|\d{1,9}[.)]([ \t]|$))\S")
+# A line of = or - under a paragraph line makes that line a heading.
+_SETEXT_UNDERLINE = re.compile(r"^ {0,3}(=+|-+)[ \t]*$")
+# A quoted paragraph continues on unmarked lines, which belong to the quote.
+_BLOCK_QUOTE = re.compile(r"^ {0,3}>")
+_QUOTE_MARKER = re.compile(r"^ {0,3}>[ \t]?")
+# A heading at an item's content column belongs to that item, not the document.
+# The marker needs whitespace after it, so `2.0` is a version, not an item.
+_LIST_ITEM = re.compile(r"^[ \t]*(?P<marker>[-*+]|\d{1,9}[.)])(?P<space>[ \t]+|$)")
+_THEMATIC_BREAK = re.compile(r"^ {0,3}(?:(?:\*[ \t]*){3,}|(?:-[ \t]*){3,}|(?:_[ \t]*){3,})$")
+# Content indented more than this after a marker is an indented code block, so
+# the item's content starts one column past the marker instead.
+_MAX_ITEM_PADDING = 4
+_HTML_BLOCK_TAGS = frozenset(
+    """
+address article aside base basefont blockquote body caption center col colgroup
+dd details dialog dir div dl dt fieldset figcaption figure footer form frame
+frameset h1 h2 h3 h4 h5 h6 head header hr html iframe legend li link main menu
+menuitem nav noframes ol optgroup option p param search section summary table
+tbody td tfoot th thead title tr track ul
+""".split()
+)
+# Type 7: any other complete tag alone on a line. It cannot interrupt a
+# paragraph, so it only counts after a break.
+_HTML_ATTRIBUTE = (
+    r"""(?:\s+[a-zA-Z_:][a-zA-Z0-9_.:-]*(?:\s*=\s*(?:[^\s"'=<>`]+|'[^']*'|"[^"]*"))?)"""
+)
+_HTML_TAG_ONLY_LINE = re.compile(
+    rf"^ {{0,3}}(?:<[a-zA-Z][a-zA-Z0-9-]*{_HTML_ATTRIBUTE}*\s*/?>|</[a-zA-Z][a-zA-Z0-9-]*\s*>)\s*$"
+)
+# Levels above studio/ are the repo root in a checkout and site-packages in an
+# install, so they are searched only when one of these markers is present.
+_CHECKOUT_ONLY_LEVELS = (3, 4)
+_CHECKOUT_MARKERS = ("pyproject.toml", ".git")
+_COMMENT_BLOCK_OPEN = re.compile(r"^ {0,3}<!--")
+_COMMENT_OPEN = "<!--"
+_COMMENT_CLOSE = "-->"
+# Stands in for a line the renderer hides. `#` is a block of its own, so list
+# tracking reads it like a comment: never a marker, never a lazy continuation.
+_HIDDEN_BLOCK = "#"
+_VERSION_TOKEN_PATTERN = re.compile(r"^[\[(]?v?(?P<version>[0-9][0-9A-Za-z.!+-]*?)[\])]?$")
+_SAFE_VERSION_PATTERN = re.compile(r"^[0-9A-Za-z][0-9A-Za-z.!+-]{0,63}$")
+
+
+@dataclass(frozen = True)
+class _ListState:
+    """The open list items, innermost last, by the column their content starts."""
+
+    columns: tuple[int, ...] = ()
+    # True while the innermost item has had no content since its marker.
+    empty_item: bool = False
+
+
+@dataclass(frozen = True)
+class ChangelogEntry:
+    """One `## <version>` section of the changelog."""
+
+    version: str
+    heading: str
+    body: str
+
+
+@dataclass(frozen = True)
+class ChangelogSource:
+    text: str | None
+    source: str | None
+    error: str | None = None
+
+
+@dataclass
+class _ChangelogCacheEntry:
+    source: ChangelogSource
+    expires_at: float
+
+
+_cache_condition = threading.Condition()
+_remote_cache: _ChangelogCacheEntry | None = None
+_remote_fetching = False
+
+
+def reset_changelog_cache() -> None:
+    """Clear the in-process changelog cache. Intended for tests."""
+    global _remote_cache, _remote_fetching
+    with _cache_condition:
+        _remote_cache = None
+        _remote_fetching = False
+        _cache_condition.notify_all()
+
+
+def is_supported_version_query(version: str) -> bool:
+    """Whether `version` is shaped like something we can look up at all.
+
+    Sections are indexed only when their version parses, so a query that does
+    not parse (`latest`, `main`) can never match and is rejected outright."""
+    candidate = version.strip()
+    if not _SAFE_VERSION_PATTERN.match(candidate):
+        return False
+    return _parse_version(candidate) is not None
+
+
+def _markdown_lines(text: str) -> list[str]:
+    """``text`` split the way CommonMark ends lines.
+
+    str.splitlines also breaks on U+2028, U+2029, NEL, vertical tab and form
+    feed, none of which end a line in Markdown. A separator sitting in prose
+    before "## 9.9.9" would otherwise index a release the renderer never shows
+    and truncate the notes above it.
+    """
+    return text.replace("\r\n", "\n").replace("\r", "\n").split("\n")
+
+
+def parse_changelog(text: str) -> list[ChangelogEntry]:
+    """Parse `## <version>` sections, in file order.
+
+    Headings whose first token is not a version (`## Unreleased`, `## Format`)
+    end the previous section but are not indexed.
+    """
+    # A Windows editor can leave a BOM on the first line, hiding a heading.
+    text = text.lstrip("")
+    entries: list[ChangelogEntry] = []
+    heading: str | None = None
+    version: str | None = None
+    body: list[str] = []
+    open_fence: str | None = None
+    # Content column of the list item the open block belongs to, 0 at document
+    # level. A fence and an HTML block are scoped to their container, so the
+    # item's end closes them. Only one of the three is ever open.
+    block_column = 0
+    in_comment = False
+    in_raw_html: int | None = None
+    in_html_block = False
+    after_paragraph = False
+    paragraph: list[str] = []
+    in_quote = False
+    quoted = False
+    lists = _ListState()
+
+    def flush() -> None:
+        if version is not None and heading is not None:
+            entries.append(
+                ChangelogEntry(
+                    version = version,
+                    heading = heading,
+                    body = "\n".join(body).strip(),
+                )
+            )
+
+    for line in _markdown_lines(text):
+        # The line as list tracking sees it: blank wherever nothing renders.
+        structural = ""
+        opened_block = False
+        in_block = open_fence is not None or in_html_block or in_raw_html is not None or in_comment
+        # A fence, comment or HTML block inside a list item runs only to the end
+        # of that item, so a line dedented out of the item closes both. Lazy
+        # continuation reaches into none of them. A raw block or comment inside an
+        # item also ends on a blank line: the item takes the break, so what
+        # follows is a block of the item's own.
+        leaves = (
+            _indent_width(line) < block_column
+            if line.strip()
+            else in_raw_html is not None or in_comment
+        )
+        if in_block and block_column and leaves:
+            open_fence = None
+            in_html_block = False
+            in_raw_html = None
+            in_comment = False
+            block_column = 0
+            # The paragraph the line could have continued is block content, so
+            # it closes the item rather than reading as more of it.
+            after_paragraph = False
+        # A fence written as a list item's first content opens inside that item, so
+        # an opener is read past a marker on the same line. Only an opener: fenced
+        # content is literal and a closer carries no marker.
+        fence_line = line if open_fence else _item_content(line, after_paragraph)
+        # Raw HTML first: its contents are literal, so a fence in it is not one.
+        if in_raw_html is not None:
+            visible, in_raw_html = _strip_raw_html(line, in_raw_html)
+        elif in_html_block:
+            # A blank line is the only thing that ends a type 6 block.
+            in_html_block = line.strip() != ""
+            visible = ""
+        elif (fence := _FENCE_PATTERN.match(fence_line)) and not in_comment:
+            was_open = open_fence
+            open_fence = _next_fence_state(open_fence, fence.group("marker"), fence.group("rest"))
+            opened_block = was_open is None and open_fence is not None
+            # Hidden from heading matching, but its indent still closes items.
+            visible = ""
+            structural = line
+        elif open_fence:
+            visible = ""
+        else:
+            # A block already open owns this line, so it is content rather than a
+            # block written at the column it happens to start in.
+            hidden = in_comment or in_raw_html is not None
+            # A comment is an HTML block too, so one written as a list item's first
+            # content opens inside it exactly as a fence does: the opener is read
+            # past a marker on the same line.
+            block_open = (
+                not in_comment
+                and _COMMENT_BLOCK_OPEN.match(_item_content(line, after_paragraph)) is not None
+            )
+            # Commented-out sections are not rendered, so they are not releases.
+            visible, in_comment = _strip_comments(line, in_comment, block_open)
+            # An HTML block written as a list item's first content opens inside
+            # that item, as a fence does, so an opener is read past a marker on the
+            # same line. The marker stays, so its item is still tracked. A comment
+            # blanks its own line, so that line is read as written: the block
+            # renders as nothing, but the item it is content of still opens.
+            source = line if block_open else visible
+            content = _item_content(source, after_paragraph)
+            marker = source[: len(source) - len(content)]
+            # Nor is anything inside a raw HTML block such as <pre>.
+            stripped, in_raw_html = _strip_raw_html(content, in_raw_html)
+            opened_block = in_raw_html is not None or (block_open and in_comment)
+            # Taken before the opener is hidden: it renders as nothing, but its
+            # indent still closes a list item it sits left of, and a marker on its
+            # line still opens one. A comment or raw block keeps only those, since
+            # the text it hides is not Markdown and must open no list.
+            if block_open or stripped != content:
+                if not hidden:
+                    structural = _hidden_structure(line, marker)
+                visible = ""
+            else:
+                visible = marker + stripped
+                if visible.strip():
+                    structural = visible
+                elif not hidden:
+                    structural = _hidden_structure(line)
+                if stripped and _opens_html_block(stripped, after_paragraph):
+                    in_html_block = True
+                    opened_block = True
+                    visible = ""
+        # A `##` inside a fenced block is sample markdown, not a real heading.
+        match = _HEADING_PATTERN.match(visible) if visible else None
+        # `1.0` over a line of dashes is the same heading written setext style.
+        setext = (
+            after_paragraph
+            and match is None
+            and paragraph != []
+            and _SETEXT_UNDERLINE.match(visible) is not None
+            and (visible.strip()[:1] == "-")
+            # Never a boundary inside a list item: dedented the dashes are a
+            # thematic break, and at the content column the heading is nested.
+            and not lists.columns
+        )
+        if setext:
+            if version is not None:
+                # The whole paragraph is the heading, read as body on arrival.
+                del body[len(body) - len(paragraph) :]
+            flush()
+            # A wrapped heading keeps every line, so token one is the version.
+            heading = "\n".join(paragraph)
+            version = _version_from_heading(heading)
+            body = []
+            paragraph = []
+            after_paragraph = False
+            continue
+        # A dashed underline is not a list marker, so track lists after setext.
+        lazy_marker = _lazy_marker(structural, lists, after_paragraph, quoted)
+        lists = _open_lists(structural, lists, after_paragraph, quoted)
+        # Taken after the opening line closed the items it is dedented out of,
+        # so the block belongs to the item it is really written inside.
+        if opened_block:
+            block_column = lists.columns[-1] if lists.columns else 0
+        elif open_fence is None and not in_html_block and in_raw_html is None and not in_comment:
+            block_column = 0
+        # At an open item's content column a heading is nested, not a boundary.
+        if lists.columns and _indent_width(visible) >= lists.columns[0]:
+            match = None
+        # The line at its own nesting level: past the container's indentation
+        # and past a marker on the same line, so `- ## 2.0` reads as a heading.
+        column = lists.columns[-1] if lists.columns else 0
+        content = _strip_indent(visible, column)
+        if (item := _LIST_ITEM.match(content)) is not None:
+            content = content[item.end() :]
+        # Only ordinary text continues a paragraph. Indented code counts four
+        # spaces past the container, so an item's own indent does not count.
+        indented_code = not after_paragraph and _indent_width(visible) - column >= 4
+        # An underline ends the paragraph it underlines, so it needs one open in
+        # its own container: the quote above owns its own, and a row left of an
+        # open item is lazy text of the item's paragraph. Three dashes are a
+        # thematic break either way, which `_INTERRUPTS` already ends on.
+        underline = (
+            _SETEXT_UNDERLINE.match(visible) is not None
+            and after_paragraph
+            and not quoted
+            and _indent_width(visible) >= column
+        )
+        after_paragraph = (
+            # Read inside its container, so an empty item and a fence written as an
+            # item's own content leave no paragraph open below them. A marker the
+            # paragraph above swallows is its text, not an item.
+            (bool(content.strip()) or lazy_marker)
+            and match is None
+            and _HEADING_PATTERN.match(content) is None
+            and _FENCE_PATTERN.match(content) is None
+            and not indented_code
+            and _INTERRUPTS.match(visible) is None
+            and (after_paragraph or _LINK_DEFINITION.match(visible) is None)
+            and not underline
+        )
+        # A quote's paragraph runs on over plain text and owns every line of it.
+        # An empty quote holds none, so the line below starts the document's.
+        flush_left = visible.lstrip(" \t")
+        quote_line = _BLOCK_QUOTE.match(visible) is not None
+        in_quote = (
+            _may_be_lazy(_quote_content(visible))
+            if quote_line
+            else in_quote and _continues_paragraph(visible, column)
+        )
+        if quote_line:
+            # The only paragraph a quote line leaves open is the quote's own,
+            # and a quote holding a heading or nothing at all leaves none.
+            after_paragraph = in_quote
+        # Whose paragraph the line below would continue. A quote owns the one its
+        # own lines hold, so a marker outside the quote is a block of its own
+        # rather than more of the text above it.
+        quoted = quote_line or in_quote
+        # The lines a later underline turns into one heading. A paragraph opens
+        # only on plain text and then runs on until something interrupts it.
+        continues = (
+            not _interrupts_paragraph(flush_left)
+            if paragraph
+            else _PARAGRAPH_TEXT.match(flush_left) is not None
+        )
+        # A paragraph inside an open item is that item's, and only one written
+        # at document level can be the heading a later underline makes of it.
+        if after_paragraph and not in_quote and not lists.columns and continues:
+            paragraph = [*paragraph, visible.strip()]
+        else:
+            paragraph = []
+        if match is None:
+            if version is not None:
+                body.append(line)
+            continue
+
+        flush()
+        # An empty heading has no title, so it ends the release above without
+        # indexing one: `_version_from_heading` finds no version and `flush` skips.
+        heading = match.group("title") or ""
+        version = _version_from_heading(heading)
+        body = []
+
+    flush()
+    return entries
+
+
+def find_release_notes(text: str, version: str) -> ChangelogEntry | None:
+    """Return the section for exactly `version`, or None.
+
+    Equality is version-aware (`2026.07.5` matches `2026.7.5`) but never fuzzy:
+    a near-miss returns None so the caller shows no notes, not the wrong ones.
+    """
+    entries = parse_changelog(text)
+    for entry in entries:
+        # An exact heading wins, so `## 1.0` is never shadowed by `## 1.0.0`.
+        if entry.version == version:
+            return entry
+
+    wanted = _parse_version(version)
+    for entry in entries:
+        if wanted is not None:
+            candidate = _parse_version(entry.version)
+            if candidate is not None and candidate == wanted:
+                return entry
+    return None
+
+
+def get_release_notes(version: str, refresh: bool = False) -> dict[str, Any]:
+    """Return release notes for exactly `version` for the update popup.
+
+    `refresh` retries a cached remote failure, so the UI's retry action is not
+    stuck behind the failure TTL once connectivity returns.
+    """
+    version = version.strip()
+    if not is_supported_version_query(version):
+        return _notes_response(version = version, error = "Unsupported version.")
+
+    local = _read_local_changelog()
+    remote = ChangelogSource(text = None, source = None)
+    if os.environ.get(DISABLE_ENV_VAR) != "1":
+        remote = get_remote_changelog(refresh = refresh)
+
+    # Remote first: the offered version is newer than the local copy.
+    for candidate in (remote, local):
+        if not candidate.text:
+            continue
+        entry = find_release_notes(candidate.text, version)
+        if entry is not None:
+            return _notes_response(
+                version = version,
+                markdown = entry.body,
+                heading = entry.heading,
+                source = candidate.source,
+            )
+
+    # Nothing matched: the bundled copy cannot know a version newer than the
+    # install, so report a remote failure and let the UI offer a retry.
+    return _notes_response(version = version, error = remote.error)
+
+
+def get_remote_changelog(refresh: bool = False) -> ChangelogSource:
+    """Fetch CHANGELOG.md from the repo using a small in-process TTL cache."""
+    global _remote_cache, _remote_fetching
+
+    if refresh:
+        # Only a cached failure is dropped, so retries cannot hammer the remote.
+        with _cache_condition:
+            if _remote_cache and _remote_cache.source.text is None:
+                _remote_cache = None
+
+    # A caller waits for an in-flight fetch only as long as it may take, then
+    # answers locally rather than holding a worker behind a stalled upstream.
+    deadline = time.monotonic() + CHANGELOG_TIMEOUT_SECONDS + 1
+    while True:
+        now = time.monotonic()
+        with _cache_condition:
+            if _remote_cache and _remote_cache.expires_at > now:
+                return _remote_cache.source
+            if not _remote_fetching:
+                _remote_fetching = True
+                break
+            if now >= deadline:
+                return ChangelogSource(
+                    text = None,
+                    source = None,
+                    error = "Release notes are still loading.",
+                )
+            _cache_condition.wait(timeout = deadline - now)
+
+    try:
+        try:
+            source = _fetch_remote_changelog()
+        except Exception:
+            source = ChangelogSource(
+                text = None,
+                source = None,
+                error = "Could not fetch release notes.",
+            )
+
+        ttl = CHANGELOG_SUCCESS_TTL_SECONDS if source.text else CHANGELOG_FAILURE_TTL_SECONDS
+        with _cache_condition:
+            _remote_cache = _ChangelogCacheEntry(source = source, expires_at = time.monotonic() + ttl)
+        return source
+    finally:
+        # Released here, not on the Exception path: stranding the single-flight
+        # flag on BaseException makes every later caller wait out the deadline.
+        with _cache_condition:
+            _remote_fetching = False
+            _cache_condition.notify_all()
+
+
+def _fetch_remote_changelog() -> ChangelogSource:
+    url = os.environ.get(CHANGELOG_URL_ENV_VAR, "").strip() or CHANGELOG_RAW_URL
+    if not url.startswith(("http://", "https://")):
+        return ChangelogSource(text = None, source = None, error = "Invalid changelog URL.")
+
+    request = urllib.request.Request(
+        url,
+        headers = {
+            "User-Agent": "unsloth-studio-update-check",
+            # Or a compressing proxy hands back bytes we would decode as notes.
+            "Accept-Encoding": "identity",
+        },
+    )
+    deadline = time.monotonic() + CHANGELOG_TIMEOUT_SECONDS
+    try:
+        with urllib.request.urlopen(request, timeout = CHANGELOG_TIMEOUT_SECONDS) as response:
+            chunks: list[bytes] = []
+            received = 0
+            while received <= CHANGELOG_MAX_BYTES:
+                remaining = deadline - time.monotonic()
+                if remaining <= 0:
+                    return ChangelogSource(
+                        text = None,
+                        source = None,
+                        error = "Release notes took too long to load.",
+                    )
+                # The socket timeout is per operation, so re-cap it each read.
+                _limit_read(response, remaining)
+                chunk = response.read1(_CHANGELOG_CHUNK_BYTES)
+                if not chunk:
+                    break
+                chunks.append(chunk)
+                received += len(chunk)
+            body = b"".join(chunks)
+        if len(body) > CHANGELOG_MAX_BYTES:
+            return ChangelogSource(
+                text = None,
+                source = None,
+                error = "Release notes response was too large.",
+            )
+        return ChangelogSource(text = body.decode("utf-8", errors = "replace"), source = "remote")
+    except TimeoutError:
+        return ChangelogSource(
+            text = None,
+            source = None,
+            error = "Release notes took too long to load.",
+        )
+    except OSError:
+        return ChangelogSource(
+            text = None,
+            source = None,
+            error = "Could not reach the changelog for release notes.",
+        )
+    except UnicodeError:
+        return ChangelogSource(text = None, source = None, error = "Malformed changelog.")
+
+
+def _limit_read(response: Any, remaining: float) -> None:
+    """Cap the next socket read at the time left in the fetch budget."""
+    sock = getattr(getattr(response, "fp", None), "raw", None)
+    sock = getattr(sock, "_sock", None)
+    if sock is None:
+        return
+    try:
+        sock.settimeout(max(remaining, _CHANGELOG_MIN_READ_SECONDS))
+    except OSError:
+        pass
+
+
+def _read_local_changelog() -> ChangelogSource:
+    """Read the CHANGELOG.md bundled with this install, if there is one."""
+    for path in _local_changelog_candidates():
+        try:
+            if not path.is_file():
+                continue
+            if path.stat().st_size > CHANGELOG_MAX_BYTES:
+                continue
+            return ChangelogSource(
+                text = path.read_text(encoding = "utf-8", errors = "replace"),
+                source = "local",
+            )
+        except OSError:
+            continue
+    return ChangelogSource(text = None, source = None)
+
+
+def _is_source_checkout(root: Path) -> bool:
+    """Whether `root` is this repository rather than an install directory."""
+    try:
+        return any((root / marker).exists() for marker in _CHECKOUT_MARKERS)
+    except OSError:
+        return False
+
+
+def _local_changelog_candidates() -> list[Path]:
+    override = os.environ.get(CHANGELOG_PATH_ENV_VAR, "").strip()
+    candidates: list[Path] = []
+    if override:
+        candidates.append(Path(override).expanduser())
+
+    # changelog.py -> utils -> backend -> studio -> repo root. Repo root first
+    # so a checkout's editable file beats the snapshot packaging writes into
+    # studio/. Installed, those outer levels are site-packages, hence the marker.
+    parents = Path(__file__).resolve().parents
+    for index in (3, 2, 1, 4):
+        if index >= len(parents):
+            continue
+        root = parents[index]
+        if index in _CHECKOUT_ONLY_LEVELS and not _is_source_checkout(root):
+            continue
+        candidates.append(root / CHANGELOG_FILENAME)
+
+    seen: set[Path] = set()
+    unique: list[Path] = []
+    for candidate in candidates:
+        if candidate not in seen:
+            seen.add(candidate)
+            unique.append(candidate)
+    return unique
+
+
+def _opens_fence(marker: str, rest: str) -> bool:
+    """A backtick fence's info string may not contain a backtick."""
+    return marker[0] != "`" or "`" not in rest
+
+
+def _next_fence_state(open_fence: str | None, marker: str, rest: str) -> str | None:
+    """Track the open fence marker.
+
+    A closer must be the same character, at least as long, and carry nothing
+    after it. So neither a ``` sample nor a ```` line with trailing text ends
+    a ```` block early, while an opening fence may still have an info string.
+    Only spaces and tabs count as nothing: other Unicode whitespace is content.
+    """
+    if open_fence is None:
+        return marker if _opens_fence(marker, rest) else None
+    closes = marker[0] == open_fence[0] and len(marker) >= len(open_fence)
+    if closes and not rest.strip(" \t"):
+        return None
+    return open_fence
+
+
+def _code_span_ranges(line: str) -> list[tuple[int, int]]:
+    """Code span bounds. A run of backticks closes only on a run of its length."""
+    # Collect the runs once: rescanning per opener is quadratic on a line of
+    # distinct unmatched runs, and notes are reparsed on every request.
+    runs: list[tuple[int, int]] = []
+    index = 0
+    while index < len(line):
+        if line[index] != "`" or _is_escaped(line, index):
+            index += 1
+            continue
+        ticks = _run_length(line, index)
+        runs.append((index, ticks))
+        index += ticks
+
+    # A run closes only on a later run of its length, so one cursor per length.
+    by_length: dict[int, list[int]] = {}
+    for position, (_, ticks) in enumerate(runs):
+        by_length.setdefault(ticks, []).append(position)
+
+    spans: list[tuple[int, int]] = []
+    cursors: dict[int, int] = {}
+    current = 0
+    while current < len(runs):
+        start, ticks = runs[current]
+        same = by_length[ticks]
+        cursor = cursors.get(ticks, 0)
+        while cursor < len(same) and same[cursor] <= current:
+            cursor += 1
+        cursors[ticks] = cursor
+        if cursor >= len(same):
+            # Nothing closes this run, so it is literal text.
+            current += 1
+            continue
+        closer = same[cursor]
+        cursors[ticks] = cursor + 1
+        spans.append((start, runs[closer][0] + ticks))
+        current = closer + 1
+    return spans
+
+
+def _run_length(line: str, index: int) -> int:
+    end = index
+    while end < len(line) and line[end] == "`":
+        end += 1
+    return end - index
+
+
+def _is_escaped(line: str, index: int) -> bool:
+    slashes = 0
+    while index - 1 - slashes >= 0 and line[index - 1 - slashes] == "\\":
+        slashes += 1
+    return slashes % 2 == 1
+
+
+def _strip_comments(line: str, in_comment: bool, block_open: bool) -> tuple[str, bool]:
+    """Return the line with HTML-comment spans removed, and the trailing state.
+
+    Only a comment that starts a line opens a block and hides the lines below
+    it. One written mid-sentence is inline HTML: it hides the rest of its own
+    line at most, so a note mentioning `<!--` cannot swallow later releases.
+    Delimiters inside inline code are literal and hide nothing.
+
+    "Starts a line" is read inside the container, so `block_open` is decided by
+    the caller from the item's content rather than from the raw line.
+    """
+    if in_comment:
+        close = line.find(_COMMENT_CLOSE)
+        # The closing line belongs to the block, tail included.
+        return ("", False) if close != -1 else ("", True)
+
+    if block_open:
+        # `<!-->` and `<!--->` are complete comments, so the closer may overlap
+        # the opener; searching past it would swallow every later release.
+        return ("", _COMMENT_CLOSE not in line)
+
+    visible: list[str] = []
+    index = 0
+    spans = _code_span_ranges(line)
+    # Spans are ordered and disjoint and each opener sits at or past the one
+    # before, so the search resumes rather than restarts: restarting per opener is
+    # quadratic, and a long line of code spans is reparsed on every request.
+    cursor = 0
+    while index < len(line):
+        opening = line.find(_COMMENT_OPEN, index)
+        if opening == -1:
+            visible.append(line[index:])
+            break
+
+        while cursor < len(spans) and spans[cursor][1] <= opening:
+            cursor += 1
+        if cursor < len(spans) and spans[cursor][0] <= opening:
+            visible.append(line[index : spans[cursor][1]])
+            index = spans[cursor][1]
+            continue
+
+        visible.append(line[index:opening])
+        close = line.find(_COMMENT_CLOSE, opening + len(_COMMENT_OPEN))
+        if close == -1:
+            # Unterminated inline comment: it hides this line and no more.
+            break
+        index = close + len(_COMMENT_CLOSE)
+    return "".join(visible), False
+
+
+def _hidden_structure(line: str, marker: str = "") -> str:
+    """`line` as list tracking sees it once the renderer hides its text.
+
+    A comment or a raw HTML block renders nothing, but it is still a block
+    written at its own column, so it closes the items it sits to the left of.
+    Only the indentation survives: what is inside the block is not Markdown and
+    must not open a list of its own. `marker` is the part of the line that opens
+    a list item the block is the content of, which survives with it."""
+    if marker:
+        return marker + _HIDDEN_BLOCK
+    if not line.strip():
+        return ""
+    return line[: len(line) - len(line.lstrip(" \t"))] + _HIDDEN_BLOCK
+
+
+def _indent_width(line: str) -> int:
+    """Columns of leading whitespace, counting a tab to the next stop of four."""
+    width = 0
+    for char in line:
+        if char == " ":
+            width += 1
+        elif char == "\t":
+            width += 4 - width % 4
+        else:
+            break
+    return width
+
+
+def _strip_indent(line: str, columns: int) -> str:
+    """`line` with up to `columns` columns of leading whitespace removed."""
+    width = 0
+    index = 0
+    while index < len(line) and width < columns and line[index] in " \t":
+        width += 1 if line[index] == " " else 4 - width % 4
+        index += 1
+    return line[index:]
+
+
+def _interrupts_paragraph(line: str) -> bool:
+    """Whether `line` starts a block that can break into an open paragraph.
+
+    A quote marker always can. A list item can only when it has content, and an
+    ordered one only when it starts at 1: anything else is text of the
+    paragraph it appears to interrupt."""
+    if _BLOCK_QUOTE.match(line):
+        return True
+    item = None if _THEMATIC_BREAK.match(line) else _LIST_ITEM.match(line)
+    if item is None:
+        return False
+    marker = item.group("marker")
+    if not line[item.end() :].strip():
+        return False
+    return marker[-1] not in ".)" or marker[:-1] == "1"
+
+
+def _item_content(line: str, after_paragraph: bool) -> str:
+    """`line` read from the content column of a list item that opens on it.
+
+    A block written as an item's first content sits inside that item, so
+    ``- ```` opens a fence even though its marker is not within three columns of
+    the container. The padding is capped the way `_open_lists` caps it, or
+    ``-     ```` would read as a fence rather than the indented code it is. A
+    marker the paragraph above swallows opens no item, so its line is returned
+    whole, as is one four columns past its container. Ported to the frontend as
+    `itemContent` in markdown-list-columns.ts."""
+    if _indent_width(line) >= 4 or (after_paragraph and not _interrupts_paragraph(line)):
+        return line
+    item = None if _THEMATIC_BREAK.match(line) else _LIST_ITEM.match(line)
+    if item is None:
+        return line
+    padding = _indent_width(item.group("space"))
+    # Over-indented content starts one column past the marker; the rest of the
+    # padding is the content's own indentation.
+    over = padding - 1 if padding > _MAX_ITEM_PADDING else 0
+    return " " * over + line[item.end() :]
+
+
+def _quote_content(line: str) -> str:
+    """What a blockquote line holds, with its markers stripped."""
+    while (marker := _QUOTE_MARKER.match(line)) is not None:
+        line = line[marker.end() :]
+    return line
+
+
+def _may_be_lazy(line: str) -> bool:
+    """Whether `line` can continue a paragraph it is indented out of.
+
+    Only plain text can: a heading, a fence, a break or an HTML block starts a
+    block of its own, which closes the item instead. An underline is not one of
+    them: it may never be lazy, so `===` written left of an open item is read as
+    more of the item's paragraph. Nor is a definition, which is a block of its
+    own but may not interrupt a paragraph. A row of dashes still closes the
+    item, as `_INTERRUPTS` reads three or more as the thematic break they are."""
+    return (
+        _PARAGRAPH_TEXT.match(line) is not None
+        and _INTERRUPTS.match(line) is None
+        and _FENCE_PATTERN.match(line) is None
+        # Types 1 to 6 interrupt a paragraph, so a `<div>` left of an open item
+        # closes it. Type 7 cannot, and is deliberately excluded.
+        and not _opens_html_block(line, True)
+    )
+
+
+def _continues_paragraph(line: str, column: int) -> bool:
+    """Whether `line` reads as more of a paragraph open in its container.
+
+    Measured from `column`, where that container's content starts: four columns
+    past it the line is an indented code block, which may not interrupt a
+    paragraph, so indentation alone never closes the one above it."""
+    inner = _strip_indent(line, column)
+    return _indent_width(inner) >= 4 or _may_be_lazy(inner)
+
+
+def _close_dedented(
+    columns: tuple[int, ...], line: str, indent: int, after_paragraph: bool
+) -> tuple[int, ...]:
+    """`columns` with every item `line` is written to the left of closed.
+
+    Read inside the container the item sits in, not from the margin: a line that
+    only looks indented there is lazy text of the item's paragraph, which leaves
+    the item open rather than closing it."""
+    while columns and indent < columns[-1]:
+        outer = columns[-2] if len(columns) > 1 else 0
+        if after_paragraph and _continues_paragraph(line, outer):
+            break
+        columns = columns[:-1]
+    return columns
+
+
+def _lazy_marker(line: str, state: _ListState, after_paragraph: bool, quoted: bool) -> bool:
+    """Whether a marker-shaped `line` is really text of the paragraph above it.
+
+    Only a marker inside the paragraph's own item interrupts it; one to the left
+    closes that item and opens a sibling. A quote owns the paragraph its lines
+    hold, so a marker written outside the quote opens a list of its own."""
+    item = None if _THEMATIC_BREAK.match(line) else _LIST_ITEM.match(line)
+    columns = state.columns
+    return (
+        item is not None
+        and after_paragraph
+        and not quoted
+        and (not columns or _indent_width(line) >= columns[-1])
+        and not _interrupts_paragraph(line)
+    )
+
+
+def _open_lists(
+    line: str,
+    state: _ListState,
+    after_paragraph: bool,
+    quoted: bool = False,
+) -> _ListState:
+    """The list items still open after `line`.
+
+    A dedented line closes an item, unless it is a lazy paragraph continuation.
+    A new marker nests under a deeper column and replaces a sibling. `quoted`
+    marks a paragraph the blockquote above owns: a marker written outside the
+    quote is not text of it, so it opens a list of its own.
+    """
+    columns = state.columns
+    if not line.strip():
+        # A blank line leaves the list open, unless the item is still empty: an
+        # item may begin with one blank line, and later content is outside it.
+        return _ListState(columns[:-1] if state.empty_item else columns)
+    indent = _indent_width(line)
+    item = None if _THEMATIC_BREAK.match(line) else _LIST_ITEM.match(line)
+    empty = item is not None and not line[item.end() :].strip()
+    if _lazy_marker(line, state, after_paragraph, quoted):
+        # A lazy continuation or an underline, so the open items are untouched.
+        return state
+    columns = _close_dedented(columns, line, indent, after_paragraph)
+    # Four columns past its container the marker is an indented code block, or
+    # lazy text of the paragraph above it, so it opens no list of its own.
+    if item is None or indent - (columns[-1] if columns else 0) >= 4:
+        return _ListState(columns)
+    marker = item.group("marker")
+    padding = _indent_width(item.group("space"))
+    if padding == 0 or padding > _MAX_ITEM_PADDING:
+        # An empty or over-indented item still holds one column of content.
+        padding = 1
+    while columns and columns[-1] > indent:
+        columns = columns[:-1]
+    return _ListState((*columns, indent + len(marker) + padding), empty_item = empty)
+
+
+def _opens_html_block(line: str, after_paragraph: bool) -> bool:
+    """True if `line` starts a CommonMark type 6 or type 7 HTML block."""
+    match = _HTML_BLOCK_OPEN.match(line)
+    if match is not None and match.group(1).lower() in _HTML_BLOCK_TAGS:
+        return True
+    return not after_paragraph and _HTML_TAG_ONLY_LINE.match(line) is not None
+
+
+def _strip_raw_html(line: str, open_block: int | None) -> tuple[str, int | None]:
+    """Drop the parts of a line inside a raw block, and return the open block.
+
+    The state is the index of the open block in `_RAW_BLOCKS`, or None."""
+    if open_block is not None:
+        close = _RAW_BLOCKS[open_block][1].search(line)
+        return ("", None) if close else ("", open_block)
+
+    # A block only opens at the start of a line; mid-line tags are inline HTML.
+    for index, (opener, closer) in enumerate(_RAW_BLOCKS):
+        opening = opener.match(line)
+        if opening is None:
+            continue
+        rest = line[opening.end() :]
+        close = closer.search(rest)
+        return ("", None) if close else ("", index)
+    return line, None
+
+
+def _version_from_heading(heading: str) -> str | None:
+    token = heading.split()[0] if heading.split() else ""
+    match = _VERSION_TOKEN_PATTERN.match(token)
+    if match is None:
+        return None
+    version = match.group("version")
+    return version if _parse_version(version) is not None else None
+
+
+def _parse_version(version: str) -> Version | None:
+    try:
+        return Version(version)
+    except InvalidVersion:
+        return None
+
+
+def _close_open_fence(markdown: str) -> str:
+    """Close a fence the truncation cut in half, so the rest still renders."""
+    open_fence: str | None = None
+    for line in _markdown_lines(markdown):
+        fence = _FENCE_PATTERN.match(line)
+        if fence:
+            open_fence = _next_fence_state(open_fence, fence.group("marker"), fence.group("rest"))
+    return f"{markdown}\n{open_fence}" if open_fence else markdown
+
+
+def _renders_visibly(markdown: str) -> bool:
+    """Whether a section body renders anything at all."""
+    in_comment = False
+    for line in _markdown_lines(markdown):
+        opens_raw = any(opener.match(line) for opener, _ in _RAW_BLOCKS)
+        if not in_comment and (_FENCE_PATTERN.match(line) or opens_raw):
+            # A code block or raw HTML block renders even when it is empty.
+            return True
+        # No containers are tracked here, so the opener is read at the margin. The
+        # answer does not turn on it: an item renders its marker whatever the block
+        # inside hides, so a commented-out item renders something either way.
+        visible, in_comment = _strip_comments(
+            line, in_comment, _COMMENT_BLOCK_OPEN.match(line) is not None
+        )
+        if visible.strip():
+            return True
+    return False
+
+
+def _notes_response(
+    *,
+    version: str,
+    markdown: str | None = None,
+    heading: str | None = None,
+    source: str | None = None,
+    error: str | None = None,
+) -> dict[str, Any]:
+    # A section that renders as nothing counts as unpublished, not as empty.
+    if markdown and not _renders_visibly(markdown):
+        markdown = None
+        source = None
+
+    truncated = False
+    if markdown and len(markdown) > RELEASE_NOTES_MAX_CHARS:
+        markdown = _close_open_fence(markdown[:RELEASE_NOTES_MAX_CHARS].rstrip())
+        truncated = True
+
+    return {
+        "version": version,
+        "markdown": markdown or None,
+        "heading": heading,
+        # False means no notes for this exact version; the UI links out.
+        "matched": bool(markdown),
+        "truncated": truncated,
+        "source": source,
+        "release_notes_url": RELEASE_NOTES_URL,
+        "error": error,
+    }
diff --git a/studio/backend/utils/child_stdio.py b/studio/backend/utils/child_stdio.py
new file mode 100644
index 0000000000..4709d650df
--- /dev/null
+++ b/studio/backend/utils/child_stdio.py
@@ -0,0 +1,22 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Make a Python child agree with the parent that its pipes are UTF-8.
+
+A child's ``sys.stdout`` uses ``locale.getpreferredencoding()``, which on
+Windows is the ANSI code page. Reading that pipe as UTF-8 would then mangle any
+non-ASCII the child prints, so the child has to be told which encoding to emit.
+Only needed for Python children; llama.cpp and node already emit UTF-8.
+"""
+
+from __future__ import annotations
+
+import os
+from typing import Mapping, Optional
+
+
+def utf8_child_env(env: Optional[Mapping[str, str]] = None) -> dict[str, str]:
+    """Copy *env* (or the current environment) with UTF-8 stdio forced."""
+    child = dict(os.environ if env is None else env)
+    child["PYTHONIOENCODING"] = "utf-8"
+    return child
diff --git a/studio/backend/utils/hardware/amd.py b/studio/backend/utils/hardware/amd.py
index 91a06c9a2a..318759f67d 100644
--- a/studio/backend/utils/hardware/amd.py
+++ b/studio/backend/utils/hardware/amd.py
@@ -144,6 +144,8 @@ def _run_amd_smi(*args: str, timeout: int = _AMD_SMI_DEFAULT_TIMEOUT) -> Optiona
             ["amd-smi", *args, "--json"],
             capture_output = True,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             timeout = timeout,
             env = _amd_env,
             **windows_hidden_subprocess_kwargs(),
diff --git a/studio/backend/utils/hardware/hardware.py b/studio/backend/utils/hardware/hardware.py
index 48ba375ec5..300d26c362 100644
--- a/studio/backend/utils/hardware/hardware.py
+++ b/studio/backend/utils/hardware/hardware.py
@@ -830,6 +830,8 @@ def _rocm_windows_perf_counter_gpu_util_pct() -> Optional[float]:
             ["powershell", "-NoProfile", "-NonInteractive", "-Command", ps],
             capture_output = True,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             timeout = 5,
         )
         if r.returncode != 0 or not r.stdout.strip():
@@ -1027,6 +1029,8 @@ def _rocm_windows_perf_counter_vram_by_adapter() -> Optional[list[tuple[str, flo
             ["powershell", "-NoProfile", "-NonInteractive", "-Command", ps],
             capture_output = True,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             timeout = 5,
         )
         if r.returncode != 0 or not r.stdout.strip():
diff --git a/studio/backend/utils/hardware/nvidia.py b/studio/backend/utils/hardware/nvidia.py
index f98ca4343e..39e3652921 100644
--- a/studio/backend/utils/hardware/nvidia.py
+++ b/studio/backend/utils/hardware/nvidia.py
@@ -55,6 +55,8 @@ def get_physical_gpu_count() -> Optional[int]:
             ["nvidia-smi", "-L"],
             capture_output = True,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             timeout = 5,
             env = child_env_without_native_path_secret(),
             **_windows_hidden_subprocess_kwargs(),
@@ -81,6 +83,8 @@ def get_primary_gpu_utilization() -> dict[str, Any]:
             ],
             capture_output = True,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             timeout = 5,
             env = child_env_without_native_path_secret(),
             **_windows_hidden_subprocess_kwargs(),
@@ -131,6 +135,8 @@ def get_visible_gpu_utilization(
             ],
             capture_output = True,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             timeout = 5,
             env = child_env_without_native_path_secret(),
             **_windows_hidden_subprocess_kwargs(),
@@ -215,6 +221,8 @@ def get_backend_visible_gpu_info(
             ],
             capture_output = True,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             timeout = 10,
             env = child_env_without_native_path_secret(),
             **_windows_hidden_subprocess_kwargs(),
diff --git a/studio/backend/utils/llama_cpp_update.py b/studio/backend/utils/llama_cpp_update.py
index dffcddb452..5c9646f4eb 100644
--- a/studio/backend/utils/llama_cpp_update.py
+++ b/studio/backend/utils/llama_cpp_update.py
@@ -121,7 +121,14 @@ def _installed_build_number(binary: Optional[str]) -> Optional[int]:
     if not binary:
         return None
     try:
-        proc = subprocess.run([binary, "--version"], capture_output = True, text = True, timeout = 20)
+        proc = subprocess.run(
+            [binary, "--version"],
+            capture_output = True,
+            text = True,
+            encoding = "utf-8",
+            errors = "replace",
+            timeout = 20,
+        )
     except Exception:  # pragma: no cover - defensive
         return None
     m = re.search(r"version:\s*(\d+)", (proc.stderr or "") + (proc.stdout or ""))
diff --git a/studio/backend/utils/mlx_repair.py b/studio/backend/utils/mlx_repair.py
index 4ea1ec62f5..8e2a6a7712 100644
--- a/studio/backend/utils/mlx_repair.py
+++ b/studio/backend/utils/mlx_repair.py
@@ -254,7 +254,7 @@ def _transformers_constraint_args() -> tuple[list[str], str | None]:
     except Exception:
         return [], None
     fd, path = tempfile.mkstemp(prefix = "mlx_repair_", suffix = ".txt")
-    with os.fdopen(fd, "w") as fh:
+    with os.fdopen(fd, "w", encoding = "utf-8") as fh:
         fh.write(f"transformers=={transformers_version}\n")
     return ["--constraint", path], path
 
@@ -290,6 +290,8 @@ def attempt_mlx_repair(*, timeout: int = _REPAIR_TIMEOUT_S) -> bool:
             stdout = subprocess.PIPE,
             stderr = subprocess.STDOUT,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             timeout = timeout,
         )
     except subprocess.TimeoutExpired:
diff --git a/studio/backend/utils/models/checkpoints.py b/studio/backend/utils/models/checkpoints.py
index 6950667bbd..eaf75140fc 100644
--- a/studio/backend/utils/models/checkpoints.py
+++ b/studio/backend/utils/models/checkpoints.py
@@ -129,7 +129,7 @@ def _read_checkpoint_loss(checkpoint_path: Path) -> Optional[float]:
     if not trainer_state.exists():
         return None
     try:
-        with open(trainer_state, encoding = "utf-8") as f:
+        with open(trainer_state, encoding = "utf-8-sig") as f:
             state = json.load(f)
         log_history = state.get("log_history", [])
         if log_history:
@@ -174,18 +174,18 @@ def scan_checkpoints(
             metadata: dict = {}
             try:
                 if adapter_config.exists():
-                    cfg = json.loads(adapter_config.read_text(encoding = "utf-8"))
+                    cfg = json.loads(adapter_config.read_text(encoding = "utf-8-sig"))
                     metadata["base_model"] = cfg.get("base_model_name_or_path")
                     metadata["peft_type"] = cfg.get("peft_type")
                     metadata["lora_rank"] = cfg.get("r")
                 elif config_file.exists():
-                    cfg = json.loads(config_file.read_text(encoding = "utf-8"))
+                    cfg = json.loads(config_file.read_text(encoding = "utf-8-sig"))
                     metadata["base_model"] = cfg.get("_name_or_path")
 
                 # Detect BNB quantization from config.json
                 if config_file.exists():
                     if "cfg" not in dir():
-                        cfg = json.loads(config_file.read_text(encoding = "utf-8"))
+                        cfg = json.loads(config_file.read_text(encoding = "utf-8-sig"))
                     quant_cfg = cfg.get("quantization_config")
                     if (
                         isinstance(quant_cfg, dict)
diff --git a/studio/backend/utils/models/model_config.py b/studio/backend/utils/models/model_config.py
index 893b842e11..6270d9e03f 100644
--- a/studio/backend/utils/models/model_config.py
+++ b/studio/backend/utils/models/model_config.py
@@ -37,6 +37,7 @@ import yaml
 
 
 from utils.native_path_leases import child_env_without_native_path_secret
+from utils.child_stdio import utf8_child_env
 from utils.hf_cache_settings import active_hf_hub_cache, get_hf_cache_paths
 from utils.subprocess_compat import (
     windows_hidden_subprocess_kwargs as _windows_hidden_subprocess_kwargs,
@@ -631,7 +632,7 @@ def _raw_config_has_vision_config(
                     cache_dir = active_hf_hub_cache(),
                 )
             )
-        config = json.loads(config_path.read_text(encoding = "utf-8"))
+        config = json.loads(config_path.read_text(encoding = "utf-8-sig"))
         architectures = config.get("architectures") or []
         model_type = config.get("model_type")
         explicit_vision = (
@@ -774,8 +775,12 @@ def _is_vision_model_subprocess(model_name: str, hf_token: Optional[str] = None)
             ],
             capture_output = True,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             timeout = 60,
-            env = get_hf_cache_paths().child_env(child_env_without_native_path_secret()),
+            env = utf8_child_env(
+                get_hf_cache_paths().child_env(child_env_without_native_path_secret())
+            ),
             **_windows_hidden_subprocess_kwargs(),
         )
 
@@ -1083,7 +1088,7 @@ def _detect_audio_from_tokenizer(
                     ]:
                         tok_file = snapshot / tok_path
                         if tok_file.exists():
-                            tok_config = json.loads(tok_file.read_text(encoding = "utf-8"))
+                            tok_config = json.loads(tok_file.read_text(encoding = "utf-8-sig"))
                             read_any = True
                             result = _check_token_patterns(tok_config)
                             if result:
@@ -2283,7 +2288,7 @@ def scan_exported_models(
                 export_meta = run_dir / "export_metadata.json"
                 try:
                     if export_meta.exists():
-                        meta = json.loads(export_meta.read_text(encoding = "utf-8"))
+                        meta = json.loads(export_meta.read_text(encoding = "utf-8-sig"))
                         base_model = meta.get("base_model")
                 except Exception:
                     pass
@@ -2312,7 +2317,7 @@ def scan_exported_models(
                 if adapter_config.exists():
                     export_type = "lora"
                     try:
-                        cfg = json.loads(adapter_config.read_text(encoding = "utf-8"))
+                        cfg = json.loads(adapter_config.read_text(encoding = "utf-8-sig"))
                         base_model = cfg.get("base_model_name_or_path")
                     except Exception:
                         pass
@@ -2321,7 +2326,7 @@ def scan_exported_models(
                     export_meta = checkpoint_dir / "export_metadata.json"
                     try:
                         if export_meta.exists():
-                            meta = json.loads(export_meta.read_text(encoding = "utf-8"))
+                            meta = json.loads(export_meta.read_text(encoding = "utf-8-sig"))
                             base_model = meta.get("base_model")
                     except Exception:
                         pass
@@ -2334,7 +2339,7 @@ def scan_exported_models(
                         export_meta = meta_dir / "export_metadata.json"
                         try:
                             if export_meta.exists():
-                                meta = json.loads(export_meta.read_text(encoding = "utf-8"))
+                                meta = json.loads(export_meta.read_text(encoding = "utf-8-sig"))
                                 base_model = meta.get("base_model")
                                 if base_model:
                                     break
@@ -2354,7 +2359,7 @@ def scan_exported_models(
                     outputs_adapter_cfg = resolve_output_dir(run_dir.name) / "adapter_config.json"
                     try:
                         if outputs_adapter_cfg.exists():
-                            cfg = json.loads(outputs_adapter_cfg.read_text(encoding = "utf-8"))
+                            cfg = json.loads(outputs_adapter_cfg.read_text(encoding = "utf-8-sig"))
                             base_model = cfg.get("base_model_name_or_path")
                     except Exception:
                         pass
@@ -2380,7 +2385,7 @@ def get_base_model_from_checkpoint(checkpoint_path: str) -> Optional[str]:
 
         adapter_config_path = checkpoint_path_obj / "adapter_config.json"
         if adapter_config_path.exists():
-            with open(adapter_config_path, "r", encoding = "utf-8") as f:
+            with open(adapter_config_path, "r", encoding = "utf-8-sig") as f:
                 config = json.load(f)
                 base_model = config.get("base_model_name_or_path")
                 if base_model:
@@ -2389,7 +2394,7 @@ def get_base_model_from_checkpoint(checkpoint_path: str) -> Optional[str]:
 
         config_path = checkpoint_path_obj / "config.json"
         if config_path.exists():
-            with open(config_path, "r", encoding = "utf-8") as f:
+            with open(config_path, "r", encoding = "utf-8-sig") as f:
                 config = json.load(f)
                 for key in ("model_name", "_name_or_path"):
                     base_model = config.get(key)
@@ -2445,7 +2450,7 @@ def get_base_model_from_lora(lora_path: str) -> Optional[str]:
         # adapter_config.json first
         adapter_config_path = lora_path_obj / "adapter_config.json"
         if adapter_config_path.exists():
-            with open(adapter_config_path, "r", encoding = "utf-8") as f:
+            with open(adapter_config_path, "r", encoding = "utf-8-sig") as f:
                 config = json.load(f)
                 base_model = config.get("base_model_name_or_path")
                 if base_model:
@@ -2535,7 +2540,7 @@ def get_base_model_from_lora_identifier(
             last_exc = exc
             continue
         try:
-            with open(cfg_path, "r", encoding = "utf-8") as f:
+            with open(cfg_path, "r", encoding = "utf-8-sig") as f:
                 base_model = json.load(f).get("base_model_name_or_path")
         except Exception as exc:
             logger.warning("Could not parse adapter_config.json for '%s': %s", identifier, exc)
@@ -2781,7 +2786,7 @@ class ModelConfig:
                 meta_path = gguf_dir / "export_metadata.json"
                 if meta_path.exists():
                     try:
-                        meta = json.loads(meta_path.read_text(encoding = "utf-8"))
+                        meta = json.loads(meta_path.read_text(encoding = "utf-8-sig"))
                         base = meta.get("base_model")
                         if base and is_vision_model(base, hf_token = hf_token):
                             base_is_vision = True
@@ -2912,7 +2917,7 @@ class ModelConfig:
                         token = hf_token,
                         cache_dir = active_hf_hub_cache(),
                     )
-                    with open(config_path, "r", encoding = "utf-8") as f:
+                    with open(config_path, "r", encoding = "utf-8-sig") as f:
                         adapter_config = json.load(f)
                     base_model = adapter_config.get("base_model_name_or_path")
                     if base_model:
diff --git a/studio/backend/utils/node_runtime.py b/studio/backend/utils/node_runtime.py
index fef2430708..697661a095 100644
--- a/studio/backend/utils/node_runtime.py
+++ b/studio/backend/utils/node_runtime.py
@@ -79,6 +79,8 @@ def _node_version_ok(executable: str) -> bool:
             [executable, "-v"],
             capture_output = True,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             timeout = _NODE_VERSION_PROBE_TIMEOUT_SECONDS,
             **windows_hidden_subprocess_kwargs(),
         )
diff --git a/studio/backend/utils/paths/storage_roots.py b/studio/backend/utils/paths/storage_roots.py
index ae1319d296..0b1398f6d2 100644
--- a/studio/backend/utils/paths/storage_roots.py
+++ b/studio/backend/utils/paths/storage_roots.py
@@ -212,7 +212,7 @@ def lmstudio_model_dirs() -> list[Path]:
     settings_path = Path.home() / ".lmstudio" / "settings.json"
     if settings_path.is_file():
         try:
-            with open(settings_path, encoding = "utf-8") as f:
+            with open(settings_path, encoding = "utf-8-sig") as f:
                 settings = json.load(f)
             downloads = settings.get("downloadsFolder", "")
             if downloads:
diff --git a/studio/backend/utils/prebuilt/update_flow.py b/studio/backend/utils/prebuilt/update_flow.py
index 74af0c18f9..69c1566fc3 100644
--- a/studio/backend/utils/prebuilt/update_flow.py
+++ b/studio/backend/utils/prebuilt/update_flow.py
@@ -24,6 +24,7 @@ from typing import Callable, Optional
 
 import structlog
 
+from utils.child_stdio import utf8_child_env
 from utils.process_lifetime import child_popen_kwargs
 
 logger = structlog.get_logger(__name__)
@@ -159,6 +160,8 @@ def resolve_prebuilt_for_host(
             cmd,
             capture_output = True,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             timeout = 60,
         )
         out = (proc.stdout or "").strip()
@@ -303,7 +306,10 @@ def stream_installer(
         stdout = subprocess.PIPE,
         stderr = subprocess.STDOUT,
         text = True,
-        env = env,
+        encoding = "utf-8",
+        errors = "replace",
+        # Make the Python child emit the UTF-8 we decode above.
+        env = utf8_child_env(env),
         **child_popen_kwargs(),
     )
     timed_out = threading.Event()
diff --git a/studio/backend/utils/security/consent.py b/studio/backend/utils/security/consent.py
index 6fee259139..9385270ee0 100644
--- a/studio/backend/utils/security/consent.py
+++ b/studio/backend/utils/security/consent.py
@@ -142,7 +142,7 @@ def _load_remote_code_configs(model_name: str, hf_token: Optional[str] = None) -
             for name in _REMOTE_CODE_CONFIG_FILES:
                 p = root / name
                 if p.is_file():
-                    configs.append(json.loads(p.read_text(encoding = "utf-8")))
+                    configs.append(json.loads(p.read_text(encoding = "utf-8-sig")))
             return configs
 
         from huggingface_hub import hf_hub_download
@@ -164,7 +164,7 @@ def _load_remote_code_configs(model_name: str, hf_token: Optional[str] = None) -
                 # Transient/auth failure is not "absent" -> fail closed to "unknown" so
                 # the caller scans (a tokenizer/processor-only auto_map must not slip by).
                 return None
-            configs.append(json.loads(Path(p).read_text(encoding = "utf-8")))
+            configs.append(json.loads(Path(p).read_text(encoding = "utf-8-sig")))
         # Every config was read or a genuine 404 -> an empty list is a definitive
         # "no auto_map", not "unknown".
         return configs
diff --git a/studio/backend/utils/security/file_security.py b/studio/backend/utils/security/file_security.py
index 7724406e8d..4588f32b90 100644
--- a/studio/backend/utils/security/file_security.py
+++ b/studio/backend/utils/security/file_security.py
@@ -199,7 +199,7 @@ def _indexed_shard_paths(
                 inconclusive = True  # transient: an index that might exist could not be read
                 continue
             try:
-                weight_map = (json.loads(open(index_path, encoding = "utf-8").read()) or {}).get(
+                weight_map = (json.loads(open(index_path, encoding = "utf-8-sig").read()) or {}).get(
                     "weight_map"
                 ) or {}
                 for shard in weight_map.values():
@@ -328,7 +328,7 @@ def _st_load_roots(snapshot: Path) -> list:
     roots = [snapshot]
     try:
         import json
-        modules = json.loads((snapshot / "modules.json").read_text(encoding = "utf-8"))
+        modules = json.loads((snapshot / "modules.json").read_text(encoding = "utf-8-sig"))
     except (OSError, ValueError):
         return roots  # no / invalid modules.json -> snapshot root is the only load root
     for module in modules or ():
@@ -355,7 +355,7 @@ def _indexed_pickle_shards(index_path: Path, root: Path, snapshot: Path) -> list
     try:
         # JSON is UTF-8 by spec; pin it so a non-ASCII index is not misdecoded (and needlessly
         # blocked) under Windows' cp1252 default.
-        parsed = json.loads(index_path.read_text(encoding = "utf-8"))
+        parsed = json.loads(index_path.read_text(encoding = "utf-8-sig"))
     except (OSError, ValueError) as exc:
         raise OSError(f"unreadable weight index: {index_path}") from exc
     weight_map = parsed.get("weight_map") if isinstance(parsed, dict) else None
diff --git a/studio/backend/utils/security/remote_code_approvals.py b/studio/backend/utils/security/remote_code_approvals.py
index f1baac6924..d6076fd2b7 100644
--- a/studio/backend/utils/security/remote_code_approvals.py
+++ b/studio/backend/utils/security/remote_code_approvals.py
@@ -69,7 +69,7 @@ def approval_target_key(targets) -> str:
 def _load() -> dict:
     """Parsed store, or an empty skeleton on any error (fail-safe = re-prompt)."""
     try:
-        with open(_store_path(), encoding = "utf-8") as f:
+        with open(_store_path(), encoding = "utf-8-sig") as f:
             data = json.load(f)
         # Validate the shape, not just the version: a hand-edited ``subjects`` that is not a
         # dict (e.g. ``[]``) would otherwise crash lookup/record instead of failing safe.
diff --git a/studio/backend/utils/security/remote_code_scan.py b/studio/backend/utils/security/remote_code_scan.py
index d4d8003252..42f9d98efe 100644
--- a/studio/backend/utils/security/remote_code_scan.py
+++ b/studio/backend/utils/security/remote_code_scan.py
@@ -454,7 +454,7 @@ def repo_remote_code_files(model_name: str, hf_token: Optional[str] = None) -> d
                 p = root / name
                 if p.is_file():
                     try:
-                        ext_refs |= _auto_map_refs(json.loads(p.read_text(encoding = "utf-8")))
+                        ext_refs |= _auto_map_refs(json.loads(p.read_text(encoding = "utf-8-sig")))
                     except Exception:
                         pass
             if not _add_external_refs(files, ext_refs, hf_token, model_name):
@@ -483,7 +483,7 @@ def repo_remote_code_files(model_name: str, hf_token: Optional[str] = None) -> d
                     f"{model_name}: config {cfg_name} could not be fetched ({exc})"
                 ) from exc
             try:
-                refs |= _auto_map_refs(json.loads(Path(cfg_path).read_text(encoding = "utf-8")))
+                refs |= _auto_map_refs(json.loads(Path(cfg_path).read_text(encoding = "utf-8-sig")))
             except Exception:
                 pass
         own_refs = {fn for repo, fn in refs if repo is None}
@@ -616,7 +616,7 @@ def external_auto_map_repos(model_name: str, hf_token: Optional[str] = None) ->
                 if not p.is_file():
                     continue
                 try:
-                    refs = _auto_map_refs(json.loads(p.read_text(encoding = "utf-8")))
+                    refs = _auto_map_refs(json.loads(p.read_text(encoding = "utf-8-sig")))
                 except Exception:
                     continue
                 repos.update(repo for repo, _fn in refs if repo)
@@ -638,7 +638,7 @@ def external_auto_map_repos(model_name: str, hf_token: Optional[str] = None) ->
             except Exception:
                 continue
             try:
-                refs = _auto_map_refs(json.loads(Path(cfg_path).read_text(encoding = "utf-8")))
+                refs = _auto_map_refs(json.loads(Path(cfg_path).read_text(encoding = "utf-8-sig")))
             except Exception:
                 continue
             repos.update(repo for repo, _fn in refs if repo)
diff --git a/studio/backend/utils/ssm_runtime.py b/studio/backend/utils/ssm_runtime.py
index ca7e2309f9..b864e78608 100644
--- a/studio/backend/utils/ssm_runtime.py
+++ b/studio/backend/utils/ssm_runtime.py
@@ -23,6 +23,7 @@ import threading
 from typing import Any, Callable, Optional
 
 from loggers import get_logger
+from utils.child_stdio import utf8_child_env
 from utils.wheel_utils import (
     direct_wheel_url,
     install_wheel,
@@ -254,6 +255,12 @@ def _install_kernel(
         "stdout": subprocess.PIPE,
         "stderr": subprocess.STDOUT,
         "text": True,
+        # pip and the compilers it drives write UTF-8 down this pipe; the Windows
+        # ANSI codepage would mojibake or raise over a fine install.
+        "encoding": "utf-8",
+        "errors": "replace",
+        # Make the Python child emit the UTF-8 we decode above.
+        "env": utf8_child_env(),
     }
     if is_hip:
         run_kwargs["timeout"] = 1800  # ROCm builds can take 10-30 min
@@ -261,7 +268,8 @@ def _install_kernel(
         if "--gcc-install-dir" not in existing:
             gcc_dir = _hipcc_gcc_install_dir()
             if gcc_dir:
-                _env = os.environ.copy()
+                # Extends the UTF-8 env above rather than replacing it.
+                _env = dict(run_kwargs["env"])
                 _env["HIPCC_COMPILE_FLAGS_APPEND"] = (
                     f"{existing} --gcc-install-dir={gcc_dir}".strip()
                 )
diff --git a/studio/backend/utils/studio_version.py b/studio/backend/utils/studio_version.py
index 82ade74bba..cfaba36a81 100644
--- a/studio/backend/utils/studio_version.py
+++ b/studio/backend/utils/studio_version.py
@@ -60,6 +60,8 @@ def _exact_git_studio_tag(repo_root: Path) -> str | None:
             stdout = subprocess.PIPE,
             stderr = subprocess.DEVNULL,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             timeout = _GIT_TIMEOUT_SECONDS,
         )
     except (OSError, subprocess.TimeoutExpired):
@@ -81,6 +83,8 @@ def _git_branch(repo_root: Path) -> str | None:
             stdout = subprocess.PIPE,
             stderr = subprocess.DEVNULL,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             timeout = _GIT_TIMEOUT_SECONDS,
         )
     except (OSError, subprocess.TimeoutExpired):
diff --git a/studio/backend/utils/transformers_version.py b/studio/backend/utils/transformers_version.py
index b0a2da0e66..3774409009 100644
--- a/studio/backend/utils/transformers_version.py
+++ b/studio/backend/utils/transformers_version.py
@@ -44,6 +44,7 @@ import time
 from pathlib import Path
 
 from utils.native_path_leases import child_env_without_native_path_secret
+from utils.child_stdio import utf8_child_env
 from utils.hf_cache_settings import get_hf_cache_paths
 from utils.subprocess_compat import (
     windows_hidden_subprocess_kwargs as _windows_hidden_subprocess_kwargs,
@@ -420,7 +421,7 @@ def _resolve_base_model(model_name: str) -> str:
     adapter_cfg_path = local_path / "adapter_config.json"
     if _safe_is_file(adapter_cfg_path):
         try:
-            with open(adapter_cfg_path, encoding = "utf-8") as f:
+            with open(adapter_cfg_path, encoding = "utf-8-sig") as f:
                 cfg = json.load(f)
             base = cfg.get("base_model_name_or_path")
             if base:
@@ -437,7 +438,7 @@ def _resolve_base_model(model_name: str) -> str:
     config_json_path = local_path / "config.json"
     if _safe_is_file(config_json_path):
         try:
-            with open(config_json_path, encoding = "utf-8") as f:
+            with open(config_json_path, encoding = "utf-8-sig") as f:
                 cfg = json.load(f)
             # Unsloth writes model_name, HF writes _name_or_path; skip a self-reference.
             for _key in ("model_name", "_name_or_path"):
@@ -544,7 +545,7 @@ def _adapter_base_from_hf_cache(model_name: str) -> str | None:
         )
         for cfg_path in candidates:
             if cfg_path.is_file():
-                base = json.loads(cfg_path.read_text(encoding = "utf-8")).get(
+                base = json.loads(cfg_path.read_text(encoding = "utf-8-sig")).get(
                     "base_model_name_or_path"
                 )
                 return base or None
@@ -616,7 +617,7 @@ def _check_tokenizer_config_needs_v5(model_name: str, hf_token: str | None = Non
     local_tc = local_path / "tokenizer_config.json"
     if _safe_is_file(local_tc):
         try:
-            with open(local_tc, encoding = "utf-8") as f:
+            with open(local_tc, encoding = "utf-8-sig") as f:
                 data = json.load(f)
             tokenizer_class = data.get("tokenizer_class", "")
             result = tokenizer_class in _TRANSFORMERS_5_TOKENIZER_CLASSES
@@ -706,7 +707,7 @@ def _config_json_from_hf_cache(model_name: str) -> dict | None:
         )
         for cfg_path in candidates:
             if cfg_path.is_file():
-                with open(cfg_path, encoding = "utf-8") as f:
+                with open(cfg_path, encoding = "utf-8-sig") as f:
                     return json.load(f)
     except Exception as exc:
         logger.debug("HF cache config.json lookup failed for '%s': %s", model_name, exc)
@@ -731,7 +732,7 @@ def _load_config_json(model_name: str, hf_token: str | None = None) -> dict | No
     local_cfg = Path(model_name) / "config.json"
     if _safe_is_file(local_cfg):
         try:
-            with open(local_cfg, encoding = "utf-8") as f:
+            with open(local_cfg, encoding = "utf-8-sig") as f:
                 cfg = json.load(f)
             _config_json_cache[cache_key] = cfg
             return cfg
@@ -1271,9 +1272,10 @@ def _probe_autoconfig(target_dir: str, model_name: str, hf_token: str | None) ->
             [sys.executable, "-c", _PROBE_CONFIG_SCRIPT, target_dir, model_name],
             capture_output = True,
             text = True,
+            encoding = "utf-8",
             errors = "replace",
             timeout = _PROBE_TIMEOUT_SECS,
-            env = env,
+            env = utf8_child_env(env),
             **_windows_hidden_subprocess_kwargs(),
         )
     except subprocess.TimeoutExpired:
@@ -1811,7 +1813,11 @@ def _install_to_dir(pkg: str, target_dir: str) -> bool:
             stdout = subprocess.PIPE,
             stderr = subprocess.STDOUT,
             text = True,
-            env = get_hf_cache_paths().child_env(child_env_without_native_path_secret()),
+            encoding = "utf-8",
+            errors = "replace",
+            env = utf8_child_env(
+                get_hf_cache_paths().child_env(child_env_without_native_path_secret())
+            ),
             **_windows_hidden_subprocess_kwargs(),
         )
         if result.returncode == 0:
@@ -1834,7 +1840,9 @@ def _install_to_dir(pkg: str, target_dir: str) -> bool:
         stdout = subprocess.PIPE,
         stderr = subprocess.STDOUT,
         text = True,
-        env = get_hf_cache_paths().child_env(child_env_without_native_path_secret()),
+        encoding = "utf-8",
+        errors = "replace",
+        env = utf8_child_env(get_hf_cache_paths().child_env(child_env_without_native_path_secret())),
         **_windows_hidden_subprocess_kwargs(),
     )
     if result.returncode != 0:
@@ -2079,7 +2087,7 @@ class SidecarSwapInProgress(RuntimeError):
 
 def _read_swap_lock(path: Path) -> dict | None:
     try:
-        data = json.loads(path.read_text(encoding = "utf-8"))
+        data = json.loads(path.read_text(encoding = "utf-8-sig"))
         return data if isinstance(data, dict) else {}
     except FileNotFoundError:
         return None
@@ -2120,7 +2128,7 @@ def try_begin_sidecar_swap(kind: str = "install") -> bool:
                 break
         if fd is not None:
             try:
-                with os.fdopen(fd, "w") as f:
+                with os.fdopen(fd, "w", encoding = "utf-8") as f:
                     f.write(
                         json.dumps(
                             {"pid": os.getpid(), "at": time.time(), "token": token, "kind": kind}
@@ -2466,7 +2474,11 @@ def _ensure_venv_llmcompressor_exists() -> bool:
             stdout = subprocess.PIPE,
             stderr = subprocess.STDOUT,
             text = True,
-            env = get_hf_cache_paths().child_env(child_env_without_native_path_secret()),
+            encoding = "utf-8",
+            errors = "replace",
+            env = utf8_child_env(
+                get_hf_cache_paths().child_env(child_env_without_native_path_secret())
+            ),
             **_windows_hidden_subprocess_kwargs(),
         )
         last_out = result.stdout or ""
diff --git a/studio/backend/utils/update_status.py b/studio/backend/utils/update_status.py
index ad9dabcf36..d4b8ca1c16 100644
--- a/studio/backend/utils/update_status.py
+++ b/studio/backend/utils/update_status.py
@@ -30,6 +30,7 @@ PYPI_SUCCESS_TTL_SECONDS = 12 * 60 * 60
 PYPI_FAILURE_TTL_SECONDS = 60 * 60
 RELEASE_NOTES_URL = "https://unsloth.ai/docs/new/changelog"
 DISABLE_ENV_VAR = "UNSLOTH_DISABLE_UPDATE_CHECK"
+FAKE_UPDATE_ENV_VAR = "UNSLOTH_STUDIO_FAKE_UPDATE"
 
 LOCAL_INSTALL_SOURCES = {"editable", "local_path", "vcs", "local_repo"}
 
@@ -107,11 +108,32 @@ def get_studio_install_source_status(current_version: str) -> dict[str, Any]:
     )
 
 
+def _is_version(value: str) -> bool:
+    try:
+        Version(value)
+    except InvalidVersion:
+        return False
+    return True
+
+
 def get_studio_update_status(current_version: str) -> dict[str, Any]:
     """Return public, read-only update status for the web UI."""
     install_source = detect_install_source()
+    disabled = os.environ.get(DISABLE_ENV_VAR) == "1"
 
-    if os.environ.get(DISABLE_ENV_VAR) == "1":
+    # Dev-only: the popup is PyPI-install-only, so fake a version to review it
+    # from a checkout. The documented opt-out still wins.
+    forced_version = os.environ.get(FAKE_UPDATE_ENV_VAR, "").strip()
+    if forced_version and not disabled and _is_version(forced_version):
+        return _status_response(
+            current_version = current_version,
+            latest_version = forced_version,
+            install_source = "pypi",
+            update_available = True,
+            can_show_web_notification = True,
+        )
+
+    if disabled:
         return _status_response(
             current_version = current_version,
             latest_version = None,
diff --git a/studio/backend/utils/utils.py b/studio/backend/utils/utils.py
index e4964b8d04..e830ea2700 100644
--- a/studio/backend/utils/utils.py
+++ b/studio/backend/utils/utils.py
@@ -114,6 +114,8 @@ def hf_cache_snapshot_dir(model_name: str) -> Optional[Path]:
                 snapshot = repo_dir / "snapshots" / commit
                 if snapshot.is_dir():
                     return snapshot
+            # UnicodeDecodeError is a ValueError, not an OSError: a torn refs
+            # file must keep meaning "not cached here", not fail the offline check.
             except (OSError, UnicodeDecodeError):
                 continue
     return None
diff --git a/studio/backend/utils/wheel_utils.py b/studio/backend/utils/wheel_utils.py
index 1b5926fd49..8ebdea3ac1 100644
--- a/studio/backend/utils/wheel_utils.py
+++ b/studio/backend/utils/wheel_utils.py
@@ -15,6 +15,7 @@ import urllib.request
 from typing import Callable
 
 from utils.native_path_leases import child_env_without_native_path_secret
+from utils.child_stdio import utf8_child_env
 from utils.subprocess_compat import windows_hidden_subprocess_kwargs
 
 _logger = logging.getLogger(__name__)
@@ -43,6 +44,8 @@ def has_blackwell_gpu() -> bool:
             stdout = subprocess.PIPE,
             stderr = subprocess.DEVNULL,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             timeout = 10,
             env = child_env_without_native_path_secret(),
         )
@@ -102,8 +105,10 @@ def probe_torch_wheel_env(*, timeout: int | None = None) -> dict[str, str] | Non
             stdout = subprocess.PIPE,
             stderr = subprocess.PIPE,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             timeout = timeout,
-            env = child_env_without_native_path_secret(),
+            env = utf8_child_env(child_env_without_native_path_secret()),
             **windows_hidden_subprocess_kwargs(),
         )
     except subprocess.TimeoutExpired:
@@ -201,6 +206,8 @@ def install_wheel(
             stdout = subprocess.PIPE,
             stderr = subprocess.STDOUT,
             text = True,
+            encoding = "utf-8",
+            errors = "replace",
             env = child_env_without_native_path_secret(),
         )
         attempts.append(("uv", result))
@@ -213,7 +220,10 @@ def install_wheel(
         stdout = subprocess.PIPE,
         stderr = subprocess.STDOUT,
         text = True,
-        env = child_env_without_native_path_secret(),
+        encoding = "utf-8",
+        errors = "replace",
+        # Make the Python child emit the UTF-8 we decode above.
+        env = utf8_child_env(child_env_without_native_path_secret()),
     )
     attempts.append(("pip", result))
     return attempts
diff --git a/studio/backend/utils/whisper_cpp_update.py b/studio/backend/utils/whisper_cpp_update.py
index cac37c25fc..45a0faf674 100644
--- a/studio/backend/utils/whisper_cpp_update.py
+++ b/studio/backend/utils/whisper_cpp_update.py
@@ -121,7 +121,14 @@ def _installed_whisper_version(binary: Optional[str]) -> Optional[str]:
     if not binary:
         return None
     try:
-        proc = subprocess.run([binary, "--version"], capture_output = True, text = True, timeout = 20)
+        proc = subprocess.run(
+            [binary, "--version"],
+            capture_output = True,
+            text = True,
+            encoding = "utf-8",
+            errors = "replace",
+            timeout = 20,
+        )
     except Exception:  # pragma: no cover - defensive
         return None
     m = re.search(r"v?(\d+\.\d+\.\d+)", (proc.stderr or "") + (proc.stdout or ""))
diff --git a/studio/frontend/public/agent-logos/pi.svg b/studio/frontend/public/agent-logos/pi.svg
new file mode 100644
index 0000000000..3f8a77bd1a
--- /dev/null
+++ b/studio/frontend/public/agent-logos/pi.svg
@@ -0,0 +1,21 @@
+<!-- Source: https://pi.dev/favicon.svg (official Pi press-kit badge) -->
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 800 800">
+  <rect width="800" height="800" rx="120" fill="#09090b"/>
+  <path fill="#fff" fill-rule="evenodd" d="
+    M165.29 165.29
+    H517.36
+    V400
+    H400
+    V517.36
+    H282.65
+    V634.72
+    H165.29
+    Z
+    M282.65 282.65
+    V400
+    H400
+    V282.65
+    Z
+  "/>
+  <path fill="#fff" d="M517.36 400 H634.72 V634.72 H517.36 Z"/>
+</svg>
diff --git a/studio/frontend/src/app/provider.tsx b/studio/frontend/src/app/provider.tsx
index d746ed952c..b076c8cf8d 100644
--- a/studio/frontend/src/app/provider.tsx
+++ b/studio/frontend/src/app/provider.tsx
@@ -214,7 +214,8 @@ function TauriUpdateLayer({
   }
 
   return (
-    <div className="pointer-events-none fixed bottom-4 right-4 z-[9998] flex w-[calc(100vw-2rem)] max-w-[400px] flex-col items-stretch gap-2">
+    // Capped like the browser stack: the download panel shares it, so both must fit.
+    <div className="pointer-events-none fixed bottom-4 right-4 z-[9998] flex max-h-[calc(100dvh_-_2rem)] flex-col items-end gap-2">
       <UpdateBanner
         status={update.status}
         info={update.info}
@@ -223,6 +224,7 @@ function TauriUpdateLayer({
         isExternalServer={isExternalServer}
         updatePolicyMode={update.updatePolicyMode}
         manualReleaseUrl={update.manualReleaseUrl}
+        releasePageUrl={update.releasePageUrl}
         positioned={false}
         onInstall={update.installUpdate}
         onDismiss={update.dismiss}
@@ -379,9 +381,11 @@ function TauriWrapper({ children }: { children: ReactNode }) {
     return (
       <>
         {children}
-        {/* One bottom-right stack so overlays never overlap; they stack with a
-            gap, download panel anchored at the corner with banners above. */}
-        <div className="pointer-events-none fixed bottom-4 right-4 z-[9998] flex w-[calc(100vw-2rem)] max-w-[400px] flex-col items-stretch gap-2">
+        {/* One bottom-right stack so overlays never overlap: download panel at the
+            corner, banners above, each owning its width. */}
+        {/* Capped to the viewport, or a long download list plus expanded notes
+            pushes the top of the stack off screen. */}
+        <div className="pointer-events-none fixed bottom-4 right-4 z-[9998] flex max-h-[calc(100dvh_-_2rem)] flex-col items-end gap-2">
           <WebUpdateBanner
             positioned={false}
             enabled={!WEB_UPDATE_HIDDEN_ROUTES.has(pathname)}
diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index d263a6a739..ecbebba326 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -1226,7 +1226,9 @@ export function AppSidebar() {
                     openNewChat(null);
                   }}
                   className={cn(
-                    "flex items-center gap-[6px] select-none transition-opacity",
+                    // min-w-0 so a narrow sidebar truncates the wordmark
+                    // instead of pushing the search icon over the logo.
+                    "flex min-w-0 items-center gap-[6px] select-none transition-opacity",
                     chatDisabled && "pointer-events-none opacity-50",
                   )}
                   aria-label={t("shell.aria.home")}
@@ -1238,17 +1240,17 @@ export function AppSidebar() {
                   <img
                     src="/circle-logo-small.png"
                     alt="Unsloth"
-                    className="h-[calc(26px+0.5rem*var(--ui-font-scale,1))] w-[calc(26px+0.5rem*var(--ui-font-scale,1))] rounded-full object-cover"
+                    className="h-[calc(26px+0.5rem*var(--ui-font-scale,1))] w-[calc(26px+0.5rem*var(--ui-font-scale,1))] shrink-0 rounded-full object-cover"
                   />
-                  <span className="font-heading text-[calc(13px+0.5rem*var(--ui-font-scale,1))] font-semibold tracking-[0em] leading-none text-black dark:text-white dark:tracking-[0.02em]">
+                  <span className="truncate font-heading text-[calc(13px+0.5rem*var(--ui-font-scale,1))] font-semibold tracking-[0em] leading-none text-black dark:text-white dark:tracking-[0.02em]">
                     unsloth
                   </span>
-                  <span className="nav-badge ml-0.5 inline-flex items-center justify-center rounded-full border border-nav-beta-border px-[5px] pt-[3px] pb-[2px] text-[calc(0.5rem*var(--ui-font-scale,1))] font-medium leading-none tracking-[0.04em] text-nav-fg-muted antialiased subpixel-antialiased shadow-[0_1px_2px_rgba(0,0,0,0.06)] dark:shadow-[0_1px_2px_rgba(0,0,0,0.35)]">
+                  <span className="nav-badge ml-0.5 inline-flex shrink-0 items-center justify-center rounded-full border border-nav-beta-border px-[5px] pt-[3px] pb-[2px] text-[calc(0.5rem*var(--ui-font-scale,1))] font-medium leading-none tracking-[0.04em] text-nav-fg-muted antialiased subpixel-antialiased shadow-[0_1px_2px_rgba(0,0,0,0.06)] dark:shadow-[0_1px_2px_rgba(0,0,0,0.35)]">
                     {t("shell.beta")}
                   </span>
                 </Link>
               )}
-              <div className="flex items-center gap-0.5">
+              <div className="flex shrink-0 items-center gap-0.25">
                 <Tooltip>
                   <TooltipPrimitive.Trigger asChild>
                     <button
@@ -1257,7 +1259,7 @@ export function AppSidebar() {
                         useChatSearchStore.getState().open();
                         closeMobileIfOpen();
                       }}
-                      className="inline-flex h-[33px] w-[32px] cursor-pointer items-center justify-center rounded-[10px] text-nav-icon-idle dark:text-nav-fg-muted transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
+                      className="inline-flex h-[33px] w-[28px] cursor-pointer items-center justify-center rounded-[10px] text-nav-icon-idle dark:text-nav-fg-muted transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
                       aria-label={t("shell.navigation.search")}
                     >
                       <HugeiconsIcon icon={Search01Icon} strokeWidth={1.75} className="size-icon" />
@@ -1281,7 +1283,7 @@ export function AppSidebar() {
                       <button
                         type="button"
                         onClick={togglePinned}
-                        className="inline-flex h-[33px] w-[32px] cursor-pointer items-center justify-center rounded-[10px] text-nav-icon-idle dark:text-nav-fg-muted transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
+                        className="inline-flex h-[33px] w-[28px] cursor-pointer items-center justify-center rounded-[10px] text-nav-icon-idle dark:text-nav-fg-muted transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
                         aria-label={t("shell.aria.closeSidebar")}
                       >
                         <HugeiconsIcon icon={LayoutAlignLeftIcon} strokeWidth={1.75} className="size-icon" />
@@ -1325,10 +1327,10 @@ export function AppSidebar() {
         )}
       </SidebarHeader>
 
-      {/* Uniform pl-1.5 pr-2 keeps every hover pill the same width, inset from the edge. */}
+      {/* Uniform pl-1.5 pr-1.75 keeps every hover pill the same width, inset from the edge. */}
       <SidebarGroup
         className={cn(
-          "group-data-[collapsible=icon]:px-0 pl-1.5 pr-2 shrink-0 transition-[padding]",
+          "group-data-[collapsible=icon]:px-0 pl-1.5 pr-1.75 shrink-0 transition-[padding]",
           showCompactMacBrand ? "pt-0" : "pt-[9px]",
           // Scrolled: New Chat is pinned, give a little gap below it.
           scrolled ? "pb-[5px]" : "pb-px",
@@ -1417,7 +1419,7 @@ export function AppSidebar() {
           scrolled && "is-scrolled",
         )}
       >
-        <SidebarGroup className="group-data-[collapsible=icon]:px-0 pl-1.5 pr-2 py-0 shrink-0">
+        <SidebarGroup className="group-data-[collapsible=icon]:px-0 pl-1.5 pr-1.75 py-0 shrink-0">
           <SidebarGroupContent>
             <SidebarMenu>
               <NavItem
@@ -1499,7 +1501,7 @@ export function AppSidebar() {
               </CollapsibleTrigger>
             </SidebarGroupLabel>
             <CollapsibleContent>
-              <SidebarGroupContent className="pl-1.5 pr-2">
+              <SidebarGroupContent className="pl-1.5 pr-1.75">
                 <SidebarMenu>
                   <NavItem
                     icon={TestTubeOutlineIcon}
@@ -1574,7 +1576,7 @@ export function AppSidebar() {
                   </CollapsibleTrigger>
                 </SidebarGroupLabel>
                 <CollapsibleContent>
-                  <SidebarGroupContent className="pl-1.5 pr-2">
+                  <SidebarGroupContent className="pl-1.5 pr-1.75">
                     <SidebarMenu>
                       {pinnedProjectRecords.map((project) => {
                         const projectChats =
@@ -1721,7 +1723,7 @@ export function AppSidebar() {
                 </CollapsibleTrigger>
               </SidebarGroupLabel>
               <CollapsibleContent>
-                <SidebarGroupContent className="pl-1.5 pr-2">
+                <SidebarGroupContent className="pl-1.5 pr-1.75">
                   <SidebarMenu>
                     {recentChatItems.map((item) =>
                       renderChatSidebarItem(item, "recent"),
@@ -1753,7 +1755,7 @@ export function AppSidebar() {
               </CollapsibleTrigger>
             </SidebarGroupLabel>
             <CollapsibleContent>
-              <SidebarGroupContent className="pl-1.5 pr-2">
+              <SidebarGroupContent className="pl-1.5 pr-1.75">
                 <SidebarMenu>
                   {runItems.map((run) => {
                     // Explicit selection wins. Otherwise highlight the active
diff --git a/studio/frontend/src/components/llama-update-banner.tsx b/studio/frontend/src/components/llama-update-banner.tsx
index 2729558630..5276eda858 100644
--- a/studio/frontend/src/components/llama-update-banner.tsx
+++ b/studio/frontend/src/components/llama-update-banner.tsx
@@ -134,7 +134,7 @@ export function LlamaUpdateBanner({
       className={cn(
         positioned
           ? "fixed bottom-4 right-4 z-[9998] w-[calc(100vw-2rem)] max-w-[400px]"
-          : "pointer-events-auto w-full",
+          : "pointer-events-auto w-[calc(100vw-2rem)] max-w-[400px]",
       )}
       data-testid="llama-update-banner"
     >
diff --git a/studio/frontend/src/components/tauri/update-banner.tsx b/studio/frontend/src/components/tauri/update-banner.tsx
index 6f5e655889..49c9c44aaa 100644
--- a/studio/frontend/src/components/tauri/update-banner.tsx
+++ b/studio/frontend/src/components/tauri/update-banner.tsx
@@ -2,6 +2,7 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { Button } from "@/components/ui/button";
+import { ReleaseNotesPanel } from "@/components/update/release-notes-panel";
 import type {
   DesktopUpdatePolicyMode,
   RetainedUpdateFailure,
@@ -22,6 +23,8 @@ interface UpdateBannerProps {
   isExternalServer?: boolean;
   updatePolicyMode: DesktopUpdatePolicyMode;
   manualReleaseUrl: string | null;
+  // Release page for this version, preferred over the generic changelog.
+  releasePageUrl?: string | null;
   // false fills a shared overlay stack; true self-anchors.
   positioned?: boolean;
   onInstall: () => void;
@@ -30,6 +33,7 @@ interface UpdateBannerProps {
 }
 
 const EASE_OUT_QUART: [number, number, number, number] = [0.165, 0.84, 0.44, 1];
+const LEADING_V = /^v/;
 
 function formatVersion(version: string | null | undefined): string {
   if (!version) return "";
@@ -44,6 +48,7 @@ export function UpdateBanner({
   isExternalServer = false,
   updatePolicyMode,
   manualReleaseUrl,
+  releasePageUrl = null,
   positioned = true,
   onInstall,
   onDismiss,
@@ -52,6 +57,8 @@ export function UpdateBanner({
   const [copying, setCopying] = useState(false);
   const [manualReport, setManualReport] = useState<string | null>(null);
   const [manualMessage, setManualMessage] = useState<string | null>(null);
+  // Version whose notes are expanded; a new offer collapses the panel.
+  const [notesVersion, setNotesVersion] = useState<string | null>(null);
   const showFailure = Boolean(lastFailure) && !dismissed;
   const showAvailable = status === "available" && !dismissed && !showFailure;
   const show = showFailure || (showAvailable && Boolean(info));
@@ -62,6 +69,11 @@ export function UpdateBanner({
   const currentVersion = formatVersion(info?.currentVersion);
   const latestVersion = formatVersion(info?.version);
   const Icon = showFailure ? CircleAlert : Download;
+  // Keyed by the backend release, not the app's SemVer; headings drop the v.
+  const notesTargetVersion =
+    (info?.pypiVersion ?? info?.version)?.replace(LEADING_V, "") ?? null;
+  const notesOpen =
+    notesTargetVersion !== null && notesVersion === notesTargetVersion;
 
   async function handleCopyDiagnostics() {
     setCopying(true);
@@ -94,13 +106,14 @@ export function UpdateBanner({
           exit={{ opacity: 0, y: 8, scale: 0.97 }}
           transition={{ duration: 0.35, ease: EASE_OUT_QUART }}
           className={cn(
+            // Wider than the other overlays: notes preview plus three buttons.
             positioned
-              ? "fixed bottom-4 right-4 z-[9999] w-[calc(100vw-2rem)] max-w-[400px]"
-              : "pointer-events-auto w-full",
+              ? "fixed bottom-4 right-4 z-[9999] w-[calc(100vw-2rem)] max-w-[448px]"
+              : "pointer-events-auto flex min-h-0 w-[calc(100vw-2rem)] max-w-[448px] flex-col",
           )}
           data-testid="tauri-update-banner"
         >
-          <div className="relative overflow-hidden rounded-[24px] bg-white px-5 pb-4 pt-5 shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:bg-card dark:shadow-[0_8px_28px_-6px_rgba(0,0,0,0.28)]">
+          <div className="relative flex max-h-[calc(100dvh_-_2rem)] flex-col overflow-hidden rounded-[24px] bg-white px-5 pb-4 pt-5 shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:bg-card dark:shadow-[0_8px_28px_-6px_rgba(0,0,0,0.28)]">
             <button
               type="button"
               onClick={onDismiss}
@@ -160,7 +173,40 @@ export function UpdateBanner({
               </p>
             )}
 
-            <div className="mt-4 flex flex-wrap items-center justify-end gap-x-1 gap-y-2">
+            {!showFailure && notesTargetVersion ? (
+              <ReleaseNotesPanel
+                version={notesTargetVersion}
+                open={notesOpen}
+                // Used only if CHANGELOG.md has no section for this version.
+                fallbackMarkdown={info?.body ?? null}
+                className="min-h-0 flex-1"
+                releaseNotesUrl={releasePageUrl ?? manualReleaseUrl}
+              />
+            ) : null}
+
+            <div
+              className={cn(
+                "mt-4 flex flex-wrap items-center gap-x-1 gap-y-2",
+                !showFailure && notesTargetVersion
+                  ? "justify-between"
+                  : "justify-end",
+              )}
+            >
+              {!showFailure && notesTargetVersion ? (
+                <Button
+                  size="sm"
+                  variant="ghost"
+                  // same type size as the action buttons
+                  className="-ml-2 h-auto whitespace-nowrap rounded-full px-2.5 py-2 text-ui-13 font-medium text-foreground"
+                  onClick={() =>
+                    setNotesVersion(notesOpen ? null : notesTargetVersion)
+                  }
+                  aria-expanded={notesOpen}
+                  data-testid="tauri-update-release-notes-toggle"
+                >
+                  {notesOpen ? "Hide release notes" : "Show release notes"}
+                </Button>
+              ) : null}
               {showFailure ? (
                 <>
                   <Button
@@ -187,28 +233,31 @@ export function UpdateBanner({
                     onClick={onInstall}
                     disabled={installDisabled}
                   >
-                    {isManualLinuxPackage ? "Open release page" : "Retry update"}
+                    {isManualLinuxPackage
+                      ? "Open release page"
+                      : "Retry update"}
                   </Button>
                 </>
               ) : (
-                <>
+                // wrap + right-align so the action pair stays together
+                <div className="flex flex-wrap items-center justify-end gap-x-1 gap-y-2">
                   <Button
                     size="sm"
                     variant="ghost"
-                    className="h-auto rounded-full px-3 py-2 text-ui-13 font-medium text-foreground"
+                    className="h-auto whitespace-nowrap rounded-full px-2.5 py-2 text-ui-13 font-medium text-foreground"
                     onClick={onDismiss}
                   >
                     Remind me later
                   </Button>
                   <Button
                     size="sm"
-                    className="-mr-1 h-auto rounded-full px-3.5 py-2 text-ui-13"
+                    className="-mr-1 h-auto whitespace-nowrap rounded-full px-3 py-2 text-ui-13"
                     onClick={onInstall}
                     disabled={installDisabled}
                   >
                     {isManualLinuxPackage ? "Open release page" : "Update"}
                   </Button>
-                </>
+                </div>
               )}
             </div>
             {manualMessage && (
diff --git a/studio/frontend/src/components/tauri/window-titlebar.tsx b/studio/frontend/src/components/tauri/window-titlebar.tsx
index 6a0ff8741a..db57cd9960 100644
--- a/studio/frontend/src/components/tauri/window-titlebar.tsx
+++ b/studio/frontend/src/components/tauri/window-titlebar.tsx
@@ -2,6 +2,7 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { useSidebarPin } from "@/hooks/use-sidebar-pin";
+import { useSidebarWidth } from "@/hooks/use-sidebar-width";
 import { isTauri } from "@/lib/api-base";
 import { cn } from "@/lib/utils";
 import {
@@ -110,9 +111,13 @@ export function WindowTitlebar({
   const [enabled] = useState(shouldUseCustomWindowTitlebar);
   const [maximized, setMaximized] = useState(false);
   const { pinned, togglePinned } = useSidebarPin();
+  // The titlebar sits outside the sidebar wrapper, so it cannot inherit
+  // --sidebar-width. Read the resized width from the same store instead.
+  const { width } = useSidebarWidth();
   const sidebarWidth = showSidebarSurface
     ? pinned
-      ? "var(--studio-sidebar-expanded-width,17.5rem)"
+      ? // The live value only exists mid-drag; otherwise the committed width.
+        `var(--studio-sidebar-live-width, ${width}px)`
       : "var(--studio-sidebar-collapsed-width,3rem)"
     : "0px";
   const contentBorderLeft = pinned ? `calc(${sidebarWidth} + 12px)` : "0px";
diff --git a/studio/frontend/src/components/ui/panel-resize-handle.tsx b/studio/frontend/src/components/ui/panel-resize-handle.tsx
new file mode 100644
index 0000000000..9c1fd70b6f
--- /dev/null
+++ b/studio/frontend/src/components/ui/panel-resize-handle.tsx
@@ -0,0 +1,317 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"use client"
+
+import * as React from "react"
+
+import { cn } from "@/lib/utils"
+import {
+  Tooltip,
+  TooltipContent,
+  TooltipTrigger,
+} from "@/components/ui/tooltip"
+import { getClientPlatform } from "@/components/tauri/window-titlebar"
+
+/** Pointer travel (px) below which a drag counts as a plain click. */
+const DRAG_SLOP = 4
+/** A compatibility click lands immediately after pointer-up. */
+const CLICK_COMPAT_WINDOW_MS = 300
+/** Arrow-key resize step for keyboard users. */
+const RESIZE_STEP = 16
+
+type DragState = {
+  startX: number
+  startWidth: number
+  moved: boolean
+}
+
+export type PanelResizeHandleProps = {
+  /** Which edge of the panel the handle sits on. */
+  edge: "left" | "right"
+  open: boolean
+  width: number
+  /** Uncapped stored preference, so a capped drag does not lower it. */
+  stored: number
+  min: number
+  max: number
+  clamp: (px: number) => number
+  setWidth: (px: number) => void
+  resetWidth: () => void
+  onToggle: () => void
+  /** Element to paint the live width onto, and the property to paint. */
+  target: () => HTMLElement | null
+  cssVar: string
+  /** Measured to start a drag from the rendered size when collapsed. */
+  measure: () => number
+  label: string
+  toggleLabel: string
+  /** Translated tooltip copy; the caller owns the translation layer. */
+  collapseHint: string
+  expandHint: string
+  dragHint: string
+  /** Shown in the tooltip when the panel has a toggle shortcut. */
+  shortcut?: string
+  dataSlot?: string
+  className?: string
+  /** Mirrors the live width onto :root for chrome outside the panel. */
+  rootVar?: string
+}
+
+/**
+ * A draggable panel edge: drag to resize, click to collapse or expand. Arrow
+ * keys resize, Home restores the default. The width is painted straight to the
+ * target while dragging and only persisted on release.
+ */
+export function PanelResizeHandle({
+  edge,
+  open,
+  width,
+  stored,
+  min,
+  max,
+  clamp,
+  setWidth,
+  resetWidth,
+  onToggle,
+  target,
+  cssVar,
+  measure,
+  label,
+  toggleLabel,
+  collapseHint,
+  expandHint,
+  dragHint,
+  shortcut,
+  dataSlot = "panel-resize-handle",
+  className,
+  rootVar,
+}: PanelResizeHandleProps) {
+  const ref = React.useRef<HTMLButtonElement>(null)
+  const dragRef = React.useRef<DragState | null>(null)
+  const [dragging, setDragging] = React.useState(false)
+  const [hovered, setHovered] = React.useState(false)
+  const [focused, setFocused] = React.useState(false)
+  const [isMacPlatform] = React.useState(() => getClientPlatform().includes("mac"))
+  const hint = shortcut ? shortcut.replace("Mod", isMacPlatform ? "⌘" : "Ctrl+") : null
+
+  // Cached on pointer down so no DOM walk per move.
+  const targetRef = React.useRef<HTMLElement | null>(null)
+  const frameRef = React.useRef(0)
+  const pendingRef = React.useRef(0)
+  // What the pointer asked for, before the viewport cap. Committing the capped
+  // value instead would quietly downgrade a stored preference on a narrow window.
+  const rawRef = React.useRef(0)
+  // When a pointer sequence last ended. The browser's compatibility click
+  // lands in the same tick, so only a click that close behind is a duplicate.
+  // A timestamp cannot go stale the way an armed flag does: a genuine cancel
+  // emits no click, and a later assistive-tech click still gets through.
+  const handledAtRef = React.useRef(0)
+  const committedRef = React.useRef(width)
+  React.useEffect(() => {
+    committedRef.current = width
+  }, [width])
+
+  const paint = React.useCallback(
+    (value: string) => {
+      targetRef.current?.style.setProperty(cssVar, value)
+      if (rootVar) {
+        document.documentElement.style.setProperty(rootVar, value)
+      }
+    },
+    [cssVar, rootVar],
+  )
+
+  // Resizing relayouts the whole shell, and pointermove fires faster than the
+  // display refreshes, so coalesce to one paint per frame.
+  const paintWidth = React.useCallback(
+    (px: number) => {
+      pendingRef.current = px
+      if (frameRef.current) return
+      frameRef.current = requestAnimationFrame(() => {
+        frameRef.current = 0
+        paint(`${pendingRef.current}px`)
+      })
+    },
+    [paint],
+  )
+
+  const endDrag = React.useCallback(() => {
+    // Only a sequence that actually started can produce a compatibility click.
+    // This also runs as the effect cleanup, where no drag happened.
+    if (dragRef.current) handledAtRef.current = Date.now()
+    dragRef.current = null
+    if (frameRef.current) {
+      cancelAnimationFrame(frameRef.current)
+      frameRef.current = 0
+    }
+    // Hand the property back to the committed value. A commit re-renders with
+    // the new width; a cancel or a no-commit drag keeps DOM and store in step.
+    paint(`${committedRef.current}px`)
+    if (rootVar) document.documentElement.style.removeProperty(rootVar)
+    targetRef.current?.removeAttribute("data-resizing")
+    document.documentElement.removeAttribute("data-panel-resizing")
+    targetRef.current = null
+    setDragging(false)
+    document.body.style.removeProperty("cursor")
+    document.body.style.removeProperty("user-select")
+  }, [paint, rootVar])
+
+  const handlePointerDown = (event: React.PointerEvent<HTMLButtonElement>) => {
+    if (event.button !== 0) return
+    event.preventDefault()
+    event.currentTarget.setPointerCapture(event.pointerId)
+    targetRef.current = target()
+    // Collapsed: grow from the rendered size so the edge tracks the pointer.
+    const start = open ? width : measure()
+    dragRef.current = { startX: event.clientX, startWidth: start, moved: false }
+    pendingRef.current = start
+    rawRef.current = start
+    targetRef.current?.setAttribute("data-resizing", "true")
+    document.documentElement.setAttribute("data-panel-resizing", "true")
+    setDragging(true)
+    document.body.style.setProperty("cursor", "col-resize")
+    document.body.style.setProperty("user-select", "none")
+  }
+
+  const handlePointerMove = (event: React.PointerEvent<HTMLButtonElement>) => {
+    const drag = dragRef.current
+    if (!drag) return
+    // A panel whose handle is on its left edge grows as the pointer moves left.
+    const delta = (edge === "left" ? -1 : 1) * (event.clientX - drag.startX)
+    if (!drag.moved && Math.abs(delta) < DRAG_SLOP) return
+    drag.moved = true
+
+    const next = drag.startWidth + delta
+    rawRef.current = next
+    if (!open) {
+      // Past the minimum, dragging the collapsed edge reopens it.
+      if (next >= min) {
+        paintWidth(clamp(next))
+        onToggle()
+      }
+      return
+    }
+    // Dragging inward stops at the minimum. Collapsing is click or the shortcut.
+    paintWidth(clamp(next))
+  }
+
+  const handlePointerUp = (event: React.PointerEvent<HTMLButtonElement>) => {
+    const drag = dragRef.current
+    if (!drag) return
+    if (event.currentTarget.hasPointerCapture(event.pointerId)) {
+      event.currentTarget.releasePointerCapture(event.pointerId)
+    }
+    endDrag()
+
+    if (!drag.moved) {
+      onToggle()
+      return
+    }
+    // A drag below the minimum leaves the stored width alone.
+    if (!open) return
+    // Capped: the visible edge is already at the cap, so an outward pull cannot
+    // express intent beyond it. Committing would silently lower the larger
+    // hidden preference. A deliberate inward drag still commits.
+    if (stored > max && rawRef.current >= max) return
+    // Commit what was asked for, not the capped paint, so a drag on a narrow
+    // window cannot shrink a larger stored preference. setWidth clamps.
+    setWidth(rawRef.current)
+  }
+
+  const handleKeyDown = (event: React.KeyboardEvent<HTMLButtonElement>) => {
+    // The collapse/expand the label advertises, for keyboard users. Pointer-up
+    // handles it for the mouse; a synthesized click never reaches it.
+    if (event.key === "Enter" || event.key === " ") {
+      // preventDefault cancels the native click, so nothing follows to guard
+      // against; arming here would swallow the next assistive-tech click.
+      event.preventDefault()
+      onToggle()
+      return
+    }
+    const outward = edge === "left" ? "ArrowLeft" : "ArrowRight"
+    const inward = edge === "left" ? "ArrowRight" : "ArrowLeft"
+    if (event.key === outward || event.key === inward) {
+      event.preventDefault()
+      if (!open) {
+        // Collapsed there is nothing to resize, so the outward arrow reopens.
+        if (event.key === outward) onToggle()
+        return
+      }
+      if (event.key === outward && stored > max && width >= max) return
+      setWidth(width + (event.key === outward ? RESIZE_STEP : -RESIZE_STEP))
+      return
+    }
+    if (event.key === "Home") {
+      event.preventDefault()
+      resetWidth()
+    }
+  }
+
+  // Clear a stuck cursor override if we unmount mid-drag.
+  React.useEffect(() => endDrag, [endDrag])
+
+  return (
+    <Tooltip open={(hovered || focused) && !dragging}>
+      <TooltipTrigger asChild>
+        <button
+          ref={ref}
+          type="button"
+          data-slot={dataSlot}
+          data-dragging={dragging || undefined}
+          aria-label={open ? label : toggleLabel}
+          {...(open ? { "aria-orientation": "vertical" as const } : {})}
+          {...(open
+            ? { "aria-valuenow": width, "aria-valuemin": min, "aria-valuemax": max }
+            : {})}
+          role={open ? "separator" : "button"}
+          onPointerDown={handlePointerDown}
+          onPointerMove={handlePointerMove}
+          onPointerUp={handlePointerUp}
+          onPointerCancel={endDrag}
+          onKeyDown={handleKeyDown}
+          onClick={() => {
+            // Switch and voice control activate by dispatching a bare click
+            // with no pointer or key events, which nothing else here catches.
+            if (Date.now() - handledAtRef.current < CLICK_COMPAT_WINDOW_MS) return
+            onToggle()
+          }}
+          onPointerEnter={() => setHovered(true)}
+          onPointerLeave={() => setHovered(false)}
+          onFocus={(event) => setFocused(event.target.matches(":focus-visible"))}
+          onBlur={() => setFocused(false)}
+          className={cn(
+            "absolute inset-y-0 z-30 hidden w-2 touch-none select-none sm:block",
+            edge === "left" ? "-left-1" : "-right-1",
+            // `!` overrides the app-wide hand cursor on buttons.
+            open
+              ? "cursor-col-resize!"
+              : edge === "left"
+                ? "cursor-w-resize!"
+                : "cursor-e-resize!",
+            // Sits exactly on the panel border so hover recolours one line.
+            "after:absolute after:inset-y-0 after:w-px after:bg-transparent after:transition-colors after:duration-150",
+            edge === "left" ? "after:left-1" : "after:right-1",
+            "hover:after:bg-sidebar-ring/25 data-dragging:after:bg-sidebar-ring/25",
+            // The app zeroes the native outline on buttons, so mark focus here.
+            "focus-visible:outline-none focus-visible:after:bg-sidebar-ring/60",
+            className,
+          )}
+        />
+      </TooltipTrigger>
+      <TooltipContent
+        side={edge === "left" ? "left" : "right"}
+        align="center"
+        className="tooltip-compact"
+      >
+        <span className="flex flex-col gap-px">
+          <span>
+            {open ? collapseHint : expandHint}
+            {hint ? ` ${hint}` : ""}
+          </span>
+          <span className="opacity-70">{dragHint}</span>
+        </span>
+      </TooltipContent>
+    </Tooltip>
+  )
+}
diff --git a/studio/frontend/src/components/ui/sidebar.tsx b/studio/frontend/src/components/ui/sidebar.tsx
index 0fe82eb428..e26a55694f 100644
--- a/studio/frontend/src/components/ui/sidebar.tsx
+++ b/studio/frontend/src/components/ui/sidebar.tsx
@@ -24,13 +24,21 @@ import {
   TooltipContent,
   TooltipTrigger,
 } from "@/components/ui/tooltip"
+import { PanelResizeHandle } from "@/components/ui/panel-resize-handle"
+import { useT } from "@/i18n"
 import { useIsMobile } from "@/hooks/use-mobile"
+import {
+  SIDEBAR_WIDTH_DEFAULT,
+  SIDEBAR_WIDTH_MIN,
+  clampSidebarWidth,
+  useSidebarWidth,
+} from "@/hooks/use-sidebar-width"
 import { HugeiconsIcon } from "@hugeicons/react"
 import { LayoutAlignLeftIcon } from "@hugeicons/core-free-icons"
 
 const noop = () => {}
 
-const SIDEBAR_WIDTH = "17.5rem"
+const SIDEBAR_WIDTH = `${SIDEBAR_WIDTH_DEFAULT}px`
 const SIDEBAR_WIDTH_ICON = "3rem"
 const SIDEBAR_KEYBOARD_SHORTCUT = "b"
 
@@ -46,6 +54,11 @@ type SidebarContextProps = {
   pinned: boolean
   setPinned: (value: boolean) => void
   togglePinned: () => void
+  width: number
+  storedWidth: number
+  maxWidth: number
+  setWidth: (value: number) => void
+  resetWidth: () => void
 }
 
 const SidebarContext = React.createContext<SidebarContextProps | null>(null)
@@ -80,6 +93,13 @@ function SidebarProvider({
 }) {
   const isMobile = useIsMobile()
   const [openMobile, setOpenMobile] = React.useState(false)
+  const {
+    width,
+    max: maxWidth,
+    stored: storedWidth,
+    setWidth,
+    resetWidth,
+  } = useSidebarWidth()
 
   const prevIsMobileRef = React.useRef(isMobile)
   React.useEffect(() => {
@@ -163,8 +183,13 @@ function SidebarProvider({
       pinned,
       setPinned,
       togglePinned,
+      width,
+      storedWidth,
+      maxWidth,
+      setWidth,
+      resetWidth,
     }),
-    [state, open, setOpen, isMobile, openMobile, setOpenMobile, toggleSidebar, hasPinMode, pinned, setPinned, togglePinned]
+    [state, open, setOpen, isMobile, openMobile, setOpenMobile, toggleSidebar, hasPinMode, pinned, setPinned, togglePinned, width, storedWidth, maxWidth, setWidth, resetWidth]
   )
 
   return (
@@ -173,7 +198,8 @@ function SidebarProvider({
         data-slot="sidebar-wrapper"
         style={
           {
-            "--sidebar-width": SIDEBAR_WIDTH,
+            // The drag handle writes this same property live while resizing.
+            "--sidebar-width": `${width}px`,
             "--sidebar-width-icon": SIDEBAR_WIDTH_ICON,
             ...style,
           } as React.CSSProperties
@@ -311,11 +337,64 @@ function Sidebar({
         >
           {children}
         </div>
+        <SidebarResizeHandle side={side} />
       </div>
     </div>
   )
 }
 
+/**
+ * The sidebar's draggable edge, over the shared panel handle.
+ */
+function SidebarResizeHandle({
+  className,
+  side = "left",
+}: {
+  className?: string
+  side?: "left" | "right"
+}) {
+  const { open, toggleSidebar, width, storedWidth, maxWidth, setWidth, resetWidth } =
+    useSidebar()
+  const ref = React.useRef<HTMLDivElement>(null)
+  const t = useT()
+
+  return (
+    <div ref={ref} className="contents">
+      <PanelResizeHandle
+        edge={side === "right" ? "left" : "right"}
+        open={open}
+        width={width}
+        stored={storedWidth}
+        min={SIDEBAR_WIDTH_MIN}
+        max={maxWidth}
+        clamp={clampSidebarWidth}
+        setWidth={setWidth}
+        resetWidth={resetWidth}
+        onToggle={toggleSidebar}
+        target={() =>
+          ref.current?.closest<HTMLElement>('[data-slot="sidebar-wrapper"]') ?? null
+        }
+        cssVar="--sidebar-width"
+        // The custom titlebar renders outside the wrapper and cannot inherit it.
+        rootVar="--studio-sidebar-live-width"
+        measure={() =>
+          ref.current
+            ?.closest<HTMLElement>('[data-slot="sidebar-container"]')
+            ?.getBoundingClientRect().width ?? SIDEBAR_WIDTH_MIN
+        }
+        label={t("shell.aria.resizeSidebar")}
+        toggleLabel={t("shell.aria.openSidebar")}
+        collapseHint={t("shell.resize.collapse")}
+        expandHint={t("shell.resize.expand")}
+        dragHint={t("shell.resize.drag")}
+        shortcut="ModB"
+        dataSlot="sidebar-resize-handle"
+        className={className}
+      />
+    </div>
+  )
+}
+
 function SidebarTrigger({
   className,
   onClick,
@@ -777,6 +856,7 @@ export {
   SidebarMenuSubItem,
   SidebarProvider,
   SidebarRail,
+  SidebarResizeHandle,
   SidebarSeparator,
   SidebarTrigger,
   useSidebar,
diff --git a/studio/frontend/src/components/update/release-notes-panel.tsx b/studio/frontend/src/components/update/release-notes-panel.tsx
new file mode 100644
index 0000000000..d98c855daa
--- /dev/null
+++ b/studio/frontend/src/components/update/release-notes-panel.tsx
@@ -0,0 +1,251 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { MarkdownPreview } from "@/components/markdown/markdown-preview";
+import { useReleaseNotes } from "@/hooks/use-release-notes";
+import { resolveChangelogLinks } from "@/lib/changelog-links";
+import { releaseNotesPreview } from "@/lib/release-notes-preview";
+import { cn } from "@/lib/utils";
+import {
+  type ReactElement,
+  type ReactNode,
+  useEffect,
+  useMemo,
+  useRef,
+} from "react";
+
+interface ReleaseNotesPanelProps {
+  // Notes are looked up for this exact version only.
+  version: string;
+  // Collapsed previews the top bullets; expanded scrolls the full notes.
+  open: boolean;
+  // Desktop updater's body, used only if CHANGELOG.md has no section here.
+  fallbackMarkdown?: string | null;
+  releaseNotesUrl?: string | null;
+  className?: string;
+}
+
+const NOTES_LINK_CLASS =
+  "shrink-0 whitespace-nowrap text-ui-11 font-medium text-foreground underline underline-offset-2";
+
+function NotesMessage({
+  children,
+  action,
+}: {
+  children: ReactNode;
+  action?: ReactNode;
+}): ReactElement {
+  return (
+    <div className="flex items-center justify-between gap-2 px-1 py-2">
+      <p className="text-ui-11 text-muted-foreground">{children}</p>
+      {action}
+    </div>
+  );
+}
+
+function ChangelogLink({ href }: { href: string }): ReactElement {
+  return (
+    <a
+      href={href}
+      target="_blank"
+      rel="noopener noreferrer"
+      className={NOTES_LINK_CLASS}
+      data-testid="update-release-notes-link"
+    >
+      Open changelog
+    </a>
+  );
+}
+
+export function ReleaseNotesPanel({
+  version,
+  open,
+  fallbackMarkdown = null,
+  releaseNotesUrl = null,
+  className,
+}: ReleaseNotesPanelProps): ReactElement | null {
+  // Fetched with the popup: the collapsed preview needs the notes too.
+  const { state, notes, retry } = useReleaseNotes({ version, enabled: true });
+  const scrollRef = useRef<HTMLElement | null>(null);
+
+  // The fallback stands in for "no section in the changelog", which the hook
+  // reports as ready. An error is retryable, and the desktop fallback is the
+  // updater's static blurb, so taking it there would hide Retry until cache expiry.
+  const source = notes?.matched
+    ? notes.markdown
+    : state === "error"
+      ? null
+      : (fallbackMarkdown ?? null);
+  // Notes target the repository, so relative links must point back at it.
+  const markdown = useMemo(
+    () => (source === null ? null : resolveChangelogLinks(source)),
+    [source],
+  );
+
+  // Notes that are only a code block or a table preview as nothing.
+  const preview = useMemo(
+    () => (markdown === null ? null : releaseNotesPreview(markdown)),
+    [markdown],
+  );
+
+  // Start at the top on expand, and again once async notes land.
+  useEffect(() => {
+    if (open && markdown && scrollRef.current) {
+      scrollRef.current.scrollTop = 0;
+    }
+  }, [open, markdown]);
+
+  // Caller's URL wins: the API returns only the generic changelog, while the
+  // desktop banner passes this version's release page.
+  const notesUrl = releaseNotesUrl ?? notes?.releaseNotesUrl;
+  const link = notesUrl ? <ChangelogLink href={notesUrl} /> : null;
+
+  // Nothing previewable yet or ever: keep the collapsed popup compact.
+  if (
+    !open &&
+    (!markdown ||
+      state === "loading" ||
+      state === "idle" ||
+      preview?.items.length === 0)
+  ) {
+    return null;
+  }
+
+  return (
+    <div
+      className={cn("mt-3 flex min-h-0 flex-col", className)}
+      data-testid="update-release-notes-panel"
+      data-notes-state={state}
+      data-notes-version={version}
+      data-notes-open={open}
+    >
+      {/* borderless fill, lighter than the card in dark mode */}
+      <div className="flex min-h-0 flex-col rounded-[14px] bg-muted/40 px-3 py-1 dark:bg-white/[0.06]">
+        {markdown ? (
+          open ? (
+            <section
+              ref={scrollRef}
+              // biome-ignore lint/a11y/noNoninteractiveTabindex: keyboard-scrollable region
+              tabIndex={0}
+              aria-label={`Release notes for version ${version}`}
+              // Long notes scroll here instead of pushing the buttons off screen.
+              className="hover-scrollbar max-h-64 min-h-0 flex-1 overflow-y-auto overscroll-contain py-3 pr-1"
+              data-testid="update-release-notes-scroll"
+            >
+              <MarkdownPreview
+                markdown={markdown}
+                // Streamdown ships headings at mt-6 and code at text-sm, and
+                // clears max-width on descendants, so rescale and re-cap both.
+                className="max-h-none overflow-visible border-0 bg-transparent p-0 text-ui-11 [&_[data-streamdown=link-safety-modal]>*]:max-w-md [&_img]:h-auto [&_img]:max-w-full [&>*:first-child]:mt-0 [&>*>*:first-child]:mt-0 [&_code]:text-[0.92em] [&_h1]:mt-4 [&_h1]:font-heading [&_h1]:text-ui-13 [&_h2]:mt-4 [&_h2]:font-heading [&_h2]:text-ui-13 [&_h3]:mt-4 [&_h3]:font-heading [&_h3]:text-ui-11 [&_pre]:text-[0.92em]"
+              />
+              {notes?.truncated ? (
+                <p className="mt-2 text-ui-10 text-muted-foreground/80">
+                  Notes truncated. See the full changelog.
+                </p>
+              ) : null}
+            </section>
+          ) : (
+            <ReleaseNotesSummary preview={preview} />
+          )
+        ) : (
+          <NotesStatus
+            state={state}
+            version={version}
+            link={link}
+            retry={retry}
+          />
+        )}
+      </div>
+      {open && markdown && link ? (
+        <div className="mt-2 flex justify-end px-1">{link}</div>
+      ) : null}
+    </div>
+  );
+}
+
+/** Collapsed view: the first few bullets, one line each where possible. */
+function ReleaseNotesSummary({
+  preview,
+}: {
+  preview: ReturnType<typeof releaseNotesPreview> | null;
+}): ReactElement | null {
+  if (preview === null || preview.items.length === 0) {
+    return null;
+  }
+  const { items, remaining } = preview;
+
+  return (
+    <ul
+      className="space-y-1 py-2 pr-1"
+      data-testid="update-release-notes-summary"
+    >
+      {items.map((item, index) => (
+        <li
+          // Two releases can carry the same bullet text, so index is the key.
+          key={`${index}-${item.lead}`}
+          className="flex gap-1.5 text-ui-11 leading-snug text-muted-foreground"
+        >
+          <span aria-hidden="true" className="text-muted-foreground/60">
+            &bull;
+          </span>
+          <span className="line-clamp-2 min-w-0">
+            {/* lead sentence carries the change */}
+            <span className="font-medium text-foreground">{item.lead}</span>
+            {item.rest ? <span> {item.rest}</span> : null}
+          </span>
+        </li>
+      ))}
+      {remaining > 0 ? (
+        <li className="pl-3 text-ui-10 text-muted-foreground/70">
+          +{remaining} more
+        </li>
+      ) : null}
+    </ul>
+  );
+}
+
+function NotesStatus({
+  state,
+  version,
+  link,
+  retry,
+}: {
+  state: ReturnType<typeof useReleaseNotes>["state"];
+  version: string;
+  link: ReactNode;
+  retry: () => void;
+}): ReactElement {
+  if (state === "loading" || state === "idle") {
+    return <NotesMessage>Loading release notes...</NotesMessage>;
+  }
+
+  if (state === "error") {
+    return (
+      <NotesMessage
+        action={
+          // The changelog page may be reachable when the lookup is not.
+          <span className="flex shrink-0 items-center gap-3">
+            <button
+              type="button"
+              onClick={retry}
+              className={NOTES_LINK_CLASS}
+              data-testid="update-release-notes-retry"
+            >
+              Retry
+            </button>
+            {link}
+          </span>
+        }
+      >
+        Could not load release notes.
+      </NotesMessage>
+    );
+  }
+
+  // Matched nothing: link out rather than show another release's notes.
+  return (
+    <NotesMessage action={link}>
+      No release notes published for {version} yet.
+    </NotesMessage>
+  );
+}
diff --git a/studio/frontend/src/components/web/update-banner.tsx b/studio/frontend/src/components/web/update-banner.tsx
index d8ae92bf5f..f36f5ec3cd 100644
--- a/studio/frontend/src/components/web/update-banner.tsx
+++ b/studio/frontend/src/components/web/update-banner.tsx
@@ -2,6 +2,7 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { Button } from "@/components/ui/button";
+import { ReleaseNotesPanel } from "@/components/update/release-notes-panel";
 import { type DeviceType, usePlatformStore } from "@/config/env";
 import { useWebUpdateCheck } from "@/hooks/use-web-update-check";
 import { isTauri } from "@/lib/api-base";
@@ -40,6 +41,7 @@ export function WebUpdateBanner({
   const deviceType = usePlatformStore((s) => s.deviceType);
   const installCmd = installCommandForDevice(deviceType);
   const [copiedVersion, setCopiedVersion] = useState<string | null>(null);
+  const [notesVersion, setNotesVersion] = useState<string | null>(null);
   const dismissTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null);
 
   useEffect(() => {
@@ -68,6 +70,8 @@ export function WebUpdateBanner({
   }
 
   const copied = status != null && copiedVersion === status.latestVersion;
+  // Keyed by version so a new offer collapses the panel.
+  const notesOpen = status != null && notesVersion === status.latestVersion;
 
   return (
     <AnimatePresence>
@@ -78,13 +82,14 @@ export function WebUpdateBanner({
           exit={{ opacity: 0, y: 8, scale: 0.97 }}
           transition={{ duration: 0.35, ease: EASE_OUT_QUART }}
           className={cn(
+            // Wider than the other overlays: notes preview plus three buttons.
             positioned
-              ? "fixed bottom-4 right-4 z-[9999] w-[calc(100vw-2rem)] max-w-[400px]"
-              : "pointer-events-auto w-full",
+              ? "fixed bottom-4 right-4 z-[9999] w-[calc(100vw-2rem)] max-w-[448px]"
+              : "pointer-events-auto flex min-h-0 w-[calc(100vw-2rem)] max-w-[448px] flex-col",
           )}
           data-testid="web-update-banner"
         >
-          <div className="relative overflow-hidden rounded-[24px] bg-white px-5 pb-4 pt-5 shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:bg-card dark:shadow-[0_8px_28px_-6px_rgba(0,0,0,0.28)]">
+          <div className="relative flex max-h-[calc(100dvh_-_2rem)] flex-col overflow-hidden rounded-[24px] bg-white px-5 pb-4 pt-5 shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:bg-card dark:shadow-[0_8px_28px_-6px_rgba(0,0,0,0.28)]">
             <button
               type="button"
               onClick={dismiss}
@@ -127,22 +132,33 @@ export function WebUpdateBanner({
               </div>
             </div>
 
+            <ReleaseNotesPanel
+              version={status.latestVersion}
+              open={notesOpen}
+              releaseNotesUrl={RELEASE_NOTES_URL}
+              className="min-h-0 flex-1"
+            />
+
+            {/* one row at one type size; wraps only on narrow viewports */}
             <div className="mt-4 flex flex-wrap items-center justify-between gap-y-2">
-              <a
-                href={RELEASE_NOTES_URL}
-                target="_blank"
-                rel="noopener noreferrer"
-                className="-ml-2 whitespace-nowrap rounded-full px-2.5 py-2 text-ui-13 font-medium text-foreground transition-colors hover:bg-muted"
-                data-testid="web-update-release-notes-link"
+              <Button
+                size="sm"
+                variant="ghost"
+                className="-ml-2 h-auto whitespace-nowrap rounded-full px-2.5 py-2 text-ui-13 font-medium text-foreground"
+                onClick={() =>
+                  setNotesVersion(notesOpen ? null : status.latestVersion)
+                }
+                aria-expanded={notesOpen}
+                data-testid="web-update-release-notes-toggle"
               >
-                Release notes
-              </a>
+                {notesOpen ? "Hide release notes" : "Show release notes"}
+              </Button>
               {/* wrap + right-align so buttons stack instead of clipping on very narrow banners */}
               <div className="flex flex-wrap items-center justify-end gap-x-1 gap-y-2">
                 <Button
                   size="sm"
                   variant="ghost"
-                  className="h-auto rounded-full px-3 py-2 text-ui-13 font-medium text-foreground"
+                  className="h-auto whitespace-nowrap rounded-full px-2.5 py-2 text-ui-13 font-medium text-foreground"
                   onClick={snooze}
                   data-testid="web-update-snooze-button"
                 >
@@ -151,7 +167,7 @@ export function WebUpdateBanner({
                 <Button
                   size="sm"
                   // -mr optically aligns the filled pill's edge with the card padding
-                  className="-mr-1 h-auto rounded-full px-3.5 py-2 text-ui-13"
+                  className="-mr-1 h-auto whitespace-nowrap rounded-full px-3 py-2 text-ui-13"
                   onClick={handleCopyCommand}
                   data-testid="web-update-copy-button"
                 >
diff --git a/studio/frontend/src/features/chat/chat-providers-dialog.tsx b/studio/frontend/src/features/chat/chat-providers-dialog.tsx
index 7e0cebea4b..4ab560e5ad 100644
--- a/studio/frontend/src/features/chat/chat-providers-dialog.tsx
+++ b/studio/frontend/src/features/chat/chat-providers-dialog.tsx
@@ -1454,7 +1454,7 @@ export function ChatProvidersSettings({
       </div>
 
       <section className="flex max-w-[760px] flex-col gap-2">
-        <div className="overflow-hidden rounded-[14px] border border-border/70 bg-muted/[0.12]">
+        <div className="overflow-hidden rounded-[10px] border border-border/70 bg-muted/[0.12]">
           <button
             type="button"
             onClick={openAddProvider}
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index 6070bd2e40..fd41e558f9 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -18,6 +18,7 @@ import {
   DropdownMenuTrigger,
 } from "@/components/ui/dropdown-menu";
 import { InfoHint } from "@/components/ui/info-hint";
+import { PanelResizeHandle } from "@/components/ui/panel-resize-handle";
 import {
   InputGroup,
   InputGroupAddon,
@@ -44,7 +45,13 @@ import { Tooltip, TooltipContent } from "@/components/ui/tooltip";
 import { NumericValueInput, snapToStep } from "@/features/model-picker";
 import { RetrievalSettingsSection } from "@/features/rag";
 import { useLlamaUpdateCheck } from "@/hooks/use-llama-update-check";
+import {
+  CHAT_SETTINGS_WIDTH_MIN,
+  clampChatSettingsWidth,
+  useChatSettingsWidth,
+} from "@/hooks/use-chat-settings-width";
 import { useIsMobile } from "@/hooks/use-mobile";
+import { useT } from "@/i18n";
 import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { toast } from "@/lib/toast";
 import { cn } from "@/lib/utils";
@@ -52,7 +59,7 @@ import { Edit03Icon, LayoutAlignRightIcon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { Braces, ChevronDown, ExternalLink } from "lucide-react";
 import { Tooltip as TooltipPrimitive } from "radix-ui";
-import { Fragment, type ReactNode } from "react";
+import { type CSSProperties, Fragment, type ReactNode } from "react";
 import { useCallback, useEffect, useMemo, useRef, useState } from "react";
 import { OpenAICodeExecSection } from "./components/openai-code-exec-section";
 import { PermissionModeDropdown } from "./permission-mode-select";
@@ -363,6 +370,15 @@ export function ChatSettingsPanel({
   onExternalProviderChange,
   externalProviderType = null,
 }: ChatSettingsPanelProps) {
+  const asideRef = useRef<HTMLElement>(null);
+  const t = useT();
+  const {
+    width: settingsWidth,
+    max: settingsMax,
+    stored: settingsStored,
+    setWidth: setSettingsWidth,
+    resetWidth: resetSettingsWidth,
+  } = useChatSettingsWidth();
   // Local models show every knob; providerCapabilities is only consulted when
   // isExternalModel. Unknown providers fall back to the OpenAI-compat shape via
   // getProviderCapabilities, so these flags never undercount support.
@@ -461,6 +477,23 @@ export function ChatSettingsPanel({
   // When the prompt overflows the inline box, clicking opens the popup editor.
   const systemPromptBoxRef = useRef<HTMLTextAreaElement>(null);
   const [systemPromptOverflows, setSystemPromptOverflows] = useState(false);
+  const promptObserverRef = useRef<ResizeObserver | null>(null);
+  const measurePromptRef = useRef<() => void>(() => {});
+  // The section unmounts its textarea when collapsed, so observe through a
+  // callback ref: a stored observer would cling to the detached node and the
+  // remounted one would never be measured.
+  const attachPromptBox = useCallback((node: HTMLTextAreaElement | null) => {
+    systemPromptBoxRef.current = node;
+    promptObserverRef.current?.disconnect();
+    promptObserverRef.current = null;
+    if (!node || typeof ResizeObserver === "undefined") return;
+    // Resizing rewraps the prompt, and a drag changes the width through a
+    // custom property without re-rendering, so watch the box itself.
+    const observer = new ResizeObserver(() => measurePromptRef.current());
+    observer.observe(node);
+    promptObserverRef.current = observer;
+    measurePromptRef.current();
+  }, []);
   const [activePresetBaseline, setActivePresetBaseline] = useState(params);
   const presets = useMemo(() => {
     return getOrderedPresets(customPresets);
@@ -746,15 +779,20 @@ export function ChatSettingsPanel({
   }, [open]);
 
   useEffect(() => {
-    const el = systemPromptBoxRef.current;
-    setSystemPromptOverflows(
-      currentSystemPrompt.length > 0 &&
-        el != null &&
-        el.clientHeight > 0 &&
-        el.scrollHeight > el.clientHeight + 1,
-    );
+    measurePromptRef.current = () => {
+      const el = systemPromptBoxRef.current;
+      setSystemPromptOverflows(
+        currentSystemPrompt.length > 0 &&
+          el != null &&
+          el.clientHeight > 0 &&
+          el.scrollHeight > el.clientHeight + 1,
+      );
+    };
+    measurePromptRef.current();
   }, [currentSystemPrompt, open]);
 
+  useEffect(() => () => promptObserverRef.current?.disconnect(), []);
+
   const settingsScrollRef = useRef<HTMLDivElement>(null);
 
   const settingsContent = (
@@ -1124,7 +1162,7 @@ export function ChatSettingsPanel({
             )}
           >
             <textarea
-              ref={systemPromptBoxRef}
+              ref={attachPromptBox}
               value={currentSystemPrompt}
               onChange={(e) => set("systemPrompt")(e.target.value)}
               onMouseDown={(e) => {
@@ -1433,17 +1471,47 @@ export function ChatSettingsPanel({
 
   return (
     <aside
+      ref={asideRef}
       data-tour="chat-settings"
+      data-slot="chat-settings-panel"
       className={cn(
-        "relative z-50 shrink-0 overflow-hidden bg-panel-surface text-panel-surface-fg font-heading",
-        open ? "w-[17rem] border-l border-sidebar-border" : "w-0",
+        "relative z-50 shrink-0 bg-panel-surface text-panel-surface-fg font-heading",
+        open
+          ? "w-(--chat-settings-width) border-l border-sidebar-border"
+          : "w-0 overflow-hidden",
       )}
-      style={{
-        height: "calc(100% - var(--studio-custom-titlebar-height, 0px))",
-        marginTop: "var(--studio-custom-titlebar-height, 0px)",
-      }}
+      style={
+        {
+          "--chat-settings-width": `${settingsWidth}px`,
+          height: "calc(100% - var(--studio-custom-titlebar-height, 0px))",
+          marginTop: "var(--studio-custom-titlebar-height, 0px)",
+        } as CSSProperties
+      }
     >
-      <div className="h-full w-full">{settingsContent}</div>
+      {open ? (
+      <PanelResizeHandle
+        edge="left"
+        open={open}
+        width={settingsWidth}
+        stored={settingsStored}
+        min={CHAT_SETTINGS_WIDTH_MIN}
+        max={settingsMax}
+        clamp={clampChatSettingsWidth}
+        setWidth={setSettingsWidth}
+        resetWidth={resetSettingsWidth}
+        onToggle={() => onOpenChange?.(!open)}
+        target={() => asideRef.current}
+        cssVar="--chat-settings-width"
+        measure={() => asideRef.current?.getBoundingClientRect().width ?? 0}
+        label={t("shell.aria.resizeRunSettings")}
+        toggleLabel={t("shell.aria.openRunSettings")}
+        collapseHint={t("shell.resize.collapse")}
+        expandHint={t("shell.resize.expand")}
+        dragHint={t("shell.resize.drag")}
+        dataSlot="chat-settings-resize-handle"
+      />
+      ) : null}
+      <div className="h-full w-full overflow-hidden">{settingsContent}</div>
     </aside>
   );
 }
diff --git a/studio/frontend/src/features/chat/components/deep-research-composer-button.tsx b/studio/frontend/src/features/chat/components/deep-research-composer-button.tsx
index 03a7d7cc5f..e4857d1603 100644
--- a/studio/frontend/src/features/chat/components/deep-research-composer-button.tsx
+++ b/studio/frontend/src/features/chat/components/deep-research-composer-button.tsx
@@ -14,7 +14,8 @@ import {
 } from "@/components/ui/dialog";
 import { Input } from "@/components/ui/input";
 import { cn } from "@/lib/utils";
-import { ChevronDownIcon, XIcon } from "lucide-react";
+import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
+import { XIcon } from "lucide-react";
 import { type KeyboardEvent, useState } from "react";
 import { useChatRuntimeStore } from "../stores/chat-runtime-store";
 import type { ResearchWebsitePolicy } from "../types/research";
@@ -24,7 +25,12 @@ function normalizeDomain(raw: string): string | null {
   if (!value || /[\\\s]/.test(value)) return null;
   try {
     const url = new URL(value.includes("://") ? value : `https://${value}`);
-    if (!/^https?:$/.test(url.protocol) || url.username || url.password || url.port) {
+    if (
+      !/^https?:$/.test(url.protocol) ||
+      url.username ||
+      url.password ||
+      url.port
+    ) {
       return null;
     }
     return url.hostname
@@ -99,7 +105,9 @@ function DomainList({
               type="button"
               className="text-muted-foreground transition-colors hover:text-foreground"
               aria-label={`Remove ${domain}`}
-              onClick={() => onChange(values.filter((value) => value !== domain))}
+              onClick={() =>
+                onChange(values.filter((value) => value !== domain))
+              }
             >
               <XIcon className="size-3" />
             </button>
@@ -129,7 +137,9 @@ export function DeepResearchComposerButton({
   onConfigure: () => void;
 }) {
   const enabled = useChatRuntimeStore((state) => state.deepResearchEnabled);
-  const setEnabled = useChatRuntimeStore((state) => state.setDeepResearchEnabled);
+  const setEnabled = useChatRuntimeStore(
+    (state) => state.setDeepResearchEnabled,
+  );
 
   if (!enabled) return null;
 
@@ -158,9 +168,12 @@ export function DeepResearchComposerButton({
         <XIcon className="composer-pill-x" />
       </span>
       <span>Deep research</span>
-      <span className="composer-pill-caret flex items-center gap-0.5 text-primary/70">
-        <ChevronDownIcon className="size-3" />
-      </span>
+      {/* Same caret as the other composer pills, so the arrows match. */}
+      <HugeiconsIcon
+        icon={ChevronDownStandardIcon}
+        strokeWidth={1.5}
+        className="composer-pill-caret size-[15px] text-primary/70"
+      />
     </button>
   );
 }
@@ -173,7 +186,9 @@ export function DeepResearchWebsiteAccessDialog({
   onOpenChange: (open: boolean) => void;
 }) {
   const policy = useChatRuntimeStore((state) => state.researchWebsitePolicy);
-  const setPolicy = useChatRuntimeStore((state) => state.setResearchWebsitePolicy);
+  const setPolicy = useChatRuntimeStore(
+    (state) => state.setResearchWebsitePolicy,
+  );
 
   return (
     <Dialog open={open} onOpenChange={onOpenChange}>
@@ -201,41 +216,40 @@ function DeepResearchWebsiteAccessContent({
 
   return (
     <DialogContent className="sm:max-w-lg">
-        <DialogHeader>
-          <DialogTitle>Website access</DialogTitle>
-          <DialogDescription>
-            Control which websites the next Deep Research run can search and
-            read. Limits are enforced by the server and shared with the research
-            model.
-          </DialogDescription>
-        </DialogHeader>
-        <div className="space-y-6">
-          <DomainList
-            label="Allow only"
-            description="When set, research can access only these domains and their subdomains."
-            values={draft.allowedDomains}
-            onChange={(allowedDomains) => setDraft({ ...draft, allowedDomains })}
-          />
-          <DomainList
-            label="Always block"
-            description="These domains and their subdomains stay blocked. Blocking takes precedence."
-            values={draft.blockedDomains}
-            onChange={(blockedDomains) => setDraft({ ...draft, blockedDomains })}
-          />
-        </div>
-        <DialogFooter>
-          <Button variant="ghost" onClick={onClose}>
-            Cancel
-          </Button>
-          <Button
-            onClick={() => {
-              setPolicy(draft);
-              onClose();
-            }}
-          >
-            Save limits
-          </Button>
-        </DialogFooter>
+      <DialogHeader>
+        <DialogTitle>Website access</DialogTitle>
+        <DialogDescription>
+          Control which websites the next Deep Research run can search and read.
+          Limits are enforced by the server and shared with the research model.
+        </DialogDescription>
+      </DialogHeader>
+      <div className="space-y-6">
+        <DomainList
+          label="Allow only"
+          description="When set, research can access only these domains and their subdomains."
+          values={draft.allowedDomains}
+          onChange={(allowedDomains) => setDraft({ ...draft, allowedDomains })}
+        />
+        <DomainList
+          label="Always block"
+          description="These domains and their subdomains stay blocked. Blocking takes precedence."
+          values={draft.blockedDomains}
+          onChange={(blockedDomains) => setDraft({ ...draft, blockedDomains })}
+        />
+      </div>
+      <DialogFooter>
+        <Button variant="ghost" onClick={onClose}>
+          Cancel
+        </Button>
+        <Button
+          onClick={() => {
+            setPolicy(draft);
+            onClose();
+          }}
+        >
+          Save limits
+        </Button>
+      </DialogFooter>
     </DialogContent>
   );
 }
diff --git a/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx b/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx
index 2b0f2c3a8d..bef2f79747 100644
--- a/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx
+++ b/studio/frontend/src/features/hub/catalog/on-device-folders-dialog.tsx
@@ -23,12 +23,21 @@ import {
   removeScanFolder,
 } from "@/features/hub";
 import { FolderBrowser } from "@/features/model-picker";
-import { openModelsDir } from "@/features/native-intents";
+import {
+  openModelsDir,
+  pickHuggingFaceCacheDir,
+} from "@/features/native-intents";
+import {
+  type HuggingFaceCacheSettings,
+  loadHuggingFaceCacheSettings,
+  updateHuggingFaceCacheSettings,
+} from "@/features/settings";
 import { isTauri } from "@/lib/api-base";
 import { toast } from "@/lib/toast";
 import { cn } from "@/lib/utils";
 import {
   Delete02Icon,
+  DownloadCircle01Icon,
   FileSearchIcon,
   FolderAddIcon,
   FolderExportIcon,
@@ -49,6 +58,12 @@ function formatError(error: unknown): string {
   return error instanceof Error ? error.message : String(error);
 }
 
+function formatFreeSpace(bytes: number | null): string | null {
+  if (bytes === null || !Number.isFinite(bytes)) return null;
+  const gb = bytes / 1024 ** 3;
+  return gb >= 10 ? `${Math.round(gb)} GB free` : `${gb.toFixed(1)} GB free`;
+}
+
 export function OnDeviceFoldersDialog({
   open,
   onOpenChange,
@@ -68,6 +83,11 @@ export function OnDeviceFoldersDialog({
   );
   const refreshIdRef = useRef(0);
   const mutationVersionRef = useRef(0);
+  const [downloadCache, setDownloadCache] =
+    useState<HuggingFaceCacheSettings | null>(null);
+  const [downloadCacheLoaded, setDownloadCacheLoaded] = useState(false);
+  const [downloadBrowserOpen, setDownloadBrowserOpen] = useState(false);
+  const [downloadSaving, setDownloadSaving] = useState(false);
 
   const sortedFolders = useMemo(
     () => [...folders].sort((a, b) => a.path.localeCompare(b.path)),
@@ -108,10 +128,66 @@ export function OnDeviceFoldersDialog({
     return () => window.clearTimeout(timer);
   }, [open, refreshFolders]);
 
+  useEffect(() => {
+    if (!open) return;
+    let cancelled = false;
+    // The dialog stays mounted between opens, so re-arm the flag or a reopen
+    // shows the previous answer as if it were fresh.
+    setDownloadCacheLoaded(false);
+    loadHuggingFaceCacheSettings()
+      // Indexed locations do not depend on this. Null drops the stale path
+      // rather than offer Change against a location we could not confirm.
+      .catch(() => null)
+      .then((settings) => {
+        if (cancelled) return;
+        setDownloadCache(settings);
+        setDownloadCacheLoaded(true);
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, [open]);
+
   const handleInventoryChanged = useCallback(() => {
     onInventoryChange?.();
   }, [onInventoryChange]);
 
+  // Relocating the cache changes which repos are on disk, but
+  // updateHuggingFaceCacheSettings already bumps the inventory version, which
+  // re-fetches every source. Refreshing here too would scan twice, since the
+  // two rounds carry different version keys and cannot be deduplicated.
+  const saveDownloadLocation = useCallback(async (nextPath: string | null) => {
+    setDownloadSaving(true);
+    try {
+      const settings = await updateHuggingFaceCacheSettings(nextPath);
+      setDownloadCache(settings);
+      toast.success("Download location updated", {
+        description: settings.cacheHome,
+      });
+    } catch (err) {
+      toast.error("Couldn't update the download location", {
+        description: formatError(err),
+      });
+    } finally {
+      setDownloadSaving(false);
+    }
+  }, []);
+
+  const changeDownloadLocation = useCallback(async () => {
+    if (!isTauri) {
+      setDownloadBrowserOpen(true);
+      return;
+    }
+    try {
+      const picked = await pickHuggingFaceCacheDir();
+      if (picked) await saveDownloadLocation(picked);
+    } catch (err) {
+      toast.error("Couldn't open the folder picker", {
+        description: formatError(err),
+      });
+    }
+  }, [saveDownloadLocation]);
+
   const handleAdd = useCallback(
     async (rawPath: string) => {
       const nextPath = rawPath.trim();
@@ -182,10 +258,10 @@ export function OnDeviceFoldersDialog({
     <>
       <Dialog open={open} onOpenChange={onOpenChange}>
         <DialogContent
-          className="gap-0 overflow-hidden p-0 sm:max-w-[620px] lg:max-w-[660px] xl:max-w-[680px] [&_[data-slot=dialog-close]]:right-3 [&_[data-slot=dialog-close]]:top-3"
+          className="flex max-h-[90dvh] flex-col gap-0 overflow-hidden p-0 sm:max-w-[620px] lg:max-w-[660px] xl:max-w-[680px] [&_[data-slot=dialog-close]]:right-3 [&_[data-slot=dialog-close]]:top-3"
           overlayClassName="bg-black/20 backdrop-blur-none"
         >
-          <DialogHeader className="border-b border-border/60 px-5 py-4">
+          <DialogHeader className="shrink-0 border-b border-border/60 px-5 py-4">
             <DialogTitle className="text-ui-15">
               On-device locations
             </DialogTitle>
@@ -195,7 +271,78 @@ export function OnDeviceFoldersDialog({
             </DialogDescription>
           </DialogHeader>
 
-          <div className="space-y-4 px-5 py-4">
+          <div className="min-h-0 flex-1 space-y-4 overflow-y-auto px-5 py-4">
+            <div className="rounded-[14px] border border-border/70 bg-muted/20 p-3">
+              <div className="mb-2 flex items-center gap-2 text-ui-12 font-medium text-foreground">
+                <HugeiconsIcon
+                  icon={DownloadCircle01Icon}
+                  strokeWidth={1.75}
+                  className="size-3.5 text-muted-foreground"
+                />
+                Download location
+              </div>
+
+              <div className="flex flex-col gap-2 sm:flex-row sm:items-center">
+                <Input
+                  readOnly={true}
+                  aria-label="Model download location"
+                  value={
+                    downloadCache?.cacheHome ??
+                    (downloadCacheLoaded ? "Unknown" : "Loading...")
+                  }
+                  title={downloadCache?.cacheHome}
+                  className="field-soft h-9 min-w-0 flex-1 rounded-full px-3 font-mono text-ui-12"
+                />
+                <div className="flex shrink-0 items-center gap-2">
+                  <Button
+                    type="button"
+                    variant="outline"
+                    size="sm"
+                    onClick={() => void changeDownloadLocation()}
+                    disabled={!downloadCache?.editable || downloadSaving}
+                    className="h-9 rounded-full px-3 text-ui-12p5"
+                  >
+                    {downloadSaving ? (
+                      <Spinner className="size-3.5" />
+                    ) : (
+                      <HugeiconsIcon
+                        icon={FolderSearchIcon}
+                        strokeWidth={1.75}
+                        data-icon="inline-start"
+                        className="size-3.5"
+                      />
+                    )}
+                    Change
+                  </Button>
+                  {downloadCache?.isCustom ? (
+                    <Button
+                      type="button"
+                      variant="ghost"
+                      size="sm"
+                      onClick={() => void saveDownloadLocation(null)}
+                      disabled={downloadSaving}
+                      className="h-9 rounded-full px-3 text-ui-12p5 text-muted-foreground"
+                    >
+                      Use default
+                    </Button>
+                  ) : null}
+                </div>
+              </div>
+
+              <p className="mt-2 text-ui-10p5 text-muted-foreground">
+                {downloadCache?.source === "environment"
+                  ? `Managed by the ${
+                      downloadCache.environmentVariable ?? "HF_HOME"
+                    } environment variable.`
+                  : [
+                      "New downloads only. Models already on disk stay where they are.",
+                      formatFreeSpace(downloadCache?.freeBytes ?? null),
+                    ]
+                      .filter(Boolean)
+                      .join(" · ")}
+              </p>
+            </div>
+
             <div className="rounded-[14px] border border-border/70 bg-muted/20 p-3">
               <div className="mb-2 flex items-center gap-2 text-ui-12 font-medium text-foreground">
                 <HugeiconsIcon
@@ -425,6 +572,16 @@ export function OnDeviceFoldersDialog({
         onOpenChange={setBrowserOpen}
         onSelect={(selectedPath) => void handleAdd(selectedPath)}
       />
+
+      <FolderBrowser
+        open={!isTauri && downloadBrowserOpen}
+        onOpenChange={setDownloadBrowserOpen}
+        onSelect={(selectedPath) => void saveDownloadLocation(selectedPath)}
+        initialPath={downloadCache?.cacheHome}
+        title="Choose model download location"
+        confirmLabel="Use for future downloads"
+        showModelHints={false}
+      />
     </>
   );
 }
diff --git a/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx b/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx
index 3e5a86a879..d68aaa5ab3 100644
--- a/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx
+++ b/studio/frontend/src/features/hub/download-manager/download-manager-panel.tsx
@@ -201,8 +201,10 @@ export function DownloadManagerPanel({
       className={cn(
         // Standalone: anchor bottom-right. In a shared stack (positioned=false)
         // flow as a right-aligned row so overlays stack instead of overlapping.
+        // min-h-0 there: a flex item's min-height defaults to auto, so the capped
+        // stack would squeeze the update card instead of this list.
         "pointer-events-none",
-        positioned ? "fixed bottom-4 right-4 z-50" : "flex justify-end",
+        positioned ? "fixed bottom-4 right-4 z-50" : "flex min-h-0 justify-end",
       )}
     >
       {collapsed ? (
@@ -229,7 +231,7 @@ export function DownloadManagerPanel({
           </TooltipContent>
         </Tooltip>
       ) : (
-        <div className="hub-download-panel pointer-events-auto w-[min(400px,calc(100vw-2rem))] overflow-hidden">
+        <div className="hub-download-panel pointer-events-auto flex min-h-0 w-[min(400px,calc(100vw-2rem))] flex-col overflow-hidden">
           <div className="flex items-center gap-2 border-b border-foreground/[0.07] py-2 pl-4 pr-3">
             <span className="min-w-0 flex-1 truncate text-ui-12p5 font-semibold text-foreground">
               {headerLabel}
diff --git a/studio/frontend/src/features/settings/api/coding-agents.ts b/studio/frontend/src/features/settings/api/coding-agents.ts
index 079938dec4..ae371b2d3a 100644
--- a/studio/frontend/src/features/settings/api/coding-agents.ts
+++ b/studio/frontend/src/features/settings/api/coding-agents.ts
@@ -16,10 +16,6 @@ type ApiCodingAgentsInfo = {
   detected: string[];
 };
 
-// Agents the CLI still supports but Studio does not list. Filtered here so
-// every consumer of this endpoint sees the same set.
-const HIDDEN_AGENTS = new Set(["pi"]);
-
 // Which CLIs are on PATH is environment state, not a persisted setting -- it
 // can change any time the user installs something new, so this only
 // de-duplicates concurrent in-flight calls (e.g. React strict-mode's double
@@ -28,8 +24,7 @@ const HIDDEN_AGENTS = new Set(["pi"]);
 let inFlightInfo: Promise<CodingAgentsInfo> | null = null;
 
 function fromApi(info: ApiCodingAgentsInfo): CodingAgentsInfo {
-  const visible = (ids: string[]) => ids.filter((id) => !HIDDEN_AGENTS.has(id));
-  return { agents: visible(info.agents), detected: visible(info.detected) };
+  return { agents: info.agents, detected: info.detected };
 }
 
 async function fetchCodingAgents(): Promise<CodingAgentsInfo> {
diff --git a/studio/frontend/src/features/settings/components/usage-examples.tsx b/studio/frontend/src/features/settings/components/usage-examples.tsx
index ad26dfb2d3..846aa0380d 100644
--- a/studio/frontend/src/features/settings/components/usage-examples.tsx
+++ b/studio/frontend/src/features/settings/components/usage-examples.tsx
@@ -106,15 +106,16 @@ const DOC_LINKS = [
   { label: "Hermes Agent", href: "https://unsloth.ai/docs/integrations/hermes-agent" },
 ];
 
-// Fallback until the backend's installed-CLI check resolves. Mirrors
-// CODING_AGENTS in studio/backend/utils/coding_agents.py, minus HIDDEN_AGENTS
-// (see ../api/coding-agents.ts).
+// Falls back to this list until the backend's installed-CLI check resolves;
+// kept in sync with the `unsloth start <agent>` subcommands and with
+// CODING_AGENTS in studio/backend/utils/coding_agents.py.
 const DEFAULT_AGENTS = [
   "claude",
   "codex",
   "openclaw",
   "opencode",
   "hermes",
+  "pi",
 ];
 // The agent selection resets to this whenever an auto-pick is no longer
 // trustworthy (leaving loopback, or the only compatible detected agent
@@ -126,6 +127,7 @@ const AGENT_LABELS: Record<string, string> = {
   openclaw: "OpenClaw",
   opencode: "OpenCode",
   hermes: "Hermes",
+  pi: "Pi",
 };
 
 const j = (s: string): string => JSON.stringify(s);
diff --git a/studio/frontend/src/features/settings/index.ts b/studio/frontend/src/features/settings/index.ts
index f27100a322..49d73dcfb3 100644
--- a/studio/frontend/src/features/settings/index.ts
+++ b/studio/frontend/src/features/settings/index.ts
@@ -3,6 +3,11 @@
 
 export { SettingsDialog } from "./settings-dialog";
 export { loadEmbeddingModelSettings } from "./api/embedding-model";
+export {
+  loadHuggingFaceCacheSettings,
+  updateHuggingFaceCacheSettings,
+} from "./api/hugging-face-cache";
+export type { HuggingFaceCacheSettings } from "./api/hugging-face-cache";
 export {
   loadPersonalization,
   savePersonalization,
diff --git a/studio/frontend/src/features/settings/settings-dialog.tsx b/studio/frontend/src/features/settings/settings-dialog.tsx
index f4ba98b1ce..2b46dbd55e 100644
--- a/studio/frontend/src/features/settings/settings-dialog.tsx
+++ b/studio/frontend/src/features/settings/settings-dialog.tsx
@@ -35,7 +35,10 @@ import {
   useRef,
   useState,
 } from "react";
-import { SETTINGS_SEARCH_INDEX } from "./settings-search";
+import {
+  SETTINGS_SEARCH_INDEX,
+  SETTINGS_SEARCH_KEYWORDS,
+} from "./settings-search";
 import {
   type SettingsTab,
   useSettingsDialogStore,
@@ -157,8 +160,12 @@ export function SettingsDialog() {
     return TABS.map((tab) => {
       const tabLabel = t(tab.labelKey);
       const entries = SETTINGS_SEARCH_INDEX[tab.id]
-        .map((key) => t(key))
-        .filter((label) => label.toLowerCase().includes(q));
+        .filter((key) => {
+          if (t(key).toLowerCase().includes(q)) return true;
+          const keywordsKey = SETTINGS_SEARCH_KEYWORDS[key];
+          return keywordsKey ? t(keywordsKey).toLowerCase().includes(q) : false;
+        })
+        .map((key) => t(key));
       const deduped = [...new Set(entries)];
       return {
         tab,
diff --git a/studio/frontend/src/features/settings/settings-search.ts b/studio/frontend/src/features/settings/settings-search.ts
index a5b008579c..582602e061 100644
--- a/studio/frontend/src/features/settings/settings-search.ts
+++ b/studio/frontend/src/features/settings/settings-search.ts
@@ -146,3 +146,15 @@ export const SETTINGS_SEARCH_INDEX: Record<SettingsTab, TranslationKey[]> = {
     "settings.about.shutDownStudio",
   ],
 };
+
+/**
+ * Extra terms a row matches on, beyond its own label. The value is a
+ * translation key holding space-separated synonyms; it is never rendered.
+ * Search matched labels only, so "models folder" or "directory" found nothing.
+ */
+export const SETTINGS_SEARCH_KEYWORDS: Partial<
+  Record<TranslationKey, TranslationKey>
+> = {
+  "settings.resources.storage.modelsFolder":
+    "settings.resources.storage.modelsFolderKeywords",
+};
diff --git a/studio/frontend/src/features/settings/tabs/agents-tab.tsx b/studio/frontend/src/features/settings/tabs/agents-tab.tsx
index a8ef07234b..0e961688c8 100644
--- a/studio/frontend/src/features/settings/tabs/agents-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/agents-tab.tsx
@@ -41,7 +41,11 @@ import { ChevronDownStandardIcon } from "@/lib/chevron-icons";
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { Tick02Icon } from "@/lib/tick-icon";
 import { cn } from "@/lib/utils";
-import { ArrowUpRight01Icon, Copy01Icon } from "@hugeicons/core-free-icons";
+import {
+  ArrowUpRight01Icon,
+  Book03Icon,
+  Copy01Icon,
+} from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useCallback, useEffect, useMemo, useRef, useState } from "react";
 import { ApiProviderLogo } from "../../chat/api-provider-logo";
@@ -62,7 +66,7 @@ const STATUS_POLL_MS = 5000;
 const HUGGING_FACE_REPO_PATTERN = /^[^/\\:\s]+\/[^/\\:\s]+$/;
 const SEARCH_TOKEN_PATTERN = /\s+/;
 const SAFE_SHELL_ARG_PATTERN = /^[A-Za-z0-9_./:@%+=,-]+$/;
-const SUBAGENT_AGENT_IDS = new Set(["claude", "codex", "opencode"]);
+const SUBAGENT_AGENT_IDS = new Set(["claude", "codex", "opencode", "pi"]);
 
 function isLoopbackBase(base: string): boolean {
   try {
@@ -161,6 +165,12 @@ const SUPPORTED_AGENTS: AgentDetails[] = [
     icon: "opencode-light.svg",
     darkIcon: "opencode-dark.svg",
   },
+  {
+    id: "pi",
+    name: "Pi Coding Agent",
+    docsUrl: DOCS_URL,
+    icon: "pi.svg",
+  },
 ];
 
 const FALLBACK_AGENT = SUPPORTED_AGENTS[0];
@@ -372,8 +382,8 @@ function AgentIcon({
 }) {
   if (logo) {
     return (
-      <span className="flex size-5 shrink-0 items-center justify-center overflow-hidden rounded">
-        <ApiProviderLogo providerType={logo} className="size-5 rounded" />
+      <span className="flex size-7 shrink-0 items-center justify-center overflow-hidden rounded-md">
+        <ApiProviderLogo providerType={logo} className="size-7 rounded-md" />
       </span>
     );
   }
@@ -383,13 +393,13 @@ function AgentIcon({
       ? `${import.meta.env.BASE_URL}agent-logos/${darkIcon}`
       : null;
     return (
-      <span className="flex size-5 shrink-0 items-center justify-center overflow-hidden rounded">
+      <span className="flex size-7 shrink-0 items-center justify-center overflow-hidden rounded-md">
         <img
           src={iconSrc}
           alt=""
           aria-hidden={true}
           className={cn(
-            "size-5 object-contain",
+            "size-7 object-contain",
             darkIconSrc && "dark:hidden",
             invertIconInDark && "dark:invert",
           )}
@@ -399,7 +409,7 @@ function AgentIcon({
             src={darkIconSrc}
             alt=""
             aria-hidden={true}
-            className="hidden size-5 object-contain dark:block"
+            className="hidden size-7 object-contain dark:block"
           />
         ) : null}
       </span>
@@ -409,7 +419,7 @@ function AgentIcon({
     <span
       aria-hidden={true}
       style={{ backgroundColor: color }}
-      className="flex size-5 shrink-0 items-center justify-center rounded font-heading text-ui-10 font-semibold text-white"
+      className="flex size-7 shrink-0 items-center justify-center rounded-md font-heading text-ui-11 font-semibold text-white"
     >
       {mark}
     </span>
@@ -463,59 +473,13 @@ const PASSTHROUGH_EXAMPLES = [
 
 const DRY_RUN_FLAGS = "--no-launch";
 
-/** Code box with the copy control inside it, top-right. Presentational: the
- *  copy state stays with the caller so existing resets still apply. */
-function CopyableCode({
-  value,
-  copyLabel,
-  copied,
-  onCopy,
-  breakAll = true,
-}: {
-  value: string;
-  copyLabel: string;
-  copied: boolean;
-  onCopy: () => void;
-  breakAll?: boolean;
-}) {
-  const t = useT();
-
-  return (
-    <div className="relative min-w-0">
-      <code
-        className={cn(
-          "block min-w-0 whitespace-pre-wrap rounded-lg border border-border bg-background/70 py-2.5 pr-9 pl-4 font-mono text-ui-11 leading-relaxed text-foreground dark:border-transparent dark:bg-white/[0.05]",
-          breakAll ? "break-all" : "break-words",
-        )}
-      >
-        {value}
-      </code>
-      <button
-        type="button"
-        onClick={onCopy}
-        aria-label={copyLabel}
-        className="absolute top-1.5 right-1.5 flex size-6 items-center justify-center rounded-md text-muted-foreground transition-colors hover:bg-accent hover:text-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
-      >
-        <HugeiconsIcon
-          icon={copied ? Tick02Icon : Copy01Icon}
-          className={cn("size-3.5", copied && "text-control-accent")}
-          strokeWidth={2}
-        />
-      </button>
-      <output className="sr-only" aria-live="polite">
-        {copied ? t("settings.agents.copied") : ""}
-      </output>
-    </div>
-  );
-}
-
 function CommandBlock({ command }: { command: string }) {
   const t = useT();
   const { copied, copy } = useCopyButton(command);
 
   return (
-    <div className="group relative overflow-hidden rounded-xl border border-border bg-muted/40 dark:border-transparent dark:bg-white/[0.04]">
-      <pre className="hover-scrollbar overflow-x-auto py-3 pr-11 pl-4 text-xs leading-relaxed text-foreground">
+    <div className="group relative">
+      <pre className="hover-scrollbar overflow-x-auto rounded-lg border border-border bg-muted/40 py-3 pl-3.5 pr-11 text-xs leading-relaxed text-foreground dark:bg-white/[0.04]">
         <code className="font-mono whitespace-pre">{command}</code>
       </pre>
       <button
@@ -571,7 +535,7 @@ function SubagentSection({
   }
 
   return (
-    <div className="flex min-w-0 flex-col gap-4">
+    <div className="flex min-w-0 flex-col gap-3 rounded-lg border border-border bg-muted/10 p-3">
       <div className="flex flex-col gap-1">
         <span
           data-settings-label={t("settings.agents.subagent.title")}
@@ -584,29 +548,60 @@ function SubagentSection({
         </p>
       </div>
 
-      <div className="flex min-w-0 flex-col gap-2">
-        <span className="text-ui-11 font-medium text-foreground">
-          {t("settings.agents.subagent.setupCommand")}
-        </span>
-        <CopyableCode
-          value={command}
-          copyLabel={t("settings.agents.subagent.copySetupCommand")}
-          copied={commandCopy.copied}
-          onCopy={commandCopy.copy}
-        />
+      <div className="flex min-w-0 flex-col gap-1.5">
+        <div className="flex items-center justify-between gap-3">
+          <span className="text-ui-11 font-medium text-foreground">
+            {t("settings.agents.subagent.setupCommand")}
+          </span>
+          <button
+            type="button"
+            onClick={commandCopy.copy}
+            aria-label={t("settings.agents.subagent.copySetupCommand")}
+            className="inline-flex h-7 items-center gap-1.5 rounded-md border border-border bg-background/70 px-2 text-ui-11 font-medium text-muted-foreground transition-colors hover:bg-accent/60 hover:text-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
+          >
+            <HugeiconsIcon
+              icon={commandCopy.copied ? Tick02Icon : Copy01Icon}
+              className={cn(
+                "size-3.5",
+                commandCopy.copied && "text-control-accent",
+              )}
+            />
+            {commandCopy.copied
+              ? t("settings.agents.copied")
+              : t("settings.agents.copy")}
+          </button>
+        </div>
+        <code className="block min-w-0 whitespace-pre-wrap break-all rounded-md border border-border bg-background/70 px-2.5 py-2 font-mono text-ui-11 leading-relaxed text-foreground">
+          {command}
+        </code>
       </div>
 
-      <div className="flex min-w-0 flex-col gap-2">
-        <span className="text-ui-11 font-medium text-foreground">
-          {t("settings.agents.subagent.usagePrompt", { agent: agent.name })}
-        </span>
-        <CopyableCode
-          value={prompt}
-          copyLabel={t("settings.agents.subagent.copyUsagePrompt")}
-          copied={promptCopy.copied}
-          onCopy={promptCopy.copy}
-          breakAll={false}
-        />
+      <div className="flex min-w-0 flex-col gap-1.5">
+        <div className="flex items-center justify-between gap-3">
+          <span className="text-ui-11 font-medium text-foreground">
+            {t("settings.agents.subagent.usagePrompt", { agent: agent.name })}
+          </span>
+          <button
+            type="button"
+            onClick={promptCopy.copy}
+            aria-label={t("settings.agents.subagent.copyUsagePrompt")}
+            className="inline-flex h-7 items-center gap-1.5 rounded-md border border-border bg-background/70 px-2 text-ui-11 font-medium text-muted-foreground transition-colors hover:bg-accent/60 hover:text-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
+          >
+            <HugeiconsIcon
+              icon={promptCopy.copied ? Tick02Icon : Copy01Icon}
+              className={cn(
+                "size-3.5",
+                promptCopy.copied && "text-control-accent",
+              )}
+            />
+            {promptCopy.copied
+              ? t("settings.agents.copied")
+              : t("settings.agents.copy")}
+          </button>
+        </div>
+        <code className="block min-w-0 whitespace-pre-wrap break-words rounded-md border border-border bg-background/70 px-2.5 py-2 font-mono text-ui-11 leading-relaxed text-foreground">
+          {prompt}
+        </code>
       </div>
     </div>
   );
@@ -670,6 +665,7 @@ export function AgentsTab() {
   const [modelSearch, setModelSearch] = useState("");
   const [modelPickerOpen, setModelPickerOpen] = useState(false);
   const [variants, setVariants] = useState<GgufVariantDetail[]>([]);
+  const [defaultVariant, setDefaultVariant] = useState<string | null>(null);
   const [selectedVariant, setSelectedVariant] = useState<string | null>(
     EXAMPLE_MODEL_VARIANT,
   );
@@ -990,6 +986,7 @@ export function AgentsTab() {
           return;
         }
         setVariants([]);
+        setDefaultVariant(null);
         setSelectedVariant(standaloneFile ? null : preferredVariant);
         setVariantsFailed(false);
         setVariantsLoading(false);
@@ -1002,6 +999,7 @@ export function AgentsTab() {
     // A programmatic model change reaches here too, so clear the previous model's
     // quants up front rather than leaving them selectable until this resolves.
     setVariants([]);
+    setDefaultVariant(null);
     setVariantsLoading(true);
     // Offer the quants from the same place the command loads from, not remote-only ones.
     listGgufVariants(selectedModel, hfToken || undefined, {
@@ -1025,6 +1023,7 @@ export function AgentsTab() {
           ).values(),
         );
         setVariants(uniqueVariants);
+        setDefaultVariant(info.default_variant);
         const available = new Set(
           uniqueVariants.map((variant) => variant.quant),
         );
@@ -1046,6 +1045,7 @@ export function AgentsTab() {
         }
         setVariantsFailed(true);
         setVariants([]);
+        setDefaultVariant(null);
         setSelectedVariant(preferredVariant);
         if (preferredVariant) {
           setVariants([
@@ -1073,7 +1073,7 @@ export function AgentsTab() {
   // picker only ever offers GGUF models.
 
   return (
-    <div className="flex min-w-0 max-w-full flex-col gap-8">
+    <div className="flex min-w-0 max-w-full flex-col gap-6">
       {/* data-settings-label lets indexed settings search scroll to these. */}
       <header className="flex min-w-0 flex-col gap-1">
         <h1
@@ -1094,265 +1094,249 @@ export function AgentsTab() {
         data-settings-label={t("settings.agents.intro")}
         className="text-sm text-muted-foreground leading-relaxed"
       >
-        {/* The chip is the docs entry point, so no separate link is needed.
-            No aria-label: it would replace the visible "unsloth start" as the
-            accessible name, leaving voice control unable to target it. */}
-        <a
-          href={DOCS_URL}
-          target="_blank"
-          rel="noopener noreferrer"
-          title={t("settings.agents.readDocs")}
-          className="rounded bg-muted px-1 py-0.5 font-mono text-[0.85em] text-foreground underline decoration-border decoration-dotted underline-offset-2 transition-colors hover:decoration-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring dark:bg-white/[0.08]"
-        >
+        <code className="rounded bg-muted px-1 py-0.5 font-mono text-[0.85em] text-foreground dark:bg-white/[0.08]">
           unsloth start
-        </a>{" "}
+        </code>{" "}
         {t("settings.agents.intro")}
       </p>
 
+      <a
+        href={DOCS_URL}
+        target="_blank"
+        rel="noopener noreferrer"
+        className="inline-flex w-fit items-center gap-1.5 text-xs font-medium text-muted-foreground hover:text-foreground"
+      >
+        <HugeiconsIcon icon={Book03Icon} className="size-3.5" />
+        {t("settings.agents.readDocs")}
+        <HugeiconsIcon icon={ArrowUpRight01Icon} className="size-3" />
+      </a>
+
       <section
         aria-label={t("settings.agents.commandBuilder")}
-        className="flex w-full flex-col gap-6"
+        className="flex w-full flex-col gap-4"
       >
-        {/* Keyed to this pane, not the viewport. The dialog leaves the tab
-            about 440px at a 768px window, where three columns crush the agent
-            and model controls; 34rem is the point all three stay usable. */}
-        <div className="@container">
-          <div className="grid grid-cols-1 items-start gap-3 @[34rem]:grid-cols-[minmax(0,0.8fr)_minmax(0,1fr)_minmax(9rem,0.5fr)]">
-            <div className="flex min-w-0 flex-col gap-1.5">
-              {/* Fixed height on every column header, so the padded docs link
-                  here cannot push this control below the other two. */}
-              <div className="flex h-5 items-center justify-between gap-3">
-                <span
-                  data-settings-label={t("settings.agents.agent")}
-                  className="text-xs font-medium text-foreground"
-                >
-                  {t("settings.agents.agent")}
+        <div className="flex flex-col gap-1.5">
+          <div className="flex items-center justify-between gap-3">
+            <span
+              data-settings-label={t("settings.agents.agent")}
+              className="text-xs font-medium text-foreground"
+            >
+              {t("settings.agents.agent")}
+            </span>
+            <a
+              href={selectedAgentDetails.docsUrl}
+              target="_blank"
+              rel="noreferrer"
+              aria-label={t("settings.agents.agentDocs", {
+                agent: selectedAgentDetails.name,
+              })}
+              className="inline-flex items-center gap-1 rounded px-1.5 py-0.5 text-ui-11 font-medium text-muted-foreground transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
+            >
+              {t("settings.agents.docs")}
+              <HugeiconsIcon icon={ArrowUpRight01Icon} className="size-3" />
+            </a>
+          </div>
+          <Select
+            value={selectedAgent}
+            onValueChange={(agent) => {
+              agentSelectionChanged.current = true;
+              setSelectedAgent(agent);
+              resetCopied();
+            }}
+          >
+            <SelectTrigger
+              aria-label={t("settings.agents.agent")}
+              className="w-full rounded-lg"
+            >
+              <SelectValue>
+                <span className="flex min-w-0 items-center gap-2">
+                  <AgentIcon
+                    logo={selectedAgentDetails.logo}
+                    icon={selectedAgentDetails.icon}
+                    darkIcon={selectedAgentDetails.darkIcon}
+                    invertIconInDark={selectedAgentDetails.invertIconInDark}
+                    color={selectedAgentDetails.color}
+                    mark={selectedAgentDetails.mark}
+                  />
+                  <span className="truncate">{selectedAgentDetails.name}</span>
                 </span>
-                <a
-                  href={selectedAgentDetails.docsUrl}
-                  target="_blank"
-                  rel="noreferrer"
-                  aria-label={t("settings.agents.agentDocs", {
-                    agent: selectedAgentDetails.name,
-                  })}
-                  className="inline-flex items-center gap-1 rounded px-1.5 py-0.5 text-ui-11 font-medium text-muted-foreground transition-colors hover:text-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
-                >
-                  {t("settings.agents.docs")}
-                  <HugeiconsIcon icon={ArrowUpRight01Icon} className="size-3" />
-                </a>
-              </div>
-              <Select
-                value={selectedAgent}
-                onValueChange={(agent) => {
-                  agentSelectionChanged.current = true;
-                  setSelectedAgent(agent);
-                  resetCopied();
-                }}
-              >
-                <SelectTrigger
-                  aria-label={t("settings.agents.agent")}
-                  className="w-full rounded-lg"
-                >
-                  <SelectValue>
+              </SelectValue>
+            </SelectTrigger>
+            <SelectContent align="start">
+              {agents.map((agentId) => {
+                const agent = detailsFor(agentId);
+                return (
+                  <SelectItem key={agent.id} value={agent.id}>
                     <span className="flex min-w-0 items-center gap-2">
                       <AgentIcon
-                        logo={selectedAgentDetails.logo}
-                        icon={selectedAgentDetails.icon}
-                        darkIcon={selectedAgentDetails.darkIcon}
-                        invertIconInDark={selectedAgentDetails.invertIconInDark}
-                        color={selectedAgentDetails.color}
-                        mark={selectedAgentDetails.mark}
+                        logo={agent.logo}
+                        icon={agent.icon}
+                        darkIcon={agent.darkIcon}
+                        invertIconInDark={agent.invertIconInDark}
+                        color={agent.color}
+                        mark={agent.mark}
                       />
-                      <span className="truncate">
-                        {selectedAgentDetails.name}
-                      </span>
-                    </span>
-                  </SelectValue>
-                </SelectTrigger>
-                <SelectContent align="start">
-                  {agents.map((agentId) => {
-                    const agent = detailsFor(agentId);
-                    return (
-                      <SelectItem key={agent.id} value={agent.id}>
-                        <span className="flex min-w-0 items-center gap-2">
-                          <AgentIcon
-                            logo={agent.logo}
-                            icon={agent.icon}
-                            darkIcon={agent.darkIcon}
-                            invertIconInDark={agent.invertIconInDark}
-                            color={agent.color}
-                            mark={agent.mark}
-                          />
-                          <span className="truncate">{agent.name}</span>
-                          {localDetection &&
-                          loaded &&
-                          detectedAgents.has(agent.id) ? (
-                            <span className="shrink-0 rounded-full bg-control-accent/10 px-2 py-1 text-ui-10 leading-none font-semibold text-control-accent">
-                              {t("settings.agents.quickstart.installed")}
-                            </span>
-                          ) : null}
+                      <span className="truncate">{agent.name}</span>
+                      {localDetection &&
+                      loaded &&
+                      detectedAgents.has(agent.id) ? (
+                        <span className="shrink-0 rounded-full bg-control-accent/10 px-2 py-1 text-ui-10 leading-none font-semibold text-control-accent">
+                          {t("settings.agents.quickstart.installed")}
                         </span>
-                      </SelectItem>
-                    );
-                  })}
-                </SelectContent>
-              </Select>
-            </div>
-
-            <div className="flex min-w-0 flex-col gap-1.5">
-              <div className="flex h-5 items-center">
-                <span
-                  data-settings-label={t("settings.agents.model")}
-                  className="text-xs font-medium text-foreground"
-                >
-                  {t("settings.agents.model")}
-                </span>
-              </div>
-              <Popover
-                open={modelPickerOpen}
-                onOpenChange={(open) => {
-                  setModelPickerOpen(open);
-                  if (!open) {
-                    setModelSearch("");
-                  }
-                }}
-              >
-                <PopoverTrigger asChild={true}>
-                  <button
-                    type="button"
-                    aria-label={t("settings.agents.model")}
-                    aria-expanded={modelPickerOpen}
-                    title={selectedModel}
-                    className="flex h-9 w-full items-center justify-between gap-2 rounded-lg border border-border bg-background px-3 text-left transition-colors hover:bg-accent/50 focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring dark:border-transparent dark:bg-white/[0.06] dark:hover:bg-white/10"
-                  >
-                    <span className="min-w-0 truncate font-mono text-xs">
-                      {labelFor(selectedModel)}
+                      ) : null}
                     </span>
-                    <HugeiconsIcon
-                      icon={ChevronDownStandardIcon}
-                      strokeWidth={2}
-                      className="size-4 shrink-0 text-muted-foreground"
-                    />
-                  </button>
-                </PopoverTrigger>
-                <PopoverContent
-                  align="start"
-                  sideOffset={4}
-                  className="w-[var(--radix-popover-trigger-width)] max-w-[calc(100vw-2rem)] gap-0 rounded-lg p-1"
-                >
-                  <Command
-                    shouldFilter={false}
-                    className="rounded-none bg-transparent p-0"
-                  >
-                    <CommandInput
-                      value={modelSearch}
-                      onValueChange={setModelSearch}
-                      aria-label={t("settings.agents.searchModels")}
-                      placeholder={t("settings.agents.searchModels")}
-                      className="font-mono text-xs"
-                    />
-                    <CommandList>
-                      <CommandEmpty>
-                        {t("settings.agents.noModels")}
-                      </CommandEmpty>
-                      {visibleModels.map((model) => (
-                        <CommandItem
-                          key={model}
-                          value={model}
-                          data-checked={model === selectedModel}
-                          onSelect={() => {
-                            modelSelectionChanged.current = true;
-                            setSelectedModel(model);
-                            setSelectedVariant(knownVariants[model] ?? null);
-                            setVariants([]);
-                            setVariantsFailed(false);
-                            setVariantsLoading(isHuggingFaceRepo(model));
-                            setModelSearch("");
-                            setModelPickerOpen(false);
-                            resetCopied();
-                          }}
-                          className="cursor-pointer font-mono text-xs"
-                        >
-                          <span className="min-w-0 truncate" title={model}>
-                            {labelFor(model)}
-                          </span>
-                        </CommandItem>
-                      ))}
-                    </CommandList>
-                    {matchingModels.length > visibleModels.length ? (
-                      <p className="border-t border-border/60 px-3 py-2 text-ui-11 text-muted-foreground">
-                        {t("settings.agents.showingModels", {
-                          shown: visibleModels.length,
-                          total: matchingModels.length,
-                        })}
-                      </p>
-                    ) : null}
-                  </Command>
-                </PopoverContent>
-              </Popover>
-            </div>
+                  </SelectItem>
+                );
+              })}
+            </SelectContent>
+          </Select>
+        </div>
 
-            <div className="flex min-w-0 flex-col gap-1.5">
-              <div className="flex h-5 items-center">
-                <span
-                  data-settings-label={t("settings.agents.quantization")}
-                  className="text-xs font-medium text-foreground"
+        <div className="grid grid-cols-[minmax(0,1fr)_minmax(10rem,0.4fr)] items-start gap-3 max-md:grid-cols-1">
+          <div className="flex min-w-0 flex-col gap-1.5">
+            <span
+              data-settings-label={t("settings.agents.model")}
+              className="text-xs font-medium text-foreground"
+            >
+              {t("settings.agents.model")}
+            </span>
+            <Popover
+              open={modelPickerOpen}
+              onOpenChange={(open) => {
+                setModelPickerOpen(open);
+                if (!open) {
+                  setModelSearch("");
+                }
+              }}
+            >
+              <PopoverTrigger asChild={true}>
+                <button
+                  type="button"
+                  aria-label={t("settings.agents.model")}
+                  aria-expanded={modelPickerOpen}
+                  title={selectedModel}
+                  className="flex h-9 w-full items-center justify-between gap-2 rounded-lg border border-border bg-background px-3 text-left transition-colors hover:bg-accent/50 focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring dark:border-transparent dark:bg-white/[0.06] dark:hover:bg-white/10"
                 >
-                  {t("settings.agents.quantization")}
-                </span>
-              </div>
-              <Select
-                value={selectedVariant ?? undefined}
-                onValueChange={(variant) => {
-                  chosenVariant.current = { model: selectedModel, variant };
-                  setSelectedVariant(variant);
-                  resetCopied();
-                }}
-                disabled={variantsLoading || variants.length === 0}
+                  <span className="min-w-0 truncate font-mono text-xs">
+                    {labelFor(selectedModel)}
+                  </span>
+                  <HugeiconsIcon
+                    icon={ChevronDownStandardIcon}
+                    strokeWidth={2}
+                    className="size-4 shrink-0 text-muted-foreground"
+                  />
+                </button>
+              </PopoverTrigger>
+              <PopoverContent
+                align="start"
+                sideOffset={4}
+                className="w-[var(--radix-popover-trigger-width)] max-w-[calc(100vw-2rem)] gap-0 rounded-lg p-1"
               >
-                <SelectTrigger
-                  aria-label={t("settings.agents.quantization")}
-                  className="w-full rounded-lg font-mono text-xs"
+                <Command
+                  shouldFilter={false}
+                  className="rounded-none bg-transparent p-0"
                 >
-                  <SelectValue
-                    placeholder={
-                      variantsLoading
-                        ? t("settings.agents.loadingQuantizations")
-                        : t("settings.agents.noQuantizations")
-                    }
-                  >
-                    {selectedVariant}
-                  </SelectValue>
-                </SelectTrigger>
-                <SelectContent align="start" className="min-w-[16rem]">
-                  {variants.map((variant) => {
-                    // Size only: the recommended/downloaded tags wrapped every
-                    // row onto two lines and made the list hard to scan.
-                    const size = formatBytes(
-                      variant.download_size_bytes ?? variant.size_bytes,
-                    );
-                    return (
-                      <SelectItem
-                        key={variant.quant}
-                        value={variant.quant}
-                        // Stretch the item text so the size can sit flush right,
-                        // giving the list a clean two-column read.
-                        className="[&>span:last-child]:w-full [&>span:last-child]:justify-between"
+                  <CommandInput
+                    value={modelSearch}
+                    onValueChange={setModelSearch}
+                    aria-label={t("settings.agents.searchModels")}
+                    placeholder={t("settings.agents.searchModels")}
+                    className="font-mono text-xs"
+                  />
+                  <CommandList>
+                    <CommandEmpty>{t("settings.agents.noModels")}</CommandEmpty>
+                    {visibleModels.map((model) => (
+                      <CommandItem
+                        key={model}
+                        value={model}
+                        data-checked={model === selectedModel}
+                        onSelect={() => {
+                          modelSelectionChanged.current = true;
+                          setSelectedModel(model);
+                          setSelectedVariant(knownVariants[model] ?? null);
+                          setVariants([]);
+                          setDefaultVariant(null);
+                          setVariantsFailed(false);
+                          setVariantsLoading(isHuggingFaceRepo(model));
+                          setModelSearch("");
+                          setModelPickerOpen(false);
+                          resetCopied();
+                        }}
+                        className="cursor-pointer font-mono text-xs"
                       >
-                        <span className="font-mono text-xs whitespace-nowrap">
-                          {variant.quant}
+                        <span className="min-w-0 truncate" title={model}>
+                          {labelFor(model)}
                         </span>
-                        {size ? (
-                          <span className="text-ui-10 whitespace-nowrap text-muted-foreground">
-                            {size}
-                          </span>
-                        ) : null}
-                      </SelectItem>
-                    );
-                  })}
-                </SelectContent>
-              </Select>
-            </div>
+                      </CommandItem>
+                    ))}
+                  </CommandList>
+                  {matchingModels.length > visibleModels.length ? (
+                    <p className="border-t border-border/60 px-3 py-2 text-ui-11 text-muted-foreground">
+                      {t("settings.agents.showingModels", {
+                        shown: visibleModels.length,
+                        total: matchingModels.length,
+                      })}
+                    </p>
+                  ) : null}
+                </Command>
+              </PopoverContent>
+            </Popover>
+          </div>
+
+          <div className="flex min-w-0 flex-col gap-1.5">
+            <span
+              data-settings-label={t("settings.agents.quantization")}
+              className="text-xs font-medium text-foreground"
+            >
+              {t("settings.agents.quantization")}
+            </span>
+            <Select
+              value={selectedVariant ?? undefined}
+              onValueChange={(variant) => {
+                chosenVariant.current = { model: selectedModel, variant };
+                setSelectedVariant(variant);
+                resetCopied();
+              }}
+              disabled={variantsLoading || variants.length === 0}
+            >
+              <SelectTrigger
+                aria-label={t("settings.agents.quantization")}
+                className="w-full rounded-lg font-mono text-xs"
+              >
+                <SelectValue
+                  placeholder={
+                    variantsLoading
+                      ? t("settings.agents.loadingQuantizations")
+                      : t("settings.agents.noQuantizations")
+                  }
+                >
+                  {selectedVariant}
+                </SelectValue>
+              </SelectTrigger>
+              <SelectContent align="start">
+                {variants.map((variant) => {
+                  const metadata = [
+                    variant.quant === defaultVariant
+                      ? t("settings.agents.recommended")
+                      : null,
+                    variant.downloaded ? t("settings.agents.downloaded") : null,
+                    formatBytes(
+                      variant.download_size_bytes ?? variant.size_bytes,
+                    ),
+                  ].filter(Boolean);
+                  return (
+                    <SelectItem key={variant.quant} value={variant.quant}>
+                      <span className="font-mono text-xs">{variant.quant}</span>
+                      {metadata.length > 0 ? (
+                        <span className="text-ui-10 text-muted-foreground">
+                          {metadata.join(" · ")}
+                        </span>
+                      ) : null}
+                    </SelectItem>
+                  );
+                })}
+              </SelectContent>
+            </Select>
           </div>
         </div>
 
@@ -1362,16 +1346,27 @@ export function AgentsTab() {
           </p>
         ) : null}
 
-        <div className="flex min-w-0 flex-col gap-2.5">
-          <span className="text-xs font-medium text-foreground">
-            {t("settings.agents.generatedCommand")}
-          </span>
-          <CopyableCode
-            value={command}
-            copyLabel={t("settings.agents.copyGeneratedCommand")}
-            copied={copied}
-            onCopy={handleCopy}
-          />
+        <div className="flex min-w-0 flex-col gap-2 rounded-lg border border-border bg-muted/20 p-3">
+          <div className="flex items-center justify-between gap-3">
+            <span className="text-xs font-medium text-foreground">
+              {t("settings.agents.generatedCommand")}
+            </span>
+            <button
+              type="button"
+              onClick={handleCopy}
+              aria-label={t("settings.agents.copyGeneratedCommand")}
+              className="inline-flex h-7 items-center gap-1.5 rounded-md border border-border bg-background/70 px-2 text-ui-11 font-medium text-muted-foreground transition-colors hover:bg-accent/60 hover:text-foreground focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-ring"
+            >
+              <HugeiconsIcon
+                icon={copied ? Tick02Icon : Copy01Icon}
+                className={cn("size-3.5", copied && "text-control-accent")}
+              />
+              {copied ? t("settings.agents.copied") : t("settings.agents.copy")}
+            </button>
+          </div>
+          <code className="block min-w-0 whitespace-pre-wrap break-all rounded-md border border-border bg-background/70 px-2.5 py-2 font-mono text-ui-11 leading-relaxed text-foreground">
+            {command}
+          </code>
         </div>
 
         <SubagentSection
@@ -1411,7 +1406,7 @@ export function AgentsTab() {
         title={t("settings.agents.remote.title")}
         description={t("settings.agents.remote.description")}
       >
-        <div className="pt-3">
+        <div className="pt-2">
           <CommandBlock command={remoteCommand} />
         </div>
       </SettingsSection>
@@ -1420,7 +1415,7 @@ export function AgentsTab() {
         title={t("settings.agents.passthrough.title")}
         description={t("settings.agents.passthrough.description")}
       >
-        <div className="flex flex-col gap-3 pt-3">
+        <div className="flex flex-col gap-2 pt-2">
           {PASSTHROUGH_EXAMPLES.map(({ agent, flags }) => (
             <CommandBlock key={flags} command={example(agent, flags)} />
           ))}
@@ -1431,7 +1426,7 @@ export function AgentsTab() {
         title={t("settings.agents.dryRun.title")}
         description={t("settings.agents.dryRun.description")}
       >
-        <div className="pt-3">
+        <div className="pt-2">
           <CommandBlock command={example("claude", DRY_RUN_FLAGS)} />
         </div>
       </SettingsSection>
diff --git a/studio/frontend/src/features/settings/tabs/general-tab.tsx b/studio/frontend/src/features/settings/tabs/general-tab.tsx
index 1d5c533a25..11606d85af 100644
--- a/studio/frontend/src/features/settings/tabs/general-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/general-tab.tsx
@@ -74,6 +74,8 @@ const PREFS_KEYS: string[] = [
   LOCALE_STORAGE_KEY,
   // UI state
   "sidebar_pinned",
+  "sidebar_width",
+  "chat_settings_width",
   "unsloth_sidebar_navigate_open",
   "unsloth_settings_active_tab",
   // Chat runtime prefs
diff --git a/studio/frontend/src/features/settings/tabs/resources-tab.tsx b/studio/frontend/src/features/settings/tabs/resources-tab.tsx
index 1a6eb5abbb..f45ddeba7c 100644
--- a/studio/frontend/src/features/settings/tabs/resources-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/resources-tab.tsx
@@ -72,10 +72,7 @@ function formatBytes(value: number | null): string | null {
 function formatGiB(value: number | null | undefined): string {
   const safe = isFiniteNumber(value) ? Math.max(0, value) : 0;
   const digits = safe >= 10 ? 1 : 2;
-  // digits is never 0, so toFixed always leaves a decimal point and trimming
-  // trailing zeros cannot reach an integer digit. "64.0" reads as "64".
-  const text = safe.toFixed(digits).replace(/\.?0+$/, "");
-  return `${text} GiB`;
+  return `${safe.toFixed(digits)} GiB`;
 }
 
 function formatMb(value: number | null | undefined): string {
@@ -119,7 +116,7 @@ function MetricTile({
   const percentKnown = isFiniteNumber(percent);
   const safePercent = clampPercent(percent);
   return (
-    <div className="flex min-w-0 flex-col gap-2.5 rounded-xl border border-border/60 bg-muted/20 p-4 dark:border-transparent dark:bg-white/[0.06]">
+    <div className="flex min-w-0 flex-col gap-2 rounded-md border border-border/60 bg-muted/20 p-3">
       <div className="flex items-center justify-between gap-3">
         <span className="truncate text-ui-11 font-semibold uppercase tracking-[0.08em] text-muted-foreground">
           {label}
@@ -127,9 +124,7 @@ function MetricTile({
         <span
           className={cn(
             "shrink-0 font-mono text-xs tabular-nums",
-            percentKnown
-              ? usageTextClass(safePercent)
-              : "text-muted-foreground",
+            percentKnown ? usageTextClass(safePercent) : "text-muted-foreground",
           )}
         >
           {percentKnown ? formatPercent(safePercent) : "--"}
@@ -146,7 +141,7 @@ function MetricTile({
       <Progress
         value={percentKnown ? safePercent : 0}
         aria-label={label}
-        className="h-1.5 rounded-full bg-muted dark:bg-black/40"
+        className="h-1.5 rounded-full bg-muted"
         indicatorClassName={usageIndicatorClass(safePercent)}
       />
     </div>
@@ -332,9 +327,7 @@ export function ResourcesTab() {
   const hasGpu =
     (displayedGpu?.available ?? false) && metrics.devices.length > 0;
   const backendLabel = (
-    displayedGpu?.backend ??
-    systemInfo.device_backend ??
-    "cpu"
+    displayedGpu?.backend ?? systemInfo.device_backend ?? "cpu"
   ).toUpperCase();
   const modelsFolderPath = hfCache
     ? hfCache.cacheHome
@@ -490,66 +483,54 @@ export function ResourcesTab() {
               ? formatPercent(safePercent)
               : unknownLabel;
             return (
-              // Name over backend on the left, figures over the meter on the
-              // right. The meter tracks the figures' width rather than the
-              // pane's, so it reads as one device's usage, not a rule.
               <div
                 key={`${device.index ?? index}-${device.name ?? "gpu"}`}
-                className="flex min-w-0 items-center justify-between gap-x-4 gap-y-2 py-3 max-[992px]:flex-col max-[992px]:items-stretch"
+                className="flex min-w-0 flex-col gap-2 py-3"
               >
-                <div className="min-w-0 flex-1">
-                  <div className="truncate text-sm font-medium text-foreground">
-                    {device.name ?? t("settings.resources.gpu.unknownDevice")}
-                  </div>
-                  <div className="mt-1 flex min-w-0 items-center gap-2">
-                    <span className="truncate text-xs text-muted-foreground">
+                <div className="flex min-w-0 items-start justify-between gap-4">
+                  <div className="min-w-0 flex-1">
+                    <div className="truncate text-sm font-medium text-foreground">
+                      {device.name ??
+                        t("settings.resources.gpu.unknownDevice")}
+                    </div>
+                    <div className="mt-0.5 truncate text-xs text-muted-foreground">
                       {ordinal === undefined
                         ? backendLabel
                         : `${t("settings.resources.gpu.deviceWithIndex", {
                             index: ordinal,
                           })}, ${backendLabel}`}
-                    </span>
-                    {/* Same accent pill as the New tags, which stays legible
-                        on the light background. */}
-                    <span className="shrink-0 rounded-full bg-control-accent/10 px-2 py-1 text-ui-10 leading-none font-semibold tabular-nums text-control-accent">
+                    </div>
+                  </div>
+                  <div className="shrink-0 font-mono text-xs tabular-nums text-muted-foreground">
+                    <span>
                       {percentText}{" "}
                       {t("settings.resources.gpu.vramUtilization")}
                     </span>
                   </div>
                 </div>
-                {/* Meter under the figures, so it spans their width instead of
-                    being squeezed into the gap beside them. Same 392px as the
-                    Model downloads control, so both blocks start on one edge.
-                    Below 992px the dialog stops filling its 960px cap and the
-                    device name would truncate, so the row stacks instead. */}
-                <div className="flex w-[392px] shrink-0 flex-col items-stretch gap-2.5 max-[992px]:w-full">
-                  {/* Ruled between the three readings: run together they are
-                      easy to misread as one number. */}
-                  {/* min-w-0 on each reading, or truncate cannot fire: a flex
-                      item defaults to min-width:auto and the longest locales
-                      would push past the block instead of ellipsizing. */}
-                  <div className="flex items-center justify-between gap-3 font-mono text-ui-11 tabular-nums text-muted-foreground">
-                    <span className="min-w-0 truncate">
-                      {t("settings.resources.gpu.used", { value: usedText })}
-                    </span>
-                    <span aria-hidden className="h-3 w-px shrink-0 bg-border" />
-                    <span className="min-w-0 truncate">
-                      {t("settings.resources.gpu.free", { value: freeText })}
-                    </span>
-                    <span aria-hidden className="h-3 w-px shrink-0 bg-border" />
-                    <span className="min-w-0 truncate">
-                      {t("settings.resources.gpu.total", {
-                        value: totalText,
-                      })}
-                    </span>
-                  </div>
-                  <Progress
-                    value={safePercent}
-                    aria-label={device.name ?? "GPU"}
-                    className="h-1.5 w-full rounded-full bg-muted dark:bg-black/40"
-                    indicatorClassName={usageIndicatorClass(safePercent)}
-                  />
+                <div className="grid gap-1 text-xs text-muted-foreground sm:grid-cols-3 sm:gap-2">
+                  <span className="min-w-0 truncate font-mono tabular-nums">
+                    {t("settings.resources.gpu.used", {
+                      value: usedText,
+                    })}
+                  </span>
+                  <span className="min-w-0 truncate font-mono tabular-nums sm:text-center">
+                    {t("settings.resources.gpu.free", {
+                      value: freeText,
+                    })}
+                  </span>
+                  <span className="min-w-0 truncate font-mono tabular-nums sm:text-right">
+                    {t("settings.resources.gpu.total", {
+                      value: totalText,
+                    })}
+                  </span>
                 </div>
+                <Progress
+                  value={safePercent}
+                  aria-label={device.name ?? "GPU"}
+                  className="h-1.5 rounded-full bg-muted"
+                  indicatorClassName={usageIndicatorClass(safePercent)}
+                />
               </div>
             );
           })
diff --git a/studio/frontend/src/hooks/use-chat-settings-width.ts b/studio/frontend/src/hooks/use-chat-settings-width.ts
new file mode 100644
index 0000000000..6b2c75fe1e
--- /dev/null
+++ b/studio/frontend/src/hooks/use-chat-settings-width.ts
@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { createPanelWidthStore } from "./use-panel-width.ts";
+
+/** The previous fixed 17rem, at a 16px root font size. */
+export const CHAT_SETTINGS_WIDTH_DEFAULT = 272;
+/** Below this the sliders and their value pills start colliding. */
+export const CHAT_SETTINGS_WIDTH_MIN = 248;
+export const CHAT_SETTINGS_WIDTH_MAX = 560;
+
+const store = createPanelWidthStore({
+  key: "chat_settings_width",
+  min: CHAT_SETTINGS_WIDTH_MIN,
+  max: CHAT_SETTINGS_WIDTH_MAX,
+  fallback: CHAT_SETTINGS_WIDTH_DEFAULT,
+});
+
+export const clampChatSettingsWidth = store.clamp;
+export const useChatSettingsWidth = store.useWidth;
diff --git a/studio/frontend/src/hooks/use-panel-width.ts b/studio/frontend/src/hooks/use-panel-width.ts
new file mode 100644
index 0000000000..d753c3a1bb
--- /dev/null
+++ b/studio/frontend/src/hooks/use-panel-width.ts
@@ -0,0 +1,138 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { useCallback, useSyncExternalStore } from "react";
+
+/** Never let one panel eat more than this share of a narrow window. */
+const MAX_VIEWPORT_FRACTION = 0.4;
+
+export type PanelWidthStore = {
+  /** Clamps to what the current viewport allows. */
+  clamp: (px: number) => number;
+  useWidth: () => {
+    width: number;
+    max: number;
+    /** The uncapped stored preference. */
+    stored: number;
+    setWidth: (value: number) => void;
+    resetWidth: () => void;
+  };
+};
+
+/**
+ * A persisted, viewport-aware width for a draggable panel. The preference is
+ * stored whole and an effective width is derived from it, so narrowing the
+ * window shrinks the panel without losing what the user picked.
+ */
+export function createPanelWidthStore({
+  key,
+  min,
+  max,
+  fallback,
+}: {
+  key: string;
+  min: number;
+  max: number;
+  fallback: number;
+}): PanelWidthStore {
+  function maxWidth(): number {
+    if (typeof window === "undefined") return max;
+    // The floor wins on a narrow window; collapsing is the escape.
+    return Math.max(min, Math.min(max, window.innerWidth * MAX_VIEWPORT_FRACTION));
+  }
+
+  /** Clamps to the absolute range, ignoring the viewport. */
+  function clampStored(px: number): number {
+    if (!Number.isFinite(px)) return fallback;
+    return Math.min(max, Math.max(min, Math.round(px)));
+  }
+
+  function clamp(px: number): number {
+    return Math.min(maxWidth(), clampStored(px));
+  }
+
+  function load(): number {
+    if (typeof window === "undefined") return fallback;
+    try {
+      const raw = window.localStorage.getItem(key);
+      if (raw === null) return fallback;
+      return clampStored(Number.parseFloat(raw));
+    } catch {
+      return fallback;
+    }
+  }
+
+  let storedWidth = load();
+  let effectiveWidth = clamp(storedWidth);
+  let effectiveMax = maxWidth();
+  const listeners = new Set<() => void>();
+
+  let lastStored = storedWidth;
+
+  function recompute() {
+    const nextWidth = clamp(storedWidth);
+    const nextMax = maxWidth();
+    if (
+      nextWidth === effectiveWidth &&
+      nextMax === effectiveMax &&
+      storedWidth === lastStored
+    ) {
+      return;
+    }
+    effectiveWidth = nextWidth;
+    effectiveMax = nextMax;
+    lastStored = storedWidth;
+    listeners.forEach((cb) => cb());
+  }
+
+  function subscribe(cb: () => void) {
+    // With no subscribers there is no resize listener, so the cache can be
+    // stale after a resize on a route that hides every panel. Refresh first;
+    // useSyncExternalStore re-reads the snapshot right after subscribing.
+    recompute();
+    listeners.add(cb);
+    if (typeof window === "undefined") {
+      return () => listeners.delete(cb);
+    }
+    // Keep tabs in sync, same as the pin flag.
+    const onStorage = (e: StorageEvent) => {
+      if (e.key === key || e.key === null) {
+        storedWidth = load();
+        effectiveWidth = clamp(storedWidth);
+        effectiveMax = maxWidth();
+        cb();
+      }
+    };
+    window.addEventListener("storage", onStorage);
+    window.addEventListener("resize", recompute);
+    return () => {
+      listeners.delete(cb);
+      window.removeEventListener("storage", onStorage);
+      window.removeEventListener("resize", recompute);
+    };
+  }
+
+  function setWidthGlobal(next: number) {
+    const stored = clampStored(next);
+    if (stored !== storedWidth) {
+      storedWidth = stored;
+      try {
+        window.localStorage.setItem(key, String(stored));
+      } catch {}
+    }
+    recompute();
+  }
+
+  function useWidth() {
+    const width = useSyncExternalStore(subscribe, () => effectiveWidth, () => fallback);
+    // What the viewport actually allows right now, for aria-valuemax.
+    const panelMax = useSyncExternalStore(subscribe, () => effectiveMax, () => max);
+    // The uncapped preference, so a capped drag can avoid lowering it.
+    const preference = useSyncExternalStore(subscribe, () => storedWidth, () => fallback);
+    const setWidth = useCallback((value: number) => setWidthGlobal(value), []);
+    const resetWidth = useCallback(() => setWidthGlobal(fallback), []);
+    return { width, max: panelMax, stored: preference, setWidth, resetWidth };
+  }
+
+  return { clamp, useWidth };
+}
diff --git a/studio/frontend/src/hooks/use-release-notes.ts b/studio/frontend/src/hooks/use-release-notes.ts
new file mode 100644
index 0000000000..7b1392fdf6
--- /dev/null
+++ b/studio/frontend/src/hooks/use-release-notes.ts
@@ -0,0 +1,146 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { authFetch, hasAuthToken } from "@/features/auth";
+import { apiUrl } from "@/lib/api-base";
+import { useCallback, useEffect, useRef, useState } from "react";
+
+// Keyed to one exact version, so a new update never pairs with older notes.
+export interface ReleaseNotes {
+  version: string;
+  markdown: string | null;
+  matched: boolean;
+  truncated: boolean;
+  source: string | null;
+  releaseNotesUrl: string | null;
+  // Set when the lookup itself failed, as opposed to a version with no notes.
+  error: string | null;
+}
+
+export type ReleaseNotesState = "idle" | "loading" | "ready" | "error";
+
+// Desktop auto-auth installs its token after first paint, so a startup popup can
+// ask before one exists. Wait briefly rather than fail.
+const AUTH_POLL_MS = 250;
+const AUTH_POLL_LIMIT = 40;
+
+interface UseReleaseNotesOptions {
+  version: string | null | undefined;
+  enabled?: boolean;
+}
+
+type ApiObject = Record<string, unknown>;
+
+function stringOrNull(value: ApiObject, key: string): string | null {
+  const field = value[key];
+  return typeof field === "string" && field.length > 0 ? field : null;
+}
+
+function toReleaseNotes(value: unknown, version: string): ReleaseNotes | null {
+  if (!value || typeof value !== "object") {
+    return null;
+  }
+  const payload = value as ApiObject;
+  const notesVersion = stringOrNull(payload, "version");
+  // A response for another version is not usable here.
+  if (notesVersion !== version) {
+    return null;
+  }
+  const markdown = stringOrNull(payload, "markdown");
+  return {
+    version,
+    markdown,
+    matched: payload.matched === true && markdown !== null,
+    truncated: payload.truncated === true,
+    source: stringOrNull(payload, "source"),
+    releaseNotesUrl: stringOrNull(payload, "release_notes_url"),
+    error: stringOrNull(payload, "error"),
+  };
+}
+
+async function fetchReleaseNotes(
+  version: string,
+  refresh = false,
+): Promise<ReleaseNotes | null> {
+  const query = `version=${encodeURIComponent(version)}${refresh ? "&refresh=true" : ""}`;
+  // authFetch, not fetch: an expired token is refreshed and retried.
+  const res = await authFetch(apiUrl(`/api/studio/release-notes?${query}`));
+  if (!res.ok) {
+    throw new Error(`Release notes request failed: ${res.status}`);
+  }
+
+  return toReleaseNotes(await res.json(), version);
+}
+
+export function useReleaseNotes({
+  version,
+  enabled = true,
+}: UseReleaseNotesOptions) {
+  const [state, setState] = useState<ReleaseNotesState>("idle");
+  const [notes, setNotes] = useState<ReleaseNotes | null>(null);
+  // Version the current state belongs to; a change invalidates it.
+  const requestedVersionRef = useRef<string | null>(null);
+  // Identifies one request, so an earlier response cannot overwrite a later one.
+  const requestIdRef = useRef(0);
+
+  const load = useCallback((target: string, refresh = false) => {
+    requestedVersionRef.current = target;
+    requestIdRef.current += 1;
+    const requestId = requestIdRef.current;
+    setState("loading");
+    setNotes(null);
+    fetchReleaseNotes(target, refresh)
+      .then((next) => {
+        // A newer request owns the state now.
+        if (requestIdRef.current !== requestId) {
+          return;
+        }
+        setNotes(next);
+        // A reported failure is retryable; "no notes for this version" is not.
+        const failed = !next || (!next.matched && next.error !== null);
+        setState(failed ? "error" : "ready");
+      })
+      .catch(() => {
+        if (requestIdRef.current === requestId) {
+          setNotes(null);
+          setState("error");
+        }
+      });
+  }, []);
+
+  useEffect(() => {
+    if (!enabled || !version || requestedVersionRef.current === version) {
+      return;
+    }
+    if (hasAuthToken()) {
+      load(version);
+      return;
+    }
+    let attempts = 0;
+    const timer = window.setInterval(() => {
+      attempts += 1;
+      if (hasAuthToken() || attempts >= AUTH_POLL_LIMIT) {
+        window.clearInterval(timer);
+        // Out of patience: load anyway so the panel settles on retry.
+        load(version);
+      }
+    }, AUTH_POLL_MS);
+    return () => window.clearInterval(timer);
+  }, [enabled, version, load]);
+
+  const retry = useCallback(() => {
+    if (version) {
+      requestedVersionRef.current = null;
+      // Bypass the cached remote failure, or retry waits for it to expire.
+      load(version, true);
+    }
+  }, [version, load]);
+
+  // Never hand back another version's notes: state lags `version` by a render.
+  const matchesVersion = notes !== null && notes.version === version;
+  return {
+    state: notes !== null && !matchesVersion ? "loading" : state,
+    notes: matchesVersion ? notes : null,
+    retry,
+  };
+}
diff --git a/studio/frontend/src/hooks/use-sidebar-width.ts b/studio/frontend/src/hooks/use-sidebar-width.ts
new file mode 100644
index 0000000000..c670e254a8
--- /dev/null
+++ b/studio/frontend/src/hooks/use-sidebar-width.ts
@@ -0,0 +1,21 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import { createPanelWidthStore } from "./use-panel-width.ts";
+
+/** The previous fixed 17.5rem, at a 16px root font size. */
+export const SIDEBAR_WIDTH_DEFAULT = 280;
+/** Narrowest width that still fits the wordmark. Firefox is the constraint:
+ * it renders the heading ~3px wider than Chromium and WebKit. */
+export const SIDEBAR_WIDTH_MIN = 260;
+export const SIDEBAR_WIDTH_MAX = 480;
+
+const store = createPanelWidthStore({
+  key: "sidebar_width",
+  min: SIDEBAR_WIDTH_MIN,
+  max: SIDEBAR_WIDTH_MAX,
+  fallback: SIDEBAR_WIDTH_DEFAULT,
+});
+
+export const clampSidebarWidth = store.clamp;
+export const useSidebarWidth = store.useWidth;
diff --git a/studio/frontend/src/hooks/use-tauri-update.ts b/studio/frontend/src/hooks/use-tauri-update.ts
index 8ebb4d2980..196e3cea2b 100644
--- a/studio/frontend/src/hooks/use-tauri-update.ts
+++ b/studio/frontend/src/hooks/use-tauri-update.ts
@@ -21,6 +21,8 @@ export type UpdateStatus =
 export interface UpdateInfo {
   version: string;
   currentVersion: string;
+  // Backend release this build pins; CHANGELOG.md is keyed by it, not the SemVer.
+  pypiVersion?: string;
   body?: string;
   date?: string;
 }
@@ -42,10 +44,17 @@ interface DesktopUpdatePolicy {
 interface ManualUpdateInfo {
   version: string;
   currentVersion: string;
+  pypiVersion?: string | null;
   body?: string;
   date?: string;
 }
 
+/** `pypi_version` from latest.json, which the updater passes through raw. */
+function rawPypiVersion(raw: Record<string, unknown>): string | undefined {
+  const value = raw.pypi_version;
+  return typeof value === "string" && value.length > 0 ? value : undefined;
+}
+
 export interface RetainedUpdateFailure {
   error: string;
   phase: UpdatePhase;
@@ -162,6 +171,7 @@ export function useTauriUpdate(isExternalServer = false) {
       setInfo({
         version: manualUpdate.version,
         currentVersion: manualUpdate.currentVersion,
+        pypiVersion: manualUpdate.pypiVersion ?? undefined,
         body: manualUpdate.body,
         date: manualUpdate.date,
       });
@@ -197,6 +207,7 @@ export function useTauriUpdate(isExternalServer = false) {
           setInfo({
             version: update.version,
             currentVersion: update.currentVersion,
+            pypiVersion: rawPypiVersion(update.rawJson),
             body: update.body,
             date: update.date,
           });
@@ -384,10 +395,13 @@ export function useTauriUpdate(isExternalServer = false) {
     });
   }
 
+  // Install target for Linux packages that cannot self-update.
   const manualReleaseUrl =
     updatePolicy.mode === "manual_linux_package" && info
       ? manualReleasePageUrl(updatePolicy, info.version)
       : null;
+  // Release page for the offered version, on every platform, for the notes link.
+  const releasePageUrl = info ? manualReleasePageUrl(updatePolicy, info.version) : null;
 
   return {
     status,
@@ -401,6 +415,7 @@ export function useTauriUpdate(isExternalServer = false) {
     isExternalServer,
     updatePolicyMode: updatePolicy.mode,
     manualReleaseUrl,
+    releasePageUrl,
     installUpdate,
     retryUpdate,
     skipAndRestart,
diff --git a/studio/frontend/src/i18n/locales/ar.ts b/studio/frontend/src/i18n/locales/ar.ts
index 927470140d..a0ea2a1ee2 100644
--- a/studio/frontend/src/i18n/locales/ar.ts
+++ b/studio/frontend/src/i18n/locales/ar.ts
@@ -27,10 +27,18 @@ export const ar = {
     product: "Unsloth Studio",
     accountMenu: "قائمة حساب {name}",
     updateAvailable: "يتوفر تحديث",
+    resize: {
+      collapse: "انقر للطي",
+      expand: "انقر للتوسيع",
+      drag: "اسحب لتغيير الحجم",
+    },
     aria: {
       home: "الصفحة الرئيسية لـ Unsloth",
       closeSidebar: "إغلاق الشريط الجانبي",
       openSidebar: "فتح الشريط الجانبي",
+      resizeSidebar: "تغيير حجم الشريط الجانبي أو طيه",
+      resizeRunSettings: "تغيير حجم إعدادات التشغيل أو إغلاقها",
+      openRunSettings: "فتح إعدادات التشغيل",
       chatOptions: "خيارات المحادثة",
       runOptions: "خيارات التدريب",
     },
@@ -317,6 +325,8 @@ export const ar = {
         diskUsage: "{used} مستخدم / {total}",
         diskFree: "{free} متاح",
         modelsFolder: "مجلد النماذج",
+        modelsFolderKeywords:
+          "النماذج مجلد دليل مسار موقع تنزيلات التنزيل ذاكرة التخزين المؤقت تخزين قرص محرك نقل تغيير models folder path hugging face",
         modelsFolderDescription: "المكان الذي تُخزَّن فيه النماذج المُنزَّلة.",
         openAction: "فتح",
         copyAction: "نسخ المسار",
@@ -451,7 +461,7 @@ export const ar = {
       codingAgents: "وكلاء البرمجة",
       codingAgentsHint:
         "شغّل وكيل برمجة مقابل هذا الخادم. يستخدم النموذج المُحمَّل؛ الخادم المحلي يُنشئ مفتاح API تلقائيًا، والخادم البعيد يُضمّنه في الأمر.",
-      codingAgentsSwap: "استبدل claude بـ codex أو openclaw أو opencode أو hermes.",
+      codingAgentsSwap: "استبدل claude بـ codex أو openclaw أو opencode أو hermes أو pi.",
       codingAgentDetected: "مثبّت على هذا الجهاز",
       codingAgentsDetectedHint: "تم اكتشافه على هذا الجهاز: {agents}.",
       relativeNever: "أبدًا",
diff --git a/studio/frontend/src/i18n/locales/de.ts b/studio/frontend/src/i18n/locales/de.ts
index bde49dd253..8a5921c623 100644
--- a/studio/frontend/src/i18n/locales/de.ts
+++ b/studio/frontend/src/i18n/locales/de.ts
@@ -27,10 +27,18 @@ export const de = {
     product: "Unsloth Studio",
     accountMenu: "Kontomenü von {name}",
     updateAvailable: "Update verfügbar",
+    resize: {
+      collapse: "Zum Einklappen klicken",
+      expand: "Zum Ausklappen klicken",
+      drag: "Zum Ändern der Größe ziehen",
+    },
     aria: {
       home: "Unsloth Startseite",
       closeSidebar: "Seitenleiste schließen",
       openSidebar: "Seitenleiste öffnen",
+      resizeSidebar: "Seitenleiste anpassen oder einklappen",
+      resizeRunSettings: "Ausführungseinstellungen anpassen oder schließen",
+      openRunSettings: "Ausführungseinstellungen öffnen",
       chatOptions: "Chat-Optionen",
       runOptions: "Trainingslauf-Optionen",
     },
@@ -330,9 +338,23 @@ export const de = {
         diskFree: "{free} frei",
         modelsFolder: "Modell-Ordner",
         modelsFolderDescription:
-          "Wo heruntergeladene Modelle gespeichert werden.",
+          "Wo heruntergeladene Modelle gespeichert werden. Ändern Sie ihn, um Modelle nicht auf dem Systemlaufwerk abzulegen.",
+        modelsFolderKeywords:
+          "Modelle Ordner Verzeichnis Pfad Speicherort Download Downloads Cache Speicher Festplatte Laufwerk verschieben ändern hugging face",
+        futureDownloads: "Nur neue Downloads",
+        environmentManaged:
+          "Wird über die Umgebungsvariable {variable} verwaltet.",
+        locationFree: "{free} frei",
         openAction: "Öffnen",
         copyAction: "Pfad kopieren",
+        changeAction: "Ändern",
+        resetAction: "Standard verwenden",
+        chooseTitle: "Speicherort für Modell-Downloads wählen",
+        chooseAction: "Für künftige Downloads verwenden",
+        cacheSaved: "Speicherort für Modell-Downloads aktualisiert",
+        cacheSaveError:
+          "Der Speicherort für Modell-Downloads konnte nicht geändert werden",
+        cachePickerError: "Die Ordnerauswahl konnte nicht geöffnet werden",
         copied: "Pfad kopiert",
         openError: "Der Ordner konnte nicht geöffnet werden",
         copyError: "Der Pfad konnte nicht kopiert werden",
@@ -469,7 +491,7 @@ export const de = {
       codingAgentsHint:
         "Starten Sie einen Coding-Agent gegen diesen Server. Er verwendet das geladene Modell; ein lokaler Server erstellt automatisch einen API-Schlüssel, ein entfernter fügt ihn dem Befehl hinzu.",
       codingAgentsSwap:
-        "Ersetzen Sie claude durch codex, openclaw, opencode oder hermes.",
+        "Ersetzen Sie claude durch codex, openclaw, opencode, hermes oder pi.",
       codingAgentDetected: "Auf diesem Gerät installiert",
       codingAgentsDetectedHint: "Auf diesem Gerät erkannt: {agents}.",
       relativeNever: "nie",
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index 642b224d2a..10e571f4da 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -24,10 +24,18 @@ export const en = {
     product: "Unsloth Studio",
     accountMenu: "{name} account menu",
     updateAvailable: "Update available",
+    resize: {
+      collapse: "Click to collapse",
+      expand: "Click to expand",
+      drag: "Drag to resize",
+    },
     aria: {
       home: "Unsloth home",
       closeSidebar: "Close sidebar",
       openSidebar: "Open sidebar",
+      resizeSidebar: "Resize or collapse sidebar",
+      resizeRunSettings: "Resize or close run settings",
+      openRunSettings: "Open run settings",
       chatOptions: "Chat options",
       runOptions: "Run options",
     },
@@ -556,8 +564,12 @@ export const en = {
         systemDisk: "System disk",
         diskUsage: "{used} used / {total}",
         diskFree: "{free} free",
-        modelsFolder: "Model downloads",
-        modelsFolderDescription: "Hugging Face cache used for model downloads.",
+        modelsFolder: "Models folder",
+        modelsFolderDescription:
+          "Where downloaded models are stored. Change it to keep models off your system drive.",
+        // Not rendered: extra terms the settings search matches this row on.
+        modelsFolderKeywords:
+          "models folder directory path location download downloads cache storage disk drive move relocate hugging face",
         futureDownloads: "New downloads only",
         environmentManaged: "Managed by the {variable} environment variable.",
         locationFree: "{free} free",
@@ -589,9 +601,9 @@ export const en = {
     agents: {
       title: "Agents",
       description:
-        "Connect coding agents like Claude Code and Codex to a local model with unsloth start.",
+        "Connect coding agents like Claude Code and Codex to a model running locally in Unsloth with unsloth start.",
       intro:
-        "connects Claude Code, Codex, Hermes, OpenClaw, OpenCode and other agents to a model served locally by Unsloth, fully offline. It runs an OpenAI-compatible server and never touches your agent's config files.",
+        "connects Claude Code, Codex, Hermes, OpenClaw, OpenCode, Pi and other agents to a model served locally by Unsloth, fully offline on your own hardware. It runs an OpenAI-compatible server for the agent and never touches your agent's config files.",
       readDocs: "Read the docs",
       copy: "Copy",
       copied: "Copied",
@@ -660,7 +672,7 @@ export const en = {
         launch: "Launch the agent, or just print the command and environment.",
         persist: "Keep Unsloth-managed agent storage between runs.",
         asSubagent:
-          "Keep the parent on its current model and register Unsloth as a local subagent (Claude Code, Codex, and OpenCode).",
+          "Keep the parent on its current model and register Unsloth as a local subagent (Claude Code, Codex, OpenCode, and Pi).",
         apiKey: "Provide your Unsloth API key (or set UNSLOTH_API_KEY).",
         yolo: "Skip approval prompts. Use only in trusted environments.",
       },
@@ -843,7 +855,8 @@ export const en = {
       codingAgents: "Coding agents",
       codingAgentsHint:
         "Launch a coding agent against this server. It uses the loaded model; a local server mints an API key automatically, a remote one includes it in the command.",
-      codingAgentsSwap: "Swap claude for codex, openclaw, opencode, or hermes.",
+      codingAgentsSwap:
+        "Swap claude for codex, openclaw, opencode, hermes, or pi.",
       codingAgentDetected: "Installed on this machine",
       codingAgentsDetectedHint: "Detected on this machine: {agents}.",
       relativeNever: "never",
diff --git a/studio/frontend/src/i18n/locales/es.ts b/studio/frontend/src/i18n/locales/es.ts
index 66d5281701..6edc33f9de 100644
--- a/studio/frontend/src/i18n/locales/es.ts
+++ b/studio/frontend/src/i18n/locales/es.ts
@@ -27,10 +27,18 @@ export const es = {
     product: "Unsloth Studio",
     accountMenu: "Menú de cuenta de {name}",
     updateAvailable: "Actualización disponible",
+    resize: {
+      collapse: "Haz clic para contraer",
+      expand: "Haz clic para expandir",
+      drag: "Arrastra para redimensionar",
+    },
     aria: {
       home: "Inicio de Unsloth",
       closeSidebar: "Cerrar barra lateral",
       openSidebar: "Abrir barra lateral",
+      resizeSidebar: "Redimensionar o contraer la barra lateral",
+      resizeRunSettings: "Redimensionar o cerrar los ajustes de ejecución",
+      openRunSettings: "Abrir los ajustes de ejecución",
       chatOptions: "Opciones de chat",
       runOptions: "Opciones de ejecución",
     },
@@ -328,6 +336,8 @@ export const es = {
         diskUsage: "{used} en uso / {total}",
         diskFree: "{free} libre",
         modelsFolder: "Carpeta de modelos",
+        modelsFolderKeywords:
+          "modelos carpeta directorio ruta ubicacion ubicación descargas descarga cache caché almacenamiento disco unidad mover cambiar models folder path hugging face",
         modelsFolderDescription:
           "Dónde se almacenan los modelos descargados.",
         openAction: "Abrir",
@@ -466,7 +476,7 @@ export const es = {
       codingAgentsHint:
         "Inicia un agente de programación contra este servidor. Usa el modelo cargado; un servidor local genera una clave de API automáticamente y uno remoto la incluye en el comando.",
       codingAgentsSwap:
-        "Reemplaza claude por codex, openclaw, opencode o hermes.",
+        "Reemplaza claude por codex, openclaw, opencode, hermes o pi.",
       codingAgentDetected: "Instalado en esta máquina",
       codingAgentsDetectedHint: "Detectados en esta máquina: {agents}.",
       relativeNever: "nunca",
diff --git a/studio/frontend/src/i18n/locales/fr.ts b/studio/frontend/src/i18n/locales/fr.ts
index 152577174c..789e89079c 100644
--- a/studio/frontend/src/i18n/locales/fr.ts
+++ b/studio/frontend/src/i18n/locales/fr.ts
@@ -27,10 +27,18 @@ export const fr = {
     product: "Unsloth Studio",
     accountMenu: "Menu du compte de {name}",
     updateAvailable: "Mise à jour disponible",
+    resize: {
+      collapse: "Cliquez pour réduire",
+      expand: "Cliquez pour développer",
+      drag: "Faites glisser pour redimensionner",
+    },
     aria: {
       home: "Accueil Unsloth",
       closeSidebar: "Fermer la barre latérale",
       openSidebar: "Ouvrir la barre latérale",
+      resizeSidebar: "Redimensionner ou réduire la barre latérale",
+      resizeRunSettings: "Redimensionner ou fermer les paramètres d'exécution",
+      openRunSettings: "Ouvrir les paramètres d'exécution",
       chatOptions: "Options de discussion",
       runOptions: "Options d'exécution",
     },
@@ -325,6 +333,8 @@ export const fr = {
         diskUsage: "{used} utilisé / {total}",
         diskFree: "{free} libre",
         modelsFolder: "Dossier des modèles",
+        modelsFolderKeywords:
+          "modeles modèles dossier repertoire répertoire chemin emplacement telechargements téléchargements cache stockage disque lecteur deplacer déplacer changer models folder path hugging face",
         modelsFolderDescription: "Emplacement de stockage des modèles téléchargés.",
         openAction: "Ouvrir",
         copyAction: "Copier le chemin",
@@ -464,7 +474,7 @@ export const fr = {
       codingAgentsHint:
         "Lancez un agent de codage sur ce serveur. Il utilise le modèle chargé ; un serveur local génère automatiquement une clé API, un serveur distant l'inclut dans la commande.",
       codingAgentsSwap:
-        "Remplacez claude par codex, openclaw, opencode ou hermes.",
+        "Remplacez claude par codex, openclaw, opencode, hermes ou pi.",
       codingAgentDetected: "Installé sur cette machine",
       codingAgentsDetectedHint: "Détecté sur cette machine : {agents}.",
       relativeNever: "jamais",
diff --git a/studio/frontend/src/i18n/locales/hi.ts b/studio/frontend/src/i18n/locales/hi.ts
index c04de206af..410318983b 100644
--- a/studio/frontend/src/i18n/locales/hi.ts
+++ b/studio/frontend/src/i18n/locales/hi.ts
@@ -27,10 +27,18 @@ export const hi = {
     product: "Unsloth Studio",
     accountMenu: "{name} खाता मेनू",
     updateAvailable: "अपडेट उपलब्ध है",
+    resize: {
+      collapse: "छोटा करने के लिए क्लिक करें",
+      expand: "विस्तार के लिए क्लिक करें",
+      drag: "आकार बदलने के लिए खींचें",
+    },
     aria: {
       home: "Unsloth होम",
       closeSidebar: "साइडबार बंद करें",
       openSidebar: "साइडबार खोलें",
+      resizeSidebar: "साइडबार का आकार बदलें या छोटा करें",
+      resizeRunSettings: "रन सेटिंग्स का आकार बदलें या बंद करें",
+      openRunSettings: "रन सेटिंग्स खोलें",
       chatOptions: "चैट विकल्प",
       runOptions: "रन विकल्प",
     },
@@ -316,6 +324,8 @@ export const hi = {
         diskUsage: "{used} उपयोग में / {total}",
         diskFree: "{free} खाली",
         modelsFolder: "मॉडल फ़ोल्डर",
+        modelsFolderKeywords:
+          "मॉडल फ़ोल्डर फोल्डर निर्देशिका पथ स्थान डाउनलोड कैश संग्रहण डिस्क ड्राइव स्थानांतरित बदलें models folder path hugging face",
         modelsFolderDescription: "जहां डाउनलोड किए गए मॉडल संग्रहीत होते हैं।",
         openAction: "खोलें",
         copyAction: "पथ कॉपी करें",
@@ -449,7 +459,7 @@ export const hi = {
       codingAgents: "कोडिंग एजेंट",
       codingAgentsHint:
         "इस सर्वर के विरुद्ध एक कोडिंग एजेंट लॉन्च करें। यह लोड किए गए मॉडल का उपयोग करता है; एक स्थानीय सर्वर स्वचालित रूप से API key बनाता है, एक रिमोट सर्वर इसे कमांड में शामिल करता है।",
-      codingAgentsSwap: "claude को codex, openclaw, opencode, या hermes से बदलें।",
+      codingAgentsSwap: "claude को codex, openclaw, opencode, hermes, या pi से बदलें।",
       codingAgentDetected: "इस मशीन पर इंस्टॉल है",
       codingAgentsDetectedHint: "इस मशीन पर पाया गया: {agents}।",
       relativeNever: "कभी नहीं",
diff --git a/studio/frontend/src/i18n/locales/ja.ts b/studio/frontend/src/i18n/locales/ja.ts
index e38dfa59e5..1653bad76e 100644
--- a/studio/frontend/src/i18n/locales/ja.ts
+++ b/studio/frontend/src/i18n/locales/ja.ts
@@ -28,10 +28,18 @@ export const ja = {
     product: "Unsloth Studio",
     accountMenu: "{name} のアカウントメニュー",
     updateAvailable: "アップデートが利用可能です",
+    resize: {
+      collapse: "クリックで折りたたむ",
+      expand: "クリックで展開",
+      drag: "ドラッグでサイズ変更",
+    },
     aria: {
       home: "Unsloth ホーム",
       closeSidebar: "サイドバーを閉じる",
       openSidebar: "サイドバーを開く",
+      resizeSidebar: "サイドバーのサイズ変更または折りたたみ",
+      resizeRunSettings: "実行設定のサイズ変更または閉じる",
+      openRunSettings: "実行設定を開く",
       chatOptions: "チャットオプション",
       runOptions: "実行オプション",
     },
@@ -393,6 +401,8 @@ export const ja = {
         diskUsage: "{used} 使用中 / {total}",
         diskFree: "{free} 空き",
         modelsFolder: "モデルフォルダ",
+        modelsFolderKeywords:
+          "モデル フォルダ ディレクトリ パス 保存先 場所 ダウンロード キャッシュ ストレージ ディスク ドライブ 移動 変更 models folder path hugging face",
         modelsFolderDescription: "ダウンロードしたモデルの保存先。",
         openAction: "開く",
         copyAction: "パスをコピー",
@@ -511,7 +521,7 @@ export const ja = {
       setupDocs: "セットアップドキュメント:",
       codingAgents: "コーディングエージェント",
       codingAgentsHint: "このサーバーに対してコーディングエージェントを起動します。読み込み済みのモデルを使用します。ローカルサーバーでは API キーが自動的に発行され、リモートサーバーではコマンドに含まれます。",
-      codingAgentsSwap: "claude を codex、openclaw、opencode、hermes に置き換えられます。",
+      codingAgentsSwap: "claude を codex、openclaw、opencode、hermes、pi に置き換えられます。",
       codingAgentDetected: "このマシンにインストール済み",
       codingAgentsDetectedHint: "このマシンで検出されました: {agents}。",
       relativeNever: "なし",
diff --git a/studio/frontend/src/i18n/locales/ko.ts b/studio/frontend/src/i18n/locales/ko.ts
index c159a8befc..b0da314896 100644
--- a/studio/frontend/src/i18n/locales/ko.ts
+++ b/studio/frontend/src/i18n/locales/ko.ts
@@ -27,10 +27,18 @@ export const ko = {
     product: "Unsloth Studio",
     accountMenu: "{name} 계정 메뉴",
     updateAvailable: "업데이트 사용 가능",
+    resize: {
+      collapse: "클릭하여 접기",
+      expand: "클릭하여 펼치기",
+      drag: "드래그하여 크기 조절",
+    },
     aria: {
       home: "Unsloth 홈",
       closeSidebar: "사이드바 닫기",
       openSidebar: "사이드바 열기",
+      resizeSidebar: "사이드바 크기 조절 또는 접기",
+      resizeRunSettings: "실행 설정 크기 조절 또는 닫기",
+      openRunSettings: "실행 설정 열기",
       chatOptions: "채팅 옵션",
       runOptions: "학습 옵션",
     },
@@ -315,6 +323,8 @@ export const ko = {
         diskUsage: "{used} 사용 중 / {total}",
         diskFree: "{free} 여유",
         modelsFolder: "모델 폴더",
+        modelsFolderKeywords:
+          "모델 폴더 디렉터리 디렉토리 경로 위치 저장 다운로드 캐시 저장소 디스크 드라이브 이동 변경 models folder path hugging face",
         modelsFolderDescription: "다운로드한 모델이 저장되는 위치입니다.",
         openAction: "열기",
         copyAction: "경로 복사",
@@ -450,7 +460,7 @@ export const ko = {
       codingAgents: "코딩 에이전트",
       codingAgentsHint:
         "이 서버를 대상으로 코딩 에이전트를 실행합니다. 로드된 모델을 사용하며, 로컬 서버는 API 키를 자동으로 발급하고 원격 서버는 명령에 포함합니다.",
-      codingAgentsSwap: "claude를 codex, openclaw, opencode 또는 hermes로 바꾸세요.",
+      codingAgentsSwap: "claude를 codex, openclaw, opencode, hermes 또는 pi로 바꾸세요.",
       codingAgentDetected: "이 컴퓨터에 설치됨",
       codingAgentsDetectedHint: "이 컴퓨터에서 감지됨: {agents}.",
       relativeNever: "없음",
diff --git a/studio/frontend/src/i18n/locales/pt-br.ts b/studio/frontend/src/i18n/locales/pt-br.ts
index 322f0af5e4..e9f23623f8 100644
--- a/studio/frontend/src/i18n/locales/pt-br.ts
+++ b/studio/frontend/src/i18n/locales/pt-br.ts
@@ -27,10 +27,18 @@ export const ptBR = {
     product: "Unsloth Studio",
     accountMenu: "Menu de conta {name}",
     updateAvailable: "Atualização disponível",
+    resize: {
+      collapse: "Clique para recolher",
+      expand: "Clique para expandir",
+      drag: "Arraste para redimensionar",
+    },
     aria: {
       home: "Início do Unsloth",
       closeSidebar: "Fechar barra lateral",
       openSidebar: "Abrir barra lateral",
+      resizeSidebar: "Redimensionar ou recolher a barra lateral",
+      resizeRunSettings: "Redimensionar ou fechar as configurações de execução",
+      openRunSettings: "Abrir as configurações de execução",
       chatOptions: "Opções de chat",
       runOptions: "Opções de execução",
     },
@@ -417,6 +425,8 @@ export const ptBR = {
         diskUsage: "{used} usados / {total}",
         diskFree: "{free} livres",
         modelsFolder: "Pasta de modelos",
+        modelsFolderKeywords:
+          "modelos pasta diretorio diretório caminho local localizacao localização downloads baixar cache armazenamento disco unidade mover alterar models folder path hugging face",
         modelsFolderDescription: "Onde os modelos baixados são armazenados.",
         openAction: "Abrir",
         copyAction: "Copiar caminho",
@@ -552,7 +562,7 @@ export const ptBR = {
       codingAgents: "Agentes de código",
       codingAgentsHint:
         "Inicie um agente de código conectado a este servidor. Ele usa o modelo carregado; um servidor local gera uma chave de API automaticamente, um remoto a inclui no comando.",
-      codingAgentsSwap: "Troque claude por codex, openclaw, opencode ou hermes.",
+      codingAgentsSwap: "Troque claude por codex, openclaw, opencode, hermes ou pi.",
       codingAgentDetected: "Instalado nesta máquina",
       codingAgentsDetectedHint: "Detectado nesta máquina: {agents}.",
       relativeNever: "nunca",
diff --git a/studio/frontend/src/i18n/locales/ru.ts b/studio/frontend/src/i18n/locales/ru.ts
index 36f29f1d22..364f224802 100644
--- a/studio/frontend/src/i18n/locales/ru.ts
+++ b/studio/frontend/src/i18n/locales/ru.ts
@@ -27,10 +27,18 @@ export const ru = {
     product: "Unsloth Studio",
     accountMenu: "Меню аккаунта {name}",
     updateAvailable: "Доступно обновление",
+    resize: {
+      collapse: "Нажмите, чтобы свернуть",
+      expand: "Нажмите, чтобы развернуть",
+      drag: "Потяните, чтобы изменить размер",
+    },
     aria: {
       home: "Главная Unsloth",
       closeSidebar: "Закрыть боковую панель",
       openSidebar: "Открыть боковую панель",
+      resizeSidebar: "Изменить размер или свернуть боковую панель",
+      resizeRunSettings: "Изменить размер или закрыть настройки запуска",
+      openRunSettings: "Открыть настройки запуска",
       chatOptions: "Параметры чата",
       runOptions: "Параметры запуска",
     },
@@ -316,6 +324,8 @@ export const ru = {
         diskUsage: "{used} использовано / {total}",
         diskFree: "{free} свободно",
         modelsFolder: "Папка моделей",
+        modelsFolderKeywords:
+          "модели папка каталог путь расположение загрузки кэш хранилище диск перенести изменить models folder path hugging face",
         modelsFolderDescription: "Где хранятся загруженные модели.",
         openAction: "Открыть",
         copyAction: "Копировать путь",
@@ -449,7 +459,7 @@ export const ru = {
       codingAgents: "Кодинг-агенты",
       codingAgentsHint:
         "Запустите кодинг-агента к этому серверу. Он использует загруженную модель; локальный сервер выпускает ключ API автоматически, удалённый включает его в команду.",
-      codingAgentsSwap: "Замените claude на codex, openclaw, opencode или hermes.",
+      codingAgentsSwap: "Замените claude на codex, openclaw, opencode, hermes или pi.",
       codingAgentDetected: "Установлен на этой машине",
       codingAgentsDetectedHint: "Обнаружены на этой машине: {agents}.",
       relativeNever: "никогда",
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index dda5dd481c..777bfbc4dd 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -27,10 +27,18 @@ export const zhCN = {
     product: "Unsloth Studio",
     accountMenu: "{name} 账号菜单",
     updateAvailable: "有可用更新",
+    resize: {
+      collapse: "点击折叠",
+      expand: "点击展开",
+      drag: "拖动调整大小",
+    },
     aria: {
       home: "Unsloth 首页",
       closeSidebar: "关闭侧边栏",
       openSidebar: "打开侧边栏",
+      resizeSidebar: "调整或折叠侧边栏",
+      resizeRunSettings: "调整或关闭运行设置",
+      openRunSettings: "打开运行设置",
       chatOptions: "聊天选项",
       runOptions: "训练选项",
     },
@@ -408,6 +416,8 @@ export const zhCN = {
         diskUsage: "已用 {used} / {total}",
         diskFree: "{free} 可用",
         modelsFolder: "模型文件夹",
+        modelsFolderKeywords:
+          "模型 文件夹 目录 路径 位置 下载 缓存 存储 磁盘 驱动器 移动 更改 models folder path hugging face",
         modelsFolderDescription: "已下载模型的存储位置。",
         openAction: "打开",
         copyAction: "复制路径",
@@ -539,7 +549,7 @@ export const zhCN = {
       codingAgents: "编程智能体",
       codingAgentsHint:
         "针对此服务器启动编程智能体。它会使用已加载的模型；本地服务器会自动生成 API 密钥，远程服务器则会将其包含在命令中。",
-      codingAgentsSwap: "可将 claude 替换为 codex、openclaw、opencode 或 hermes。",
+      codingAgentsSwap: "可将 claude 替换为 codex、openclaw、opencode、hermes 或 pi。",
       codingAgentDetected: "已安装在本机",
       codingAgentsDetectedHint: "本机检测到：{agents}。",
       relativeNever: "从未",
diff --git a/studio/frontend/src/index.css b/studio/frontend/src/index.css
index e298dbbe8c..4ffd29ad7c 100644
--- a/studio/frontend/src/index.css
+++ b/studio/frontend/src/index.css
@@ -664,20 +664,6 @@ html[data-chat-font] .aui-root {
    in the same style pass that swaps the color tokens. Driving it from
    React state paints one frame late while the whole page recalculates,
    which briefly leaves the ring on the previous card. */
-
-/* Dark mode drops the resting border. Same specificity as, and before, the
-   hover/selected rules below so those still paint a visible border. */
-.dark .palette-card {
-	border-color: transparent;
-}
-
-/* The card removes its outline, so the border is the only focus indicator.
-   This rule is unlayered like the one above, which would otherwise beat
-   Tailwind's layered focus-visible:border-ring and leave nothing visible. */
-.dark .palette-card:focus-visible {
-	border-color: var(--ring);
-}
-
 .palette-card:hover {
 	border-color: var(--ring-soft);
 }
@@ -1377,6 +1363,23 @@ html[data-chat-font] .aui-root {
 		cursor: pointer;
 	}
 
+	/* While a panel edge is dragged, keep the resize cursor even as the pointer
+	   travels over buttons and text that would claim their own. */
+	html[data-panel-resizing],
+	html[data-panel-resizing] * {
+		cursor: col-resize !important;
+		user-select: none !important;
+	}
+
+	html[data-panel-resizing]
+		:is(
+			[data-slot="sidebar-inner"],
+			[data-slot="sidebar-inset"],
+			[data-slot="chat-settings-panel"] > div
+		) {
+		pointer-events: none;
+	}
+
 	/* Model selector: pointer cursor on every clickable element. */
 	.unsloth-model-selector-trigger,
 	.unsloth-model-selector-menu button {
diff --git a/studio/frontend/src/lib/changelog-links.ts b/studio/frontend/src/lib/changelog-links.ts
new file mode 100644
index 0000000000..16b3d3c8bc
--- /dev/null
+++ b/studio/frontend/src/lib/changelog-links.ts
@@ -0,0 +1,664 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+/**
+ * A relative link in CHANGELOG.md means "somewhere in the Unsloth repository",
+ * but inside Studio it would resolve against Studio's own origin. Rewriting to
+ * absolute repository URLs makes them behave the way GitHub renders the file.
+ */
+
+import {
+  type CodeSpan,
+  codeSpans,
+  insideSpan,
+} from "@/lib/markdown-code-spans";
+import { commentClosesBelow } from "@/lib/markdown-inline-comments";
+import {
+  EMPTY_LIST_STATE,
+  type ListState,
+  NO_QUOTE,
+  type QuoteState,
+  containerContent,
+  hiddenStructure,
+  indentWidth,
+  itemContent,
+  openLists,
+  quoteDepth,
+  quoteState,
+} from "@/lib/markdown-list-columns";
+
+const LINK_BASE = "https://github.com/unslothai/unsloth/blob/main/";
+const IMAGE_BASE = "https://raw.githubusercontent.com/unslothai/unsloth/main/";
+
+// Inline `](dest)` plus the `[label]: dest` reference form. The destination is
+// either <bracketed> or runs to whitespace or the closing paren.
+const NESTED_LABEL = String.raw`((?:[^[\]\\]|\\.|\[(?:[^[\]\\]|\\.)*\])*)`;
+// Only ASCII punctuation is escapable, so the backslash in `a\ b.md` is an
+// ordinary character of the destination and the space still ends it.
+const ESCAPABLE = String.raw`[!-/:-@[-\`{-~]`;
+const DESTINATION_CHAR = String.raw`\\${ESCAPABLE}|[^\s()]`;
+// A destination may hold balanced parentheses, and a path may nest them, so
+// `[x](((draft)).md)` points at `((draft)).md`. An expression cannot count, so
+// pairs are unrolled to the depth cmark stops at, which is what GitHub renders.
+const MAX_DESTINATION_NESTING = 32;
+
+/** A balanced parenthesised run nested up to `depth` levels deep. */
+function nestedParens(depth: number): string {
+  let group = String.raw`\((?:${DESTINATION_CHAR})*\)`;
+  for (let left = depth - 1; left > 0; left -= 1) {
+    group = String.raw`\((?:${DESTINATION_CHAR}|${group})*\)`;
+  }
+  return group;
+}
+
+const BALANCED_DESTINATION = String.raw`(?:${DESTINATION_CHAR}|${nestedParens(MAX_DESTINATION_NESTING)})*`;
+const PLAIN_DESTINATION = String.raw`(?:${DESTINATION_CHAR})*`;
+// A balanced pair counts only while a `)` or a title still closes the link
+// after it, or swallowing it would invent a link across lines.
+const CLOSES_LINK = String.raw`(?=[ \t]*[)'"])`;
+// A destination that runs out of line has its closer below it, the line being
+// only part of the link. One stopping short of a closer is no destination at all,
+// so `[x](a b.md)` and `[x](a(b.md)` stay plain text and keep the paths they name.
+const CLOSES_OR_ENDS_LINE = String.raw`(?=[ \t]*(?:[)'"]|$))`;
+const INLINE_TARGET = new RegExp(
+  String.raw`(!?)\[${NESTED_LABEL}\]\(\s*(<[^<>\n]*>|${BALANCED_DESTINATION}${CLOSES_LINK}|${PLAIN_DESTINATION}${CLOSES_OR_ENDS_LINE})`,
+  "g",
+);
+const REFERENCE_TARGET = /^( {0,3}\[((?:[^[\]\\]|\\.)*)\]:\s*)(<[^<>\n]*>|\S+)/;
+// `![alt][label]`, `![label][]` and `![label]`: a definition they point at
+// has to resolve to the raw file, not to its page on GitHub.
+const IMAGE_REFERENCE =
+  /!\[((?:[^[\]\\]|\\.)*)\](?:\[((?:[^[\]\\]|\\.)*)\]|(?!\())/g;
+const FENCE = /^ {0,3}(`{3,}|~{3,})(.*)$/;
+// Four columns past the container start indented code, unless a paragraph is
+// open. Inside a list item that is measured from the item's content column, so a
+// link indented under a bullet is prose and still resolves.
+const INDENTED_CODE_INDENT = 4;
+// CommonMark type 1 HTML blocks show their contents verbatim.
+const RAW_HTML_OPEN = /^ {0,3}<(pre|script|style|textarea)(?=[\s>]|$)/i;
+const RAW_HTML_CLOSE = /<\/(pre|script|style|textarea)\s*>/i;
+// Type 6 and 7 blocks are literal too and run to the next blank line, not to a
+// closing tag, so `<details>` holds Markdown only after a blank line. Type 7 (any
+// other complete tag alone on a line) cannot interrupt a paragraph.
+const HTML_BLOCK_OPEN = /^ {0,3}<\/?([a-zA-Z][a-zA-Z0-9-]*)(?=[\s/>]|$)/;
+const HTML_ATTRIBUTE =
+  "(?:\\s+[a-zA-Z_:][a-zA-Z0-9_.:-]*(?:\\s*=\\s*(?:[^\\s\"'=<>`]+|'[^']*'|\"[^\"]*\"))?)";
+const HTML_TAG_ONLY_LINE = new RegExp(
+  `^ {0,3}(?:<[a-zA-Z][a-zA-Z0-9-]*${HTML_ATTRIBUTE}*\\s*/?>|</[a-zA-Z][a-zA-Z0-9-]*\\s*>)\\s*$`,
+);
+const HTML_BLOCK_TAGS = new Set(
+  `address article aside base basefont blockquote body caption center col colgroup
+   dd details dialog dir div dl dt fieldset figcaption figure footer form frame
+   frameset h1 h2 h3 h4 h5 h6 head header hr html iframe legend li link main menu
+   menuitem nav noframes ol optgroup option p param search section summary table
+   tbody td tfoot th thead title tr track ul`.split(/\s+/),
+);
+// Lines that are blocks in their own right, so no paragraph is open after.
+const BLOCK_LINE =
+  /^ {0,3}(?:#{1,6}([ \t]|$)|(?:\*[ \t]*){3,}$|(?:-[ \t]*){3,}$|(?:_[ \t]*){3,}$|>|=+[ \t]*$)/;
+// A definition is a block of its own but may not interrupt a paragraph, so it
+// ends the one above only when there is none to continue. It opens none either,
+// or consecutive definitions could never start (spec 0.31.2 section 4.7). Same
+// rule as `_LINK_DEFINITION` in the backend's `after_paragraph`.
+const LINK_DEFINITION = /^ {0,3}\[(?:[^[\]\\]|\\.)+\]:/;
+const LINE_ENDINGS = /\r\n?/g;
+// A scheme, a protocol-relative host, or a fragment: already absolute enough.
+// `//` needs a host after it, so `///docs` stays a repository path.
+const ABSOLUTE = /^(?:[a-zA-Z][a-zA-Z0-9+.-]*:|\/\/[^/]|#)/;
+
+const COMMENT_OPEN = "<!--";
+const COMMENT_CLOSE = "-->";
+const COMMENT_BLOCK_OPEN = /^ {0,3}<!--/;
+
+/**
+ * `line` with its commented spans blanked, and whether a comment block is still
+ * open below it. Commented content renders as nothing, so it holds no fence,
+ * block or code span. Lengths are preserved so offsets still line up.
+ *
+ * Only a comment that starts a line opens a block (CommonMark type 2), and only
+ * that runs on to the line holding `-->`, tail included. One written mid-sentence
+ * is inline raw HTML belonging to its paragraph, so its `-->` may arrive on a
+ * later line and only the text up to it is hidden. `closesBelow` says one does;
+ * without it the opener is ordinary text, so a note merely mentioning `<!--` must
+ * not hide the links below it.
+ *
+ * "Starts a line" is read inside the container, so `blockOpen` comes from the
+ * item's content rather than the raw line.
+ */
+function maskComments(
+  line: string,
+  inComment: boolean,
+  runOn: boolean,
+  closesBelow: boolean,
+  blockOpen: boolean,
+): [string, boolean, boolean] {
+  if (inComment) {
+    // The closing line belongs to the block, tail included.
+    return [" ".repeat(line.length), !line.includes(COMMENT_CLOSE), false];
+  }
+  if (runOn) {
+    const closed = line.indexOf(COMMENT_CLOSE);
+    if (closed < 0) {
+      return [" ".repeat(line.length), false, true];
+    }
+    // Only up to the closer: the tail is the paragraph's own text again.
+    const resumed = closed + COMMENT_CLOSE.length;
+    return maskInline(line, resumed, closesBelow);
+  }
+  if (blockOpen) {
+    // `<!-->` and `<!--->` are complete comments, so the closer may overlap the
+    // opener; searching past it would blank the rest of the file.
+    return [" ".repeat(line.length), !line.includes(COMMENT_CLOSE), false];
+  }
+  return maskInline(line, 0, closesBelow);
+}
+
+/** `maskComments` from `from`, where no comment block is open. */
+function maskInline(
+  line: string,
+  from: number,
+  closesBelow: boolean,
+): [string, boolean, boolean] {
+  let out = " ".repeat(from);
+  let index = from;
+  // Scanned only once an opener turns up. Spans are ordered and disjoint and each
+  // opener sits at or past the last, so the search resumes rather than restarts.
+  let spans: CodeSpan[] | null = null;
+  let cursor = 0;
+  while (index < line.length) {
+    const start = line.indexOf(COMMENT_OPEN, index);
+    if (start < 0) {
+      return [out + line.slice(index), false, false];
+    }
+    spans ??= codeSpans(line);
+    while (cursor < spans.length && (spans[cursor]?.end ?? 0) <= start) {
+      cursor += 1;
+    }
+    // A delimiter inside inline code is literal, not a comment opener.
+    const span = spans[cursor];
+    if (span !== undefined && span.start <= start) {
+      out += line.slice(index, span.end);
+      index = span.end;
+      continue;
+    }
+    // `<!-->` and `<!--->` are complete comments, so the closer may overlap.
+    const close = line.indexOf(COMMENT_CLOSE, start + 2);
+    if (close < 0) {
+      if (closesBelow) {
+        // The paragraph carries the comment on, so the line from the opener is
+        // inside it, and so is the line below.
+        return [
+          out + line.slice(index, start) + " ".repeat(line.length - start),
+          false,
+          true,
+        ];
+      }
+      // Nothing closes it at all, so the renderer shows it as ordinary text.
+      return [out + line.slice(index), false, false];
+    }
+    out += line.slice(index, start);
+    out += " ".repeat(close + COMMENT_CLOSE.length - start);
+    index = close + COMMENT_CLOSE.length;
+  }
+  return [out, false, false];
+}
+
+/**
+ * Whether `line` is written outside the container an open block belongs to. A
+ * fence and an HTML block hold no lazy continuation line, so content left of the
+ * item, or outside the quote, ends the block with its container. A raw block or
+ * comment inside a list item ends on a blank line too: the item takes the break,
+ * so what follows is a block of the item's own.
+ */
+function leavesContainer(
+  line: string,
+  quotes: number,
+  column: number,
+  blockQuotes: number,
+  rawInItem: boolean,
+): boolean {
+  if (quotes < blockQuotes) {
+    return true;
+  }
+  if (!line.trim()) {
+    return rawInItem;
+  }
+  return column > 0 && indentWidth(line) < column;
+}
+
+/** True if `line` starts a CommonMark type 6 or type 7 HTML block. */
+function opensHtmlBlock(line: string, afterParagraph: boolean): boolean {
+  const named = HTML_BLOCK_OPEN.exec(line);
+  if (named && HTML_BLOCK_TAGS.has((named[1] ?? "").toLowerCase())) {
+    return true;
+  }
+  return !afterParagraph && HTML_TAG_ONLY_LINE.test(line);
+}
+
+/** A reference label as CommonMark compares them. */
+function label(text: string): string {
+  return text.trim().replace(/\s+/g, " ").toLowerCase();
+}
+
+const NEEDS_BRACKETS = /[()\s]/;
+// `\(` in a destination is a literal paren. Only ASCII punctuation is escapable,
+// so the backslash in `docs\alpha.md` is part of the path and has to survive.
+const ESCAPE = new RegExp(String.raw`\\(${ESCAPABLE})`, "g");
+// A URL parser reads a backslash as a path separator, so `docs\a.md` would
+// resolve to `docs/a.md`. Encode it first, the way a renderer normalises it.
+const BACKSLASH = /\\/g;
+// Only spaces and tabs may follow a closing fence.
+const NON_SPACE = /[^ \t]/;
+const LEADING_SLASHES = /^\/+/;
+
+function absolute(target: string, image: boolean): string {
+  const base = image ? IMAGE_BASE : LINK_BASE;
+  const trimmed = target.trim().replace(ESCAPE, "$1");
+  if (!trimmed || ABSOLUTE.test(trimmed)) {
+    return target;
+  }
+  try {
+    // A leading slash means the repository root, not the site root, so append
+    // it to the base instead of replacing the base path.
+    const resolved = new URL(
+      trimmed.replace(LEADING_SLASHES, "").replace(BACKSLASH, "%5C"),
+      base,
+    ).toString();
+    // `../` can climb out of the repository: leave those alone.
+    return resolved.startsWith(base) ? resolved : target;
+  } catch {
+    return target;
+  }
+}
+
+/** True when `index` is escaped by an odd run of backslashes. */
+function isEscaped(line: string, index: number): boolean {
+  let slashes = 0;
+  while (line[index - 1 - slashes] === "\\") {
+    slashes += 1;
+  }
+  return slashes % 2 === 1;
+}
+
+function unwrap(target: string): string {
+  return target.startsWith("<") && target.endsWith(">")
+    ? target.slice(1, -1)
+    : target;
+}
+
+/** The destination as it goes back into the line. */
+function wrap(resolved: string, original: string): string {
+  const bracketed = original.startsWith("<") && original.endsWith(">");
+  return bracketed || (resolved !== original && NEEDS_BRACKETS.test(resolved))
+    ? `<${resolved}>`
+    : resolved;
+}
+
+/** Rewrites one line's link and image targets, leaving code spans alone. */
+function rewriteLine(
+  line: string,
+  imageLabels: Set<string>,
+  spans: CodeSpan[],
+  base: number,
+  isDefinition: boolean,
+): string {
+  const reference = isDefinition ? REFERENCE_TARGET.exec(line) : null;
+  if (reference) {
+    const target = reference[3] ?? "";
+    const resolved = absolute(
+      unwrap(target),
+      imageLabels.has(label(reference[2] ?? "")),
+    );
+    const rest = line.slice(reference[0].length);
+    return `${reference[1]}${wrap(resolved, target)}${rest}`;
+  }
+
+  INLINE_TARGET.lastIndex = 0;
+  return line.replace(INLINE_TARGET, (match, bang, text, target, offset) => {
+    // `\\[` is a literal bracket, so the expression is not a link.
+    const opener = offset + (bang ? 1 : 0);
+    if (insideSpan(spans, base + offset) || isEscaped(line, opener)) {
+      return match;
+    }
+    // `\\!` is a literal mark, so what follows is a link, not an image.
+    const image = bang === "!" && !isEscaped(line, offset);
+    const resolved = absolute(unwrap(target), image);
+    // A badge nests an image inside a link, so the label is rewritten too.
+    const inner = text.includes("](")
+      ? rewriteLine(text, imageLabels, codeSpans(text), 0, false)
+      : text;
+    return `${bang}[${inner}](${wrap(resolved, target)}`;
+  });
+}
+
+interface Classified {
+  // Lines the renderer shows as Markdown, by index.
+  text: number[];
+  // Same lines, blanked where the renderer shows code, for span scanning.
+  masked: string;
+  // Lines where a `[label]: dest` definition can start.
+  definition: Set<number>;
+  // Document ranges the renderer hides inside HTML comments.
+  comments: CodeSpan[];
+}
+
+/**
+ * Sorts lines into Markdown and code, masking the code so a span cannot pair
+ * across it. Offsets are preserved, so a mask span sits where it does in the doc.
+ */
+function classify(lines: string[]): Classified {
+  const text: number[] = [];
+  const definition = new Set<number>();
+  const masked: string[] = [];
+  let openFence: string | null = null;
+  let inRawHtml = false;
+  let inHtmlBlock = false;
+  // Where the open block was written: the content column of the item it belongs
+  // to, 0 at document level, plus the blockquotes it sits inside. Only one is ever
+  // open, and none holds a lazy continuation line, so a line left of the item or
+  // outside the quote ends the block with its container.
+  let blockColumn = 0;
+  let blockQuotes = 0;
+  let inComment = false;
+  // True while an inline comment opened above runs on into this line, carried by
+  // the paragraph holding it.
+  let runOn = false;
+  const closesBelow = commentClosesBelow(lines);
+  let inCode = false;
+  let afterParagraph = false;
+  let quote: QuoteState = NO_QUOTE;
+  let lists: ListState = EMPTY_LIST_STATE;
+  const comments: CodeSpan[] = [];
+  let offset = 0;
+
+  // The line as list tracking sees it: blank wherever nothing renders. Taken
+  // with the paragraph state from the line above, as the renderer would.
+  const track = (structural: string, above: QuoteState): void => {
+    lists = openLists(structural, lists, afterParagraph, above.quoted);
+  };
+  // Where a block just opened sits, read after the opener closed the items it
+  // is dedented out of, so it belongs to the container it is really in.
+  const startBlock = (quotes: number): void => {
+    blockColumn = lists.columns.at(-1) ?? 0;
+    blockQuotes = quotes;
+  };
+  const endBlock = (): void => {
+    blockColumn = 0;
+    blockQuotes = 0;
+  };
+
+  lines.forEach((original, index) => {
+    const start = offset;
+    offset += original.length + 1;
+    // The quote state from the line above, which is what list tracking asks
+    // about. Only plain text below rewrites it, so every block returning early
+    // leaves no quoted paragraph open behind it.
+    const above = quote;
+    quote = NO_QUOTE;
+    // A fence, comment or HTML block runs only to the end of the container it was
+    // written in, so a line dedented out of that item or outside that quote
+    // closes both.
+    const quotes = quoteDepth(original);
+    let inBlock = openFence !== null || inRawHtml || inHtmlBlock || inComment;
+    if (
+      inBlock &&
+      leavesContainer(
+        original,
+        quotes,
+        blockColumn,
+        blockQuotes,
+        (inRawHtml || inComment) && blockColumn > 0 && blockQuotes === 0,
+      )
+    ) {
+      openFence = null;
+      inRawHtml = false;
+      inHtmlBlock = false;
+      inComment = false;
+      endBlock();
+      inBlock = false;
+    }
+    // Read from the container the line is written in, so a fence three columns
+    // past a nested bullet or behind a quote marker still opens one. A block
+    // already open keeps only its own quote stripped, or a deeper marker in it
+    // would read as a closer.
+    const container = containerContent(
+      original,
+      lists,
+      inBlock ? blockQuotes : quotes,
+    );
+    // A comment cannot open a fence and a fence hides a comment opener, so resolve
+    // them in that order or a hidden delimiter opens a phantom fence. An opener is
+    // read past a marker on the same line too, since a fence written as an item's
+    // first content opens inside it. Only an opener: fenced content is literal and
+    // a closer carries no marker.
+    const fenceSource = inComment
+      ? null
+      : FENCE.exec(
+          openFence === null
+            ? itemContent(container, afterParagraph)
+            : container,
+        );
+    if (inRawHtml) {
+      track("", above);
+      inRawHtml = !RAW_HTML_CLOSE.test(container);
+      if (!inRawHtml) {
+        endBlock();
+      }
+      masked.push(" ".repeat(original.length));
+      afterParagraph = false;
+      return;
+    }
+    if (inHtmlBlock) {
+      track("", above);
+      // Only a blank line ends a type 6 or 7 block, so nothing inside one is a
+      // fence or a link. A bare quote marker holds nothing, so it ends one too.
+      inHtmlBlock = !!container.trim();
+      if (!inHtmlBlock) {
+        endBlock();
+      }
+      masked.push(" ".repeat(original.length));
+      afterParagraph = false;
+      return;
+    }
+    const fence = fenceSource;
+    if (fence) {
+      // A fence renders as nothing, but its indent still closes an item.
+      track(original, above);
+      const marker = fence[1] ?? "";
+      if (openFence === null) {
+        // A backtick fence's info string may not contain a backtick.
+        openFence =
+          marker[0] !== "`" || !(fence[2] ?? "").includes("`") ? marker : null;
+        if (openFence === null) {
+          text.push(index);
+          masked.push(original);
+          afterParagraph = true;
+          return;
+        }
+        startBlock(quotes);
+      } else if (
+        // A closer matches the opening character and carries nothing after it.
+        marker[0] === openFence[0] &&
+        marker.length >= openFence.length &&
+        !NON_SPACE.test(fence[2] ?? "")
+      ) {
+        openFence = null;
+        endBlock();
+      }
+      masked.push(" ".repeat(original.length));
+      afterParagraph = false;
+      return;
+    }
+    if (openFence !== null) {
+      track("", above);
+      // Fenced content is literal, so a comment opener in it is not one.
+      masked.push(" ".repeat(original.length));
+      return;
+    }
+    // A block already open owns this line, so it is content rather than a block
+    // written at the column it happens to start in.
+    const hidden = inComment;
+    const carried = runOn;
+    // A comment is an HTML block too, so one written as a list item's first
+    // content opens inside that item exactly as a fence does: read past a marker
+    // on the same line and from its container's column, not the line's margin.
+    const opensComment =
+      !(hidden || carried) &&
+      COMMENT_BLOCK_OPEN.test(itemContent(container, afterParagraph));
+    // Only now, outside every fence, does a comment hide what follows.
+    const [line, stillInComment, stillRunOn] = maskComments(
+      original,
+      inComment,
+      runOn,
+      closesBelow[index + 1] ?? false,
+      opensComment,
+    );
+    inComment = stillInComment;
+    runOn = stillRunOn;
+    // A line an inline comment runs on into is still a line of the paragraph
+    // that carries it: only its text is hidden, never its block structure.
+    const structure = carried ? original : line;
+    // The same container reading as above, now the comments are masked. A comment
+    // blanks its own line, so that line is read as written: the block renders as
+    // nothing, but the item it is the content of still opens.
+    const source = opensComment ? original : line;
+    const visible = containerContent(source, lists, quotes);
+    // An HTML block written as a list item's first content opens inside that item,
+    // as a fence does, so an opener is read past a marker on the same line. The
+    // marker survives into the structural line, so its item is still tracked.
+    const content = itemContent(visible, afterParagraph);
+    const marker =
+      content === visible
+        ? ""
+        : source.slice(0, source.length - content.length);
+    // Taken before an HTML opener is hidden: it renders as nothing, but its indent
+    // still closes a list item it sits left of. A comment or a <pre> keeps only its
+    // column and marker, since the text it hides is not Markdown and opens no list.
+    const opensRaw = !carried && RAW_HTML_OPEN.test(content);
+    track(
+      !(hidden || carried) && (opensRaw || !line.trim())
+        ? hiddenStructure(original, marker)
+        : structure,
+      above,
+    );
+    // Read once the opener has closed the items it is dedented out of, so the
+    // comment block belongs to the item it is really written inside.
+    if (inComment !== hidden) {
+      if (inComment) {
+        startBlock(quotes);
+      } else {
+        endBlock();
+      }
+    }
+    for (let at = 0; at < line.length; at += 1) {
+      if (line[at] === " " && original[at] !== " ") {
+        const from = at;
+        while (at < line.length && line[at] === " " && original[at] !== " ") {
+          at += 1;
+        }
+        comments.push({ start: start + from, end: start + at, content: "" });
+      }
+    }
+    if (opensRaw) {
+      inRawHtml = !RAW_HTML_CLOSE.test(content.replace(RAW_HTML_OPEN, ""));
+      if (inRawHtml) {
+        startBlock(quotes);
+      }
+      masked.push(" ".repeat(line.length));
+      afterParagraph = false;
+      return;
+    }
+    if (!carried && content.trim() && opensHtmlBlock(content, afterParagraph)) {
+      inHtmlBlock = true;
+      startBlock(quotes);
+      masked.push(" ".repeat(line.length));
+      afterParagraph = false;
+      return;
+    }
+    const blank = !structure.trim();
+    // Measured from the innermost open item's content column, not the margin:
+    // four spaces under "- Details:" is a paragraph, not a code block.
+    const column = lists.columns.at(-1) ?? 0;
+    const indented = indentWidth(structure) - column >= INDENTED_CODE_INDENT;
+    // Indented code starts only outside a paragraph and runs to a dedent.
+    if (inCode) {
+      inCode = blank || indented;
+    } else {
+      inCode = !afterParagraph && !blank && indented;
+    }
+    if (inCode) {
+      masked.push(" ".repeat(line.length));
+      afterParagraph = false;
+      return;
+    }
+    // A definition cannot interrupt a paragraph.
+    if (!afterParagraph) {
+      definition.add(index);
+    }
+    text.push(index);
+    masked.push(line);
+    afterParagraph =
+      !blank &&
+      !BLOCK_LINE.test(structure) &&
+      (afterParagraph || !LINK_DEFINITION.test(structure));
+    quote = quoteState(structure, above.inQuote);
+  });
+
+  return { text, masked: masked.join("\n"), definition, comments };
+}
+
+/** Absolute repository URLs for every relative link and image in `markdown`. */
+export function resolveChangelogLinks(markdown: string): string {
+  // The desktop updater body arrives with CRLF, which would hide fences.
+  const lines = markdown.replace(LINE_ENDINGS, "\n").split("\n");
+  const { text, masked, definition, comments } = classify(lines);
+  // Scanned over the whole document, so a span may cross a line break. Commented
+  // ranges join them: the renderer shows neither, so a link in one is not
+  // followable and rewriting it would only mutate hidden text.
+  const spans = [...codeSpans(masked), ...comments].sort(
+    (a, b) => a.start - b.start,
+  );
+
+  // Offset of each line in the document, to place matches inside it.
+  const offsets: number[] = [];
+  let cursor = 0;
+  for (const line of lines) {
+    offsets.push(cursor);
+    cursor += line.length + 1;
+  }
+
+  // Only images resolve against the raw host, so collect the image labels
+  // before rewriting any definition.
+  const imageLabels = new Set<string>();
+  for (const index of text) {
+    const line = lines[index] ?? "";
+    IMAGE_REFERENCE.lastIndex = 0;
+    for (
+      let match = IMAGE_REFERENCE.exec(line);
+      match !== null;
+      match = IMAGE_REFERENCE.exec(line)
+    ) {
+      // An escaped mark makes it a link, so its definition stays a page URL.
+      if (
+        insideSpan(spans, (offsets[index] ?? 0) + match.index) ||
+        isEscaped(line, match.index)
+      ) {
+        continue;
+      }
+      const explicit = match[2] ?? "";
+      imageLabels.add(label(explicit.trim() ? explicit : (match[1] ?? "")));
+    }
+  }
+
+  const rewritten = [...lines];
+  for (const index of text) {
+    rewritten[index] = rewriteLine(
+      lines[index] ?? "",
+      imageLabels,
+      spans,
+      offsets[index] ?? 0,
+      definition.has(index),
+    );
+  }
+  return rewritten.join("\n");
+}
diff --git a/studio/frontend/src/lib/markdown-code-spans.ts b/studio/frontend/src/lib/markdown-code-spans.ts
new file mode 100644
index 0000000000..537aabb1ab
--- /dev/null
+++ b/studio/frontend/src/lib/markdown-code-spans.ts
@@ -0,0 +1,123 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+/**
+ * CommonMark code spans: a backtick run closes only on an equal-length run.
+ * That needs lookbehind, which older Safari rejects, so runs are scanned by hand.
+ */
+
+export interface CodeSpan {
+  // Offsets of the whole span, delimiters included.
+  start: number;
+  end: number;
+  // Between the delimiters, with the one space of padding removed.
+  content: string;
+}
+
+function runLength(text: string, index: number): number {
+  let end = index;
+  while (text[end] === "`") {
+    end += 1;
+  }
+  return end - index;
+}
+
+/** True when `index` is escaped by an odd run of backslashes. */
+function escaped(text: string, index: number): boolean {
+  let slashes = 0;
+  while (text[index - 1 - slashes] === "\\") {
+    slashes += 1;
+  }
+  return slashes % 2 === 1;
+}
+
+/** CommonMark drops one space of padding, so `` ` a ` `` renders as "a". */
+function stripPadding(content: string): string {
+  if (
+    content.length > 1 &&
+    content.startsWith(" ") &&
+    content.endsWith(" ") &&
+    content.trim() !== ""
+  ) {
+    return content.slice(1, -1);
+  }
+  return content;
+}
+
+/** Every code span in `text`, in order. Unclosed runs are ordinary text. */
+export function codeSpans(text: string): CodeSpan[] {
+  const spans: CodeSpan[] = [];
+  let index = 0;
+
+  while (index < text.length) {
+    if (text[index] !== "`" || escaped(text, index)) {
+      index += 1;
+      continue;
+    }
+    const ticks = runLength(text, index);
+    const contentStart = index + ticks;
+
+    let cursor = contentStart;
+    let closed = false;
+    while (cursor < text.length) {
+      // Escapes do not apply inside a span, so a run after a backslash closes it.
+      if (text[cursor] !== "`") {
+        cursor += 1;
+        continue;
+      }
+      const candidate = runLength(text, cursor);
+      if (candidate === ticks) {
+        spans.push({
+          start: index,
+          end: cursor + ticks,
+          content: stripPadding(text.slice(contentStart, cursor)),
+        });
+        index = cursor + ticks;
+        closed = true;
+        break;
+      }
+      cursor += candidate;
+    }
+    if (!closed) {
+      // Nothing closes this run: it is literal text, carry on after it.
+      index = contentStart;
+    }
+  }
+  return spans;
+}
+
+/** Replaces every code span with `park(content)`, leaving the rest as is. */
+export function parkCodeSpans(
+  text: string,
+  park: (content: string) => string,
+): string {
+  const spans = codeSpans(text);
+  if (spans.length === 0) {
+    return text;
+  }
+  let out = "";
+  let cursor = 0;
+  for (const span of spans) {
+    out += text.slice(cursor, span.start) + park(span.content);
+    cursor = span.end;
+  }
+  return out + text.slice(cursor);
+}
+
+/** True when `index` falls inside one of `spans`, which are in order. */
+export function insideSpan(spans: CodeSpan[], index: number): boolean {
+  let low = 0;
+  let high = spans.length - 1;
+  while (low <= high) {
+    const mid = (low + high) >> 1;
+    const span = spans[mid];
+    if (span === undefined || index < span.start) {
+      high = mid - 1;
+    } else if (index >= span.end) {
+      low = mid + 1;
+    } else {
+      return true;
+    }
+  }
+  return false;
+}
diff --git a/studio/frontend/src/lib/markdown-inline-comments.ts b/studio/frontend/src/lib/markdown-inline-comments.ts
new file mode 100644
index 0000000000..33bbfddc31
--- /dev/null
+++ b/studio/frontend/src/lib/markdown-inline-comments.ts
@@ -0,0 +1,62 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+/**
+ * An HTML comment written mid-sentence is inline raw HTML, not a block, so it
+ * belongs to its paragraph: the `-->` may arrive on a later line of that same
+ * paragraph and everything between renders as nothing, while past the paragraph
+ * the `<!--` is ordinary text. Both changelog scanners share that answer here.
+ *
+ * The backend needs none of it: a heading closes the paragraph it sits under, so
+ * no heading can ever land inside one of these comments.
+ */
+
+import { interruptsParagraph } from "@/lib/markdown-list-columns";
+
+const COMMENT_CLOSE = "-->";
+// A line that cannot be more of the paragraph above it: blank, or a block that
+// may interrupt one. Leading punctuation is not one: `-->` alone is the ordinary
+// multiline close and a continuation may open with emphasis, so reading either as
+// a break leaves the comment unclosed and its text on show. Indented code and link
+// definitions are absent: neither may interrupt a paragraph (spec 0.31.2 4.4, 4.7).
+const BLANK = /^[ \t]*$/;
+const ATX_HEADING = /^ {0,3}#{1,6}([ \t]|$)/;
+const FENCE = /^ {0,3}(?:`{3,}|~{3,})/;
+const THEMATIC_BREAK =
+  /^ {0,3}(?:(?:\*[ \t]*){3,}|(?:-[ \t]*){3,}|(?:_[ \t]*){3,})$/;
+// A row of `=` or `-` alone makes the paragraph above it a setext heading, ending it.
+const SETEXT_UNDERLINE = /^ {0,3}(?:=+|-+)[ \t]*$/;
+// A tag, comment or declaration at the start of a line. HTML block types 1 to 6
+// interrupt a paragraph; type 7 does not, but reading one as a break only leaves
+// the opener as plain text, which is what a leading `<` has always meant here.
+const HTML_LINE = /^ {0,3}</;
+
+/** Whether `line` starts a block of its own rather than continuing a paragraph. */
+function startsBlock(line: string): boolean {
+  return (
+    BLANK.test(line) ||
+    ATX_HEADING.test(line) ||
+    FENCE.test(line) ||
+    THEMATIC_BREAK.test(line) ||
+    SETEXT_UNDERLINE.test(line) ||
+    HTML_LINE.test(line) ||
+    // Blockquote, or a list item with content: the rule the other scanners share.
+    interruptsParagraph(line)
+  );
+}
+
+/**
+ * For each line, whether a `-->` is reachable without leaving the paragraph it
+ * starts in. Read at `index + 1` it answers whether an inline comment opened on
+ * `index` and left unclosed there is a comment at all.
+ */
+export function commentClosesBelow(lines: string[]): boolean[] {
+  const closes: boolean[] = new Array(lines.length + 1).fill(false);
+  for (let at = lines.length - 1; at >= 0; at -= 1) {
+    const line = lines[at] ?? "";
+    closes[at] =
+      !startsBlock(line) &&
+      (line.includes(COMMENT_CLOSE) || (closes[at + 1] ?? false));
+  }
+  return closes;
+}
diff --git a/studio/frontend/src/lib/markdown-list-columns.ts b/studio/frontend/src/lib/markdown-list-columns.ts
new file mode 100644
index 0000000000..761cdfac54
--- /dev/null
+++ b/studio/frontend/src/lib/markdown-list-columns.ts
@@ -0,0 +1,357 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+/**
+ * CommonMark measures a block's indentation from its container, not the left
+ * margin: four spaces at document level and four under a bullet mean different
+ * things. Tracking the open items lets both changelog scanners ask "is this
+ * indented code?" the way a renderer would.
+ *
+ * Ported from `_open_lists` in studio/backend/utils/changelog.py so the three
+ * scanners classify a line the same way.
+ */
+
+/** The open list items, innermost last, by the column their content starts. */
+export interface ListState {
+  columns: number[];
+  // True while the innermost item has had no content since its marker.
+  emptyItem: boolean;
+}
+
+export const EMPTY_LIST_STATE: ListState = { columns: [], emptyItem: false };
+
+// The marker needs whitespace after it, so `2.0` is a version, not an item.
+const LIST_ITEM = /^[ \t]*([-*+]|\d{1,9}[.)])([ \t]+|$)/;
+const THEMATIC_BREAK =
+  /^ {0,3}(?:(?:\*[ \t]*){3,}|(?:-[ \t]*){3,}|(?:_[ \t]*){3,})$/;
+const BLOCK_QUOTE = /^ {0,3}>/;
+const QUOTE_MARKER = /^ {0,3}>[ \t]?/;
+// Blocks that are not paragraph text, so they cannot continue one lazily.
+const PARAGRAPH_TEXT = /^ {0,3}(?![-*+>]([ \t]|$)|\d{1,9}[.)]([ \t]|$))\S/;
+// Blocks that break into an open paragraph, closing it rather than continuing
+// it. A link reference definition is not one of them.
+const INTERRUPTS =
+  /^ {0,3}(?:#{1,6}([ \t]|$)|(?:\*[ \t]*){3,}$|(?:-[ \t]*){3,}$|(?:_[ \t]*){3,}$)/;
+const FENCE = /^ {0,3}(?:`{3,}|~{3,})/;
+const HTML_BLOCK_OPEN = /^ {0,3}<\/?([a-zA-Z][a-zA-Z0-9-]*)(?=[\s/>]|$)/;
+const HTML_BLOCK_TAGS = new Set(
+  `address article aside base basefont blockquote body caption center col colgroup
+   dd details dialog dir div dl dt fieldset figcaption figure footer form frame
+   frameset h1 h2 h3 h4 h5 h6 head header hr html iframe legend li link main menu
+   menuitem nav noframes ol optgroup option p param search section summary table
+   tbody td tfoot th thead title tr track ul`.split(/\s+/),
+);
+// Content indented more than this after a marker is an indented code block, so
+// the item's content starts one column past the marker instead.
+const MAX_ITEM_PADDING = 4;
+// Columns past its container at which a line becomes an indented code block.
+const INDENTED_CODE = 4;
+// Stands in for a line the renderer hides. `#` is a block of its own, so list
+// tracking reads it like a comment: never a marker, never a lazy continuation.
+const HIDDEN_BLOCK = "#";
+const LEADING_SPACE = /^[ \t]*/;
+
+/**
+ * `line` as list tracking sees it once the renderer hides its text. A comment or
+ * raw HTML block renders nothing but is still a block at its own column, so it
+ * closes the items it sits left of. Only the indentation survives: what the block
+ * hides is not Markdown and must not open a list. `marker` is the part opening
+ * the item the block is content of, which survives too. Ported from
+ * `_hidden_structure` on the backend.
+ */
+export function hiddenStructure(line: string, marker = ""): string {
+  if (marker) {
+    return `${marker}${HIDDEN_BLOCK}`;
+  }
+  const indent = LEADING_SPACE.exec(line)?.[0] ?? "";
+  return line.trim() ? `${indent}${HIDDEN_BLOCK}` : "";
+}
+
+/** Columns of leading whitespace, counting a tab to the next stop of four. */
+export function indentWidth(line: string): number {
+  let width = 0;
+  for (const char of line) {
+    if (char === " ") {
+      width += 1;
+    } else if (char === "\t") {
+      width += 4 - (width % 4);
+    } else {
+      break;
+    }
+  }
+  return width;
+}
+
+/**
+ * Whether `line` starts a block that can break into an open paragraph. A quote
+ * marker always can; a list item only with content, an ordered one only at 1.
+ * Anything else is text of the paragraph it appears to interrupt.
+ */
+export function interruptsParagraph(line: string): boolean {
+  if (BLOCK_QUOTE.test(line)) {
+    return true;
+  }
+  const item = THEMATIC_BREAK.test(line) ? null : LIST_ITEM.exec(line);
+  if (item === null) {
+    return false;
+  }
+  const marker = item[1] ?? "";
+  if (!line.slice(item[0].length).trim()) {
+    return false;
+  }
+  const ordered = marker.endsWith(".") || marker.endsWith(")");
+  return !ordered || marker.slice(0, -1) === "1";
+}
+
+/**
+ * Whether a marker-shaped `line` is really text of the paragraph above. Only a
+ * marker inside the paragraph's own item interrupts it; one to the left closes
+ * that item and opens a sibling. A quote owns the paragraph its lines hold, so a
+ * marker outside the quote opens a list of its own.
+ */
+export function lazyMarker(
+  line: string,
+  state: ListState,
+  afterParagraph: boolean,
+  quoted: boolean,
+): boolean {
+  const item = THEMATIC_BREAK.test(line) ? null : LIST_ITEM.exec(line);
+  const columns = state.columns;
+  const inside =
+    columns.length === 0 || indentWidth(line) >= (columns.at(-1) ?? 0);
+  return (
+    item !== null &&
+    afterParagraph &&
+    !quoted &&
+    inside &&
+    !interruptsParagraph(line)
+  );
+}
+
+/** `columns` with every item whose content starts past `indent` closed. */
+function dropDeeper(columns: number[], indent: number): number[] {
+  let open = columns.length;
+  while (open > 0 && (columns[open - 1] ?? 0) > indent) {
+    open -= 1;
+  }
+  return open === columns.length ? columns : columns.slice(0, open);
+}
+
+/** `line` with up to `columns` columns of leading whitespace removed. */
+function stripIndent(line: string, columns: number): string {
+  let width = 0;
+  let index = 0;
+  while (index < line.length && width < columns) {
+    const char = line[index];
+    if (char !== " " && char !== "\t") {
+      break;
+    }
+    width += char === " " ? 1 : 4 - (width % 4);
+    index += 1;
+  }
+  return line.slice(index);
+}
+
+/**
+ * Whether `line` can continue a paragraph it is indented out of. Only plain text
+ * can: a heading, fence, break or HTML block starts a block of its own, closing
+ * the item instead. An underline is not one: it may never be lazy, so `===` left
+ * of an open item is more of the item's paragraph. Nor is a definition, a block
+ * of its own that may not interrupt a paragraph. A row of dashes still closes the
+ * item: `INTERRUPTS` reads three or more as the thematic break they are.
+ */
+function mayBeLazy(line: string): boolean {
+  const named = HTML_BLOCK_OPEN.exec(line);
+  // Types 1 to 6 interrupt a paragraph, so a `<div>` left of an open item closes
+  // it. Type 7 cannot, and is deliberately excluded.
+  const htmlBlock =
+    named !== null && HTML_BLOCK_TAGS.has((named[1] ?? "").toLowerCase());
+  return (
+    PARAGRAPH_TEXT.test(line) &&
+    !INTERRUPTS.test(line) &&
+    !FENCE.test(line) &&
+    !htmlBlock
+  );
+}
+
+/**
+ * Whether `line` reads as more of a paragraph open in its container, measured
+ * from `column` where that container's content starts: four columns past it the
+ * line is indented code, which may not interrupt a paragraph, so indentation
+ * alone never closes the one above.
+ */
+export function continuesParagraph(line: string, column: number): boolean {
+  const inner = stripIndent(line, column);
+  return indentWidth(inner) >= INDENTED_CODE || mayBeLazy(inner);
+}
+
+/** `line` with up to `depth` blockquote markers removed, and how many went. */
+function stripQuotes(line: string, depth: number): [string, number] {
+  let rest = line;
+  let removed = 0;
+  let marker = removed < depth ? QUOTE_MARKER.exec(rest) : null;
+  while (marker !== null) {
+    rest = rest.slice(marker[0].length);
+    removed += 1;
+    marker = removed < depth ? QUOTE_MARKER.exec(rest) : null;
+  }
+  return [rest, removed];
+}
+
+/** What a blockquote line holds, with its markers stripped. */
+function quoteContent(line: string): string {
+  return stripQuotes(line, Number.POSITIVE_INFINITY)[0];
+}
+
+/** How many blockquotes `line` is written inside. */
+export function quoteDepth(line: string): number {
+  return stripQuotes(line, Number.POSITIVE_INFINITY)[1];
+}
+
+/**
+ * `line` as the container it is written in sees it, with `quotes` blockquote
+ * markers and the open item's content column removed. CommonMark measures a block
+ * from its container, not the margin (spec 0.31.2 sections 5.1, 5.2), so `> ~~~`
+ * and a fence under a nested bullet are openers despite sitting more than three
+ * columns in.
+ */
+export function containerContent(
+  line: string,
+  state: ListState,
+  quotes: number,
+): string {
+  const [inner] = stripQuotes(line, quotes);
+  if (quotes > 0) {
+    // A list inside a quote is the quote's own; this tracker follows document
+    // level only, so its columns do not apply here.
+    return inner;
+  }
+  const columns = dropDeeper(state.columns, indentWidth(inner));
+  return stripIndent(inner, columns.at(-1) ?? 0);
+}
+
+/**
+ * `line` read from the content column of a list item that opens on it. A block
+ * written as an item's first content sits inside that item, so ``- ``` `` opens a
+ * fence even though its marker is not within three columns of the container (spec
+ * 0.31.2 section 5.2). Padding is capped the way `openLists` caps it, or
+ * ``-     ``` `` would read as a fence rather than the indented code it is. A
+ * marker the paragraph above swallows opens no item, so its line is returned
+ * whole, as is one four columns past its container.
+ */
+export function itemContent(line: string, afterParagraph: boolean): string {
+  if (
+    indentWidth(line) >= INDENTED_CODE ||
+    (afterParagraph && !interruptsParagraph(line))
+  ) {
+    return line;
+  }
+  const item = THEMATIC_BREAK.test(line) ? null : LIST_ITEM.exec(line);
+  if (item === null) {
+    return line;
+  }
+  const padding = indentWidth(item[2] ?? "");
+  // Over-indented content starts one column past the marker; the rest of the
+  // padding is the content's own indentation.
+  const over = padding > MAX_ITEM_PADDING ? padding - 1 : 0;
+  return `${" ".repeat(over)}${line.slice(item[0].length)}`;
+}
+
+/** Whether a blockquote owns the paragraph the line below could continue. */
+export interface QuoteState {
+  // True while a quoted paragraph is open, so plain text below is more of it.
+  inQuote: boolean;
+  // True whenever that paragraph is the quote's rather than the document's.
+  quoted: boolean;
+}
+
+export const NO_QUOTE: QuoteState = { inQuote: false, quoted: false };
+
+/**
+ * The quote state after `line`, given the state after the line above and the
+ * content column of the item `line` sits in. A quote owns the paragraph its own
+ * lines hold, so a marker written outside the quote opens a list of its own
+ * rather than reading as more of that paragraph. Ported from `in_quote` tracking
+ * in changelog.py.
+ */
+export function quoteState(
+  line: string,
+  inQuote: boolean,
+  column = 0,
+): QuoteState {
+  if (BLOCK_QUOTE.test(line)) {
+    // An empty quote holds no paragraph, so the line below starts a new one.
+    return { inQuote: mayBeLazy(quoteContent(line)), quoted: true };
+  }
+  const open = inQuote && continuesParagraph(line, column);
+  return { inQuote: open, quoted: open };
+}
+
+/**
+ * `columns` with every item `line` is written to the left of closed. Read inside
+ * the container the item sits in, not from the margin: a line that only looks
+ * dedented there is lazy text of the item's paragraph, leaving the item open.
+ */
+function closeDedented(
+  columns: number[],
+  line: string,
+  indent: number,
+  afterParagraph: boolean,
+): number[] {
+  let open = columns.length;
+  while (open > 0 && (columns[open - 1] ?? 0) > indent) {
+    const outer = open > 1 ? (columns[open - 2] ?? 0) : 0;
+    if (afterParagraph && continuesParagraph(line, outer)) {
+      break;
+    }
+    open -= 1;
+  }
+  return open === columns.length ? columns : columns.slice(0, open);
+}
+
+/**
+ * The list items still open after `line`. A dedented line closes an item unless
+ * it is a lazy paragraph continuation. A new marker nests under a deeper column
+ * and replaces a sibling. `quoted` marks a paragraph the blockquote above owns:
+ * a marker outside the quote is not text of it, so it opens a list of its own.
+ */
+export function openLists(
+  line: string,
+  state: ListState,
+  afterParagraph: boolean,
+  quoted = false,
+): ListState {
+  let columns = state.columns;
+  if (!line.trim()) {
+    // A blank line leaves the list open, unless the item is still empty: an
+    // item may begin with one blank line, and later content is outside it.
+    return {
+      columns: state.emptyItem ? columns.slice(0, -1) : columns,
+      emptyItem: false,
+    };
+  }
+  const indent = indentWidth(line);
+  const item = THEMATIC_BREAK.test(line) ? null : LIST_ITEM.exec(line);
+  const empty = item !== null && !line.slice(item[0].length).trim();
+  if (lazyMarker(line, state, afterParagraph, quoted)) {
+    // A lazy continuation or an underline, so the open items are untouched.
+    return state;
+  }
+  columns = closeDedented(columns, line, indent, afterParagraph);
+  // Four columns past its container the marker is an indented code block, or
+  // lazy text of the paragraph above it, so it opens no list of its own.
+  if (item === null || indent - (columns.at(-1) ?? 0) >= INDENTED_CODE) {
+    return { columns, emptyItem: false };
+  }
+  const marker = item[1] ?? "";
+  let padding = indentWidth(item[2] ?? "");
+  if (padding === 0 || padding > MAX_ITEM_PADDING) {
+    // An empty or over-indented item still holds one column of content.
+    padding = 1;
+  }
+  // A sibling marker replaces the item it lines up with.
+  return {
+    columns: [...dropDeeper(columns, indent), indent + marker.length + padding],
+    emptyItem: empty,
+  };
+}
diff --git a/studio/frontend/src/lib/release-notes-preview.ts b/studio/frontend/src/lib/release-notes-preview.ts
new file mode 100644
index 0000000000..97441b7cbb
--- /dev/null
+++ b/studio/frontend/src/lib/release-notes-preview.ts
@@ -0,0 +1,1005 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+// Top changelog bullets, shown in the collapsed update popup.
+import { codeSpans, parkCodeSpans } from "@/lib/markdown-code-spans";
+import { commentClosesBelow } from "@/lib/markdown-inline-comments";
+import {
+  EMPTY_LIST_STATE,
+  type ListState,
+  NO_QUOTE,
+  type QuoteState,
+  hiddenStructure,
+  indentWidth,
+  itemContent,
+  openLists,
+  quoteState,
+} from "@/lib/markdown-list-columns";
+
+export const RELEASE_NOTES_PREVIEW_ITEMS = 4;
+const PREVIEW_ITEM_MAX_CHARS = 120;
+// Bullets indented past the shallowest one are nested detail, not headlines.
+const NESTED_INDENT_TOLERANCE = 1;
+const TAB_WIDTH = 4;
+// Four spaces starts an indented code block in Markdown.
+const INDENTED_CODE_INDENT = 4;
+
+// At most three leading spaces: deeper is indented code, not a fence.
+const FENCE = /^ {0,3}(`{3,}|~{3,})(.*)$/;
+// An ATX heading needs a space, tab or line end after the marker, as in
+// _HEADING_PATTERN. `\s` would match a non-breaking space and eat prose, and a
+// bare `##` is an empty heading that still ends a bullet.
+const HEADING = /^#{1,6}(?:[ \t]|$)/;
+const BULLET = /^(?:[-*+]|(\d{1,9})[.)])[ \t]+(.*)$/;
+// At most three leading spaces, as everywhere else: deeper is indented code,
+// so a quoted line inside a code sample cannot reach the collector.
+const BLOCKQUOTE = /^ {0,3}>[ \t]?/;
+// A GFM delimiter cell is hyphens with an optional alignment colon each side.
+const TABLE_DELIMITER_CELL = /^:?-+:?$/;
+// "- - -" and "***" are horizontal rules, not bullets and not notes.
+const THEMATIC_BREAK =
+  /^ {0,3}(?:(?:\*[ \t]*){3,}|(?:-[ \t]*){3,}|(?:_[ \t]*){3,})$/;
+// Destinations may escape or balance parentheses, and labels may nest one
+// level so `[![alt](img)](link)` still resolves.
+const DESTINATION = "\\((?:\\\\.|[^()\\\\]|\\([^()]*\\))*\\)";
+const LABEL = "((?:[^\\[\\]\\\\]|\\\\.|\\[(?:[^\\[\\]\\\\]|\\\\.)*\\])*)";
+const IMAGE = new RegExp(`!\\[${LABEL}\\]${DESTINATION}`, "g");
+const LINK = new RegExp(`\\[${LABEL}\\]${DESTINATION}`, "g");
+// Reference forms: `[text][label]`, `[text][]` and the shortcut `[text]`.
+const IMAGE_REFERENCE = new RegExp(`!\\[${LABEL}\\](?:\\[([^\\]]*)\\])?`, "g");
+const LINK_REFERENCE = new RegExp(`\\[${LABEL}\\](?:\\[([^\\]]*)\\])?`, "g");
+// A definition line renders as nothing at all.
+const DEFINITION = /^ {0,3}\[((?:[^\[\]\\]|\\.)+)\]:/;
+// CommonMark: a backslash escapes ASCII punctuation.
+const ESCAPE = /\\([!-/:-@[-`{-~])/g;
+// Private-use sentinels park code spans, so document text cannot contain them.
+const SENTINELS = /[\uE000\uE001]/g;
+const LINE_ENDINGS = /\r\n?/g;
+const TABS = /\t/g;
+// Real tags only: a name character must follow "<", so a version constraint
+// like "Support Python <3.15 and >3.9" keeps its operators.
+const HTML_TAG = /<\/?[a-zA-Z][^>]*>/g;
+// <https://x> and <a@b.c> are Markdown autolinks: keep the text they render.
+const AUTOLINK = /<([a-zA-Z][a-zA-Z0-9+.-]*:[^\s<>]*|[^\s<>@]+@[^\s<>@]+)>/g;
+// CommonMark type 1 HTML blocks render literally until a closing tag, which
+// the spec says need not be the one that opened the block.
+const RAW_HTML_OPEN = /^ {0,3}<(pre|script|style|textarea)(?=[\s>]|$)/i;
+const RAW_HTML_CLOSE = /<\/(pre|script|style|textarea)\s*>/i;
+// Types 3 to 5 (processing instructions, declarations, CDATA) are literal too,
+// each ending on its own delimiter. Comments open mid-line, handled separately.
+const RAW_BLOCKS: [RegExp, RegExp][] = [
+  [RAW_HTML_OPEN, RAW_HTML_CLOSE],
+  [/^ {0,3}<\?/, /\?>/],
+  [/^ {0,3}<!\[CDATA\[/, /\]\]>/],
+  // A declaration needs an uppercase letter, so `<!note` stays ordinary text.
+  [/^ {0,3}<![A-Z]/, />/],
+];
+// Type 6 and 7 blocks run to the next blank line, so `<details>` holds Markdown
+// only after one. Type 7 (any complete tag alone) cannot interrupt a paragraph.
+const HTML_BLOCK_OPEN = /^ {0,3}<\/?([a-zA-Z][a-zA-Z0-9-]*)(?=[\s/>]|$)/;
+const HTML_ATTRIBUTE =
+  "(?:\\s+[a-zA-Z_:][a-zA-Z0-9_.:-]*(?:\\s*=\\s*(?:[^\\s\"'=<>`]+|'[^']*'|\"[^\"]*\"))?)";
+const HTML_TAG_ONLY_LINE = new RegExp(
+  `^ {0,3}(?:<[a-zA-Z][a-zA-Z0-9-]*${HTML_ATTRIBUTE}*\\s*/?>|</[a-zA-Z][a-zA-Z0-9-]*\\s*>)\\s*$`,
+);
+const HTML_BLOCK_TAGS = new Set(
+  `address article aside base basefont blockquote body caption center col colgroup
+   dd details dialog dir div dl dt fieldset figcaption figure footer form frame
+   frameset h1 h2 h3 h4 h5 h6 head header hr html iframe legend li link main menu
+   menuitem nav noframes ol optgroup option p param search section summary table
+   tbody td tfoot th thead title tr track ul`.split(/\s+/),
+);
+// Only spaces and tabs may follow a closing fence.
+const NON_SPACE = /[^ \t]/;
+const HEADING_LINE = /^ {0,3}#{1,6}(?:[ \t]|$)/;
+const COMMENT_BLOCK_OPEN = /^ {0,3}<!--/;
+const COMMENT_OPEN = "<!--";
+const COMMENT_CLOSE = "-->";
+// Paired emphasis only. Underscores inside identifiers are literal, so
+// UNSLOTH_DISABLE_UPDATE_CHECK keeps its name.
+const BOLD_STAR = /\*\*(?=\S)([\s\S]*?\S)\*\*/g;
+const BOLD_UNDERSCORE = /(^|[^\w])__(?=\S)([\s\S]*?\S)__(?=[^\w]|$)/g;
+const ITALIC_STAR = /\*(?=\S)([^*\n]*?\S)\*/g;
+const ITALIC_UNDERSCORE = /(^|[^\w])_(?=\S)([^_\n]*?\S)_(?=[^\w]|$)/g;
+const BACKTICK = /`/g;
+// A closer is a run of the same length, so `` `x` `` keeps its backticks.
+// Streamdown renders `AT&amp;T` as "AT&T", so the preview decodes entities too.
+const NAMED_ENTITIES: Record<string, string> = {
+  amp: "&",
+  lt: "<",
+  gt: ">",
+  quot: '"',
+  apos: "'",
+  nbsp: "\u00a0",
+};
+const ENTITY = /&(#\d{1,7}|#[xX][0-9a-fA-F]{1,6}|[a-zA-Z][a-zA-Z0-9]{1,31});/g;
+const PARKED = /\uE000(\d+)\uE001/g;
+const WHITESPACE = /\s+/g;
+// Sentence end followed by something that actually starts a sentence.
+const SENTENCE_BREAK = /[.!?]\s+(?=["'“‘]?[A-Z0-9])/g;
+const TRAILING_WORD = /(\S+)$/;
+// A period here ends an abbreviation, not the sentence.
+const ABBREVIATIONS = new Set([
+  "e.g.",
+  "i.e.",
+  "etc.",
+  "vs.",
+  "cf.",
+  "approx.",
+  "no.",
+  "fig.",
+  "al.",
+  "dr.",
+  "mr.",
+  "mrs.",
+  "ms.",
+  "prof.",
+  "inc.",
+  "ltd.",
+  "st.",
+  "jr.",
+  "sr.",
+]);
+const INITIAL = /^[A-Za-z]\.$/;
+const MIN_LEAD_CHARS = 12;
+
+/** Strip tags until stable, so a removal cannot re-form a tag. */
+function stripHtmlTags(text: string): string {
+  let out = text;
+  let previous: string;
+  do {
+    previous = out;
+    out = out.replace(HTML_TAG, "");
+  } while (out !== previous);
+  return out;
+}
+
+export interface ReleaseNotesPreviewItem {
+  // Leading sentence, highlighted in the preview.
+  lead: string;
+  // Rest of the bullet, de-emphasised. Empty for single-sentence bullets.
+  rest: string;
+}
+
+export interface ReleaseNotesPreview {
+  items: ReleaseNotesPreviewItem[];
+  // Bullets past the preview limit, for a "+N more" affordance.
+  remaining: number;
+}
+
+interface Bullet {
+  text: string;
+  indent: number;
+}
+
+/** Whether a reference points at a definition the document actually has. */
+function definedLabel(
+  labels: Set<string> | undefined,
+  reference: string | undefined,
+  text: string,
+): boolean {
+  if (labels === undefined) {
+    return false;
+  }
+  const label = (reference?.trim() ? reference : text)
+    .trim()
+    .replace(WHITESPACE, " ")
+    .toLowerCase();
+  return labels.has(label);
+}
+
+/** One entity as the character it renders as, or unchanged if unknown. */
+function decodeEntity(match: string, body: string): string {
+  if (body.startsWith("#")) {
+    const hex = body[1] === "x" || body[1] === "X";
+    const code = Number.parseInt(
+      hex ? body.slice(2) : body.slice(1),
+      hex ? 16 : 10,
+    );
+    return Number.isFinite(code) && code > 0 && code <= 0x10ffff
+      ? String.fromCodePoint(code)
+      : match;
+  }
+  return NAMED_ENTITIES[body.toLowerCase()] ?? match;
+}
+
+/** Inline markdown stripped to plain text. */
+function toPlainText(markdown: string, labels?: Set<string>): string {
+  // Park code spans first: their contents are literal and must survive below.
+  const codes: string[] = [];
+  const park = (text: string): string => {
+    codes.push(text);
+    return `\uE000${codes.length - 1}\uE001`;
+  };
+  // Escaped punctuation is literal too, so `\*not italic\*` keeps its stars.
+  const parked = parkCodeSpans(markdown, park).replace(ESCAPE, (_match, char) =>
+    park(char),
+  );
+
+  return stripHtmlTags(
+    parked
+      .replace(AUTOLINK, "$1")
+      .replace(IMAGE, "")
+      .replace(LINK, "$1")
+      .replace(IMAGE_REFERENCE, (match, text, ref) =>
+        definedLabel(labels, ref, text) ? "" : match,
+      )
+      .replace(LINK_REFERENCE, (match, text, ref) =>
+        definedLabel(labels, ref, text) ? text : match,
+      ),
+  )
+    .replace(BOLD_STAR, "$1")
+    .replace(BOLD_UNDERSCORE, "$1$2")
+    .replace(ITALIC_STAR, "$1")
+    .replace(ITALIC_UNDERSCORE, "$1$2")
+    .replace(BACKTICK, "")
+    .replace(ENTITY, decodeEntity)
+    .replace(PARKED, (_match, index: string) => codes[Number(index)] ?? "")
+    .replace(WHITESPACE, " ")
+    .trim();
+}
+
+function truncate(text: string): string {
+  if (text.length <= PREVIEW_ITEM_MAX_CHARS) {
+    return text;
+  }
+  const clipped = text.slice(0, PREVIEW_ITEM_MAX_CHARS);
+  const lastSpace = clipped.lastIndexOf(" ");
+  return `${(lastSpace > 40 ? clipped.slice(0, lastSpace) : clipped).trimEnd()}...`;
+}
+
+interface ContentLine {
+  text: string;
+  indent: number;
+  // Blockquoted lines are quoted examples, not the release's own bullets.
+  quoted: boolean;
+  // Content column of the innermost open list item. CommonMark measures
+  // indentation from here, so `indent - column` is the real depth.
+  column: number;
+}
+
+/**
+ * `line` with its comments removed, whether a comment block stays open, and
+ * whether an inline comment runs on into the line below.
+ *
+ * Only a comment starting a line opens a block, which hides whole lines to the
+ * one holding `-->`. One written mid-sentence is inline HTML belonging to its
+ * paragraph, so its `-->` may arrive on a later line and only the text up to it
+ * is hidden. `closesBelow` says one does; without it the opener is ordinary text
+ * and hides nothing below.
+ *
+ * "Starting a line" is read inside the container, so `blockOpen` comes from the
+ * item's content rather than the raw line.
+ */
+function stripCommentSpans(
+  line: string,
+  startInComment: boolean,
+  runOn: boolean,
+  closesBelow: boolean,
+  blockOpen: boolean,
+): [string, boolean, boolean] {
+  if (startInComment) {
+    // The closing line belongs to the block, tail included.
+    return ["", !line.includes(COMMENT_CLOSE), false];
+  }
+
+  let visible = "";
+  let index = 0;
+  if (runOn) {
+    const closed = line.indexOf(COMMENT_CLOSE);
+    if (closed === -1) {
+      return ["", false, true];
+    }
+    // Only up to the closer: the tail is the paragraph's own text again.
+    index = closed + COMMENT_CLOSE.length;
+  } else if (blockOpen) {
+    // `<!-->` and `<!--->` are complete comments, so the closer may overlap the
+    // opener; searching past it would hide every later release.
+    return ["", !line.includes(COMMENT_CLOSE), false];
+  }
+
+  const spans = codeSpans(line);
+  while (index < line.length) {
+    const open = line.indexOf(COMMENT_OPEN, index);
+    if (open === -1) {
+      visible += line.slice(index);
+      break;
+    }
+    // A delimiter inside inline code is literal, not a comment opener.
+    const span = spans.find(
+      (candidate) => candidate.start <= open && candidate.end > open,
+    );
+    if (span) {
+      visible += line.slice(index, span.end);
+      index = span.end;
+      continue;
+    }
+    const close = line.indexOf(COMMENT_CLOSE, open + COMMENT_OPEN.length);
+    if (close === -1) {
+      if (closesBelow) {
+        // The paragraph carries the comment on, so this line and the next are in it.
+        return [visible + line.slice(index, open), false, true];
+      }
+      // Nothing closes it at all, so the renderer shows it as text.
+      visible += line.slice(index);
+      break;
+    }
+    visible += line.slice(index, open);
+    index = close + COMMENT_CLOSE.length;
+  }
+  return [visible, false, false];
+}
+
+/** Strips raw block content. State is the open block's index, or null. */
+function stripRawHtml(
+  line: string,
+  openBlock: number | null,
+): [string, number | null] {
+  if (openBlock !== null) {
+    return RAW_BLOCKS[openBlock]?.[1].test(line) ? ["", null] : ["", openBlock];
+  }
+  // A block only opens at the start of a line; mid-line tags are inline HTML.
+  for (const [index, [opener, closer]] of RAW_BLOCKS.entries()) {
+    const open = opener.exec(line);
+    if (!open) {
+      continue;
+    }
+    const rest = line.slice(open[0].length);
+    return closer.test(rest) ? ["", null] : ["", index];
+  }
+  return [line, null];
+}
+
+/** True if `line` starts a CommonMark type 6 or type 7 HTML block. */
+function opensHtmlBlock(line: string, afterParagraph: boolean): boolean {
+  const named = HTML_BLOCK_OPEN.exec(line);
+  if (named && HTML_BLOCK_TAGS.has((named[1] ?? "").toLowerCase())) {
+    return true;
+  }
+  return !afterParagraph && HTML_TAG_ONLY_LINE.test(line);
+}
+
+/**
+ * The line as list tracking sees it. A comment or raw block renders nothing, but
+ * the line opening one is still a block at its own column, so it closes a list
+ * item it sits left of. Only the column survives, since the text it hides is not
+ * Markdown. A line inside a block already open is that block's content, so it
+ * keeps neither. A marker the hidden block is the content of survives with the
+ * column, so the item it opens is still tracked.
+ */
+function structuralLine(
+  line: string,
+  visible: string,
+  hidden: boolean,
+  marker: string,
+): string {
+  if (visible.trim() || hidden) {
+    return visible;
+  }
+  return hiddenStructure(line, marker);
+}
+
+interface ScanState {
+  openFence: string | null;
+  // Content column of the list item the open block belongs to, 0 at document
+  // level. A fence and an HTML block are scoped to their container, so the item's
+  // end closes them. Only one of the three is ever open.
+  blockColumn: number;
+  inComment: boolean;
+  // True while an inline comment opened above runs on into this line, carried by
+  // the paragraph holding it.
+  runOn: boolean;
+  inRawHtml: number | null;
+  inHtmlBlock: boolean;
+  afterParagraph: boolean;
+}
+
+interface ScannedLine {
+  // What a reader would see: "" for structure and hidden blocks, null for
+  // fenced content, which is skipped so it cannot split a bullet.
+  text: string | null;
+  // The same line as list tracking sees it: blank wherever nothing renders,
+  // but kept whole where an indent still closes an open item.
+  structural: string;
+}
+
+function visibleText(
+  line: string,
+  state: ScanState,
+  closesBelow: boolean,
+): ScannedLine {
+  // Raw HTML first: its contents are literal, so a fence inside it is not one.
+  if (state.inRawHtml !== null) {
+    const [after, stillInRaw] = stripRawHtml(line, state.inRawHtml);
+    state.inRawHtml = stillInRaw;
+    return { text: after, structural: "" };
+  }
+  if (state.inHtmlBlock) {
+    // A blank line is the only thing that ends a type 6 or 7 block.
+    state.inHtmlBlock = line.trim() !== "";
+    return { text: "", structural: "" };
+  }
+  // An opener is read past a marker on the same line, since a fence written as a
+  // list item's first content opens inside it. Only an opener: fenced content is
+  // literal and a closer carries no marker.
+  const commented = state.inComment || state.runOn;
+  const fence = commented
+    ? null
+    : FENCE.exec(
+        state.openFence === null
+          ? itemContent(line, state.afterParagraph)
+          : line,
+      );
+  // A backtick fence whose info string holds a backtick is prose, not a fence.
+  if (
+    fence &&
+    (state.openFence !== null || opensFence(fence[1] ?? "", fence[2] ?? ""))
+  ) {
+    state.openFence = nextFence(
+      state.openFence,
+      fence[1] ?? "",
+      fence[2] ?? "",
+    );
+    // Hidden from the collector, but its indent still closes an item.
+    return { text: "", structural: line };
+  }
+  if (state.openFence !== null) {
+    return { text: null, structural: "" };
+  }
+  return visibleContent(line, state, closesBelow);
+}
+
+/** `visibleText` for a line no fence or HTML block already owns. */
+function visibleContent(
+  line: string,
+  state: ScanState,
+  closesBelow: boolean,
+): ScannedLine {
+  // A block already open owns this line, so it is content rather than a block
+  // written at the column it happens to start in.
+  const hidden = state.inComment || state.inRawHtml !== null;
+  const carried = state.runOn;
+  // A comment is an HTML block too, so one written as a list item's first content
+  // opens inside that item exactly as a fence does: read past a marker on the
+  // same line rather than from the margin.
+  const content = itemContent(line, state.afterParagraph);
+  const opensComment =
+    !(state.inComment || carried) && COMMENT_BLOCK_OPEN.test(content);
+  // Commented-out notes are not rendered, so they are not previewed either.
+  const [uncommented, stillInComment, stillRunOn] = stripCommentSpans(
+    line,
+    state.inComment,
+    state.runOn,
+    closesBelow,
+    opensComment,
+  );
+  state.inComment = stillInComment;
+  state.runOn = stillRunOn;
+  const [visible, stillInRaw] = stripRawHtml(uncommented, state.inRawHtml);
+  state.inRawHtml = stillInRaw;
+  // Taken before the opener is hidden: it renders as nothing, but its indent still
+  // closes a list item it sits left of, and a marker on its line still opens one.
+  // A line an inline comment runs on into is still a line of the paragraph that
+  // carries it, so only its text is hidden, never its block structure.
+  const marker = opensComment
+    ? line.slice(0, line.length - content.length)
+    : "";
+  const structural = carried
+    ? line
+    : structuralLine(line, visible, hidden, marker);
+  if (
+    !carried &&
+    stillInRaw === null &&
+    visible.trim() &&
+    opensHtmlBlock(visible, state.afterParagraph)
+  ) {
+    state.inHtmlBlock = true;
+    return { text: "", structural };
+  }
+  return { text: visible, structural };
+}
+
+/**
+ * Marker of a fence the line scanner skipped because it is indented. Only a line
+ * within three columns of its item's content column is one: deeper than that it
+ * is an indented code block, which a dedented bullet ends.
+ */
+function opensDeepFence(line: ContentLine): string | null {
+  if (
+    line.indent < INDENTED_CODE_INDENT ||
+    line.indent - line.column >= INDENTED_CODE_INDENT
+  ) {
+    return null;
+  }
+  const fence = FENCE.exec(line.text);
+  return fence ? (fence[1] ?? null) : null;
+}
+
+/**
+ * True when `line` is the first one outside the deep fence opened with `marker`
+ * at `column`. A fence inside a list item runs only to the end of that item, so a
+ * line left of the item's content column closes both, as `fence_column` does on
+ * the backend.
+ */
+function endsDeepFence(
+  marker: string,
+  column: number,
+  line: ContentLine,
+): boolean {
+  return line.indent < column || closesDeepFence(marker, line);
+}
+
+/** True when `line` closes the deep fence opened with `marker`. */
+function closesDeepFence(marker: string, line: ContentLine): boolean {
+  const fence = FENCE.exec(line.text);
+  if (!fence) {
+    return false;
+  }
+  const closer = fence[1] ?? "";
+  return (
+    closer[0] === marker[0] &&
+    closer.length >= marker.length &&
+    !NON_SPACE.test(fence[2] ?? "")
+  );
+}
+
+/**
+ * Cells of a GFM table row, or null when the line holds no pipe at all. The
+ * optional leading and trailing pipes are delimiters, not empty cells, and a
+ * `\|` is literal text inside one.
+ */
+function tableCells(text: string): string[] | null {
+  if (!text.includes("|")) {
+    return null;
+  }
+  const cells: string[] = [];
+  let cell = "";
+  for (let at = 0; at < text.length; at += 1) {
+    const char = text[at];
+    if (char === "\\") {
+      cell += char + (text[at + 1] ?? "");
+      at += 1;
+      continue;
+    }
+    if (char === "|") {
+      cells.push(cell);
+      cell = "";
+      continue;
+    }
+    cell += char;
+  }
+  cells.push(cell);
+  if (cells.length > 1 && text.startsWith("|")) {
+    cells.shift();
+  }
+  if (cells.length > 1 && text.endsWith("|")) {
+    cells.pop();
+  }
+  return cells;
+}
+
+/** Width of a GFM delimiter row such as `| --- |:-:|`, or null if not one. */
+function delimiterWidth(text: string): number | null {
+  const cells = tableCells(text);
+  if (cells === null || cells.length === 0) {
+    return null;
+  }
+  return cells.every((cell) => TABLE_DELIMITER_CELL.test(cell.trim()))
+    ? cells.length
+    : null;
+}
+
+/**
+ * Line indices that belong to a GFM table. A table needs a header row and a
+ * delimiter row of the same width, and runs to a blank line or another block. Its
+ * cells render as a grid, not prose, so the preview drops them like a code block.
+ */
+function opensTable(
+  header: ContentLine | undefined,
+  delimiter: ContentLine | undefined,
+): boolean {
+  if (header === undefined || delimiter === undefined) {
+    return false;
+  }
+  if (!header.text || header.quoted) {
+    return false;
+  }
+  if (header.indent - header.column >= INDENTED_CODE_INDENT) {
+    return false;
+  }
+  const width = delimiterWidth(delimiter.text);
+  const cells = tableCells(header.text);
+  return width !== null && cells !== null && cells.length === width;
+}
+
+/** A blank line, a heading or a list marker: where GFM breaks a table. */
+function breaksTable(line: ContentLine | undefined): boolean {
+  return (
+    !line?.text ||
+    line.quoted ||
+    HEADING.test(line.text) ||
+    BULLET.test(line.text) ||
+    line.indent - line.column >= INDENTED_CODE_INDENT
+  );
+}
+
+function tableLines(lines: ContentLine[]): Set<number> {
+  const rows = new Set<number>();
+  let at = 0;
+  while (at + 1 < lines.length) {
+    if (!opensTable(lines[at], lines[at + 1])) {
+      at += 1;
+      continue;
+    }
+    rows.add(at);
+    rows.add(at + 1);
+    let row = at + 2;
+    while (row < lines.length && !breaksTable(lines[row])) {
+      rows.add(row);
+      row += 1;
+    }
+    at = row;
+  }
+  return rows;
+}
+
+/** A backtick fence's info string may not contain a backtick. */
+function opensFence(marker: string, rest: string): boolean {
+  return marker[0] !== "`" || !rest.includes("`");
+}
+
+function nextFence(
+  open: string | null,
+  marker: string,
+  rest: string,
+): string | null {
+  if (open === null) {
+    return opensFence(marker, rest) ? marker : null;
+  }
+  const closes =
+    marker[0] === open[0] &&
+    marker.length >= open.length &&
+    // Only spaces or tabs may follow a closer, per CommonMark.
+    !NON_SPACE.test(rest);
+  return closes ? null : open;
+}
+
+/** Whether a fence, a raw block, a comment or an HTML block is open. */
+function inBlock(state: ScanState): boolean {
+  return (
+    state.openFence !== null ||
+    state.inRawHtml !== null ||
+    state.inHtmlBlock ||
+    state.inComment
+  );
+}
+
+/**
+ * A fence, comment or HTML block inside a list item runs only to the end of that
+ * item, so a line dedented out of the item closes both. Lazy continuation reaches
+ * into none of them, so any content left of the item ends it.
+ */
+function closeDedentedBlock(line: string, state: ScanState): void {
+  if (state.blockColumn === 0 || !inBlock(state)) {
+    return;
+  }
+  if (line.trim() && indentWidth(line) < state.blockColumn) {
+    state.openFence = null;
+    state.inRawHtml = null;
+    state.inHtmlBlock = false;
+    state.inComment = false;
+    state.blockColumn = 0;
+  }
+}
+
+/** Ties a block just opened to the list item it is written inside. */
+function scopeBlock(
+  state: ScanState,
+  wasInBlock: boolean,
+  lists: ListState,
+): void {
+  if (!inBlock(state)) {
+    state.blockColumn = 0;
+    return;
+  }
+  if (!wasInBlock) {
+    // The opener closed the items it is dedented out of first, so this is the
+    // column of the item the block really sits in.
+    state.blockColumn = lists.columns.at(-1) ?? 0;
+  }
+}
+
+function contentLines(markdown: string): ContentLine[] {
+  const lines: ContentLine[] = [];
+  const state: ScanState = {
+    openFence: null,
+    blockColumn: 0,
+    inComment: false,
+    runOn: false,
+    inRawHtml: null,
+    inHtmlBlock: false,
+    afterParagraph: false,
+  };
+  let lists: ListState = EMPTY_LIST_STATE;
+  let quote: QuoteState = NO_QUOTE;
+
+  const rawLines = markdown
+    .split("\n")
+    .map((raw) => raw.replace(TABS, " ".repeat(TAB_WIDTH)));
+  const closesBelow = commentClosesBelow(rawLines);
+  for (const [index, line] of rawLines.entries()) {
+    closeDedentedBlock(line, state);
+    const wasInBlock = inBlock(state);
+    const carried = state.runOn;
+    const { text: visible, structural } = visibleText(
+      line,
+      state,
+      closesBelow[index + 1] ?? false,
+    );
+    // The quote state from the line above, which is what list tracking asks about.
+    // Only a line of text below rewrites it, so a fenced, blank or hidden line
+    // leaves no quoted paragraph open behind it.
+    const above = quote;
+    quote = NO_QUOTE;
+    // Taken with the paragraph state from the line above, as a renderer would.
+    lists = openLists(structural, lists, state.afterParagraph, above.quoted);
+    scopeBlock(state, wasInBlock, lists);
+    if (visible === null) {
+      continue;
+    }
+    if (carried && !visible.trim()) {
+      // Wholly inside a comment its paragraph carries: no text, and no break.
+      continue;
+    }
+    if (!visible.trim() || THEMATIC_BREAK.test(visible)) {
+      // A rule separates notes, so it breaks a bullet just like a blank line.
+      state.afterParagraph = false;
+      lines.push({ text: "", indent: 0, quoted: false, column: 0 });
+      continue;
+    }
+    const quoted = BLOCKQUOTE.test(visible);
+    const stripped = visible.replace(BLOCKQUOTE, "");
+    const indent = stripped.length - stripped.trimStart().length;
+    // A quoted line is measured inside its quote, where the document's open
+    // list items do not reach.
+    const column = quoted ? 0 : (lists.columns.at(-1) ?? 0);
+    // Only ordinary text continues a paragraph; a heading or indented code line
+    // (four columns past its container, outside a paragraph) ends one.
+    const startsCode =
+      !state.afterParagraph && indent - column >= INDENTED_CODE_INDENT;
+    state.afterParagraph = !HEADING_LINE.test(stripped) && !startsCode;
+    quote = quoteState(visible, above.inQuote);
+    lines.push({ text: stripped.trim(), indent, quoted, column });
+  }
+  return lines;
+}
+
+/**
+ * Split a bullet at its first sentence boundary. Conservative: the next
+ * sentence must start like one, so "CHANGELOG.md in the repo" is not a break.
+ */
+function splitLeadSentence(text: string): ReleaseNotesPreviewItem {
+  SENTENCE_BREAK.lastIndex = 0;
+  let match = SENTENCE_BREAK.exec(text);
+  while (match) {
+    const cut = match.index + 1;
+    const word =
+      TRAILING_WORD.exec(text.slice(0, cut))?.[1]?.toLowerCase() ?? "";
+    const isAbbreviation = ABBREVIATIONS.has(word) || INITIAL.test(word);
+    if (!isAbbreviation && cut >= MIN_LEAD_CHARS) {
+      return { lead: text.slice(0, cut).trim(), rest: text.slice(cut).trim() };
+    }
+    match = SENTENCE_BREAK.exec(text);
+  }
+  return { lead: text, rest: "" };
+}
+
+/** Bullets in document order, plus prose for changelogs written as paragraphs. */
+interface Collector {
+  bullets: Bullet[];
+  prose: string[];
+  // Wrapped bullets continue on following lines and belong to one item.
+  current: Bullet | null;
+  paragraph: string;
+  // True while the open paragraph is a quote's, which owns its own text: a
+  // marker written outside the quote opens a list rather than continuing it.
+  quotedParagraph: boolean;
+}
+
+function flush(collector: Collector): void {
+  if (collector.current?.text) {
+    collector.bullets.push({
+      text: truncate(collector.current.text),
+      indent: collector.current.indent,
+    });
+  }
+  collector.current = null;
+  if (collector.paragraph) {
+    collector.prose.push(truncate(collector.paragraph));
+    collector.paragraph = "";
+  }
+  collector.quotedParagraph = false;
+}
+
+function takeBullet(
+  collector: Collector,
+  text: string,
+  line: ContentLine,
+  labels: Set<string>,
+): void {
+  flush(collector);
+  const item = toPlainText(text, labels);
+  // A quoted list is example output: prose at best, never a headline bullet.
+  if (!line.quoted) {
+    collector.current = { text: item, indent: line.indent };
+  } else if (item) {
+    collector.prose.push(truncate(item));
+  }
+}
+
+function takeText(
+  collector: Collector,
+  text: string,
+  labels: Set<string>,
+  quoted: boolean,
+): void {
+  const plain = toPlainText(text, labels);
+  if (!plain) {
+    return;
+  }
+  if (collector.current === null) {
+    // Wrapped paragraphs render as one block, so preview them as one item.
+    collector.paragraph = collector.paragraph
+      ? `${collector.paragraph} ${plain}`
+      : plain;
+    collector.quotedParagraph = quoted;
+    return;
+  }
+  collector.current = {
+    text: `${collector.current.text} ${plain}`,
+    indent: collector.current.indent,
+  };
+}
+
+function collectBullets(markdown: string): {
+  bullets: Bullet[];
+  prose: string[];
+} {
+  const collector: Collector = {
+    bullets: [],
+    prose: [],
+    current: null,
+    paragraph: "",
+    quotedParagraph: false,
+  };
+
+  const lines = contentLines(markdown);
+  const labels = new Set<string>();
+  // Skips the same code the pass below skips: a definition-shaped line inside
+  // code is literal, and a real definition never indents past three spaces.
+  let labelFence: string | null = null;
+  let labelColumn = 0;
+  for (const line of lines) {
+    if (labelFence !== null && !endsDeepFence(labelFence, labelColumn, line)) {
+      continue;
+    }
+    if (labelFence !== null) {
+      const dedented = line.indent < labelColumn;
+      labelFence = null;
+      // Its own closing line is code too; only a dedented one is a new block.
+      if (!dedented) {
+        continue;
+      }
+    }
+    const opener = opensDeepFence(line);
+    if (opener !== null) {
+      labelFence = opener;
+      labelColumn = line.column;
+      continue;
+    }
+    if (line.indent - line.column >= INDENTED_CODE_INDENT) {
+      continue;
+    }
+    const definition = DEFINITION.exec(line.text);
+    if (definition) {
+      labels.add(
+        (definition[1] ?? "").trim().replace(WHITESPACE, " ").toLowerCase(),
+      );
+    }
+  }
+
+  const tables = tableLines(lines);
+  let deepFence: string | null = null;
+  let deepColumn = 0;
+  for (const [index, line] of lines.entries()) {
+    if (!line.text || HEADING.test(line.text)) {
+      flush(collector);
+      continue;
+    }
+    // A table renders as a grid, no more previewable than a code block, and it
+    // ends whatever came before it.
+    if (tables.has(index)) {
+      flush(collector);
+      continue;
+    }
+    // A link reference definition renders as nothing.
+    if (collector.current === null && DEFINITION.test(line.text)) {
+      continue;
+    }
+    // A fence indented past three spaces belongs to a list item, so the line
+    // scanner missed it. Its contents are code either way.
+    if (deepFence !== null && !endsDeepFence(deepFence, deepColumn, line)) {
+      continue;
+    }
+    if (deepFence !== null) {
+      const dedented = line.indent < deepColumn;
+      deepFence = null;
+      // Its own closing line is code too; only a dedented one is a new block.
+      if (!dedented) {
+        continue;
+      }
+    }
+    const opener = opensDeepFence(line);
+    if (opener !== null) {
+      deepFence = opener;
+      deepColumn = line.column;
+      continue;
+    }
+    // An indented code block renders as code, so a "- cmd" line in one is not
+    // a bullet. Inside an open bullet or paragraph it is just a wrapped line.
+    const insideBlock =
+      collector.current !== null || collector.paragraph !== "";
+    if (!insideBlock && line.indent - line.column >= INDENTED_CODE_INDENT) {
+      continue;
+    }
+    const bullet = BULLET.exec(line.text);
+    // Only an ordered list starting at 1 may interrupt a paragraph, so "2. Restart
+    // Studio" under prose is prose. A list item is not a paragraph.
+    const interrupts =
+      collector.current === null &&
+      collector.paragraph !== "" &&
+      !collector.quotedParagraph;
+    if (
+      bullet &&
+      !(interrupts && bullet[1] !== undefined && bullet[1] !== "1")
+    ) {
+      takeBullet(collector, bullet[2] ?? "", line, labels);
+      continue;
+    }
+    takeText(collector, line.text, labels, line.quoted);
+  }
+  flush(collector);
+
+  return { bullets: collector.bullets, prose: collector.prose };
+}
+
+/**
+ * Top-level bullets of a release section, in document order. Nested bullets are
+ * detail and are skipped; prose is used when a release has no bullets.
+ */
+export function releaseNotesPreview(
+  markdown: string | null | undefined,
+  limit: number = RELEASE_NOTES_PREVIEW_ITEMS,
+): ReleaseNotesPreview {
+  if (!markdown) {
+    return { items: [], remaining: 0 };
+  }
+
+  // The updater body arrives with CRLF; sentinels would collide with parking.
+  const text = markdown.replace(LINE_ENDINGS, "\n").replace(SENTINELS, "");
+  const { bullets, prose } = collectBullets(text);
+  // Shallowest bullet defines top level, so a uniformly indented list previews.
+  const baseIndent = bullets.reduce(
+    (min, bullet) => Math.min(min, bullet.indent),
+    Number.POSITIVE_INFINITY,
+  );
+  const topLevel = bullets
+    .filter((bullet) => bullet.indent <= baseIndent + NESTED_INDENT_TOLERANCE)
+    .map((bullet) => bullet.text);
+
+  const source = topLevel.length > 0 ? topLevel : prose;
+  return {
+    items: source.slice(0, limit).map(splitLeadSentence),
+    remaining: Math.max(source.length - limit, 0),
+  };
+}
diff --git a/studio/frontend/tests/sidebar-width.test.ts b/studio/frontend/tests/sidebar-width.test.ts
new file mode 100644
index 0000000000..861bd45e3f
--- /dev/null
+++ b/studio/frontend/tests/sidebar-width.test.ts
@@ -0,0 +1,72 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import assert from "node:assert/strict";
+import test from "node:test";
+import { readFile } from "node:fs/promises";
+
+// Every localStorage key written by a panel width store.
+const PANEL_WIDTH_KEYS = ["sidebar_width", "chat_settings_width"];
+
+// The store reads window at import time, so stub it before importing.
+const stubWindow = {
+  innerWidth: 1440,
+  localStorage: {
+    getItem: () => null,
+    setItem: () => {},
+  },
+  addEventListener: () => {},
+  removeEventListener: () => {},
+};
+(globalThis as { window?: unknown }).window = stubWindow;
+
+const {
+  clampSidebarWidth,
+  SIDEBAR_WIDTH_DEFAULT,
+  SIDEBAR_WIDTH_MAX,
+  SIDEBAR_WIDTH_MIN,
+} = await import("../src/hooks/use-sidebar-width.ts");
+
+test("clamps to the absolute range on a roomy window", () => {
+  stubWindow.innerWidth = 1440;
+  assert.equal(clampSidebarWidth(320), 320);
+  assert.equal(clampSidebarWidth(SIDEBAR_WIDTH_MAX + 200), SIDEBAR_WIDTH_MAX);
+  assert.equal(clampSidebarWidth(10), SIDEBAR_WIDTH_MIN);
+  assert.equal(clampSidebarWidth(Number.NaN), SIDEBAR_WIDTH_DEFAULT);
+});
+
+test("caps at 40% of a narrow window", () => {
+  stubWindow.innerWidth = 800;
+  assert.equal(clampSidebarWidth(SIDEBAR_WIDTH_MAX), 320);
+  assert.equal(clampSidebarWidth(300), 300);
+});
+
+test("the floor still wins when 40% falls below it", () => {
+  stubWindow.innerWidth = 500;
+  assert.equal(clampSidebarWidth(SIDEBAR_WIDTH_MAX), SIDEBAR_WIDTH_MIN);
+});
+
+test("re-evaluates the cap per call, so a resize can re-clamp", () => {
+  stubWindow.innerWidth = 1440;
+  assert.equal(clampSidebarWidth(SIDEBAR_WIDTH_MAX), SIDEBAR_WIDTH_MAX);
+  stubWindow.innerWidth = 900;
+  assert.equal(clampSidebarWidth(SIDEBAR_WIDTH_MAX), 360);
+  stubWindow.innerWidth = 1440;
+  assert.equal(clampSidebarWidth(SIDEBAR_WIDTH_MAX), SIDEBAR_WIDTH_MAX);
+});
+
+// The reset action promises to clear every stored preference, so a persisted
+// panel width that is missing from the list survives the reload.
+test("persisted panel widths are cleared by the preference reset", async () => {
+  const source = await readFile(
+    new URL("../src/features/settings/tabs/general-tab.tsx", import.meta.url),
+    "utf8",
+  );
+  const keys = source.slice(
+    source.indexOf("const PREFS_KEYS"),
+    source.indexOf("];", source.indexOf("const PREFS_KEYS")),
+  );
+  for (const key of PANEL_WIDTH_KEYS) {
+    assert.ok(keys.includes(`"${key}"`), `${key} missing from PREFS_KEYS`);
+  }
+});
diff --git a/studio/install_node_prebuilt.py b/studio/install_node_prebuilt.py
index 82ca1d2c68..1f42729c80 100644
--- a/studio/install_node_prebuilt.py
+++ b/studio/install_node_prebuilt.py
@@ -707,18 +707,55 @@ def existing_install_usable(install_dir: Path, host: HostInfo) -> bool:
     return npm_major is not None and npm_major >= NPM_MIN_MAJOR
 
 
+def _replace_with_retry(
+    src: Path,
+    dst: Path,
+    *,
+    attempts: int = 8,
+) -> None:
+    """os.replace, retried against transient Windows sharing violations.
+
+    A directory rename fails with WinError 5/32 while any process holds a handle inside
+    it, and Defender or the indexer routinely does right after extraction (seen in CI on
+    a fresh install, with no existing directory to conflict with). Handles clear in a
+    second or two, so a bounded backoff turns the failure into a pause; other errors
+    raise immediately rather than stalling on a real problem.
+    """
+    delay = 0.25
+    for attempt in range(attempts):
+        try:
+            os.replace(src, dst)
+            return
+        except OSError as exc:
+            transient = os.name == "nt" and getattr(exc, "winerror", None) in (5, 32, 145)
+            if not transient or attempt == attempts - 1:
+                raise
+            log(
+                f"rename blocked ({exc.winerror}), retrying in {delay:.2f}s "
+                f"-- a scanner is likely still holding the extracted files"
+            )
+            time.sleep(delay)
+            delay = min(delay * 2, 4.0)
+
+
 def _swap_into_place(extracted_root: Path, install_dir: Path) -> None:
     """Atomically replace install_dir with extracted_root (same filesystem)."""
     install_dir.parent.mkdir(parents = True, exist_ok = True)
     backup: Path | None = None
     if install_dir.exists():
         backup = install_dir.parent / f".{install_dir.name}.old-{os.getpid()}"
-        os.replace(install_dir, backup)
+        _replace_with_retry(install_dir, backup)
     try:
-        os.replace(extracted_root, install_dir)
+        _replace_with_retry(extracted_root, install_dir)
     except OSError:
+        # The forward rename retries ~16s, ample time for a scanner to grab the backup too.
+        # A plain os.replace would then raise over the original error and leave no
+        # install_dir at all, so the rollback gets the same backoff and never masks it.
         if backup is not None and not install_dir.exists():
-            os.replace(backup, install_dir)
+            try:
+                _replace_with_retry(backup, install_dir)
+            except OSError as rollback_exc:
+                log(f"could not restore the previous Node install from {backup}: {rollback_exc}")
         raise
     if backup is not None:
         shutil.rmtree(backup, ignore_errors = True)
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index a4eb54a9ef..0b6cf292c2 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -869,12 +869,22 @@ function Ensure-BuildToolsForLlamaSourceBuild {
     }
 }
 
-# Detect the VC++ 2015-2022 Redistributable that the prebuilt llama-server and
-# PyTorch need (they link VCRUNTIME140_1.dll etc., which the Universal CRT lacks).
-# Signal is System32\vcruntime140_1.dll (VS 2019+), registry as fallback.
+# Machine arch: PROCESSOR_ARCHITECTURE describes this PROCESS, so an emulated x64 shell on
+# ARM64 reports AMD64; PROCESSOR_ARCHITEW6432 is ARM64 in exactly that case.
+function Get-HostMachineArch {
+    $osArch = ""
+    try { $osArch = [System.Runtime.InteropServices.RuntimeInformation]::OSArchitecture.ToString() } catch { }
+    foreach ($s in @([string]$env:PROCESSOR_ARCHITEW6432, [string]$env:PROCESSOR_ARCHITECTURE, $osArch)) {
+        if ($s.ToLowerInvariant() -eq "arm64") { return "arm64" }
+    }
+    return "other"
+}
+
+# Detect the VC++ 2015-2022 Redistributable prebuilt llama-server and PyTorch need (they
+# link VCRUNTIME140_1.dll, absent from the Universal CRT). Registry first: Runtimes\x64 is
+# the only x64-specific proof; System32\vcruntime140_1.dll is arch-blind and on ARM64 may
+# be the ARM64-only package, unloadable under x64 emulation.
 function Test-VCRedistInstalled {
-    $sys = $env:SystemRoot
-    if ($sys -and (Test-Path (Join-Path $sys 'System32\vcruntime140_1.dll'))) { return $true }
     foreach ($k in @(
         'HKLM:\SOFTWARE\Microsoft\VisualStudio\14.0\VC\Runtimes\x64',
         'HKLM:\SOFTWARE\WOW6432Node\Microsoft\VisualStudio\14.0\VC\Runtimes\x64'
@@ -884,10 +894,14 @@ function Test-VCRedistInstalled {
             if ($r.Installed -eq 1 -and [int]$r.Major -ge 14 -and [int]$r.Minor -ge 20) { return $true }
         } catch { }
     }
+    if ((Get-HostMachineArch) -eq "arm64") { return $false }
+    $sys = $env:SystemRoot
+    if ($sys -and (Test-Path (Join-Path $sys 'System32\vcruntime140_1.dll'))) { return $true }
     return $false
 }
 
-# Install the VC++ 2015-2022 runtime if missing (non-fatal; usually a no-op).
+# Install the VC++ 2015-2022 runtime if missing (non-fatal; usually a no-op). Unlike CMake
+# and Build Tools torch cannot import without it, and winget is absent on LTSC/Server images.
 function Ensure-VCRedist {
     if (Test-VCRedistInstalled) { step "vcredist" "present"; return }
     Write-Host "Microsoft Visual C++ Redistributable (2015-2022) is missing; the prebuilt llama.cpp and PyTorch need it. Installing the runtime..." -ForegroundColor Yellow
@@ -897,6 +911,45 @@ function Ensure-VCRedist {
             Refresh-Environment
         } catch { substep "VCRedist install failed: $($_.Exception.Message)" "Yellow" }
     }
+    if (-not (Test-VCRedistInstalled)) {
+        # Evergreen link; /quiet /norestart so it never blocks or reboots an unattended run.
+        # Always the x64 package, deliberately: Microsoft ships it as the Arm64X superset of
+        # both ARM64 and X64 binaries and documents it as the one for ARM64 devices, while
+        # the arm64 package is ARM64-only (learn.microsoft.com/cpp/windows/latest-supported-vc-redist).
+        # PROCESSOR_ARCHITECTURE is wrong twice here: it reports the process, and the runtime
+        # must match the interpreter loading the DLLs, an emulated x64 Python not yet created.
+        $url = "https://aka.ms/vs/17/release/vc_redist.x64.exe"
+        $dst = Join-Path ([System.IO.Path]::GetTempPath()) "vc_redist.x64.exe"
+        substep "winget unavailable or failed; downloading the runtime directly..."
+        # Windows PowerShell 5.1 on an old image can carry a .NET default protocol set that
+        # predates TLS 1.2, which aka.ms refuses -- exactly the no-winget host this fallback
+        # exists for. SystemDefault (0) means "let the OS choose" and already covers TLS 1.2+,
+        # so only an explicit legacy set is upgraded, and it is restored afterwards.
+        $_prevProtocol = $null
+        try {
+            $_cur = [System.Net.ServicePointManager]::SecurityProtocol
+            if ([int]$_cur -ne 0 -and ([int]$_cur -band [int][System.Net.SecurityProtocolType]::Tls12) -eq 0) {
+                [System.Net.ServicePointManager]::SecurityProtocol = $_cur -bor [System.Net.SecurityProtocolType]::Tls12
+                $_prevProtocol = $_cur
+            }
+        } catch { $_prevProtocol = $null }
+        try {
+            Invoke-WebRequest -Uri $url -OutFile $dst -UseBasicParsing -TimeoutSec 300
+            $p = Start-Process -FilePath $dst -ArgumentList '/quiet', '/norestart' -Wait -PassThru
+            # 3010 = success, reboot required; usable either way.
+            if ($p.ExitCode -notin @(0, 3010)) {
+                substep "VC++ runtime installer exited $($p.ExitCode)" "Yellow"
+            }
+            Refresh-Environment
+        } catch {
+            substep "Direct VC++ runtime download failed: $($_.Exception.Message)" "Yellow"
+        } finally {
+            if ($null -ne $_prevProtocol) {
+                try { [System.Net.ServicePointManager]::SecurityProtocol = $_prevProtocol } catch { }
+            }
+            Remove-Item -LiteralPath $dst -Force -ErrorAction SilentlyContinue
+        }
+    }
     if (Test-VCRedistInstalled) { step "vcredist" "installed" }
     else {
         substep "Could not install the VC++ Redistributable automatically." "Yellow"
@@ -1650,11 +1703,42 @@ if ($LongPathsEnabled) {
 }
 
 # ============================================
-# 1b. Git (required by pip for git+https:// deps and by npm)
+# 1b. Git (only required for --local / source installs)
 # ============================================
+# Was fatal as "required by pip and npm", but the consumer path uses neither: the
+# unsloth-zoo git+https URL is STUDIO_LOCAL_INSTALL only, node is a pinned prebuilt, and the
+# frontend lockfile has no VCS deps. Being fatal blocked clean no-winget Windows boxes.
 $HasGit = $null -ne (Get-Command git -ErrorAction SilentlyContinue)
 if (-not $HasGit) {
-    Write-Host "Git not found -- installing via winget..." -ForegroundColor Yellow
+    # Fatal only where git is used: --local and the opt-in llama.cpp source build. A local
+    # llama.cpp dir overrides those opt-ins, but only once it holds a reusable binary:
+    # pointing at the canonical install location with nothing built there falls through to
+    # the normal install, so an explicit source build still needs git. The automatic
+    # fallback after a failed prebuilt download is not knowable here; Phase 4 handles it.
+    $gitNeeded = ($env:STUDIO_LOCAL_INSTALL -eq '1')
+    $_localLlamaDir = if ($env:UNSLOTH_LOCAL_LLAMA_CPP_DIR) { $env:UNSLOTH_LOCAL_LLAMA_CPP_DIR.Trim() } else { "" }
+    $_localLlamaBuilt = $false
+    if ($_localLlamaDir) {
+        # Same layout candidates as the reuse check in Phase 4.
+        foreach ($_c in @("llama-server.exe", "build\bin\llama-server.exe", "build\bin\Release\llama-server.exe")) {
+            if (Test-Path -LiteralPath (Join-Path $_localLlamaDir $_c)) { $_localLlamaBuilt = $true; break }
+        }
+    }
+    if (-not $_localLlamaBuilt) {
+        $_prForce = if ($env:UNSLOTH_LLAMA_PR_FORCE) { $env:UNSLOTH_LLAMA_PR_FORCE.Trim() } else { $DefaultLlamaPrForce }
+        $_llamaSrc = $DefaultLlamaSource -replace '\.git$', ''
+        # Same tag resolution as Phase 4. "master" is a branch, never a release, so the
+        # prebuilt lookup always misses and Phase 4 rebuilds it from source.
+        $_llamaTag = if ($env:UNSLOTH_LLAMA_TAG) { $env:UNSLOTH_LLAMA_TAG } else { $DefaultLlamaTag }
+        if ($_llamaTag -eq "master") { $gitNeeded = $true }
+        if ($env:UNSLOTH_LLAMA_FORCE_COMPILE -eq '1') { $gitNeeded = $true }
+        if (-not [string]::IsNullOrWhiteSpace($env:UNSLOTH_LLAMA_PR)) { $gitNeeded = $true }
+        # Same positive-integer predicate as the PR_FORCE promotion below: 0 or non-numeric
+        # never forces a source build, so it must not demand git.
+        if ($_prForce -match '^\d+$' -and [int]$_prForce -gt 0) { $gitNeeded = $true }
+        if ($_llamaSrc -ne "https://github.com/ggml-org/llama.cpp") { $gitNeeded = $true }
+    }
+    Write-Host "Git not found -- attempting install via winget..." -ForegroundColor Yellow
     $HasWinget = $null -ne (Get-Command winget -ErrorAction SilentlyContinue)
     if ($HasWinget) {
         try {
@@ -1664,11 +1748,18 @@ if (-not $HasGit) {
         } catch { }
     }
     if (-not $HasGit) {
-        Write-Host "[ERROR] Git is required but could not be installed automatically." -ForegroundColor Red
-        Write-Host "        Install Git from https://git-scm.com/download/win and re-run." -ForegroundColor Red
-        Exit-SetupFailure "Git is required but could not be installed automatically"
+        if ($gitNeeded) {
+            Write-Host "[ERROR] Git is required for --local and llama.cpp source-build installs but could not be installed." -ForegroundColor Red
+            Write-Host "        --local clones unsloth-zoo, and a source build clones llama.cpp." -ForegroundColor Red
+            Write-Host "        Install Git from https://git-scm.com/download/win and re-run." -ForegroundColor Red
+            Exit-SetupFailure "Git is required for --local / source-build installs but could not be installed"
+        }
+        step "git" "not found (not required)" "Yellow"
+        substep "Unsloth installs prebuilt binaries and wheels, so git is not needed."
+        substep "Install it only for --local/source installs: https://git-scm.com/download/win"
+    } else {
+        step "git" "$(git --version)"
     }
-    step "git" "$(git --version)"
 } else {
     step "git" "$(git --version)"
 }
@@ -3275,18 +3366,32 @@ $PyTorchWhlBase = if ($env:UNSLOTH_PYTORCH_MIRROR) { $env:UNSLOTH_PYTORCH_MIRROR
 $TorchInstallIndexUrl = if ($ROCmIndexUrl) { "$PyTorchWhlBase/cpu" } elseif ($PinnedTorchIndexUrl) { $PinnedTorchIndexUrl } else { "$PyTorchWhlBase/$CuTag" }
 
 if (-not $NoTorchMode) {
+# Windows on ARM has win_arm64 torch and torchvision wheels but no torchaudio on any index,
+# so every branch below drops it. Ask the interpreter uv resolves for, not
+# PROCESSOR_ARCHITECTURE, which describes the host process. Inside the no-torch guard
+# because all three uses are, and no-torch installs nothing to skip.
+$_setupPlatform = ""
+try {
+    $_setupPlatform = (& python -c "import sysconfig; print(sysconfig.get_platform())" 2>$null | Out-String).Trim().ToLowerInvariant()
+} catch { $_setupPlatform = "" }
+$WinArm64NoAudio = ($_setupPlatform -eq "win-arm64")
+if ($WinArm64NoAudio) { substep "windows on arm: skipping torchaudio (no win_arm64 wheel upstream)" }
+
 $ROCmCpuFallback = $false
 if ($ROCmIndexUrl) {
     substep "installing PyTorch (AMD ROCm, $ROCmGfxArch)..."
     if ($ROCmTorchSpec -ne "torch") {
         substep "  enforcing $ROCmTorchSpec $ROCmVisionSpec $ROCmAudioSpec (known _grouped_mm bug in older wheels)" "Cyan"
     }
+    # Built above the verbose branch: a splat assigned inside it is unset on the other.
+    $_rocmTrio = @($ROCmTorchSpec, $ROCmVisionSpec, $ROCmAudioSpec)
+    if ($WinArm64NoAudio) { $_rocmTrio = @($ROCmTorchSpec, $ROCmVisionSpec) }
     if ($script:UnslothVerbose) {
-        Fast-Install $ROCmTorchSpec $ROCmVisionSpec $ROCmAudioSpec --force-reinstall --index-url $ROCmIndexUrl | ForEach-Object { Redact-InstallOutput "$_" } | Out-Host
+        Fast-Install @_rocmTrio --force-reinstall --index-url $ROCmIndexUrl | ForEach-Object { Redact-InstallOutput "$_" } | Out-Host
         $torchInstallExit = $LASTEXITCODE
         $output = ""
     } else {
-        $output = Fast-Install $ROCmTorchSpec $ROCmVisionSpec $ROCmAudioSpec --force-reinstall --index-url $ROCmIndexUrl | Out-String
+        $output = Fast-Install @_rocmTrio --force-reinstall --index-url $ROCmIndexUrl | Out-String
         $torchInstallExit = $LASTEXITCODE
     }
     if ($torchInstallExit -ne 0) {
@@ -3322,12 +3427,14 @@ if (-not $ROCmIndexUrl -and ($CuTag -eq "cpu" -or $ROCmCpuFallback)) {
         $cpuVisionSpec = "torchvision>=0.19,<0.27.0"
         $cpuAudioSpec  = "torchaudio>=2.4,<2.12.0"
     }
+    $_torchTrio = @($cpuTorchSpec, $cpuVisionSpec, $cpuAudioSpec)
+    if ($WinArm64NoAudio) { $_torchTrio = @($cpuTorchSpec, $cpuVisionSpec) }
     if ($script:UnslothVerbose) {
-        Fast-Install $cpuTorchSpec $cpuVisionSpec $cpuAudioSpec @cpuForce --index-url $TorchInstallIndexUrl | ForEach-Object { Redact-InstallOutput "$_" } | Out-Host
+        Fast-Install @_torchTrio @cpuForce --index-url $TorchInstallIndexUrl | ForEach-Object { Redact-InstallOutput "$_" } | Out-Host
         $torchInstallExit = $LASTEXITCODE
         $output = ""
     } else {
-        $output = Fast-Install $cpuTorchSpec $cpuVisionSpec $cpuAudioSpec @cpuForce --index-url $TorchInstallIndexUrl | Out-String
+        $output = Fast-Install @_torchTrio @cpuForce --index-url $TorchInstallIndexUrl | Out-String
         $torchInstallExit = $LASTEXITCODE
     }
     if ($torchInstallExit -ne 0) {
@@ -3354,12 +3461,16 @@ if (-not $ROCmIndexUrl -and ($CuTag -eq "cpu" -or $ROCmCpuFallback)) {
         $cudaVisionSpec = "torchvision>=0.19,<0.26.0"
         $cudaAudioSpec = "torchaudio>=2.4,<2.11.0"
     }
+    # A custom pin whose leaf is not cpu (a corporate /simple mirror) lands an ARM64 host
+    # here, so this branch drops torchaudio too.
+    $_cudaTrio = @($cudaTorchSpec, $cudaVisionSpec, $cudaAudioSpec)
+    if ($WinArm64NoAudio) { $_cudaTrio = @($cudaTorchSpec, $cudaVisionSpec) }
     if ($script:UnslothVerbose) {
-        Fast-Install $cudaTorchSpec $cudaVisionSpec $cudaAudioSpec @cudaForce --index-url $TorchInstallIndexUrl | ForEach-Object { Redact-InstallOutput "$_" } | Out-Host
+        Fast-Install @_cudaTrio @cudaForce --index-url $TorchInstallIndexUrl | ForEach-Object { Redact-InstallOutput "$_" } | Out-Host
         $torchInstallExit = $LASTEXITCODE
         $output = ""
     } else {
-        $output = Fast-Install $cudaTorchSpec $cudaVisionSpec $cudaAudioSpec @cudaForce --index-url $TorchInstallIndexUrl | Out-String
+        $output = Fast-Install @_cudaTrio @cudaForce --index-url $TorchInstallIndexUrl | Out-String
         $torchInstallExit = $LASTEXITCODE
     }
     if ($torchInstallExit -ne 0) {
@@ -4048,6 +4159,7 @@ $BuildDir = Join-Path $LlamaCppDir "build"
 $LlamaServerBin = Join-Path $BuildDir "bin\Release\llama-server.exe"
 
 $HasCmakeForBuild = $null -ne (Get-Command cmake -ErrorAction SilentlyContinue)
+$HasGitForBuild = $null -ne (Get-Command git -ErrorAction SilentlyContinue)
 
 # Check if existing llama-server matches current GPU mode. A CUDA-built binary
 # on a now-CPU-only machine (or vice versa) needs to be rebuilt.
@@ -4073,9 +4185,27 @@ if (Test-Path -LiteralPath $LlamaServerBin) {
 $WillBuildLlamaFromSource = $NeedLlamaSourceBuild -and `
     -not ((Test-Path -LiteralPath $LlamaServerBin) -and -not $NeedRebuild -and $RequestedLlamaTag -ne "master")
 if ($WillBuildLlamaFromSource) {
-    Ensure-BuildToolsForLlamaSourceBuild
-    # refresh so the chain below sees a newly installed cmake
-    $HasCmakeForBuild = $null -ne (Get-Command cmake -ErrorAction SilentlyContinue)
+    if (-not $HasGitForBuild) {
+        # Phase 1 keeps git optional, so only the automatic fallback after a failed prebuilt
+        # download arrives here without it. Last chance to install: Invoke-SetupCommand
+        # returns 0 for command-not-found, so a git-less clone misreports as a cmake failure.
+        if ($null -ne (Get-Command winget -ErrorAction SilentlyContinue)) {
+            try {
+                Invoke-SetupCommand { winget install Git.Git --source winget --accept-package-agreements --accept-source-agreements } | Out-Null
+                Refresh-Environment
+            } catch { }
+        }
+        $HasGitForBuild = $null -ne (Get-Command git -ErrorAction SilentlyContinue)
+    }
+    # Git first, then the toolchain: Ensure-BuildToolsForLlamaSourceBuild exits setup when
+    # Build Tools cannot be installed, so running it first made the degraded path below
+    # unreachable on a no-winget box, and elsewhere spent a multi-GB download on a clone
+    # that cannot happen.
+    if ($HasGitForBuild) {
+        Ensure-BuildToolsForLlamaSourceBuild
+        # refresh so the chain below sees a newly installed cmake
+        $HasCmakeForBuild = $null -ne (Get-Command cmake -ErrorAction SilentlyContinue)
+    }
 }
 
 if ($LocalLlamaCppLinked) {
@@ -4093,6 +4223,16 @@ if ($LocalLlamaCppLinked) {
     # up new model architecture support (e.g. Gemma 4).
     Write-Host ""
     step "llama.cpp" "already built"
+} elseif (-not $HasGitForBuild) {
+    # Before cmake: the toolchain install is skipped without git, so cmake may be missing
+    # purely as a consequence. Degrade rather than abort; the opt-in source triggers already
+    # required git in Phase 1, so only the automatic fallback lands here.
+    Write-Host ""
+    step "llama.cpp" "build skipped (git not available)" "Yellow"
+    substep "The prebuilt download failed and a source build clones llama.cpp." "Yellow"
+    substep "GGUF inference and export will not be available." "Yellow"
+    substep "Install Git from https://git-scm.com/download/win and re-run setup." "Yellow"
+    $script:LlamaCppDegraded = $true
 } elseif (-not $HasCmakeForBuild) {
     Write-Host ""
     if (-not $HasNvidiaSmi) {
diff --git a/studio/src-tauri/src/desktop_update_policy.rs b/studio/src-tauri/src/desktop_update_policy.rs
index c83f847eda..d186c2d01d 100644
--- a/studio/src-tauri/src/desktop_update_policy.rs
+++ b/studio/src-tauri/src/desktop_update_policy.rs
@@ -27,6 +27,8 @@ pub(crate) struct DesktopUpdatePolicy {
 pub(crate) struct ManualUpdateInfo {
     version: String,
     current_version: String,
+    // Backend release this desktop build pins; CHANGELOG.md is keyed by it.
+    pypi_version: Option<String>,
     body: Option<String>,
     date: Option<String>,
 }
@@ -34,8 +36,12 @@ pub(crate) struct ManualUpdateInfo {
 #[derive(Debug, serde::Deserialize)]
 struct ChannelMetadata {
     version: String,
-    body: Option<String>,
-    date: Option<String>,
+    // latest.json publishes Tauri's `notes`/`pub_date`; aliases keep older metadata working.
+    pypi_version: Option<String>,
+    #[serde(alias = "body")]
+    notes: Option<String>,
+    #[serde(alias = "date")]
+    pub_date: Option<String>,
     platforms: HashMap<String, ChannelPlatform>,
 }
 
@@ -99,8 +105,9 @@ pub(crate) async fn check_desktop_manual_update() -> Result<Option<ManualUpdateI
     Ok(Some(ManualUpdateInfo {
         version: latest_version,
         current_version: current_version.to_string(),
-        body: metadata.body,
-        date: metadata.date,
+        pypi_version: metadata.pypi_version,
+        body: metadata.notes,
+        date: metadata.pub_date,
     }))
 }
 
diff --git a/tests/python/test_bitsandbytes_kernel_readiness.py b/tests/python/test_bitsandbytes_kernel_readiness.py
new file mode 100644
index 0000000000..fe595f150d
--- /dev/null
+++ b/tests/python/test_bitsandbytes_kernel_readiness.py
@@ -0,0 +1,177 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""`ALLOW_BITSANDBYTES` must follow the kernels, not the mere presence of the module.
+
+From bitsandbytes 0.46 a wheel whose native library never loaded still imports and
+resolves every ctypes handle to a `throw_on_call` closure, so a probe made of attribute
+reads alone sees a healthy wheel, the loader selects a 4bit checkpoint, and the failure
+lands inside a kernel mid-run instead of degrading to 16bit.
+"""
+
+from __future__ import annotations
+
+import ast
+import importlib.util
+import types
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+
+
+def _load_probe():
+    """Import by path, not as ``unsloth.bnb_availability``, which would run the package
+    __init__ and pull in torch. Works only because the module is a leaf - the property
+    that lets device_type.py, imported very early, use it without a cycle."""
+    path = REPO_ROOT / "unsloth" / "bnb_availability.py"
+    spec = importlib.util.spec_from_file_location("_unsloth_bnb_availability", path)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def _fake_bnb(lib):
+    functional = types.ModuleType("bitsandbytes.functional")
+    functional.get_ptr = lambda tensor: None
+    functional.lib = lib
+    bnb = types.ModuleType("bitsandbytes")
+    bnb.__version__ = "0.50.0"
+    bnb.functional = functional
+    return bnb
+
+
+class _DeferredFailureLib:
+    """What bitsandbytes >= 0.46 hands back when the native library is dead."""
+
+    def __getattr__(self, name):
+        def throw_on_call(*args, **kwargs):
+            raise RuntimeError(f"Method '{name}' not available in CPU-only version")
+
+        return throw_on_call
+
+
+class _RealHandleLib:
+    """ctypes caches the function object on first lookup; its handles carry restype."""
+
+    def __getattr__(self, name):
+        def handle(*args, **kwargs):
+            return None
+
+        handle.restype = None
+        setattr(self, name, handle)
+        return handle
+
+
+def test_probe_covers_every_module_scope_ctypes_bind():
+    """A probe that misses one of the import-time binds lets a dead wheel through."""
+    tree = ast.parse((REPO_ROOT / "unsloth" / "kernels" / "utils.py").read_text(encoding = "utf-8"))
+    bound = {
+        node.attr
+        for node in ast.walk(tree)
+        if isinstance(node, ast.Attribute)
+        and isinstance(node.value, ast.Attribute)
+        and node.value.attr == "lib"
+    }
+    probe = _load_probe()
+    xpu = set(probe.bitsandbytes_symbols("xpu"))
+    cuda = set(probe.bitsandbytes_symbols("cuda"))
+    assert bound == xpu | cuda, f"probe and module-scope binds differ: {bound ^ (xpu | cuda)}"
+    # xpu probes the gemv pair, every other device the naive gemm pair, never both.
+    assert xpu - cuda and cuda - xpu, "the device split collapsed"
+
+
+def test_a_deferred_failure_handle_is_not_ready():
+    probe = _load_probe()
+    bnb = _fake_bnb(_DeferredFailureLib())
+    for device in ("cuda", "xpu"):
+        assert probe.native_kernels_ready(bnb, device) is False, device
+
+
+def test_a_real_ctypes_handle_is_ready():
+    probe = _load_probe()
+    bnb = _fake_bnb(_RealHandleLib())
+    for device in ("cuda", "xpu"):
+        assert probe.native_kernels_ready(bnb, device) is True, device
+
+
+def test_a_lib_that_never_loaded_is_not_ready():
+    """bitsandbytes 0.45.5, the floor in pyproject.toml, sets ``functional.lib = None``."""
+    probe = _load_probe()
+    assert probe.native_kernels_ready(_fake_bnb(None), "cuda") is False
+
+
+def test_a_partially_exporting_library_is_not_ready():
+    """One resolvable symbol is not enough: the same verdict gates the module-scope
+    binds, so a partial library would pass here and raise `AttributeError` at the bind."""
+
+    class _MissingOne(_RealHandleLib):
+        def __getattr__(self, name):
+            if name == "cgemm_4bit_inference_naive_bf16":
+                raise AttributeError(name)
+            return super().__getattr__(name)
+
+    probe = _load_probe()
+    assert probe.native_kernels_ready(_fake_bnb(_MissingOne()), "cuda") is False
+
+
+def test_one_dead_handle_among_live_ones_is_not_ready():
+    """The realistic partial shape: the library loaded but one symbol is a closure."""
+
+    class _OneDeferred(_RealHandleLib):
+        def __getattr__(self, name):
+            if name == "cdequantize_blockwise_bf16_nf4":
+                return lambda *a, **k: None
+            return super().__getattr__(name)
+
+    probe = _load_probe()
+    assert probe.native_kernels_ready(_fake_bnb(_OneDeferred()), "cuda") is False
+
+
+def test_absent_bitsandbytes_is_not_ready():
+    probe = _load_probe()
+    assert probe.native_kernels_ready(None, "cuda") is False
+
+
+def test_device_type_gates_the_flags_on_the_kernels():
+    """The flags must follow ``native_kernels_ready``, not the bare import."""
+    head = (REPO_ROOT / "unsloth" / "device_type.py").read_text(encoding = "utf-8")
+    head = head.split('if DEVICE_TYPE == "hip":')[0]
+    assert "import bitsandbytes as _bnb_probe" in head
+    assert 'find_spec("bitsandbytes")' not in head, "find_spec cannot see a broken wheel"
+    assert "native_kernels_ready(_bnb_probe, DEVICE_TYPE)" in head
+    assert (
+        head.count("ALLOW_BITSANDBYTES = False") >= 2
+    ), "both the failed-import path and the dead-kernels path must clear the flag"
+
+
+def test_the_ctypes_binds_are_gated_on_the_same_verdict():
+    """Clearing the flag is not enough on its own: ``bnb is None`` alone let an
+    importable-but-dead wheel reach the binds, and 0.45.5 sets ``functional.lib = None``
+    on a native-load failure, so they killed ``import unsloth`` outright instead of
+    degrading to 16bit."""
+    source = (REPO_ROOT / "unsloth" / "kernels" / "utils.py").read_text(encoding = "utf-8")
+    assert "from ..bnb_availability import native_kernels_ready" in source
+    assert (
+        "if bnb is None or not native_kernels_ready(bnb, DEVICE_TYPE):" in source
+    ), "the ctypes bind block must take the _bnb_required branch on a dead library too"
+    guarded = source.split("if bnb is None or not native_kernels_ready(bnb, DEVICE_TYPE):")[1]
+    # Anchor on the symbol, not the module alias: #7580 renamed the binding from
+    # `bnb.functional.lib` to `bnb_functional.lib`, which is exactly the kind of rename
+    # this assertion should survive.
+    assert "lib.cdequantize_blockwise_fp32" in guarded, "the binds must sit under that guard"
+
+
+def test_the_kernel_check_reads_the_submodule_not_the_parent_attribute():
+    """A part-initialised bitsandbytes leaves the parent without ``functional`` while
+    the submodule stays in sys.modules, which ``import bitsandbytes.functional`` reads
+    directly."""
+    probe = _load_probe()
+    bnb = types.ModuleType("bitsandbytes")  # zombie: parent has no `functional`
+    bnb.__version__ = "0.50.0"
+    import sys
+
+    real = sys.modules.get("bitsandbytes.functional")
+    if real is None:
+        return  # bitsandbytes not importable here; the fallback has nothing to read
+    # Falls back to the cached submodule instead of raising on the missing attribute.
+    assert probe.native_kernels_ready(bnb, "cuda") in (True, False)
diff --git a/tests/python/test_cross_platform_parity.py b/tests/python/test_cross_platform_parity.py
index b20e715ebc..06e444314b 100644
--- a/tests/python/test_cross_platform_parity.py
+++ b/tests/python/test_cross_platform_parity.py
@@ -454,9 +454,12 @@ class TestKnown211SetParity:
             "$_pinCuLeaf" not in text
         ), "install.ps1 must bound companions on every index (no cu-family exemption)"
         # The bounded companions must actually be passed to the install command.
-        assert re.search(
-            r'"torch>=2\.4,<2\.11\.0" \$_pinVisionSpec \$_pinAudioSpec --default-index \$TorchIndexUrl',
-            text,
+        # Specs are splatted, so check both halves: the list is built, and it is passed.
+        assert (
+            '$_torchSpecs = @("torch>=2.4,<2.11.0", $_pinVisionSpec, $_pinAudioSpec)' in text
+        ), "install.ps1 custom-pin install must build the bounded spec list"
+        assert (
+            "@_torchSpecs --default-index $TorchIndexUrl" in text
         ), "install.ps1 custom-pin install must pass the bounded companion specs to uv"
 
     def test_gfx_allowlist_matches_across_installers(self):
@@ -704,9 +707,13 @@ class TestPinnedIndexClearsUvEnvParity:
         assert (
             "if ($TorchIndexPinned -and -not (Test-CudaFamilyLeaf $CuTag)) {" in text
         ), "the custom-leaf trio bounds must be gated on a pinned non-cu-family leaf"
+        # Specs are splatted, so check both halves: the list is built, and it is passed.
         assert (
-            "Fast-Install $cudaTorchSpec $cudaVisionSpec $cudaAudioSpec" in text
-        ), "setup.ps1's CUDA branch must install via the bounded spec variables"
+            "$_cudaTrio = @($cudaTorchSpec, $cudaVisionSpec, $cudaAudioSpec)" in text
+        ), "setup.ps1's CUDA branch must build the trio from the bounded spec variables"
+        assert (
+            "Fast-Install @_cudaTrio @cudaForce" in text
+        ), "setup.ps1's CUDA branch must install the trio it built"
 
     def test_setup_ps1_bounds_pinned_cpu_torch(self):
         """setup.ps1's CPU branch must bound the trio under an explicit pin (parity with
@@ -724,8 +731,11 @@ class TestPinnedIndexClearsUvEnvParity:
             "if ($TorchIndexPinned) {" in text
         ), "the CPU trio bounds must be gated on an explicit pin"
         assert (
-            "Fast-Install $cpuTorchSpec $cpuVisionSpec $cpuAudioSpec @cpuForce" in text
-        ), "setup.ps1's CPU branch must install via the spec variables"
+            "$_torchTrio = @($cpuTorchSpec, $cpuVisionSpec, $cpuAudioSpec)" in text
+        ), "setup.ps1's CPU branch must build the trio from the spec variables"
+        assert (
+            "Fast-Install @_torchTrio @cpuForce" in text
+        ), "setup.ps1's CPU branch must install the trio it built"
         # The ceilings mirror the Python repair spec exactly.
         stack = STACK_PY.read_text(encoding = "utf-8")
         spec_block = re.search(r"_CUDA_TORCH_PKG_SPEC[^(]*\(\s*(.*?)\)", stack, re.DOTALL)
diff --git a/tests/python/test_torchcodec_torch_compat.py b/tests/python/test_torchcodec_torch_compat.py
index 6ad16a73f4..728a51a321 100644
--- a/tests/python/test_torchcodec_torch_compat.py
+++ b/tests/python/test_torchcodec_torch_compat.py
@@ -11,12 +11,21 @@ import sys
 import types
 from pathlib import Path
 
+import pytest
+
 
 REPO_ROOT = Path(__file__).resolve().parents[2]
 PYPROJECT = REPO_ROOT / "pyproject.toml"
 IMPORT_FIXES_PATH = REPO_ROOT / "unsloth" / "import_fixes.py"
 
 
+def _tomllib():
+    if sys.version_info >= (3, 11):
+        import tomllib
+        return tomllib
+    return pytest.importorskip("tomli")
+
+
 def _load_import_fixes_module():
     spec = importlib.util.spec_from_file_location(
         "unsloth_import_fixes_under_test",
@@ -127,3 +136,41 @@ def test_import_fixes_loads_on_python39_syntax():
     """Regression: module must import on 3.9 (postponed annotations for str | None)."""
     fixes = _load_import_fixes_module()
     assert callable(fixes._torchcodec_version_mismatch_hint)
+
+
+def test_audio_extras_are_gated_to_platforms_with_a_torchcodec_wheel():
+    """torchcodec publishes no sdist and no wheel for Linux aarch64, Windows ARM64 or
+    Intel Mac, so an ungated pin makes pip fail the whole install on those hosts instead
+    of just skipping audio -- and the cu*/rocm*/intel torch 2.10 extras pull it in.
+    The marker must match PLATFORM_LACKS_TORCHCODEC_WHEEL in install_python_stack.py.
+    """
+    markers = pytest.importorskip("packaging.markers")
+    tomllib = _tomllib()
+    extras = tomllib.loads(PYPROJECT.read_text(encoding = "utf-8"))["project"][
+        "optional-dependencies"
+    ]
+    audio = {n: d for n, d in extras.items() if n.startswith("audio-torch")}
+    assert audio, "expected audio-torch* extras"
+
+    supported = [
+        {"sys_platform": "linux", "platform_machine": "x86_64"},
+        {"sys_platform": "win32", "platform_machine": "AMD64"},
+        {"sys_platform": "darwin", "platform_machine": "arm64"},
+    ]
+    unsupported = [
+        {"sys_platform": "linux", "platform_machine": "aarch64"},
+        {"sys_platform": "win32", "platform_machine": "ARM64"},
+        {"sys_platform": "darwin", "platform_machine": "x86_64"},
+    ]
+    for name, deps in audio.items():
+        for dep in deps:
+            _, _, marker_text = dep.partition(";")
+            assert marker_text.strip(), f"{name}: {dep!r} has no marker"
+            marker = markers.Marker(marker_text.strip())
+            env = {"python_version": "3.12"}
+            for case in supported:
+                assert marker.evaluate({**env, **case}), f"{name} must install on {case}"
+            for case in unsupported:
+                assert not marker.evaluate(
+                    {**env, **case}
+                ), f"{name} has no wheel for {case} and must not be resolved there"
diff --git a/tests/python/test_windows_arm64_python_choice.py b/tests/python/test_windows_arm64_python_choice.py
new file mode 100644
index 0000000000..89546e7ac4
--- /dev/null
+++ b/tests/python/test_windows_arm64_python_choice.py
@@ -0,0 +1,143 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Windows on ARM: install.ps1 must not settle for a native ARM64 interpreter.
+
+pyarrow (via datasets) and hf-transfer publish no win_arm64 wheels, so an ARM64
+Python source-builds both and dies minutes into the run. The resolver prefers an
+x64 build of the requested minor and bootstraps one otherwise; the case pinned
+here is the recovery path, where nothing can be downloaded but an x64 build of a
+lower-priority supported minor is already installed.
+"""
+
+from __future__ import annotations
+
+import os
+import re
+import shutil
+import subprocess
+from pathlib import Path
+
+import pytest
+
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+INSTALL_PS1 = REPO_ROOT / "install.ps1"
+
+
+def _extract(pattern: str, source: str) -> str:
+    match = re.search(pattern, source, flags = re.DOTALL)
+    assert match is not None, f"install.ps1 block not found: {pattern}"
+    return match.group(0)
+
+
+def _resolver_script(installed: list[tuple[str, str]], can_download: bool) -> str:
+    """Both production functions verbatim, over a fake set of interpreters.
+
+    Extracted rather than reimplemented so the test cannot drift away from the
+    text install.ps1 actually runs. `installed` is (minor, arch) in py-launcher
+    order, so the first entry for a minor is what a bare `py -3.13` resolves to.
+    The fake interpreters are named `*.exe` and invoked through the call operator,
+    which resolves a string to a function, so no real binary is needed.
+    """
+    source = INSTALL_PS1.read_text(encoding = "utf-8")
+    finder = _extract(r"    function Find-CompatiblePython \{.*?\n    \}\n", source)
+    installer = _extract(r"    function Install-X64Python \{.*?\n    \}\n", source)
+
+    names = [f"Py{minor.replace('.', '')}{arch}.exe" for minor, arch in installed]
+    table = ", ".join(
+        f'@{{ Minor = "{minor}"; Arch = "{arch}"; Name = "{name}" }}'
+        for (minor, arch), name in zip(installed, names)
+    )
+    downloaded = (
+        '@{ Version = "3.13"; Path = "Downloaded.exe"; Arch = "x86_64" }'
+        if can_download
+        else "$null"
+    )
+    version_stubs = "\n".join(
+        f"function {name} {{ param([Parameter(ValueFromRemainingArguments = $true)]$Rest)\n"
+        f'    if ($Rest -contains "--version") {{ return "Python {minor}.0" }}\n'
+        f'    return "{name}" }}'
+        for (minor, _arch), name in zip(installed, names)
+    )
+    return f"""
+$ErrorActionPreference = "Stop"
+$PythonVersion = "3.13"
+$script:WingetAvailable = $false
+$script:CondaSkipPattern = 'conda'
+$Interpreters = @({table})
+{version_stubs}
+# `py -0p` lists every registration; `py -3.x` runs the launcher's preferred build
+# for that minor, which on an ARM64 host is normally the native one.
+function FakePy {{
+    param([Parameter(ValueFromRemainingArguments = $true)]$Rest)
+    if ($Rest -contains "-0p") {{
+        return @($Interpreters | ForEach-Object {{ "  -V:$($_.Minor) *        $($_.Name)" }})
+    }}
+    $minor = ([string]$Rest[0]).TrimStart('-')
+    $hit = @($Interpreters | Where-Object {{ $_.Minor -eq $minor }})
+    if ($hit.Count -eq 0) {{ return "" }}
+    if ($Rest -contains "--version") {{ return "Python $minor.0" }}
+    return $hit[0].Name
+}}
+function substep {{ param($a, $b) }}
+function Get-HostMachineArch {{ return "arm64" }}
+function Get-Command {{
+    param([Parameter(Position = 0)][string]$Name,
+          [Parameter(ValueFromRemainingArguments = $true)]$Rest)
+    if ($Name -eq "py") {{ return @([pscustomobject]@{{ Source = "FakePy" }}) }}
+    return @()
+}}
+function Test-Path {{ param([Parameter(ValueFromRemainingArguments = $true)]$Rest) return $true }}
+function Test-IsCondaPython {{ param([string]$Exe) return $false }}
+function Get-PythonPlatformTag {{
+    param([string]$Exe)
+    foreach ($i in $Interpreters) {{
+        if ($i.Name -eq $Exe) {{
+            if ($i.Arch -eq "x86_64") {{ return "win-amd64" }} else {{ return "win-arm64" }}
+        }}
+    }}
+    return "win-amd64"
+}}
+function Refresh-SessionPath {{ }}
+function Install-PythonFromPythonOrg {{ param([string]$Arch = "") return {downloaded} }}
+{finder}
+{installer}
+# The caller's ARM64 swap, condensed to what decides the interpreter.
+$found = Find-CompatiblePython
+if ($found -and $found.Arch -ne "x86_64") {{
+    $x64 = Install-X64Python
+    if ($x64) {{ $found = $x64 }}
+}}
+if ($found) {{ Write-Output "$($found.Version)|$($found.Arch)" }} else {{ Write-Output "none" }}
+"""
+
+
+def _pwsh(script: str) -> str:
+    result = subprocess.run(
+        ["pwsh", "-NoProfile", "-NonInteractive", "-Command", script],
+        check = True,
+        capture_output = True,
+        text = True,
+        env = os.environ.copy(),
+    )
+    return result.stdout.strip()
+
+
+@pytest.mark.skipif(shutil.which("pwsh") is None, reason = "PowerShell is unavailable")
+@pytest.mark.parametrize(
+    ("installed", "can_download", "expected"),
+    [
+        # An x64 build of the requested minor wins outright, downloads irrelevant.
+        ([("3.13", "arm64"), ("3.13", "x86_64")], False, "3.13|x86_64"),
+        # Requested minor is ARM64-only: bootstrap x64 rather than take the native one.
+        ([("3.13", "arm64")], True, "3.13|x86_64"),
+        # Offline, but an x64 build of a lower-priority minor is here. Use it: the native
+        # 3.13 cannot resolve pyarrow or hf-transfer, and this one can.
+        ([("3.13", "arm64"), ("3.11", "x86_64")], False, "3.11|x86_64"),
+        # ARM64 everywhere: still returned, and the caller warns.
+        ([("3.13", "arm64"), ("3.11", "arm64")], False, "3.13|arm64"),
+    ],
+)
+def test_arm64_host_prefers_an_x64_interpreter(installed, can_download, expected):
+    assert _pwsh(_resolver_script(installed, can_download)) == expected
diff --git a/tests/python/test_windows_git_gate.py b/tests/python/test_windows_git_gate.py
new file mode 100644
index 0000000000..60de430191
--- /dev/null
+++ b/tests/python/test_windows_git_gate.py
@@ -0,0 +1,117 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Git is optional on the consumer Windows path, but still required for source builds."""
+
+from __future__ import annotations
+
+import os
+import shutil
+import subprocess
+from pathlib import Path
+
+import pytest
+
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+SETUP_PS1 = REPO_ROOT / "studio" / "setup.ps1"
+
+_START = "$gitNeeded = ($env:STUDIO_LOCAL_INSTALL -eq '1')"
+_TAIL = "if (-not $_localLlamaBuilt) {"
+
+
+def _git_gate_block() -> str:
+    """Slice the real $gitNeeded computation out of setup.ps1 so the test cannot drift."""
+    source = SETUP_PS1.read_text(encoding = "utf-8")
+    start = source.index(_START)
+    brace = source.index("{", source.index(_TAIL, start))
+    depth = 0
+    for index in range(brace, len(source)):
+        if source[index] == "{":
+            depth += 1
+        elif source[index] == "}":
+            depth -= 1
+            if depth == 0:
+                return source[start : index + 1]
+    raise AssertionError("Unclosed git gate block in setup.ps1")
+
+
+def _script() -> str:
+    return f"""
+$DefaultLlamaPrForce = "0"
+$DefaultLlamaSource = "https://github.com/ggml-org/llama.cpp"
+$DefaultLlamaTag = "latest"
+{_git_gate_block()}
+Write-Output $gitNeeded
+"""
+
+
+def _needs_git(env: dict[str, str]) -> bool:
+    merged = {k: v for k, v in os.environ.items() if not k.startswith(("UNSLOTH_", "STUDIO_"))}
+    merged.update(env)
+    result = subprocess.run(
+        ["pwsh", "-NoProfile", "-NonInteractive", "-Command", _script()],
+        check = True,
+        capture_output = True,
+        text = True,
+        env = merged,
+    )
+    return result.stdout.strip() == "True"
+
+
+pwsh_only = pytest.mark.skipif(shutil.which("pwsh") is None, reason = "PowerShell is unavailable")
+
+
+@pwsh_only
+@pytest.mark.parametrize(
+    ("env", "expected"),
+    [
+        # The consumer install: prebuilt wheels and a prebuilt llama.cpp, so no git.
+        ({}, False),
+        # --local clones unsloth-zoo.
+        ({"STUDIO_LOCAL_INSTALL": "1"}, True),
+        # Opt-in source builds clone llama.cpp.
+        ({"UNSLOTH_LLAMA_FORCE_COMPILE": "1"}, True),
+        ({"UNSLOTH_LLAMA_PR": "1234"}, True),
+        # PR_FORCE only forces a build for a positive integer.
+        ({"UNSLOTH_LLAMA_PR_FORCE": "0"}, False),
+        ({"UNSLOTH_LLAMA_PR_FORCE": "not-a-number"}, False),
+        ({"UNSLOTH_LLAMA_PR_FORCE": "1234"}, True),
+        # "master" is a branch with no release, so Phase 4 always builds it from source.
+        ({"UNSLOTH_LLAMA_TAG": "master"}, True),
+        # A release tag resolves to a prebuilt bundle.
+        ({"UNSLOTH_LLAMA_TAG": "latest"}, False),
+        ({"UNSLOTH_LLAMA_TAG": "b8635"}, False),
+    ],
+)
+def test_git_is_required_only_for_local_and_source_builds(env, expected):
+    assert _needs_git(env) is expected
+
+
+@pwsh_only
+def test_a_built_local_llama_dir_drops_the_source_build_git_requirement(tmp_path):
+    (tmp_path / "llama-server.exe").write_text("", encoding = "utf-8")
+    env = {
+        "UNSLOTH_LOCAL_LLAMA_CPP_DIR": str(tmp_path),
+        "UNSLOTH_LLAMA_FORCE_COMPILE": "1",
+    }
+    # Reusing an existing binary skips both the prebuilt download and the source build.
+    assert _needs_git(env) is False
+
+
+@pwsh_only
+@pytest.mark.parametrize("trigger", ["UNSLOTH_LLAMA_FORCE_COMPILE", "UNSLOTH_LLAMA_PR"])
+def test_an_unbuilt_local_llama_dir_still_requires_git(tmp_path, trigger):
+    # Nothing built at the canonical install location falls through to the normal install,
+    # so the source build still runs and still needs git. Suppressing the requirement here
+    # let a no-git host silently degrade to a prebuilt instead.
+    env = {
+        "UNSLOTH_LOCAL_LLAMA_CPP_DIR": str(tmp_path),
+        trigger: "1",
+    }
+    assert _needs_git(env) is True
+
+
+@pwsh_only
+def test_an_unbuilt_local_llama_dir_alone_does_not_require_git(tmp_path):
+    assert _needs_git({"UNSLOTH_LOCAL_LLAMA_CPP_DIR": str(tmp_path)}) is False
diff --git a/tests/python/test_windows_vcredist_download_tls.py b/tests/python/test_windows_vcredist_download_tls.py
new file mode 100644
index 0000000000..9fb1c697e2
--- /dev/null
+++ b/tests/python/test_windows_vcredist_download_tls.py
@@ -0,0 +1,80 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""The direct VC++ runtime download must negotiate TLS 1.2 on legacy protocol defaults."""
+
+from __future__ import annotations
+
+import shutil
+import subprocess
+from pathlib import Path
+
+import pytest
+
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+SETUP_PS1 = REPO_ROOT / "studio" / "setup.ps1"
+
+_START = '$url = "https://aka.ms/vs/17/release/vc_redist.x64.exe"'
+_END = "Remove-Item -LiteralPath $dst -Force -ErrorAction SilentlyContinue\n        }"
+
+
+def _download_block() -> str:
+    """Slice the real download block out of setup.ps1 so the test cannot drift."""
+    source = SETUP_PS1.read_text(encoding = "utf-8")
+    start = source.index(_START)
+    end = source.index(_END, start) + len(_END)
+    return source[start:end]
+
+
+def _script(starting_protocol: str) -> str:
+    # Start from a non-zero set that lacks Tls12. Tls13 is the only such value modern .NET
+    # accepts, and it stands in for the legacy Ssl3/Tls default of Windows PowerShell 5.1.
+    return f"""
+function substep {{ param($a, $b) }}
+function Refresh-Environment {{ }}
+function Invoke-WebRequest {{
+    param($Uri, $OutFile, [switch]$UseBasicParsing, $TimeoutSec)
+    Write-Output "DURING=$([System.Net.ServicePointManager]::SecurityProtocol)"
+    throw "stop before Start-Process"
+}}
+[System.Net.ServicePointManager]::SecurityProtocol = [System.Net.SecurityProtocolType]::{starting_protocol}
+{_download_block()}
+Write-Output "AFTER=$([System.Net.ServicePointManager]::SecurityProtocol)"
+"""
+
+
+def _run(starting_protocol: str) -> dict[str, str]:
+    result = subprocess.run(
+        ["pwsh", "-NoProfile", "-NonInteractive", "-Command", _script(starting_protocol)],
+        check = True,
+        capture_output = True,
+        text = True,
+    )
+    out = {}
+    for line in result.stdout.splitlines():
+        if "=" in line:
+            key, _, value = line.partition("=")
+            out[key.strip()] = value.strip()
+    return out
+
+
+pwsh_only = pytest.mark.skipif(shutil.which("pwsh") is None, reason = "PowerShell is unavailable")
+
+
+@pwsh_only
+def test_tls12_is_added_for_the_download_and_restored_after():
+    seen = _run("Tls13")
+    during = {part.strip() for part in seen["DURING"].split(",")}
+    assert "Tls12" in during, "the download must negotiate TLS 1.2 or aka.ms refuses it"
+    assert "Tls13" in during, "adding TLS 1.2 must not drop protocols the host already allowed"
+    assert seen["AFTER"] == "Tls13", "the process-wide protocol must be restored"
+
+
+@pwsh_only
+def test_system_default_is_left_alone():
+    # SystemDefault means "let the OS choose" and already covers TLS 1.2+; pinning it to
+    # Tls12 would strip TLS 1.3 from every later request in the process.
+    seen = _run("SystemDefault")
+    assert seen["DURING"] == "SystemDefault"
+    assert seen["AFTER"] == "SystemDefault"
diff --git a/tests/sh/test_install_pipe_safety.sh b/tests/sh/test_install_pipe_safety.sh
new file mode 100755
index 0000000000..be479dd5a5
--- /dev/null
+++ b/tests/sh/test_install_pipe_safety.sh
@@ -0,0 +1,89 @@
+#!/bin/bash
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+#
+# Guards that `curl ... | sh` cannot report a bogus transport error.
+#
+# History: install.sh was ~150KB of top-level statements. A top-level `exit` left most
+# of the file unread, the write end failed, and curl appended "(56) Failure writing
+# output to destination" (or "(23) Failed writing body") after our own error message,
+# so users read a real diagnosis as a broken download. The fix is structural: the body
+# lives in _unsloth_main, so sh parses the whole file before running anything.
+#
+# This pins both halves of that contract: the writer must not be killed, AND the
+# installer's own exit code must still reach the caller.
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+INSTALL_SH="$SCRIPT_DIR/../../install.sh"
+PASS=0
+FAIL=0
+
+assert_eq() {
+    _label="$1"; _expected="$2"; _actual="$3"
+    if [ "$_actual" = "$_expected" ]; then
+        echo "  PASS: $_label"
+        PASS=$((PASS + 1))
+    else
+        echo "  FAIL: $_label (expected '$_expected', got '$_actual')"
+        FAIL=$((FAIL + 1))
+    fi
+}
+
+echo "=== structure ==="
+
+# The wrapper must be invoked on the LAST executable line, or sh starts executing
+# before it has drained the pipe.
+if grep -q '^_unsloth_main() {' "$INSTALL_SH"; then
+    echo "  PASS: _unsloth_main is defined at top level"
+    PASS=$((PASS + 1))
+else
+    echo "  FAIL: install.sh is not wrapped in _unsloth_main -- curl-pipe safety is gone"
+    FAIL=$((FAIL + 1))
+fi
+
+_last="$(grep -vE '^\s*(#|$)' "$INSTALL_SH" | tail -1)"
+assert_eq "last statement invokes the wrapper" '_unsloth_main "$@"' "$_last"
+
+# Below one pipe buffer the file would fit in the kernel's buffer and this test would
+# prove nothing, so fail loudly instead of passing vacuously.
+_bytes="$(wc -c < "$INSTALL_SH" | tr -d ' ')"
+if [ "$_bytes" -gt 65536 ]; then
+    echo "  PASS: install.sh ($_bytes bytes) exceeds a 64KiB pipe buffer, so this matters"
+    PASS=$((PASS + 1))
+else
+    echo "  FAIL: install.sh is only $_bytes bytes; re-derive whether pipe safety still applies"
+    FAIL=$((FAIL + 1))
+fi
+
+echo "=== behaviour: an early exit must not kill the writer ==="
+
+# `--python` with no argument exits 1 from argument validation having done no work: no
+# venv, no downloads, no filesystem writes. Deterministic and safe to run for real.
+#
+# PIPESTATUS must be read on the very next line, so drop errexit around the pipeline
+# rather than appending `|| true`, which would clobber it with the status of `true`.
+set +e
+cat "$INSTALL_SH" | sh -s -- --python >/dev/null 2>&1
+_pipe=("${PIPESTATUS[@]}")
+set -e
+_writer_rc="${_pipe[0]}"
+_reader_rc="${_pipe[1]}"
+
+# A writer rc of 141 (128 + SIGPIPE) is the failure mode curl reports as (56)/(23).
+assert_eq "writer survives the early exit (not SIGPIPE)" "0" "$_writer_rc"
+assert_eq "installer's own exit code still propagates" "1" "$_reader_rc"
+
+echo "=== behaviour: the same holds for a mid-file exit ==="
+# `--package '-evil'` exits from a later validation block, still before any filesystem
+# work, so the property is not specific to one early branch.
+set +e
+cat "$INSTALL_SH" | sh -s -- --package '-evil' >/dev/null 2>&1
+_pipe2=("${PIPESTATUS[@]}")
+set -e
+assert_eq "writer survives a later exit" "0" "${_pipe2[0]}"
+assert_eq "later exit code propagates" "1" "${_pipe2[1]}"
+
+echo ""
+echo "=== $PASS passed, $FAIL failed ==="
+[ "$FAIL" -eq 0 ] || exit 1
diff --git a/tests/studio/install/test_install_node_prebuilt_logic.py b/tests/studio/install/test_install_node_prebuilt_logic.py
index 5476702d65..5bcc9733cf 100644
--- a/tests/studio/install/test_install_node_prebuilt_logic.py
+++ b/tests/studio/install/test_install_node_prebuilt_logic.py
@@ -762,3 +762,94 @@ def test_pinned_target_wrong_sha_not_kept_when_download_fails(tmp_path: Path, mo
     monkeypatch.setattr(M, "download_file_verified", _offline)  # transient download failure
     with pytest.raises(OSError):
         M.install_prebuilt(install_dir, channel = "pinned", min_major = 24, force = False)
+
+
+# ── _replace_with_retry: transient Windows sharing violations ──────────────────
+# Seen in CI: WinError 5 renaming extracted Node into place on a FRESH install, a scanner
+# still holding handles inside the new files.
+
+
+def _oserror(winerror: int) -> OSError:
+    exc = OSError(winerror, "mock")
+    exc.winerror = winerror
+    return exc
+
+
+@pytest.mark.parametrize("winerror", [5, 32, 145])
+def test_replace_retries_transient_windows_errors(monkeypatch, tmp_path, winerror):
+    monkeypatch.setattr(M.os, "name", "nt")
+    monkeypatch.setattr(M.time, "sleep", lambda _s: None)  # no real backoff in tests
+    calls = {"n": 0}
+
+    def flaky(src, dst):
+        calls["n"] += 1
+        if calls["n"] < 3:
+            raise _oserror(winerror)
+
+    monkeypatch.setattr(M.os, "replace", flaky)
+    M._replace_with_retry(tmp_path / "src", tmp_path / "dst")
+    assert calls["n"] == 3, "should have retried until the handle was released"
+
+
+def test_replace_gives_up_and_reports_the_real_error(monkeypatch, tmp_path):
+    monkeypatch.setattr(M.os, "name", "nt")
+    monkeypatch.setattr(M.time, "sleep", lambda _s: None)
+    monkeypatch.setattr(M.os, "replace", lambda s, d: (_ for _ in ()).throw(_oserror(5)))
+    # A scanner that never lets go must surface as a failure, not a hang.
+    with pytest.raises(OSError) as excinfo:
+        M._replace_with_retry(tmp_path / "src", tmp_path / "dst", attempts = 3)
+    assert excinfo.value.winerror == 5
+
+
+def test_replace_does_not_retry_a_genuine_error(monkeypatch, tmp_path):
+    # A cross-device move or real permissions problem must fail immediately.
+    monkeypatch.setattr(M.os, "name", "nt")
+    monkeypatch.setattr(M.time, "sleep", lambda _s: None)
+    calls = {"n": 0}
+
+    def hard_fail(src, dst):
+        calls["n"] += 1
+        raise _oserror(17)  # ERROR_NOT_SAME_DEVICE
+
+    monkeypatch.setattr(M.os, "replace", hard_fail)
+    with pytest.raises(OSError):
+        M._replace_with_retry(tmp_path / "src", tmp_path / "dst")
+    assert calls["n"] == 1
+
+
+def test_replace_is_a_plain_rename_on_posix(monkeypatch, tmp_path):
+    # POSIX has no sharing violations, so the retry must add no latency there.
+    monkeypatch.setattr(M.os, "name", "posix")
+    calls = {"n": 0}
+
+    def once(src, dst):
+        calls["n"] += 1
+        raise _oserror(5)
+
+    monkeypatch.setattr(M.os, "replace", once)
+    with pytest.raises(OSError):
+        M._replace_with_retry(tmp_path / "src", tmp_path / "dst")
+    assert calls["n"] == 1
+
+
+def test_swap_into_place_survives_a_transient_lock(monkeypatch, tmp_path):
+    # End-to-end through the function the installer actually calls.
+    monkeypatch.setattr(M.os, "name", "nt")
+    monkeypatch.setattr(M.time, "sleep", lambda _s: None)
+    extracted = tmp_path / "extracted" / "node-v24"
+    extracted.mkdir(parents = True)
+    (extracted / "marker.txt").write_text("node", encoding = "utf-8")
+    install_dir = tmp_path / "node"
+
+    real_replace = os.replace
+    state = {"failed": False}
+
+    def flaky(src, dst):
+        if not state["failed"]:
+            state["failed"] = True
+            raise _oserror(32)
+        real_replace(src, dst)
+
+    monkeypatch.setattr(M.os, "replace", flaky)
+    M._swap_into_place(extracted, install_dir)
+    assert (install_dir / "marker.txt").read_text(encoding = "utf-8") == "node"
diff --git a/tests/studio/test_cli_studio_stop_windows.py b/tests/studio/test_cli_studio_stop_windows.py
index 2267d7feda..cef7cc6db7 100644
--- a/tests/studio/test_cli_studio_stop_windows.py
+++ b/tests/studio/test_cli_studio_stop_windows.py
@@ -44,9 +44,12 @@ def _load_pid_alive(platform: str, fake_run = None):
 # ── AST: stop() must not use the broken bare liveness probe ──────────────────
 
 
-def test_stop_does_not_use_bare_oskill_liveness_probe():
-    """stop() must not call os.kill(pid, 0) -- it crashes on Windows."""
-    stop_src = _func_source("stop")
+# `stop` delegates signalling to `_signal_stop`, so guarding only `stop` would
+# let os.kill(pid, 0) come back one function along and still pass.
+@pytest.mark.parametrize("func", ["stop", "_signal_stop"])
+def test_stop_does_not_use_bare_oskill_liveness_probe(func):
+    """The signalling path must not call os.kill(pid, 0) -- WinError 87 on Windows."""
+    stop_src = _func_source(func)
     tree = ast.parse(stop_src)
     for call in ast.walk(tree):
         if not isinstance(call, ast.Call):
@@ -62,14 +65,17 @@ def test_stop_does_not_use_bare_oskill_liveness_probe():
             sig = call.args[1]
             if isinstance(sig, ast.Constant) and sig.value == 0:
                 raise AssertionError(
-                    "stop() still uses os.kill(pid, 0); it raises WinError 87 on "
-                    "Windows. Use the cross-platform _pid_alive() helper instead."
+                    f"{func}() still uses os.kill(pid, 0); it raises WinError 87 "
+                    "on Windows. Use the cross-platform _pid_alive() helper."
                 )
 
 
 def test_pid_alive_helper_is_defined_and_used_by_stop():
     assert "def _pid_alive(" in _SOURCE, "_pid_alive helper missing"
     assert "_pid_alive(pid)" in _func_source("stop"), "stop() must use _pid_alive"
+    # The kill itself moved into _signal_stop; keep both ends of the path pinned.
+    assert "def _signal_stop(" in _SOURCE, "_signal_stop helper missing"
+    assert "taskkill" in _func_source("_signal_stop")
     # The helper must special-case Windows via tasklist (os.kill(pid,0) is invalid there).
     helper = _func_source("_pid_alive")
     assert 'sys.platform == "win32"' in helper
diff --git a/tests/studio/test_studio_pid_file_contract.py b/tests/studio/test_studio_pid_file_contract.py
new file mode 100644
index 0000000000..23ace706b5
--- /dev/null
+++ b/tests/studio/test_studio_pid_file_contract.py
@@ -0,0 +1,73 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""run.py writes the Studio PID files; `unsloth studio stop` globs for them.
+
+Nothing else ties the writer's filename to the reader's glob, and each side's own
+tests hardcode the names they expect, so a rename on either side alone leaves both
+suites green while `stop` silently finds nothing. `unsloth_cli/tests/` also runs
+in no workflow, so this lives here, where the repo CPU job discovers it.
+
+AST + exec of the writer, so no backend dependency stack is imported.
+"""
+
+import ast
+import os
+import sys
+from pathlib import Path
+
+_ROOT = Path(__file__).resolve().parents[2]
+if str(_ROOT) not in sys.path:
+    sys.path.insert(0, str(_ROOT))
+
+_RUN_SRC = (_ROOT / "studio" / "backend" / "run.py").read_text(encoding = "utf-8")
+
+
+def _func_source(source: str, name: str) -> str:
+    for node in ast.walk(ast.parse(source)):
+        if isinstance(node, ast.FunctionDef) and node.name == name:
+            return ast.get_source_segment(source, node)
+    raise AssertionError(f"function {name!r} not found")
+
+
+def _backend_pid_path(root: Path, port: int) -> Path:
+    """The path run.py's own _pid_file_for_port builds, without importing run.py."""
+    ns = {"os": os, "Path": Path, "_studio_root": lambda: root}
+    exec(_func_source(_RUN_SRC, "_pid_file_for_port"), ns)
+    return ns["_pid_file_for_port"](port)
+
+
+def test_stop_finds_a_pid_file_named_the_way_the_backend_writes_it(tmp_path, monkeypatch):
+    from unsloth_cli.commands import studio as cli
+
+    path = _backend_pid_path(tmp_path, 8901)
+    # The same three-line body _write_pid_file emits (create_time is blank when
+    # psutil is unavailable, and the CLI must tolerate that).
+    path.write_text(f"{os.getpid()}\n\n127.0.0.1", encoding = "utf-8")
+
+    monkeypatch.setattr(cli, "STUDIO_HOME", tmp_path)
+    monkeypatch.setattr(cli, "_PID_FILE", tmp_path / "studio.pid")
+
+    assert [pid for pid, _times, _files in cli._pid_file_entries()] == [os.getpid()]
+
+
+def test_the_legacy_file_stays_a_bare_pid_an_older_cli_can_parse(tmp_path, monkeypatch):
+    # An older `unsloth studio stop` reads studio.pid and requires str.isdigit(),
+    # so the compatibility file must never gain the extra metadata lines.
+    ns = {
+        "os": os,
+        "Path": Path,
+        "_studio_root": lambda: tmp_path,
+        "_PID_FILE": tmp_path / "studio.pid",
+        "_pid_file_for_port": lambda port: _backend_pid_path(tmp_path, port),
+        "_process_create_time": lambda pid: None,
+        "_bind_addresses": lambda host, port: {host},
+        # _write_pid_file consults these before taking over studio.pid.
+        "_read_pid_record": lambda path: None,
+        "_pid_alive": lambda pid: False,
+        "_OWN_PID_FILE": None,
+    }
+    exec(_func_source(_RUN_SRC, "_write_pid_file"), ns)
+    ns["_write_pid_file"](8901, "127.0.0.1")
+
+    assert (tmp_path / "studio.pid").read_text(encoding = "utf-8").strip().isdigit()
diff --git a/tests/studio/test_update_release_notes.py b/tests/studio/test_update_release_notes.py
new file mode 100644
index 0000000000..765ad55321
--- /dev/null
+++ b/tests/studio/test_update_release_notes.py
@@ -0,0 +1,1906 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Contracts for the update popup's release-notes preview.
+
+The popup renders CHANGELOG.md notes for the exact version it is offering. The
+risk this file guards is showing notes from a different release: a near-miss
+lookup must return nothing rather than the newest section it can find."""
+
+from __future__ import annotations
+
+import http.server
+import json
+import os
+import re
+import shutil
+import subprocess
+import sys
+import threading
+import time
+from pathlib import Path
+
+import pytest
+
+REPO = Path(__file__).resolve().parents[2]
+BACKEND = REPO / "studio/backend"
+FRONTEND = REPO / "studio/frontend/src"
+CHANGELOG = REPO / "CHANGELOG.md"
+PANEL = FRONTEND / "components/update/release-notes-panel.tsx"
+NOTES_HOOK = FRONTEND / "hooks/use-release-notes.ts"
+PREVIEW = FRONTEND / "lib/release-notes-preview.ts"
+CODE_SPANS = FRONTEND / "lib/markdown-code-spans.ts"
+LINKS = FRONTEND / "lib/changelog-links.ts"
+LIST_COLUMNS = FRONTEND / "lib/markdown-list-columns.ts"
+INLINE_COMMENTS = FRONTEND / "lib/markdown-inline-comments.ts"
+WEB_BANNER = FRONTEND / "components/web/update-banner.tsx"
+TAURI_BANNER = FRONTEND / "components/tauri/update-banner.tsx"
+
+# The scanners are the frontend half of the contract the parser implements, so they are
+# run rather than read. Node strips the types and nothing imports a package: no install.
+_TS_ALIAS = re.compile(r'"@/lib/([a-z-]+)"')
+_TS_RUNNER = """
+import { resolveChangelogLinks } from "./changelog-links.ts";
+import { releaseNotesPreview } from "./release-notes-preview.ts";
+
+const chunks: Buffer[] = [];
+process.stdin.on("data", (chunk: Buffer) => chunks.push(chunk));
+process.stdin.on("end", () => {
+  const markdown = Buffer.concat(chunks).toString("utf8");
+  const result =
+    process.argv[2] === "links"
+      ? resolveChangelogLinks(markdown)
+      : releaseNotesPreview(markdown);
+  process.stdout.write(JSON.stringify(result));
+});
+"""
+
+SAMPLE = """# Changelog
+
+Intro prose that belongs to no release.
+
+## Format
+
+```md
+## 9999.9.9 - fenced sample, not a real section
+```
+
+## Unreleased
+
+- staged note
+
+## 2026.7.6 - 2026-07-22
+
+### What's Changed
+
+- newer thing
+
+## 2026.7.5
+
+### What's Changed
+
+- older thing
+"""
+
+
+@pytest.fixture(scope = "module")
+def changelog_module():
+    sys.path.insert(0, str(BACKEND))
+    try:
+        from utils import changelog
+    finally:
+        sys.path.pop(0)
+    changelog.reset_changelog_cache()
+    yield changelog
+    changelog.reset_changelog_cache()
+
+
+@pytest.fixture
+def isolated_changelog(changelog_module, tmp_path, monkeypatch):
+    """Point the module at a temp file and away from the network."""
+    monkeypatch.setenv(changelog_module.DISABLE_ENV_VAR, "1")
+    path = tmp_path / "CHANGELOG.md"
+    path.write_text(SAMPLE, encoding = "utf-8")
+    monkeypatch.setenv(changelog_module.CHANGELOG_PATH_ENV_VAR, str(path))
+    changelog_module.reset_changelog_cache()
+    yield changelog_module
+    changelog_module.reset_changelog_cache()
+
+
+def test_only_real_release_headings_become_sections(changelog_module):
+    versions = [entry.version for entry in changelog_module.parse_changelog(SAMPLE)]
+    # "Format"/"Unreleased" are not versions, and 9999.9.9 is fenced sample.
+    assert versions == ["2026.7.6", "2026.7.5"]
+
+
+def test_section_body_stops_at_the_next_release(changelog_module):
+    entry = changelog_module.find_release_notes(SAMPLE, "2026.7.6")
+    assert entry is not None
+    assert "newer thing" in entry.body
+    assert "older thing" not in entry.body
+
+
+def test_unknown_version_returns_no_notes_instead_of_a_nearby_release(changelog_module):
+    assert changelog_module.find_release_notes(SAMPLE, "2026.7.7") is None
+    assert changelog_module.find_release_notes(SAMPLE, "2026.7") is None
+
+
+def test_version_equality_is_normalized_not_fuzzy(changelog_module):
+    entry = changelog_module.find_release_notes(SAMPLE, "2026.07.6")
+    assert entry is not None and entry.version == "2026.7.6"
+
+
+def test_response_reports_no_match_without_markdown(isolated_changelog):
+    payload = isolated_changelog.get_release_notes("2026.7.7")
+    assert payload["matched"] is False
+    assert payload["markdown"] is None
+    assert payload["version"] == "2026.7.7"
+    # The UI still needs somewhere to send the user.
+    assert payload["release_notes_url"]
+
+
+def test_response_matches_local_changelog_when_offline(isolated_changelog):
+    payload = isolated_changelog.get_release_notes("2026.7.6")
+    assert payload["matched"] is True
+    assert payload["source"] == "local"
+    assert "newer thing" in payload["markdown"]
+
+
+def test_unsupported_version_query_is_rejected(isolated_changelog):
+    assert isolated_changelog.is_supported_version_query("2026.7.6") is True
+    for bad in ("../etc/passwd", "2026.7.6 OR 1", "", "a" * 80):
+        assert isolated_changelog.is_supported_version_query(bad) is False
+    assert isolated_changelog.get_release_notes("../etc/passwd")["matched"] is False
+
+
+def test_remote_changelog_wins_over_bundled_copy(changelog_module, tmp_path, monkeypatch):
+    """The offered version is newer than the installed checkout, so the repo
+    copy has to be able to describe versions the local file has never heard of."""
+    monkeypatch.delenv(changelog_module.DISABLE_ENV_VAR, raising = False)
+    local = tmp_path / "CHANGELOG.md"
+    local.write_text(SAMPLE, encoding = "utf-8")
+    monkeypatch.setenv(changelog_module.CHANGELOG_PATH_ENV_VAR, str(local))
+
+    remote_body = "# Changelog\n\n## 2026.8.0\n\n- shipped after this install\n"
+
+    class Handler(http.server.BaseHTTPRequestHandler):
+        def do_GET(self):  # noqa: N802 - stdlib naming
+            payload = remote_body.encode("utf-8")
+            self.send_response(200)
+            self.send_header("Content-Length", str(len(payload)))
+            self.end_headers()
+            self.wfile.write(payload)
+
+        def log_message(self, *_args):
+            pass
+
+    server = http.server.HTTPServer(("127.0.0.1", 0), Handler)
+    thread = threading.Thread(target = server.serve_forever, daemon = True)
+    thread.start()
+    try:
+        monkeypatch.setenv(
+            changelog_module.CHANGELOG_URL_ENV_VAR,
+            f"http://127.0.0.1:{server.server_port}/CHANGELOG.md",
+        )
+        changelog_module.reset_changelog_cache()
+        payload = changelog_module.get_release_notes("2026.8.0")
+        assert payload["matched"] is True
+        assert payload["source"] == "remote"
+        assert "shipped after this install" in payload["markdown"]
+    finally:
+        server.shutdown()
+        server.server_close()
+        changelog_module.reset_changelog_cache()
+
+
+def test_repo_changelog_exists_and_parses(changelog_module):
+    assert CHANGELOG.is_file(), "CHANGELOG.md is the editable source of release notes"
+    entries = changelog_module.parse_changelog(CHANGELOG.read_text(encoding = "utf-8"))
+    assert entries, "CHANGELOG.md needs at least one `## <version>` section"
+
+
+def test_longer_outer_fence_does_not_leak_a_fake_section(changelog_module):
+    """A ``` sample inside a ```` block must not close the block and let the
+    sample's heading be indexed as a real release."""
+    text = "## 1.0\n\n````md\n```\n## 9.9.9\n```\n````\n\n- real note\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0"]
+    assert changelog_module.find_release_notes(text, "9.9.9") is None
+
+
+def test_tilde_fence_is_not_closed_by_backticks(changelog_module):
+    text = "## 1.0\n\n~~~\n```\n## 9.9.9\n~~~\n\n- real\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0"]
+
+
+def test_utf8_bom_does_not_hide_the_first_section(changelog_module):
+    """Editors on Windows can leave a BOM on the first line."""
+    assert [e.version for e in changelog_module.parse_changelog("\ufeff## 1.0\n\n- x\n")] == ["1.0"]
+
+
+@pytest.mark.parametrize("newline", ["\r\n", "\r"])
+def test_non_unix_line_endings(changelog_module, newline):
+    text = f"## 1.0{newline}{newline}- windows note{newline}"
+    entry = changelog_module.find_release_notes(text, "1.0")
+    assert entry is not None and "windows note" in entry.body
+    assert "\r" not in entry.body
+
+
+def test_closing_fence_must_carry_nothing_after_it(changelog_module):
+    """CommonMark: a closer is the delimiter plus whitespace only. A ```` line
+    with trailing text inside a ```` block is content, not the end."""
+    text = "## 1.0\n\n````md\n```` not a closer\n## 9.9.9\n````\n\n- real\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0"]
+    # An opening fence may still carry an info string.
+    info = "## 1.0\n\n```python\n## 9.9.9\n```\n\n- real\n"
+    assert [e.version for e in changelog_module.parse_changelog(info)] == ["1.0"]
+
+
+@pytest.mark.parametrize(
+    "text",
+    [
+        "## 1.0\n\n- real\n\n<!--\n## 9.9.9\n\n- unpublished\n-->\n",
+        "## 1.0\n\n- real\n\n<!-- ## 9.9.9 -->\n",
+    ],
+)
+def test_commented_out_sections_are_not_releases(changelog_module, text):
+    """Markdown does not render them, so they are not published notes."""
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0"]
+    assert changelog_module.find_release_notes(text, "9.9.9") is None
+
+
+def test_repo_root_changelog_is_preferred_over_the_build_snapshot(changelog_module):
+    """The build backend writes studio/CHANGELOG.md; the root file must win."""
+    # Resolved paths, not name suffixes: a checkout may be renamed and Windows uses "\".
+    paths = [Path(p).resolve() for p in changelog_module._local_changelog_candidates()]
+    root = paths.index((REPO / changelog_module.CHANGELOG_FILENAME).resolve())
+    packaged = paths.index((REPO / "studio" / changelog_module.CHANGELOG_FILENAME).resolve())
+    assert root < packaged
+    build = (REPO / "build.sh").read_text(encoding = "utf-8")
+    assert "rm -f studio/CHANGELOG.md" in build, "snapshot must not linger after a build"
+
+
+def test_preview_keeps_identifier_underscores():
+    """UNSLOTH_DISABLE_UPDATE_CHECK must not render as UNSLOTHDISABLEUPDATECHECK."""
+    src = PREVIEW.read_text(encoding = "utf-8")
+    assert "BOLD_UNDERSCORE" in src and "ITALIC_UNDERSCORE" in src
+    assert "parkCodeSpans" in src, "code spans are parked so their underscores survive"
+    assert "const EMPHASIS" not in src, "the blanket emphasis strip is gone"
+
+
+def test_panel_prefers_the_callers_release_url():
+    """The API only returns the generic changelog; the desktop banner passes
+    the exact release page for the version being offered."""
+    src = PANEL.read_text(encoding = "utf-8")
+    assert "releaseNotesUrl ?? notes?.releaseNotesUrl" in src
+
+
+def test_remote_failure_is_reported_so_the_ui_can_retry(changelog_module, tmp_path, monkeypatch):
+    """A bundled changelog cannot know a version newer than the install, so a
+    failed remote lookup must not read as "no notes were published"."""
+    monkeypatch.delenv(changelog_module.DISABLE_ENV_VAR, raising = False)
+    local = tmp_path / "CHANGELOG.md"
+    local.write_text("## 1.0\n\n- old release\n", encoding = "utf-8")
+    monkeypatch.setenv(changelog_module.CHANGELOG_PATH_ENV_VAR, str(local))
+    # Port 9 (discard) refuses fast, standing in for an unreachable host.
+    monkeypatch.setenv(changelog_module.CHANGELOG_URL_ENV_VAR, "http://127.0.0.1:9/CHANGELOG.md")
+    changelog_module.reset_changelog_cache()
+    try:
+        payload = changelog_module.get_release_notes("2.0")
+        assert payload["matched"] is False
+        assert payload["error"], "remote failure must reach the UI"
+    finally:
+        changelog_module.reset_changelog_cache()
+
+
+def test_preview_keeps_comparison_operators():
+    """ "Support Python <3.15 and >3.9" must not lose its operators to the tag
+    strip, which would turn it into "Support Python 3.9"."""
+    src = PREVIEW.read_text(encoding = "utf-8")
+    assert "/<\\/?[a-zA-Z][^>]*>/g" in src, "tag strip must require a name character"
+
+
+def test_preview_hides_commented_out_notes():
+    """Unpublished notes inside <!-- --> are not rendered, so not previewed."""
+    src = PREVIEW.read_text(encoding = "utf-8")
+    assert "stripCommentSpans" in src and "COMMENT_OPEN" in src
+
+
+def test_hook_treats_a_reported_failure_as_retryable():
+    src = NOTES_HOOK.read_text(encoding = "utf-8")
+    assert "next.error !== null" in src
+
+
+def test_comment_delimiter_in_inline_code_is_literal(changelog_module):
+    """A note documenting `<!--` used to put the parser into comment state,
+    swallowing every release below it."""
+    text = "## 2.0\n\n- Type `<!--` to begin a comment\n\n## 1.0\n\n- older\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["2.0", "1.0"]
+    assert changelog_module.find_release_notes(text, "1.0") is not None
+    assert "older" not in changelog_module.find_release_notes(text, "2.0").body
+
+
+def test_refresh_retries_a_cached_remote_failure(changelog_module, tmp_path, monkeypatch):
+    """Retry must reach the network again once connectivity returns, rather
+    than replaying the cached failure until its TTL expires."""
+    monkeypatch.delenv(changelog_module.DISABLE_ENV_VAR, raising = False)
+    local = tmp_path / "CHANGELOG.md"
+    local.write_text("## 1.0\n\n- old\n", encoding = "utf-8")
+    monkeypatch.setenv(changelog_module.CHANGELOG_PATH_ENV_VAR, str(local))
+
+    hits = {"count": 0}
+
+    class Handler(http.server.BaseHTTPRequestHandler):
+        def do_GET(self):  # noqa: N802 - stdlib naming
+            hits["count"] += 1
+            self.send_response(500)
+            self.send_header("Content-Length", "0")
+            self.end_headers()
+
+        def log_message(self, *_args):
+            pass
+
+    server = http.server.HTTPServer(("127.0.0.1", 0), Handler)
+    threading.Thread(target = server.serve_forever, daemon = True).start()
+    try:
+        monkeypatch.setenv(
+            changelog_module.CHANGELOG_URL_ENV_VAR,
+            f"http://127.0.0.1:{server.server_port}/CHANGELOG.md",
+        )
+        changelog_module.reset_changelog_cache()
+        changelog_module.get_release_notes("2.0")
+        changelog_module.get_release_notes("2.0")
+        assert hits["count"] == 1, "the failure should be cached"
+        changelog_module.get_release_notes("2.0", refresh = True)
+        assert hits["count"] == 2, "refresh must bypass the cached failure"
+    finally:
+        server.shutdown()
+        server.server_close()
+        changelog_module.reset_changelog_cache()
+
+
+def test_hook_never_returns_another_versions_notes():
+    """On the render where the offered version changes, state still describes
+    the previous one until the effect runs."""
+    src = NOTES_HOOK.read_text(encoding = "utf-8")
+    assert "notes.version === version" in src
+    assert "refresh" in src, "retry must ask the backend to bypass its cache"
+
+
+@pytest.mark.parametrize("indent", ["", " ", "  ", "   "])
+def test_headings_and_fences_allow_commonmark_indentation(changelog_module, indent):
+    """Markdown renders up to three leading spaces, so the parser must agree
+    or an indented release is unreachable and its notes join the one above."""
+    text = f"## 1.0\n\nOne.\n\n{indent}## 2.0\n\nTwo.\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0", "2.0"]
+    fenced = f"## 1.0\n\n{indent}```\n{indent}## 9.9.9\n{indent}```\n\n- real\n"
+    assert [e.version for e in changelog_module.parse_changelog(fenced)] == ["1.0"]
+
+
+def test_four_space_indentation_is_code_not_structure(changelog_module):
+    """At four spaces Markdown switches to indented code, for both forms."""
+    assert [
+        e.version for e in changelog_module.parse_changelog("    ## 9.9.9\n\n## 1.0\n\n- real\n")
+    ] == ["1.0"]
+    assert [
+        e.version
+        for e in changelog_module.parse_changelog(
+            "## 1.0\n\n    ```\n    sample\n\n## 2.0\n\n- two\n"
+        )
+    ] == ["1.0", "2.0"]
+
+
+def test_desktop_notes_link_to_the_release_page_on_every_platform():
+    """manualReleaseUrl is Linux-package only, so in-app updates on macOS,
+    Windows and AppImage would otherwise link to the generic changelog."""
+    hook = (FRONTEND / "hooks/use-tauri-update.ts").read_text(encoding = "utf-8")
+    assert "const releasePageUrl = info ?" in hook
+    banner = TAURI_BANNER.read_text(encoding = "utf-8")
+    assert "releaseNotesUrl={releasePageUrl ?? manualReleaseUrl}" in banner
+    provider = (FRONTEND / "app/provider.tsx").read_text(encoding = "utf-8")
+    assert "releasePageUrl={update.releasePageUrl}" in provider
+
+
+def test_preview_matches_how_markdown_renders_prose_and_links():
+    """Three rendering mismatches the preview must not reintroduce: wrapped
+    paragraphs split into fragments, autolinks eaten as tags, and a lead cut
+    at an abbreviation."""
+    src = PREVIEW.read_text(encoding = "utf-8")
+    # Contiguous prose lines accumulate and flush at a paragraph boundary.
+    assert "collector.paragraph = collector.paragraph" in src
+    # <https://x> renders as link text, so it is not a tag.
+    assert "AUTOLINK" in src
+    # "e.g. GGUF" is not a sentence boundary.
+    assert "ABBREVIATIONS" in src and "INITIAL" in src
+
+
+def test_preview_treats_code_as_literal():
+    """Inside a code span, and inside an indented code block, Markdown renders
+    the text literally, so the preview must not transform or promote it."""
+    src = PREVIEW.read_text(encoding = "utf-8")
+    # Code spans are parked before any other inline transformation.
+    park = src.index("parkCodeSpans(markdown")
+    assert park < src.index("stripHtmlTags(\n    parked")
+    # A "- cmd" line inside an indented code block is not a headline bullet.
+    assert "INDENTED_CODE_INDENT" in src
+
+
+def test_desktop_updater_metadata_maps_published_field_names():
+    """latest.json publishes Tauri's `notes`/`pub_date`; the manual Linux path
+    must read those, not `body`/`date`, or its release notes are always empty."""
+    rust = (REPO / "studio/src-tauri/src/desktop_update_policy.rs").read_text(encoding = "utf-8")
+    assert 'alias = "body"' in rust and "notes: Option<String>" in rust
+    assert 'alias = "date"' in rust and "pub_date: Option<String>" in rust
+    assert "body: metadata.notes" in rust and "date: metadata.pub_date" in rust
+    workflow = (REPO / ".github/workflows/release-desktop.yml").read_text(encoding = "utf-8")
+    assert "'notes': notes," in workflow, "workflow no longer publishes `notes`"
+
+
+def test_backend_exposes_release_notes_route():
+    src = (BACKEND / "main.py").read_text(encoding = "utf-8")
+    assert '@app.get("/api/studio/release-notes")' in src
+    assert "is_supported_version_query" in src
+
+
+def test_panel_is_scrollable_and_version_scoped():
+    src = PANEL.read_text(encoding = "utf-8")
+    assert "overflow-y-auto" in src, "release notes must scroll inside the popup"
+    assert "max-h-" in src, "the scroller needs a bounded height"
+    # Falls back to the payload's own body only, never to another version.
+    assert "fallbackMarkdown" in src
+
+
+def test_notes_surface_is_borderless_and_lifts_in_dark_mode():
+    src = PANEL.read_text(encoding = "utf-8")
+    assert "border border-border" not in src, "the notes box is a fill, not a bordered box"
+    # Lighter than the card behind it, rather than a darker inset.
+    assert "dark:bg-white/[0.06]" in src
+    # Streamdown's mt-6 clips the first heading against the scroller edge.
+    assert "[&>*>*:first-child]:mt-0" in src
+    # Shared utility: thumb hidden until the notes are hovered.
+    assert "hover-scrollbar" in src
+    # Streamdown renders code at text-sm, twice this panel's body size.
+    assert "[&_code]:text-[0.92em]" in src
+
+
+def test_hook_discards_notes_for_a_different_version():
+    src = NOTES_HOOK.read_text(encoding = "utf-8")
+    assert "notesVersion !== version" in src
+
+
+def test_collapsed_panel_previews_the_top_bullets():
+    """Collapsed popups show the headline changes without an extra click."""
+    preview = PREVIEW.read_text(encoding = "utf-8")
+    assert "RELEASE_NOTES_PREVIEW_ITEMS = 4" in preview
+    # Wrapped bullets join into one item, or a preview ends mid-sentence.
+    assert "collectBullets" in preview and "flush" in preview
+    # Nested list items are detail, not headline changes.
+    assert "NESTED_INDENT_TOLERANCE" in preview
+    # Tag stripping repeats: one pass turns `<<b>b>` back into a live tag.
+    assert "while (out !== previous)" in preview
+
+    panel = PANEL.read_text(encoding = "utf-8")
+    assert "releaseNotesPreview" in panel
+    assert 'data-testid="update-release-notes-summary"' in panel
+    # Fetched when the popup appears: the collapsed preview needs them too.
+    assert "enabled: true" in panel
+
+
+def test_preview_highlights_the_leading_sentence():
+    """Each bullet leads with its headline sentence, emphasised over the rest."""
+    preview = PREVIEW.read_text(encoding = "utf-8")
+    assert "splitLeadSentence" in preview
+    # A period inside "CHANGELOG.md" or "e.g." must not read as a break.
+    assert "SENTENCE_BREAK" in preview and "(?=" in preview
+
+    panel = PANEL.read_text(encoding = "utf-8")
+    assert '<span className="font-medium text-foreground">{item.lead}</span>' in panel
+    assert "item.rest" in panel
+
+
+@pytest.mark.parametrize("banner", [WEB_BANNER, TAURI_BANNER])
+def test_update_popup_is_wider_than_the_other_overlays(banner):
+    """The card is sized for three same-size buttons on one row.
+
+    Width moved from the shared overlay stack onto each overlay, so widening
+    the update popup does not widen the llama.cpp banner or download panel."""
+    assert "max-w-[448px]" in banner.read_text(encoding = "utf-8")
+    provider = (FRONTEND / "app/provider.tsx").read_text(encoding = "utf-8")
+    assert "max-w-[400px]" not in provider, "stack must not cap overlay width"
+    llama = (FRONTEND / "components/llama-update-banner.tsx").read_text(encoding = "utf-8")
+    assert "max-w-[400px]" in llama, "unrelated overlays keep their width"
+
+
+@pytest.mark.parametrize("banner", [WEB_BANNER, TAURI_BANNER])
+def test_banners_toggle_inline_release_notes(banner):
+    src = banner.read_text(encoding = "utf-8")
+    assert "ReleaseNotesPanel" in src
+    assert "Show release notes" in src and "Hide release notes" in src
+    # Keyed by version, so a new offer cannot leave old notes on screen.
+    assert "notesVersion" in src
+
+
+@pytest.mark.parametrize(
+    "banner,toggle,action",
+    [
+        (WEB_BANNER, "web-update-release-notes-toggle", "web-update-snooze-button"),
+        (TAURI_BANNER, "tauri-update-release-notes-toggle", "Remind me later"),
+    ],
+)
+def test_notes_toggle_shares_the_action_row(banner, toggle, action):
+    """The toggle sits in the same row as the actions, not on its own line."""
+    src = banner.read_text(encoding = "utf-8")
+    row = src.index("mt-4 flex")
+    assert row < src.index(toggle) < src.index(action)
+    # Same type size as the actions beside it; nowrap keeps labels on one line.
+    toggle_line = next(line for line in src.splitlines() if toggle in line)
+    toggle_block = src[src.index("Button", row) : src.index(toggle_line)]
+    assert "text-ui-13" in toggle_block and "whitespace-nowrap" in toggle_block
+
+
+def test_headings_inside_a_raw_html_block_are_not_releases(changelog_module):
+    """<pre> content is literal, so a sample heading in it must not become a
+    section and must not cut the real section's body short."""
+    text = "## 1.0\n\n<pre>\n## 9.9.9\n</pre>\n\n- real note\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0"]
+    assert "real note" in changelog_module.find_release_notes(text, "1.0").body
+    assert changelog_module.find_release_notes(text, "9.9.9") is None
+
+
+def test_details_blocks_still_contain_markdown(changelog_module):
+    """<details> is a CommonMark type 6 block: headings inside it still count,
+    so collapsible sections keep working."""
+    text = "## 2.0\n\n<details>\n<summary>More</summary>\n\n- note\n\n</details>\n\n## 1.0\n\n- older\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["2.0", "1.0"]
+
+
+def test_inline_raw_html_tag_does_not_open_a_block(changelog_module):
+    """A block opens only at the start of a line. A tag named mid-sentence is
+    inline HTML and must not swallow the releases below it."""
+    text = "## 2.0\n\n- Warn when a <script> tag is pasted\n\n## 1.0\n\n- older\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["2.0", "1.0"]
+
+
+def test_preview_skips_raw_html_blocks():
+    src = PREVIEW.read_text(encoding = "utf-8")
+    assert "stripRawHtml" in src
+    # Anchored: only a line-leading tag opens a block, matching the parser.
+    assert "/^ {0,3}<(pre|script|style|textarea)" in src
+
+
+def test_fence_inside_a_raw_html_block_is_literal(changelog_module):
+    """Raw HTML contents are literal, so a stray ``` in a <pre> sample is not a
+    fence. Treating it as one left a block open and hid every later release."""
+    text = "## 2.0\n\n<pre>\n```\n</pre>\n\n## 1.0\n\n- older\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["2.0", "1.0"]
+
+
+def test_raw_html_block_closes_on_any_of_the_four_tags(changelog_module):
+    """CommonMark ends a type 1 block at the first `</pre>`, `</script>`,
+    `</style>` or `</textarea>`: the closer need not match the opener."""
+    text = '## 1.0\n\n<script>\nconst sample = "</pre>";\n## 9.9.9\n</script>\n'
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0", "9.9.9"]
+
+
+@pytest.mark.parametrize("tag", ["details", "div", "table"])
+def test_type_6_blocks_run_until_a_blank_line(changelog_module, tag):
+    """`<details>` holds Markdown only after a blank line closes the block, so
+    a heading pressed against the opening tag is not a release."""
+    packed = f"## 1.0\n\n<{tag}>\n## 9.9.9\n</{tag}>\n\n- note\n"
+    assert [e.version for e in changelog_module.parse_changelog(packed)] == ["1.0"]
+    spaced = f"## 1.0\n\n<{tag}>\n\n## 2.0\n\n- note\n"
+    assert [e.version for e in changelog_module.parse_changelog(spaced)] == ["1.0", "2.0"]
+
+
+def test_a_tag_only_line_cannot_interrupt_a_paragraph(changelog_module):
+    """Type 7 blocks do not interrupt a paragraph, so prose followed by a bare
+    tag keeps the releases below it reachable."""
+    text = "## 2.0\n\nSome prose.\n<span>\n\n## 1.0\n\n- older\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["2.0", "1.0"]
+
+
+def test_preview_joins_an_indented_continuation_line():
+    """Four spaces only start code outside a paragraph. Inside one the line is
+    a wrapped continuation, so it must not be dropped from the preview."""
+    src = PREVIEW.read_text(encoding = "utf-8")
+    # Measured from the line's container, so an item's own indent does not count.
+    assert "!insideBlock && line.indent - line.column >= INDENTED_CODE_INDENT" in src
+    # A fence indented into a list item is a block, not a wrapped line.
+    assert "opensDeepFence" in src
+
+
+def test_every_packaging_path_snapshots_the_changelog():
+    """`python -m build` and `pip install .` must ship the offline copy too,
+    so the snapshot is made by the build backend rather than by build.sh."""
+    pyproject = (REPO / "pyproject.toml").read_text(encoding = "utf-8")
+    assert 'build_py = "_changelog_build.build_py"' in pyproject
+    hook = (REPO / "_changelog_build.py").read_text(encoding = "utf-8")
+    assert "studio" in hook and "CHANGELOG.md" in hook
+    # The hook has to reach the sdist, or building from one loses the snapshot.
+    manifest = (REPO / "MANIFEST.in").read_text(encoding = "utf-8")
+    assert "include _changelog_build.py" in manifest
+    assert "include CHANGELOG.md" in manifest
+
+
+def test_preview_code_spans_need_a_matching_closer():
+    """A closer is a run of the same length, so ``Use `` `x` `` `` keeps the
+    inner backticks the expanded notes show."""
+    src = CODE_SPANS.read_text(encoding = "utf-8")
+    assert "candidate === ticks" in src, "a closer is a run of the same length"
+    assert "stripPadding" in src, "one space of padding is dropped, as in Markdown"
+
+
+def test_preview_skips_thematic_breaks():
+    """`- - -` renders as a rule, so it must not take a preview slot."""
+    src = PREVIEW.read_text(encoding = "utf-8")
+    assert "THEMATIC_BREAK" in src
+    assert "THEMATIC_BREAK.test(visible)" in src
+
+
+def test_preview_keeps_quoted_examples_out_of_the_headlines():
+    """A quoted list is example output, not a change, so it never competes
+    with the release's own bullets."""
+    src = PREVIEW.read_text(encoding = "utf-8")
+    assert "quoted: boolean" in src
+    assert "if (!line.quoted)" in src, "quoted bullets never become headlines"
+
+
+def test_notes_panel_keeps_the_link_when_the_lookup_fails():
+    """Retry is not the only route: the changelog page can be reachable even
+    when the backend lookup is not."""
+    src = PANEL.read_text(encoding = "utf-8")
+    error_branch = src[src.index('if (state === "error")') :]
+    retry = error_branch.index("update-release-notes-retry")
+    assert error_branch.index("{link}") > retry, "link sits beside retry"
+
+
+def test_hook_waits_for_the_desktop_auth_token():
+    """The desktop popup can render before auto-auth installs its token, so a
+    missing token must not be recorded as a failed lookup."""
+    src = NOTES_HOOK.read_text(encoding = "utf-8")
+    assert "hasAuthToken()" in src and "AUTH_POLL_LIMIT" in src
+
+
+def test_installed_layout_prefers_the_bundled_changelog(tmp_path):
+    """Installed, the levels above studio/ are site-packages. A stray
+    CHANGELOG.md left there by another package must not outrank the bundled
+    snapshot, so those levels are only searched in a source checkout."""
+    site_packages = tmp_path / "site-packages"
+    package = site_packages / "studio/backend/utils"
+    package.mkdir(parents = True)
+    for name in ("changelog.py", "update_status.py"):
+        shutil.copy(BACKEND / "utils" / name, package / name)
+    for parent in (site_packages / "studio", package.parent, package):
+        (parent / "__init__.py").write_text("", encoding = "utf-8")
+    (site_packages / CHANGELOG.name).write_text("## 2.0\n\n- stray\n", encoding = "utf-8")
+    bundled = site_packages / "studio" / CHANGELOG.name
+    bundled.write_text("## 2.0\n\n- bundled\n", encoding = "utf-8")
+
+    env = {**os.environ, "PYTHONPATH": str(site_packages)}
+    env.pop("UNSLOTH_CHANGELOG_PATH", None)
+
+    def served() -> str:
+        # cwd is outside the checkout, so this imports the installed copy.
+        return subprocess.run(
+            [
+                sys.executable,
+                "-c",
+                "from studio.backend.utils import changelog\n"
+                "print(changelog._read_local_changelog().text)",
+            ],
+            capture_output = True,
+            text = True,
+            env = env,
+            cwd = tmp_path,
+            check = True,
+        ).stdout
+
+    assert "bundled" in served() and "stray" not in served()
+
+    # A checkout marker there means it really is a repo root, so it wins again.
+    (site_packages / "pyproject.toml").write_text("", encoding = "utf-8")
+    assert "stray" in served()
+
+
+def test_a_section_staged_as_a_comment_reads_as_unpublished(
+    changelog_module, tmp_path, monkeypatch
+):
+    """Notes staged inside <!-- --> render as nothing, so the popup must say
+    no notes were published rather than show an empty surface."""
+    monkeypatch.setenv(changelog_module.DISABLE_ENV_VAR, "1")
+    local = tmp_path / "CHANGELOG.md"
+    local.write_text("## 2.0\n\n<!-- not ready -->\n\n## 1.0\n\n- shipped\n", encoding = "utf-8")
+    monkeypatch.setenv(changelog_module.CHANGELOG_PATH_ENV_VAR, str(local))
+    changelog_module.reset_changelog_cache()
+    try:
+        staged = changelog_module.get_release_notes("2.0")
+        assert staged["matched"] is False and staged["markdown"] is None
+        assert changelog_module.get_release_notes("1.0")["matched"] is True
+    finally:
+        changelog_module.reset_changelog_cache()
+
+
+@pytest.mark.parametrize(
+    "body,visible",
+    [
+        ("- note", True),
+        ("<!-- staged -->", False),
+        ("```\n```", True),
+        ("<pre>\n</pre>", True),
+        ("  ", False),
+    ],
+)
+def test_visibility_check_only_hides_comments(changelog_module, body, visible):
+    assert changelog_module._renders_visibly(body) is visible
+
+
+@pytest.mark.parametrize(
+    "block",
+    [
+        "<?php\n## 9.9.9\n?>",
+        "<![CDATA[\n## 9.9.9\n]]>",
+        "<!DOCTYPE\n## 9.9.9\n>",
+    ],
+)
+def test_processing_instructions_and_declarations_are_literal(changelog_module, block):
+    """Raw block types 3 to 5 render literally, like <pre>, so a heading inside
+    one is a sample and not a release."""
+    text = f"## 1.0\n\n{block}\n\n- real note\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0"]
+    assert "real note" in changelog_module.find_release_notes(text, "1.0").body
+
+
+def test_headings_need_a_space_or_tab_after_the_hashes(changelog_module):
+    """A non-breaking space pasted from rich text renders as ordinary text, so
+    the line must not end the release above it."""
+    text = "## 1.0\n\n- real note\n\n## 9.9.9\n\n- not a release\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0"]
+    assert changelog_module.find_release_notes(text, "9.9.9") is None
+    # A tab is valid and still opens a heading.
+    tabbed = "## 1.0\n\n- one\n\n##\t2.0\n\n- two\n"
+    assert [e.version for e in changelog_module.parse_changelog(tabbed)] == ["1.0", "2.0"]
+
+
+def test_preview_skips_every_raw_block_form():
+    """The extractor tracks the same block forms as the parser, so a sample
+    bullet inside one cannot become the collapsed headline."""
+    src = PREVIEW.read_text(encoding = "utf-8")
+    assert "RAW_BLOCKS" in src
+    assert "CDATA" in src and "[A-Za-z]" in src
+
+
+@pytest.mark.parametrize("banner", [WEB_BANNER, TAURI_BANNER])
+def test_expanded_popup_fits_a_short_viewport(banner):
+    """A window under roughly 430px high used to push the card's title and
+    dismiss control above the top of the screen."""
+    panel = PANEL.read_text(encoding = "utf-8")
+    # The notes region shrinks inside the capped card, so header and actions stay on screen.
+    assert "min-h-0 flex-1" in panel, "notes height must follow the viewport"
+    src = banner.read_text(encoding = "utf-8")
+    assert "max-h-[calc(100dvh_-_2rem)]" in src, "card is the backstop on tiny viewports"
+
+
+def test_relative_changelog_links_point_at_the_repository():
+    """CHANGELOG.md links are repository-relative. Rendered as-is they resolve
+    against Studio's origin, so the renderer blocks them."""
+    src = LINKS.read_text(encoding = "utf-8")
+    assert "https://github.com/unslothai/unsloth/blob/main/" in src
+    assert "https://raw.githubusercontent.com/unslothai/unsloth/main/" in src
+    # Absolute targets, fragments, fenced code and code spans stay untouched.
+    assert "ABSOLUTE" in src and "codeSpans" in src and "FENCE" in src
+    panel = PANEL.read_text(encoding = "utf-8")
+    assert "resolveChangelogLinks" in panel
+
+
+@pytest.mark.parametrize("query", ["latest", "main", "not-a-version", "abc"])
+def test_unparseable_versions_are_rejected(changelog_module, query):
+    """Sections are indexed only when their version parses, so a query that
+    cannot parse can never match and is a bad request, not an empty result."""
+    assert changelog_module.is_supported_version_query(query) is False
+
+
+@pytest.mark.parametrize("query", ["2026.7.5", "v2026.7.5", "2026.07.5", "1.0.0rc1"])
+def test_real_versions_are_still_accepted(changelog_module, query):
+    assert changelog_module.is_supported_version_query(query) is True
+
+
+def test_reference_style_images_resolve_to_the_raw_host():
+    """`![alt][arch]` with `[arch]: docs/arch.png` needs the raw file: the blob
+    URL is an HTML page, so the image would not load."""
+    src = LINKS.read_text(encoding = "utf-8")
+    assert "IMAGE_REFERENCE" in src
+    assert "imageLabels" in src
+
+
+def test_collapsed_notes_surface_is_hidden_when_nothing_previews():
+    """Notes that are only a fenced command block preview as nothing, and an
+    empty muted strip is worse than no strip."""
+    src = PANEL.read_text(encoding = "utf-8")
+    assert "preview?.items.length === 0" in src
+
+
+def test_a_fence_closer_accepts_only_spaces_and_tabs(changelog_module):
+    """A delimiter followed by a non-breaking space is code content, so it must
+    not close the block and let a sample heading through."""
+    text = "## 1.0\n\n```\n```\u00a0\n## 9.9.9\n```\n\n- real note\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0"]
+    plain = "## 1.0\n\n```\nx\n```\t\n\n## 2.0\n\n- two\n"
+    assert [e.version for e in changelog_module.parse_changelog(plain)] == ["1.0", "2.0"]
+    # The same rule in both frontend scanners.
+    for source in (PREVIEW, LINKS):
+        assert "/[^ \\t]/" in source.read_text(encoding = "utf-8")
+
+
+def test_code_spans_close_on_a_run_of_equal_length():
+    """`a``b [x](y.md)` is one code span, so the link inside it is literal."""
+    src = CODE_SPANS.read_text(encoding = "utf-8")
+    assert "candidate === ticks" in src, "closer length must match the opener"
+    # Shared, so the preview and the link resolver cannot drift apart.
+    assert "markdown-code-spans" in PREVIEW.read_text(encoding = "utf-8")
+    assert "markdown-code-spans" in LINKS.read_text(encoding = "utf-8")
+
+
+def test_preview_decodes_entities_like_the_renderer():
+    """Streamdown renders `AT&amp;T` as AT&T, so the collapsed preview must
+    not show the raw entity."""
+    src = PREVIEW.read_text(encoding = "utf-8")
+    assert "NAMED_ENTITIES" in src and "decodeEntity" in src
+    # Decoded before code spans are restored, so code keeps the literal text.
+    assert src.index(".replace(ENTITY, decodeEntity)") < src.index(".replace(PARKED")
+
+
+def test_release_notes_request_refreshes_an_expired_token():
+    """A direct fetch cannot recover from a 401; authFetch refreshes first."""
+    src = NOTES_HOOK.read_text(encoding = "utf-8")
+    assert "authFetch(" in src
+    assert "getAuthToken" not in src
+
+
+def test_preview_handles_the_desktop_updater_line_endings():
+    """The updater body arrives with CRLF, which used to hide fences from the
+    extractor and promote a code sample to a headline."""
+    src = PREVIEW.read_text(encoding = "utf-8")
+    assert "LINE_ENDINGS" in src
+    assert "LINE_ENDINGS" in LINKS.read_text(encoding = "utf-8")
+
+
+def test_preview_renders_reference_links_as_text():
+    """`[text][label]` and `![alt][label]` render as a link and an image, so
+    the preview must not show their raw markup."""
+    src = PREVIEW.read_text(encoding = "utf-8")
+    assert "LINK_REFERENCE" in src and "IMAGE_REFERENCE" in src
+    # A definition line renders as nothing, so it is not a preview item.
+    assert "DEFINITION" in src
+
+
+def test_preview_treats_escaped_punctuation_as_literal():
+    """`\\*not italic\\*` keeps its stars and an escaped backtick does not open
+    a code span."""
+    assert "ESCAPE" in PREVIEW.read_text(encoding = "utf-8")
+    assert "escaped(" in CODE_SPANS.read_text(encoding = "utf-8")
+
+
+def test_link_resolver_skips_every_code_form():
+    """Indented code and code spans crossing a line render as code, so their
+    contents must not be rewritten."""
+    src = LINKS.read_text(encoding = "utf-8")
+    assert "INDENTED_CODE" in src
+    # Spans are scanned over the whole document, not line by line.
+    assert "codeSpans(masked)" in src
+    # A definition cannot interrupt a paragraph.
+    assert "definition.has(index)" in src
+
+
+def test_badge_links_resolve_both_targets():
+    """`[![alt](img)](link)` is the badge idiom: the outer link used to stay
+    relative because the label was not allowed to nest."""
+    assert "NESTED_LABEL" in LINKS.read_text(encoding = "utf-8")
+
+
+def test_in_flight_requests_are_identified_not_just_versioned():
+    """Two requests for the same version could resolve out of order and leave
+    the panel showing the older result."""
+    assert "requestIdRef" in NOTES_HOOK.read_text(encoding = "utf-8")
+
+
+def test_notes_repair_the_shared_previews_width_reset():
+    """MarkdownPreview clears max-width on every descendant, so a wide image
+    and the renderer's own link dialog escape the card."""
+    src = PANEL.read_text(encoding = "utf-8")
+    assert "[&_img]:max-w-full" in src
+    assert "[&_[data-streamdown=link-safety-modal]>*]:max-w-md" in src
+
+
+@pytest.mark.parametrize("banner", [WEB_BANNER, TAURI_BANNER])
+def test_only_the_notes_region_scrolls(banner):
+    """The dismiss control sits inside the card, so scrolling the card itself
+    carried it off screen on a short viewport."""
+    src = banner.read_text(encoding = "utf-8")
+    assert "flex max-h-[calc(100dvh_-_2rem)] flex-col overflow-hidden" in src
+    assert 'className="min-h-0 flex-1"' in src
+    panel = PANEL.read_text(encoding = "utf-8")
+    assert "max-h-64 min-h-0 flex-1 overflow-y-auto" in panel
+
+
+def test_a_comment_marker_in_prose_cannot_swallow_later_releases(changelog_module):
+    """A note that mentions `<!--` used to put the parser into comment state
+    for the rest of the file: the releases below it disappeared and their
+    notes were served under the newer version's heading."""
+    text = (
+        "## 2026.8.0\n\n- Studio strips <!-- markers from pasted prompts.\n\n"
+        "## 2026.7.5\n\n- SECRET: an older release\n"
+    )
+    assert [e.version for e in changelog_module.parse_changelog(text)] == [
+        "2026.8.0",
+        "2026.7.5",
+    ]
+    assert "SECRET" not in changelog_module.find_release_notes(text, "2026.8.0").body
+    assert changelog_module.find_release_notes(text, "2026.7.5") is not None
+    # A comment that starts a line is still a block and still hides its body.
+    hidden = "## 2.0\n\n<!--\n## 9.9.9\n-->\n\n- note\n"
+    assert [e.version for e in changelog_module.parse_changelog(hidden)] == ["2.0"]
+
+
+def test_unmatched_backtick_runs_stay_linear(changelog_module):
+    """Rescanning the suffix for every opener was quadratic: a line of runs of
+    1, 2, 3 ... backticks, none of which ever closes, took 7.7s at 321 KB and
+    is reparsed on every popup request, so one malformed remote changelog could
+    tie up backend workers."""
+    line = "".join("`" * (i + 1) + "x" for i in range(800))
+    assert len(line) > 300_000
+    started = time.monotonic()
+    assert changelog_module._code_span_ranges(line) == []
+    assert time.monotonic() - started < 2.0
+
+
+def test_a_base_exception_releases_the_single_flight_flag(changelog_module, monkeypatch):
+    """The flag was cleared only after `except Exception`, so a BaseException
+    (KeyboardInterrupt, SystemExit, CancelledError) stranded it and every later
+    caller then waited out the full deadline for the life of the process."""
+    changelog_module.reset_changelog_cache()
+
+    def explode():
+        raise KeyboardInterrupt
+
+    monkeypatch.setattr(changelog_module, "_fetch_remote_changelog", explode)
+    with pytest.raises(KeyboardInterrupt):
+        changelog_module.get_remote_changelog()
+    assert changelog_module._remote_fetching is False
+    changelog_module.reset_changelog_cache()
+
+
+@pytest.mark.parametrize("marker", ["<!-->", "<!--->"])
+def test_an_empty_comment_does_not_swallow_later_releases(changelog_module, marker):
+    """`<!-->` and `<!--->` are complete comments in CommonMark: the closer
+    overlaps the opener. Searching for `-->` past the opener missed them, so an
+    empty comment used as a section marker hid every release below it."""
+    text = f"## 2.0\n\n- new stuff\n\n{marker}\n\n## 1.0\n\n- old stuff\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["2.0", "1.0"]
+    assert changelog_module.find_release_notes(text, "1.0") is not None
+    assert "old stuff" not in changelog_module.find_release_notes(text, "2.0").body
+    # The frontend scanner has to agree, or the preview and the body disagree.
+    assert "!line.includes(COMMENT_CLOSE)" in PREVIEW.read_text(encoding = "utf-8")
+
+
+def test_an_unterminated_comment_still_hides_the_rest(changelog_module):
+    """The fix must not turn every `<!--` line into a no-op block."""
+    text = "## 2.0\n\n<!-- never closed\n\n## 1.0\n\n- old stuff\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["2.0"]
+
+
+def test_a_closing_delimiter_takes_its_whole_line(changelog_module):
+    """CommonMark keeps the closing line inside the block, so a heading glued
+    after `-->` or `</pre>` is not a release."""
+    for text in (
+        "## 1.0\n\n<!-- hidden -->## 9.9.9\n\n- note\n",
+        "## 1.0\n\n<pre>\nx\n</pre>## 9.9.9\n\n- note\n",
+    ):
+        assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0"]
+
+
+def test_an_exact_heading_is_never_shadowed(changelog_module):
+    """PEP 440 says 1.0 == 1.0.0, so the normalised match used to win even
+    when the file had a section spelled exactly as asked."""
+    text = "## 1.0.0\n\n- padded\n\n## 1.0\n\n- exact\n"
+    assert changelog_module.find_release_notes(text, "1.0").body == "- exact"
+    assert changelog_module.find_release_notes(text, "1.0.0").body == "- padded"
+    # Normalised matching still applies when there is no exact heading.
+    assert changelog_module.find_release_notes("## 2026.7.6\n\n- x\n", "2026.07.6") is not None
+
+
+def test_setext_headings_are_release_boundaries(changelog_module):
+    """A version over a line of dashes is the same heading in setext form."""
+    text = "2.0\n---\n\n- new\n\n1.0\n---\n\n- old\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["2.0", "1.0"]
+    assert changelog_module.find_release_notes(text, "2.0").body == "- new"
+    # A rule between sections is still a rule, and a setext h1 is not a release.
+    assert [
+        e.version
+        for e in changelog_module.parse_changelog("## 2.0\n\n- a\n\n---\n\n## 1.0\n\n- b\n")
+    ] == ["2.0", "1.0"]
+
+
+def test_a_long_backtick_run_does_not_stall_the_parser(changelog_module):
+    """The code-span guard used to backtrack: 20k backticks took over a minute
+    and every request re-parsed the file."""
+    import time
+
+    text = "## 1.0\n\n- " + "`" * 20_000 + " <!--\n"
+    started = time.perf_counter()
+    changelog_module.parse_changelog(text)
+    assert time.perf_counter() - started < 1.0
+
+
+def test_the_remote_fetch_has_a_total_deadline(changelog_module):
+    """The socket timeout resets on every read, so a trickling server could
+    hold a worker for minutes and still be treated as a success."""
+    source = (BACKEND / "utils/changelog.py").read_text(encoding = "utf-8")
+    assert "deadline = time.monotonic() + CHANGELOG_TIMEOUT_SECONDS" in source
+    # read1 returns after one socket read, so the deadline is actually checked.
+    assert "response.read1(" in source
+    # Waiters give up rather than queue behind a stalled fetch.
+    assert "Release notes are still loading." in source
+
+
+def test_truncated_notes_close_their_fence(changelog_module):
+    """A blind slice could end inside a code block and break the rendering."""
+    body = "```\n" + "x\n" * 20_000 + "```\n"
+    payload = changelog_module._notes_response(version = "1.0", markdown = body, source = "local")
+    assert payload["truncated"] is True
+    assert payload["markdown"].rstrip().endswith("```")
+
+
+def test_the_opt_out_beats_the_developer_override():
+    """UNSLOTH_STUDIO_FAKE_UPDATE is a dev switch; the documented kill switch
+    still wins, and the value has to parse as a version."""
+    source = (BACKEND / "utils/update_status.py").read_text(encoding = "utf-8")
+    assert "forced_version and not disabled and _is_version(forced_version)" in source
+
+
+def test_a_list_item_over_dashes_is_not_a_setext_heading(changelog_module):
+    """`- first` followed by `---` is a list and a rule. Reading it as a
+    heading discarded the bullet and the rest of the section with it."""
+    text = "## 1.0\n\n- first\n---\n\n- second\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0"]
+    body = changelog_module.find_release_notes(text, "1.0").body
+    assert "first" in body and "second" in body
+    # Real setext headings still work.
+    setext = "2.0\n---\n\n- new\n\n1.0\n---\n\n- old\n"
+    assert [e.version for e in changelog_module.parse_changelog(setext)] == ["2.0", "1.0"]
+
+
+def test_a_backtick_in_a_fence_info_string_is_not_a_fence(changelog_module):
+    """CommonMark forbids backticks in a backtick fence's info string, so such
+    a line is prose and must not swallow the releases below it."""
+    text = "## 2.0\n\n```bad`info\n\n## 1.0\n\n- old\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["2.0", "1.0"]
+    # A tilde fence may hold backticks, and a normal fence still hides samples.
+    assert [
+        e.version
+        for e in changelog_module.parse_changelog(
+            "## 2.0\n\n```md\n## 9.9.9\n```\n\n## 1.0\n\n- old\n"
+        )
+    ] == ["2.0", "1.0"]
+    for source in (PREVIEW, LINKS):
+        assert "info string" in source.read_text(encoding = "utf-8")
+
+
+def test_preview_follows_commonmark_paragraph_rules():
+    """Only an ordered list starting at 1 may interrupt a paragraph, and an
+    unresolved reference keeps its brackets. A quote owns the paragraph its own
+    lines hold, so a marker written outside the quote interrupts nothing."""
+    src = " ".join(PREVIEW.read_text(encoding = "utf-8").split())
+    assert "const interrupts = collector.current === null" in src
+    assert "!collector.quotedParagraph;" in src
+    assert "definedLabel" in src, "a reference only renders as text when defined"
+    # A comment written mid-sentence hides its own line at most.
+    assert "COMMENT_BLOCK_OPEN" in src
+
+
+def test_link_resolver_leaves_raw_blocks_and_escapes_alone():
+    src = LINKS.read_text(encoding = "utf-8")
+    assert "RAW_HTML_OPEN" in src and "inRawHtml" in src
+    assert "isEscaped(line, opener)" in src
+    # A heading ends a paragraph, so a definition under one is a definition.
+    assert "BLOCK_LINE.test(structure)" in src
+
+
+def test_code_span_closers_ignore_backslashes():
+    """Escapes are not processed inside a code span, so a run after a
+    backslash still closes it."""
+    src = CODE_SPANS.read_text(encoding = "utf-8")
+    body = src[src.index("export function codeSpans") :]
+    assert body.count("escaped(text") == 1, "only an opener can be escaped"
+
+
+def test_the_overlay_stack_fits_the_viewport():
+    """The update card's own cap does not account for a long download list
+    stacked beneath it."""
+    provider = (FRONTEND / "app/provider.tsx").read_text(encoding = "utf-8")
+    assert "max-h-[calc(100dvh_-_2rem)]" in provider
+    panel = (FRONTEND / "features/hub/download-manager/download-manager-panel.tsx").read_text(
+        encoding = "utf-8"
+    )
+    # Both overlays scroll internally, so they can give up height.
+    assert "flex min-h-0" in panel
+    assert "flex min-h-0" in WEB_BANNER.read_text(encoding = "utf-8")
+
+
+def test_the_desktop_stack_is_capped_like_the_browser_one():
+    """The download panel shares the desktop stack, so the update card's own
+    cap is not enough there either."""
+    provider = (FRONTEND / "app/provider.tsx").read_text(encoding = "utf-8")
+    assert provider.count("max-h-[calc(100dvh_-_2rem)]") == 2, "both stacks are capped"
+    assert "flex min-h-0" in TAURI_BANNER.read_text(encoding = "utf-8")
+
+
+def test_desktop_notes_are_looked_up_by_the_backend_version():
+    """latest.json's `version` is the app SemVer while CHANGELOG.md is keyed by
+    the backend release, so the desktop popup used to find no section at all
+    and fall back to the updater's generic text."""
+    workflow = (REPO / ".github/workflows/release-desktop.yml").read_text(encoding = "utf-8")
+    assert "'pypi_version': os.environ['PYPI_VERSION']" in workflow
+    assert "PYPI_VERSION: ${{ needs.prepare-version.outputs.pypi_version }}" in workflow
+    rust = (REPO / "studio/src-tauri/src/desktop_update_policy.rs").read_text(encoding = "utf-8")
+    assert "pypi_version: Option<String>" in rust
+    hook = NOTES_HOOK.parent.joinpath("use-tauri-update.ts").read_text(encoding = "utf-8")
+    # Both desktop paths carry it: the plugin exposes the raw metadata.
+    assert "rawPypiVersion(update.rawJson)" in hook
+    assert "manualUpdate.pypiVersion" in hook
+    banner = TAURI_BANNER.read_text(encoding = "utf-8")
+    assert "info?.pypiVersion ?? info?.version" in banner
+
+
+def test_one_slow_read_cannot_outlast_the_fetch_budget(changelog_module):
+    """The socket timeout is per operation, so slow headers followed by a slow
+    body could hold a worker for twice the advertised deadline."""
+    source = (BACKEND / "utils/changelog.py").read_text(encoding = "utf-8")
+    assert "_limit_read(response, remaining)" in source
+    assert "sock.settimeout(max(remaining, _CHANGELOG_MIN_READ_SECONDS))" in source
+
+
+def test_a_heading_indented_into_a_list_item_is_not_a_release(changelog_module):
+    """CommonMark keeps a heading at the item's content column inside the item.
+    Treating it as a boundary truncated the real release and indexed a version
+    that does not exist. Checked against markdown-it (commonmark preset)."""
+    text = "## 1.0\n\n- Example:\n  ## 9.9.9\n\n- after\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0"]
+    body = changelog_module.find_release_notes(text, "1.0").body
+    assert "9.9.9" in body and "after" in body
+    # One space short of the content column, the list ends and it is a release.
+    left = "## 1.0\n\n- Example:\n ## 2.0\n"
+    assert [e.version for e in changelog_module.parse_changelog(left)] == ["1.0", "2.0"]
+
+
+def test_a_closed_list_stops_holding_headings(changelog_module):
+    """Only an open item nests a heading, so a dedented paragraph, heading,
+    break or fence hands the following indentation back to the document."""
+
+    def versions(text):
+        return [e.version for e in changelog_module.parse_changelog(text)]
+
+    assert versions("## 1.0\n\n- Example:\n\nText.\n\n  ## 2.0\n") == ["1.0", "2.0"]
+    assert versions("## 1.0\n\n- Example:\n## 2.0\n  ## 3.0\n") == ["1.0", "2.0", "3.0"]
+    assert versions("## 1.0\n\n- Example:\n  Text.\n---\n  ## 2.0\n") == ["1.0", "2.0"]
+    assert versions("## 1.0\n\n- Example:\n```\n```\n  ## 2.0\n") == ["1.0", "2.0"]
+    # An item may begin with one blank line; content after that is outside it.
+    assert versions("## 1.0\n\n-\n\n  ## 2.0\n") == ["1.0", "2.0"]
+
+
+def test_a_version_line_is_not_an_ordered_list_marker(changelog_module):
+    """`2.` needs whitespace after it to be a marker, or list tracking would
+    read every setext version as a list item and lose the heading."""
+    text = "2.0\n---\n\n- new\n\n1.0\n---\n\n- old\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["2.0", "1.0"]
+    # An ordered item interrupts a paragraph only when it starts at 1.
+    assert [
+        e.version for e in changelog_module.parse_changelog("## 1.0\n\nText.\n9) one\n   ## 2.0\n")
+    ] == ["1.0", "2.0"]
+
+
+def test_a_wrapped_setext_heading_is_still_a_release(changelog_module):
+    """CommonMark promotes the whole paragraph, so a heading that wraps keeps
+    the version in its first token. Reading only the last line left the release
+    unindexed and its notes unreachable."""
+    text = "2026.7.5 - Release\nJuly 25\n---\n\n- note\n"
+    entries = changelog_module.parse_changelog(text)
+    assert [e.version for e in entries] == ["2026.7.5"]
+    # The heading lines are the heading, not the body.
+    assert entries[0].body == "- note"
+    assert "July 25" not in entries[0].body
+
+
+def test_a_lowercase_declaration_is_not_a_raw_block(changelog_module):
+    """Only `<!` plus an uppercase letter opens one, so prose that mentions
+    `<!note` must not hide every release under it."""
+    assert [
+        e.version for e in changelog_module.parse_changelog("<!note\n\n## 1.0\n\n- real\n")
+    ] == ["1.0"]
+    # A real declaration still hides its own block.
+    assert [
+        e.version for e in changelog_module.parse_changelog("<!DOCTYPE\n## 9.9.9\n>\n\n## 1.0\n")
+    ] == ["1.0"]
+    # The collapsed preview needs the same rule or it drops visible bullets.
+    assert "<![A-Z]" in PREVIEW.read_text(encoding = "utf-8")
+
+
+def test_link_resolver_reads_html_containers_the_way_the_others_do():
+    """A `<details>` or `<div>` with no blank line inside is a type 6 block, so
+    its contents render literally. Rewriting a link there mutates text the
+    reader sees verbatim, and a fence inside such a block was being taken for a
+    real fence, which stopped every link below it from resolving at all. The
+    backend parser and the collapsed preview already apply the type 6 and 7
+    rules, so the resolver has to share them or the three disagree on the same
+    notes."""
+    links = LINKS.read_text(encoding = "utf-8")
+    for source in (PREVIEW, LINKS):
+        text = source.read_text(encoding = "utf-8")
+        assert "HTML_BLOCK_TAGS" in text and "HTML_TAG_ONLY_LINE" in text
+    # A blank line ends the block, not the closing tag, and a bare quote marker counts as blank.
+    assert "inHtmlBlock = !!container.trim()" in links
+    # Type 7 cannot interrupt a paragraph, so prose above it keeps its links.
+    assert "return !afterParagraph && HTML_TAG_ONLY_LINE.test(line);" in links
+
+
+def test_an_escaped_mark_makes_an_image_a_link():
+    """`\\![alt](path)` renders as a link, so it resolves to the file's page on
+    GitHub rather than to the raw-content host."""
+    links = LINKS.read_text(encoding = "utf-8")
+    assert 'const image = bang === "!" && !isEscaped(line, offset);' in links
+    # The reference pre-scan has to skip it too, or the definition flips host.
+    assert "isEscaped(line, match.index)" in links
+
+
+def test_only_markdown_line_endings_split_the_changelog(changelog_module):
+    """str.splitlines also breaks on U+2028, U+2029, NEL, vertical tab and form
+    feed, none of which end a line in CommonMark. A separator sitting in prose
+    ahead of "## 9.9.9" made the parser index a release the renderer never shows
+    and truncate the notes above it."""
+    text = "## 2.0\n\nnote with a separator 
## 9.9.9\n\n## 1.0\n\n- old\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["2.0", "1.0"]
+    # The prose stays whole rather than being cut at the separator.
+    entry = changelog_module.find_release_notes(text, "2.0")
+    assert entry is not None and "9.9.9" in entry.body
+    for separator in ("
", "\x85", "\x0b", "\x0c"):
+        broken = f"## 2.0\n\nnote{separator}## 9.9.9\n\n## 1.0\n\n- old\n"
+        assert [e.version for e in changelog_module.parse_changelog(broken)] == ["2.0", "1.0"]
+    # The three real line endings still split.
+    for ending in ("\n", "\r\n", "\r"):
+        real = f"## 2.0{ending}{ending}- new{ending}{ending}## 1.0{ending}{ending}- old{ending}"
+        assert [e.version for e in changelog_module.parse_changelog(real)] == ["2.0", "1.0"]
+
+
+def test_the_build_does_not_require_a_writable_source_tree():
+    """A PEP 517 build may run against an immutable checkout (Nix, Bazel, a
+    read-only container mount). Writing the snapshot beside the sources raised
+    PermissionError before build_py started, so no wheel could be built at all.
+    """
+    src = (REPO / "_changelog_build.py").read_text(encoding="utf-8")
+    # The source-tree copy is best effort.
+    assert "except OSError:" in src
+    # The wheel gets its copy from the staging directory either way.
+    assert 'Path(self.build_lib) / "studio" / "CHANGELOG.md"' in src
+
+
+def test_link_resolver_reads_comments_before_fences():
+    """A fence delimiter hidden inside an HTML comment is not a fence. Reading
+    it as one left the fence open, so every visible line below was classified as
+    code and none of its links were resolved, which is far worse than the
+    mutated-text case: the whole rest of the notes silently stops working. The
+    order matters both ways, so a comment opener inside a real fence is not a
+    comment either."""
+    links = LINKS.read_text(encoding="utf-8")
+    # Fence state is read before comments are masked, the order the collapsed preview uses.
+    assert "const fenceSource = inComment\n      ? null\n      : FENCE.exec(" in links
+    # Masking happens only after the in-fence early return.
+    fence_return = links.index("// Fenced content is literal")
+    assert links.index("const [line, stillInComment, stillRunOn] = maskComments(") > fence_return
+    # Commented ranges join the code spans, so a hidden link is left alone.
+    assert "const spans = [...codeSpans(masked), ...comments].sort(" in links
+
+
+def test_preview_heading_and_quote_markers_follow_the_backend_rule():
+    """An ATX heading needs an ASCII space, a tab or the end of the line after
+    the marker, which is what _HEADING_PATTERN requires; `\\s` also matches a
+    non-breaking space, so prose beginning "## Important change" with one was
+    read as a heading and dropped, leaving a prose-only release with no
+    collapsed preview at all. A blockquote marker takes at most three leading
+    spaces for the same reason every other marker here does: accepting any run
+    let an indented code sample containing "> - sample output" shed its
+    indentation and be shown as the summary."""
+    src = PREVIEW.read_text(encoding="utf-8")
+    assert "const HEADING = /^#{1,6}(?:[ \\t]|$)/;" in src
+    assert "const HEADING_LINE = /^ {0,3}#{1,6}(?:[ \\t]|$)/;" in src
+    assert "const BLOCKQUOTE = /^ {0,3}>[ \\t]?/;" in src
+    # The backend rule this mirrors.
+    backend = (BACKEND / "utils" / "changelog.py").read_text(encoding="utf-8")
+    assert "^ {0,3}##(?:[ \\t]+(?P<title>.*?))?[ \\t]*$" in backend
+
+
+def test_preview_collects_labels_only_from_real_definitions():
+    """A definition-shaped line inside an indented code block or a deep fence is
+    literal text, so CommonMark leaves a later "[Beta] support" unresolved with
+    its brackets showing. Recording the label anyway made toPlainText strip them
+    in the collapsed preview, so it disagreed with the expanded view. The
+    pre-scan skips the same code the collector pass skips; a real definition
+    takes at most three spaces of indentation, so the indent test cannot reject
+    one."""
+    src = PREVIEW.read_text(encoding="utf-8")
+    scan = src.index("const labels = new Set<string>();")
+    collect = src.index("let deepFence: string | null = null;")
+    prescan = " ".join(src[scan:collect].split())
+    assert "let labelFence: string | null = null;" in prescan
+    assert "if (line.indent - line.column >= INDENTED_CODE_INDENT) { continue; }" in prescan
+    assert "endsDeepFence(labelFence, labelColumn, line)" in prescan
+
+
+def test_an_html_block_to_the_left_of_a_list_item_closes_it(changelog_module):
+    """Types 1 to 6 interrupt a paragraph, so an unindented <div> after "- item"
+    closes the item and a following one-to-three-space-indented "## 2.0" is a
+    real document heading. It was read as a lazy paragraph continuation, so the
+    item stayed open and the release below the block was swallowed."""
+    text = "## 3.0\n\n- item\n<div>\nhidden\n</div>\n\n  ## 2.0\n\n- two\n\n## 1.0\n\n- one\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["3.0", "2.0", "1.0"]
+    # Without the block the heading really is nested, so it stays suppressed.
+    nested = "## 3.0\n\n- item\n\n  ## 2.0\n\n- two\n\n## 1.0\n\n- one\n"
+    assert [e.version for e in changelog_module.parse_changelog(nested)] == ["3.0", "1.0"]
+    # Ordinary lazy continuation is untouched.
+    lazy = "## 3.0\n\n- item\ncontinued\n\n  ## 2.0\n\n## 1.0\n\n- one\n"
+    assert [e.version for e in changelog_module.parse_changelog(lazy)] == ["3.0", "1.0"]
+
+
+def test_the_download_panel_can_shrink_inside_the_capped_stack():
+    """The bottom-right stack is capped to the viewport, and a flex item defaults
+    to min-height:auto, so this wrapper could not shrink below its own content.
+    On a short viewport the cap was then absorbed by the update card, whose
+    header and actions are fixed, rather than by the download list, which
+    scrolls. Only the shared-stack branch needs it; standalone is positioned
+    fixed and is not a flex item at all."""
+    panel = (FRONTEND / "features/hub/download-manager/download-manager-panel.tsx").read_text(
+        encoding="utf-8"
+    )
+    assert 'positioned ? "fixed bottom-4 right-4 z-50" : "flex min-h-0 justify-end"' in panel
+    provider = (FRONTEND / "app/provider.tsx").read_text(encoding="utf-8")
+    assert "max-h-[calc(100dvh_-_2rem)]" in provider, "the cap this has to absorb"
+
+
+@pytest.fixture(scope="module")
+def run_scanner(tmp_path_factory):
+    """Run the frontend's markdown scanners under node.
+
+    Their job is to classify a line the way a CommonMark renderer would, which
+    only a real run can show. The sources are copied with their "@/lib" aliases
+    rewritten, because that alias resolves through Vite and not through node."""
+    node = shutil.which("node")
+    if node is None:
+        pytest.skip("node is needed to run the TypeScript scanners")
+    work = tmp_path_factory.mktemp("release-notes-scanners")
+    for source in (PREVIEW, CODE_SPANS, LINKS, LIST_COLUMNS, INLINE_COMMENTS):
+        rewritten = _TS_ALIAS.sub(r'"./\1.ts"', source.read_text(encoding="utf-8"))
+        (work / source.name).write_text(rewritten, encoding="utf-8")
+    (work / "run.ts").write_text(_TS_RUNNER, encoding="utf-8")
+
+    def run(kind: str, markdown: str):
+        result = subprocess.run(
+            [node, "--experimental-strip-types", "--no-warnings", str(work / "run.ts"), kind],
+            input=markdown,
+            capture_output=True,
+            text=True,
+        )
+        if result.returncode != 0:
+            pytest.skip(f"node could not run the scanners: {result.stderr.strip()[:200]}")
+        return json.loads(result.stdout)
+
+    return run
+
+
+def preview_leads(preview) -> list[str]:
+    return [item["lead"] for item in preview["items"]]
+
+
+def test_a_link_indented_under_a_bullet_still_resolves(run_scanner):
+    """CommonMark measures indentation from the container, not the margin
+    (spec 0.31.2 section 5.2, list items). Under "- Details:" the content column
+    is 2, so a four-space line is only two columns in: a paragraph holding a
+    link, which GitHub renders and follows. The scanner measured from the margin
+    instead, called it an indented code block (section 4.4) and left the
+    destination relative, so the link resolved against Studio's own origin."""
+    resolved = run_scanner("links", "- Details:\n\n    [guide](docs/a.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/a.md" in resolved
+    # The same prose one column further in really is code, and stays untouched.
+    code = run_scanner("links", "- Added.\n\n      [guide](docs/a.md)\n")
+    assert "[guide](docs/a.md)" in code and "github.com" not in code
+    # At document level four spaces is code, so that link is still left alone.
+    top = run_scanner("links", "Intro.\n\n    [guide](docs/a.md)\n")
+    assert "[guide](docs/a.md)" in top and "github.com" not in top
+
+
+def test_an_indented_fence_does_not_swallow_the_bullets_below_it(run_scanner):
+    """A four-space line at document level is an indented code block, and a
+    top-level bullet is not indented enough to continue it, so the block ends
+    and the list renders. Promoting the line to a list-contained fence left a
+    block open with no closer, so every bullet after it was skipped and the
+    collapsed popup lost its summary."""
+    swallowed = "Example:\n\n    ```\n\n- Added the exporter\n- Fixed the crash\n"
+    assert preview_leads(run_scanner("preview", swallowed)) == [
+        "Added the exporter",
+        "Fixed the crash",
+    ]
+    # With nothing else to fall back on the summary disappeared entirely.
+    assert preview_leads(run_scanner("preview", "    ```\n\n- Added the exporter\n")) == [
+        "Added the exporter"
+    ]
+    # A fence that really is inside an item still hides that item's code.
+    nested = "- a\n  - b\n    ```\n    - not a bullet\n    ```\n\n- Added tests\n"
+    assert preview_leads(run_scanner("preview", nested)) == ["a", "Added tests"]
+
+
+def test_a_table_only_release_previews_as_nothing(run_scanner):
+    """A release written as a GFM table renders as a grid, and the panel treats
+    notes that preview as nothing by staying collapsed rather than showing an
+    empty strip. Falling through to the prose collector put the raw
+    "| Change | Detail | | --- | --- |" delimiters in the popup instead."""
+    table = "| Change | Detail |\n| --- | --- |\n| Exporter | Added GGUF |\n"
+    assert run_scanner("preview", table)["items"] == []
+    # A table after prose is dropped too, rather than joined onto it.
+    assert preview_leads(run_scanner("preview", f"Some prose.\n\n{table}")) == ["Some prose."]
+    # A bullet right after the rows ends the table, so it still previews.
+    assert preview_leads(run_scanner("preview", f"{table}- Added tests\n")) == ["Added tests"]
+    # Mismatched header and delimiter widths are no table, as on GitHub, so both lines are prose.
+    assert preview_leads(run_scanner("preview", "| a | b |\n| --- |\n")) == ["| a | b | | --- |"]
+
+
+def test_a_fence_inside_a_list_item_ends_with_the_item(changelog_module):
+    """A fence is scoped to its container: with no closer it runs to the end of
+    the containing block, not the document (spec 0.31.2 section 4.5). A
+    dedented "## 2.0" closes the list item, so it is a real release heading.
+    Document-wide fence state kept the block open and hid every release below
+    it, so one missing closing line emptied the rest of the changelog."""
+    text = "## 1.0\n\n- item\n  ```\n\n## 2.0\n\n- two\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0", "2.0"]
+    # A fence at document level still runs to the end of the file.
+    top = "## 1.0\n\n```\n\n## 2.0\n\n- two\n"
+    assert [e.version for e in changelog_module.parse_changelog(top)] == ["1.0"]
+    # A closed fence inside an item is unaffected, and its sample stays hidden.
+    closed = "## 1.0\n\n- Run:\n  ```bash\n  ## 9.9.9\n  ```\n\n## 2.0\n\n- two\n"
+    assert [e.version for e in changelog_module.parse_changelog(closed)] == ["1.0", "2.0"]
+    # Content dedented out of the item ends the item and the fence with it.
+    assert changelog_module.find_release_notes(text, "2.0").body == "- two"
+
+
+def test_stripping_comments_stays_linear_in_the_code_spans(changelog_module):
+    """The comment scanner restarted its code-span search at the first span for
+    every opener, so a line of N spans and N openers cost N squared. A 203 KiB
+    line is well inside the 2 MiB the fetcher accepts, and notes are reparsed on
+    every request, so one such line held a worker for over ten seconds."""
+    line = "`a` <!--x--> " * 16_000
+    assert len(line) < changelog_module.CHANGELOG_MAX_BYTES
+    started = time.monotonic()
+    visible, in_comment = changelog_module._strip_comments(line, False, False)
+    elapsed = time.monotonic() - started
+    # Roughly 40ms scanning forward against roughly 11s restarting each time.
+    assert elapsed < 2.0, f"comment stripping took {elapsed:.1f}s"
+    # Same result as before: the spans survive and the comments are gone.
+    assert in_comment is False
+    assert "<!--" not in visible and visible.count("`a`") == 16_000
+
+
+def test_the_three_scanners_share_one_list_column_rule():
+    """The parser and both frontend scanners have to classify a line the same
+    way, and drifting apart on indentation is what put a paragraph link inside a
+    code block. The frontend pair reads its list columns from one module, ported
+    from the backend's own tracker."""
+    shared = LIST_COLUMNS.read_text(encoding="utf-8")
+    assert "export function openLists(" in shared
+    assert "_open_lists" in shared, "the backend function this mirrors"
+    for source in (PREVIEW, LINKS):
+        src = source.read_text(encoding="utf-8")
+        assert 'from "@/lib/markdown-list-columns"' in src
+        assert "openLists(" in src
+    # Both sides measure indented code from the container, not from the margin.
+    backend = (BACKEND / "utils" / "changelog.py").read_text(encoding="utf-8")
+    assert "_indent_width(visible) - column >= 4" in backend
+    assert "indentWidth(structure) - column >= INDENTED_CODE_INDENT" in LINKS.read_text(
+        encoding="utf-8"
+    )
+
+
+def test_a_failed_fetch_keeps_retry_reachable():
+    """The fallback stands in for "no section for this version", which the hook
+    reports as ready. A failed fetch is reported as error and is retryable, and on
+    desktop the fallback is the updater's static install blurb, so taking it there
+    replaced the Retry button with generic text until the cache expired."""
+    src = " ".join(PANEL.read_text(encoding="utf-8").split())
+    assert 'notes?.matched ? notes.markdown : state === "error" ? null' in src
+    # Only NotesStatus renders retry, in the else of the markdown branch: an error has no markdown.
+    assert "{markdown ? (" in src
+    assert "retry={retry}" in src
+
+    hook = " ".join(
+        (FRONTEND / "hooks" / "use-release-notes.ts").read_text(encoding="utf-8").split()
+    )
+    assert (
+        "const failed = !next || (!next.matched && next.error !== null);" in hook
+    ), "the distinction this relies on"
+
+
+def test_an_unclosed_comment_in_prose_cannot_hide_later_links(run_scanner):
+    """CommonMark opens an HTML block (spec 0.31.2 section 4.6, type 2) only
+    when the line itself begins with `<!--`; one written mid-sentence is inline
+    raw HTML and cannot outlive the block it sits in. The link resolver carried
+    the unclosed state to every following line instead, so a note that merely
+    mentions the delimiter masked the relative links under it and they resolved
+    against Studio's own origin."""
+    repo = "https://github.com/unslothai/unsloth/blob/main/docs/a.md"
+    # A separate list item is a separate block, so the link below still renders.
+    item = run_scanner("links", "- Type <!-- to begin a comment\n- See [docs](docs/a.md)\n")
+    assert repo in item
+    # So does a paragraph the blank line already ended.
+    paragraph = run_scanner("links", "Type <!-- to begin\n\nSee [docs](docs/a.md)\n")
+    assert repo in paragraph
+    # A delimiter inside inline code is literal, as it is for the parser.
+    spanned = run_scanner("links", "- Wrap in `<!--` and `-->`\n- See [docs](docs/a.md)\n")
+    assert repo in spanned
+    # A comment starting a line is a block: it hides down to the closer's line, that line included.
+    block = run_scanner("links", "<!-- staged\n- See [docs](docs/a.md)\n-->\n")
+    assert repo not in block
+    closer = run_scanner("links", "<!-- staged\n--> See [docs](docs/a.md)\n")
+    assert repo not in closer
+
+
+def test_a_bare_level_two_marker_ends_the_release(changelog_module, run_scanner):
+    """An ATX heading's opening sequence may be followed by the end of the line
+    (spec 0.31.2 section 4.2), so a bare `##` is an empty level-two heading. The
+    scanners required whitespace after the hashes, so everything below such a
+    line stayed inside the release above it and the popup showed unrelated notes
+    under that version."""
+    text = "## 2.0\n\n- new thing\n\n##\n\n- SECRET: not part of 2.0\n"
+    entry = changelog_module.find_release_notes(text, "2.0")
+    assert "new thing" in entry.body
+    assert "SECRET" not in entry.body
+    # An empty heading has no version, so it ends a release without indexing one.
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["2.0"]
+    # Prose still needs a space or a tab: `##x` is a paragraph, not a heading.
+    prose = "## 2.0\n\n- new thing\n\n##x\n\n- still 2.0\n"
+    assert "still 2.0" in changelog_module.find_release_notes(prose, "2.0").body
+    # The preview agrees: an empty heading renders as nothing, so it ends the bullet.
+    preview = run_scanner("preview", "- new thing\n##\nUnrelated scratch notes\n")
+    assert preview_leads(preview) == ["new thing"]
+
+
+def test_a_comment_between_bullets_closes_the_list(changelog_module, run_scanner):
+    """A comment is an HTML block (spec 0.31.2 section 4.6, type 2), so one
+    written at the margin under a bullet is not indented enough to continue that
+    item and closes the list. The scanners blanked the line before list tracking
+    saw it, which reads as a blank line and leaves the item open, so the release
+    heading below it looked like nested item content and the new release was
+    merged into the one above."""
+    text = "## 1.0\n\n- old item\n<!-- separator -->\n  ## 2.0\n\n- new item\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0", "2.0"]
+    assert "new item" not in changelog_module.find_release_notes(text, "1.0").body
+    assert "new item" in changelog_module.find_release_notes(text, "2.0").body
+    # At the item's content column the comment stays inside it, so the heading under it is nested.
+    nested = "## 1.0\n\n- old item\n  <!-- separator -->\n  ## 2.0\n\n- new item\n"
+    assert [e.version for e in changelog_module.parse_changelog(nested)] == ["1.0"]
+    # The link resolver reads the same column: list closed, four spaces is code, left untouched.
+    code = run_scanner("links", "- old item\n<!-- separator -->\n    [guide](docs/a.md)\n")
+    assert "[guide](docs/a.md)" in code and "github.com" not in code
+    # Inside the item those four spaces are two columns in, so it is prose and the link resolves.
+    prose = run_scanner("links", "- old item\n  <!-- separator -->\n    [guide](docs/a.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/a.md" in prose
+    # The preview agrees: the fence is indented code, not a fence swallowing the bullet below.
+    preview = run_scanner(
+        "preview",
+        "- Details:\n<!-- separator -->\n    ```\n    - hidden sample\n- Real second item\n",
+    )
+    assert preview_leads(preview) == ["Details:", "Real second item"]
+
+
+def test_a_parenthesised_link_destination_still_resolves(run_scanner):
+    """A destination may hold parentheses while they balance (spec 0.31.2
+    section 6.3), so `[x]((draft).md)` points at `(draft).md`. The resolver's
+    destination expression stopped at the first paren, matched an empty
+    destination and left the markdown alone, so the link resolved against
+    Studio's own origin instead of the repository."""
+    leading = run_scanner("links", "[details]((draft).md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/(draft).md" in leading
+    # An image resolves against the raw host the same way.
+    image = run_scanner("links", "![shield]((badge).png)\n")
+    assert "https://raw.githubusercontent.com/unslothai/unsloth/main/(badge).png" in image
+    # A pair in the middle of a path balances too.
+    middle = run_scanner("links", "[api](docs/(v2)/api.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/(v2)/api.md" in middle
+    # An unbalanced paren makes the destination invalid, so `[x](a(b.md)` is plain text, not a link.
+    unbalanced = run_scanner("links", "[x](a(b.md)\n")
+    assert unbalanced == "[x](a(b.md)\n"
+    # One more closer balances the pair, and then it is a link again.
+    closed = run_scanner("links", "[x](a(b.md))\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/a(b.md)" in closed
+    # Pairs nest, and one level was all the expression allowed, so a path with two stayed relative.
+    nested = run_scanner("links", "[x](((draft)).md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/((draft)).md" in nested
+    deep = run_scanner("links", "![shot](((((v2))))).png)\n")
+    assert "https://raw.githubusercontent.com/unslothai/unsloth/main/((((v2))))" in deep
+    # The closer must still be there: an unbalanced run below a nested pair is not a link.
+    across = run_scanner("links", "[x](((a).md\n[y](docs/y.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/y.md" in across
+    assert "[x](((a).md" in across
+
+
+def test_a_fence_inside_a_container_still_hides_its_sample(run_scanner):
+    """A fence is measured from its container and not from the margin (spec
+    0.31.2 section 4.5), so `> ~~~` and a fence three columns under a nested
+    bullet open one. Reading the margin instead never saw them, so the sample
+    inside was treated as prose and a relative link written in a code block was
+    rewritten into the text the reader sees verbatim."""
+    quoted = run_scanner("links", "> ~~~\n> [guide](docs/a.md)\n> ~~~\n")
+    assert "[guide](docs/a.md)" in quoted and "github.com" not in quoted
+    nested = run_scanner("links", "- a\n  - b\n    ~~~\n    [x](docs/x.md)\n    ~~~\n")
+    assert "[x](docs/x.md)" in nested and "github.com" not in nested
+    # A longer closer is still a closer, so the pair is not something a code span hid.
+    uneven = run_scanner("links", "> ```\n> [guide](docs/a.md)\n> ````\n")
+    assert "[guide](docs/a.md)" in uneven and "github.com" not in uneven
+    # The fence ends with its container: a line outside the quote, or left of the item, is Markdown.
+    left = run_scanner("links", "> ~~~\n[guide](docs/a.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/a.md" in left
+    dedented = run_scanner("links", "- a\n  ~~~\n[guide](docs/a.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/a.md" in dedented
+    # A document-level fence owns the quoted lines below, so the marker does not undo it.
+    document = run_scanner("links", "~~~\n> [guide](docs/a.md)\n~~~\n")
+    assert "[guide](docs/a.md)" in document and "github.com" not in document
+    # Four columns past the item's content column it is indented code, not a fence: still literal.
+    code = run_scanner("links", "- Details:\n\n      ~~~\n      [guide](docs/a.md)\n")
+    assert "[guide](docs/a.md)" in code and "github.com" not in code
+
+
+def test_an_html_block_inside_a_container_is_literal_too(run_scanner):
+    """Type 1 and type 6 blocks are measured from their container the same way,
+    so a `<details>` under a nested bullet and a `<pre>` inside a quote both
+    show their contents verbatim. Missing the opener treated the body as
+    Markdown and rewrote the literal examples in it."""
+    nested = run_scanner("links", "- a\n  - b\n    <details>\n    [x](docs/x.md)\n    </details>\n")
+    assert "[x](docs/x.md)" in nested and "github.com" not in nested
+    quoted = run_scanner("links", "> <pre>\n> [x](docs/x.md)\n> </pre>\n")
+    assert "[x](docs/x.md)" in quoted and "github.com" not in quoted
+    # The block ends with its container, so a line dedented out of the item is Markdown again.
+    dedented = run_scanner("links", "- a\n  - b\n    <details>\n[x](docs/x.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/x.md" in dedented
+    # Inside a quote a bare marker holds nothing, the blank line that ends a type 6 block.
+    blank = run_scanner("links", "> <details>\n>\n> [x](docs/x.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/x.md" in blank
+
+
+def test_an_underline_left_of_an_item_is_lazy_text_of_it(changelog_module, run_scanner):
+    """A setext underline may never be a lazy continuation line (spec 0.31.2
+    section 4.3), so `===` written left of an open list item is read as more of
+    the item's paragraph rather than as a block that closes it. Rejecting every
+    underline-shaped line ended the list there, which promoted the nested
+    "## 2.0" below it to a document-level heading and indexed a release the
+    renderer never shows."""
+    nested = "## 1.0\n- old note\n===\n  ## 2.0\n- new\n"
+    assert [e.version for e in changelog_module.parse_changelog(nested)] == ["1.0"]
+    # A row of dashes is a thematic break, closing the item, so the heading is the next release.
+    broken = "## 1.0\n- old note\n---\n  ## 2.0\n"
+    assert [e.version for e in changelog_module.parse_changelog(broken)] == ["1.0", "2.0"]
+    # With no paragraph above it the underline opens one, so the blank line closes the item.
+    apart = "## 1.0\n- old note\n\n===\n  ## 2.0\n"
+    assert [e.version for e in changelog_module.parse_changelog(apart)] == ["1.0", "2.0"]
+    # The link scanner keeps the item open, so the four-space line is a paragraph and resolves.
+    resolved = run_scanner("links", "- Details:\n===\n\n    [guide](docs/a.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/a.md" in resolved
+
+
+def test_a_quote_keeps_its_paragraph_to_itself(changelog_module, run_scanner):
+    """Lazy continuation runs the other way too: a marker written outside a
+    blockquote is not text of the quote's paragraph, so `2. item` under
+    `> quote` opens a list even though an ordered marker past 1 may not
+    interrupt a paragraph (spec 0.31.2 section 5.2). Lending the quote's
+    paragraph to the document left the list closed, so the heading indented to
+    the item's content column read as a release of its own."""
+    quoted = "## 1.0\n> quote\n2. item\n   ## 2.0\n- new\n"
+    assert [e.version for e in changelog_module.parse_changelog(quoted)] == ["1.0"]
+    # A quote holding a heading leaves no paragraph, nor does an empty one, so the list opens.
+    heading = "## 1.0\n> # inner\n2. item\n   ## 2.0\n"
+    assert [e.version for e in changelog_module.parse_changelog(heading)] == ["1.0"]
+    # An unquoted line the quote's paragraph swallows keeps it open, the marker still outside.
+    lazy = "## 1.0\n> quote\ntext\n2. item\n   ## 2.0\n"
+    assert [e.version for e in changelog_module.parse_changelog(lazy)] == ["1.0"]
+    # Under an ordinary paragraph the marker is its text, so no list opens and the heading is real.
+    prose = "## 1.0\nprose\n2. item\n   ## 2.0\n"
+    assert [e.version for e in changelog_module.parse_changelog(prose)] == ["1.0", "2.0"]
+    # The preview reads the marker as a bullet for the same reason.
+    assert preview_leads(run_scanner("preview", "> quote\n2. item\n")) == ["item"]
+
+
+def test_indented_code_before_an_ordered_marker_still_opens_a_list(changelog_module):
+    """An indented code block ends at the first line that is not indented enough
+    to continue it, and no paragraph is open for the marker below to continue,
+    so `2. item` opens a list whatever its start number. Reading it as text of
+    the code block instead would leave the list closed and index the heading at
+    the item's content column as a release."""
+    joined = "## 1.0\n\n    code\n2. item\n   ## 2.0\n- new\n"
+    assert [e.version for e in changelog_module.parse_changelog(joined)] == ["1.0"]
+    # A blank line between the two changes nothing: the list opens either way.
+    apart = "## 1.0\n\n    code\n\n2. item\n   ## 2.0\n- new\n"
+    assert [e.version for e in changelog_module.parse_changelog(apart)] == ["1.0"]
+    # Four columns past its container the marker is code, so no list opens and the heading stands.
+    inside = "## 1.0\n\n    code\n    - item\n  ## 2.0\n"
+    assert [e.version for e in changelog_module.parse_changelog(inside)] == ["1.0", "2.0"]
+
+
+def test_a_fence_written_as_an_item_first_content_opens_in_that_item(run_scanner):
+    """A block written straight after a list marker is the item's own first
+    content, measured from the column that content starts (spec 0.31.2 section
+    5.2), so "- ```md" opens a fence. Reading the whole line instead never saw
+    one, so the code sample below it was treated as prose: the resolver rewrote
+    a destination the reader sees verbatim, and the preview offered the info
+    string as a headline bullet."""
+    sample = run_scanner("links", "- ```md\n  [example](docs/a.md)\n  ```\n")
+    assert "[example](docs/a.md)" in sample and "github.com" not in sample
+    ordered = run_scanner("links", "1. ~~~\n   [example](docs/a.md)\n   ~~~\n")
+    assert "[example](docs/a.md)" in ordered and "github.com" not in ordered
+    # The preview agrees: an item of only a code block previews as nothing; the next is a bullet.
+    preview = run_scanner("preview", "- ```md\n  sample text\n  ```\n- Added tests\n")
+    assert preview_leads(preview) == ["Added tests"]
+    # One column further in it is indented code inside the item, so the link is prose and resolves.
+    padded = run_scanner("links", "-     ```\n  [example](docs/a.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/a.md" in padded
+    # A marker the paragraph above swallows opens no item, so no fence: ordered items open at 1.
+    lazy = run_scanner("links", "Intro.\n2. ```\n[guide](docs/a.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/a.md" in lazy
+
+
+def test_an_html_block_ends_with_the_item_it_was_written_in(changelog_module, run_scanner):
+    """An HTML block holds no lazy continuation line, so one opened on a list
+    item's continuation line ends where the item does, exactly as a fence there
+    does. Ending it only on a blank line let it run past the item and swallow
+    the next release heading, so those notes could never be found, and the
+    collapsed preview lost every bullet below it."""
+    text = "## 1.0\n\n- item\n\n  <div>\n## 2.0\n\n- new thing\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0", "2.0"]
+    assert "new thing" in changelog_module.find_release_notes(text, "2.0").body
+    # A raw block such as <pre> is scoped the same way.
+    raw = "## 1.0\n\n- item\n\n  <pre>\n## 2.0\n\n- new thing\n"
+    assert [e.version for e in changelog_module.parse_changelog(raw)] == ["1.0", "2.0"]
+    # At the item's content column the block holds the heading, which is nested and indexes nothing.
+    nested = "## 1.0\n\n- item\n\n  <div>\n  ## 2.0\n"
+    assert [e.version for e in changelog_module.parse_changelog(nested)] == ["1.0"]
+    # The preview reads it the same way: the bullet below the block is a bullet.
+    preview = run_scanner("preview", "- item\n\n  <div>\n- Added tests\n")
+    assert preview_leads(preview) == ["item", "Added tests"]
+    # An opener straight after a marker opens in that item, so the dedented heading is a release.
+    marked = "## 1.0\n\n- <div>\n## 2.0\n\n- new thing\n"
+    assert [e.version for e in changelog_module.parse_changelog(marked)] == ["1.0", "2.0"]
+
+
+def test_a_comment_may_close_on_a_later_line_of_its_paragraph(run_scanner):
+    """A comment written mid-sentence is inline raw HTML belonging to the
+    paragraph around it, so its `-->` may arrive on a later line of that same
+    paragraph and everything between renders as nothing. Ending the comment at
+    its own line left a backtick inside it pairing with a real one below, which
+    hid a following link from the resolver, and left the collapsed preview
+    quoting text the popup body does not show."""
+    carried = run_scanner("links", "Note <!-- ` open\nstill --> see [d](docs/a.md) and `x`\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/a.md" in carried
+    # Text inside the comment renders as nothing, so it is left alone.
+    inside = run_scanner("links", "Note <!-- see [c](docs/c.md)\nmore --> end\n")
+    assert "[c](docs/c.md)" in inside and "github.com" not in inside
+    # The preview hides it too, rather than quoting the comment at the reader.
+    preview = run_scanner(
+        "preview", "- Added X <!-- TODO: rewrite\n  this properly -->\n- Second\n"
+    )
+    assert preview_leads(preview) == ["Added X", "Second"]
+    # An opener cannot outlive its paragraph: with it closed the `<!--` is text and hides nothing.
+    broken = run_scanner("links", "Note <!-- open\n\nsecret --> end [d](docs/a.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/a.md" in broken
+    # A heading breaks into the paragraph, so it ends the comment's reach too.
+    headed = run_scanner("links", "Note <!-- open\n## 2.0 --> end [d](docs/a.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/a.md" in headed
+    assert preview_leads(run_scanner("preview", "Note <!-- open\n\n- Second\n")) == ["Second"]
+
+
+def test_only_punctuation_is_escapable_in_a_link_destination(run_scanner):
+    """CommonMark escapes ASCII punctuation and nothing else (spec 0.31.2
+    section 2.4), so the backslash in `docs\\alpha.md` is a character of the
+    path. Dropping every backslash rewrote it to a path that does not exist,
+    and a URL parser reads what is left as a separator, so a Windows or
+    namespaced path pointed at the wrong file either way."""
+    kept = run_scanner("links", "[guide](docs\\alpha.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs%5Calpha.md" in kept
+    # An escaped backslash is one literal backslash, which survives the same.
+    escaped = run_scanner("links", "[guide](docs\\\\alpha.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs%5Calpha.md" in escaped
+    # A real escape is still an escape: `\\(` is a paren of the path.
+    paren = run_scanner("links", "[guide](a\\(b.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/a(b.md" in paren
+    # A space still ends the destination, escaped or not, so there is no link.
+    spaced = run_scanner("links", "[guide](a\\ b.md)\n")
+    assert spaced == "[guide](a\\ b.md)\n"
+
+
+def test_one_definition_does_not_hide_the_next(run_scanner):
+    """Definitions may run consecutively (spec 0.31.2 section 4.7): a block of
+    them is how a changelog collects its link targets. A definition is not
+    paragraph text, so it opens no paragraph for the next one to be unable to
+    interrupt. The resolver counted one as prose, which left every definition
+    after the first outside the set of lines a definition may start on, so only
+    the first was rewritten and the rest resolved against Studio's own origin.
+    The backend already reads the line this way."""
+    text = (
+        "- AMD support is here, see [the AMD guide][amd] and the\n"
+        "  [Intel notes][xpu].\n\n"
+        "[amd]: docs/basics/amd.md\n"
+        "[xpu]: docs/basics/xpu.md\n"
+    )
+    resolved = run_scanner("links", text)
+    base = "https://github.com/unslothai/unsloth/blob/main/docs/basics/"
+    assert f"[amd]: {base}amd.md" in resolved
+    assert f"[xpu]: {base}xpu.md" in resolved
+    # A run of them stays a run however long it is.
+    run = run_scanner("links", "[a]: docs/a.md\n[b]: docs/b.md\n[c]: docs/c.md\n")
+    assert run.count("https://github.com/unslothai/unsloth/blob/main/docs/") == 3
+    # Prose between them opens a paragraph the next line may not interrupt, so it is not one.
+    prose = run_scanner("links", "[a]: docs/a.md\nintro\n[b]: docs/b.md\n")
+    assert "[b]: docs/b.md" in prose
+
+
+def test_a_comment_closed_on_its_own_line_still_closes(run_scanner):
+    """A multiline comment is ordinarily closed by a `-->` written on a line of
+    its own, and a wrapped line may open with emphasis. The guard asking whether
+    the closer is reachable read any line whose first character was punctuation
+    as the start of a new block, so neither shape counted as more of the
+    paragraph carrying the comment. The comment then never closed, and the
+    collapsed popup showed the author's internal note to the user."""
+    closer = run_scanner(
+        "preview",
+        "- DoRA training is available in Studio. <!-- TODO confirm the exact\n"
+        "  flag name before release\n-->\n",
+    )
+    assert preview_leads(closer) == ["DoRA training is available in Studio."]
+    # A continuation may open with emphasis, which is text and not a block.
+    starred = run_scanner(
+        "preview",
+        "- DoRA training is available. <!-- TODO confirm the\n  *before* release -->\n",
+    )
+    assert preview_leads(starred) == ["DoRA training is available."]
+    underscored = run_scanner(
+        "preview",
+        "- DoRA training is available. <!-- TODO confirm the\n  _draft_ note -->\n",
+    )
+    assert preview_leads(underscored) == ["DoRA training is available."]
+    # A real block still ends the paragraph, so the opener below one is text and hides nothing.
+    broken = run_scanner("links", "Note <!-- open\n## 2.0\nsecret --> [d](docs/a.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/a.md" in broken
+    # So does a list item with content, which may interrupt a paragraph.
+    item = run_scanner("links", "Note <!-- open\n- bullet\nsecret --> [d](docs/a.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/a.md" in item
+
+
+def test_a_comment_written_as_an_item_first_content_is_a_block(changelog_module, run_scanner):
+    """A comment is an HTML block (spec 0.31.2 section 4.6, type 2), so one
+    written as a list item's first content opens inside that item, exactly as a
+    fence written there does. The scanners looked for the opener at the margin
+    of the line as written, so a marker in front of it hid the block: the
+    resolver rewrote a destination inside raw HTML, which Streamdown then shows
+    the reader as a literal URL, and the preview quoted the hidden note back at
+    them as though the bullet were Markdown."""
+    item = run_scanner("links", "- <!-- new --> AMD support, see [the guide](docs/amd.md)\n")
+    assert item == "- <!-- new --> AMD support, see [the guide](docs/amd.md)\n"
+    # Every marker opens an item, and a nested one is still an item.
+    for text in (
+        "* <!-- new --> see [the guide](docs/amd.md)\n",
+        "1. <!-- new --> see [the guide](docs/amd.md)\n",
+        "- outer\n  - <!-- new --> see [the guide](docs/amd.md)\n",
+    ):
+        assert "github.com" not in run_scanner("links", text)
+    # The multiline form hides lines to the closer, as a comment at the item's content column did.
+    multiline = run_scanner("links", "- <!-- hidden\n  [a](docs/x.md)\n  -->\n")
+    assert "[a](docs/x.md)" in multiline and "github.com" not in multiline
+    # Still scoped to the item it was written in, so a line dedented out of it ends the block.
+    dedented = run_scanner("links", "- <!-- hidden\n[a](docs/x.md)\n")
+    assert "https://github.com/unslothai/unsloth/blob/main/docs/x.md" in dedented
+    # The preview agrees: an item of only the block previews as nothing; the next is a bullet.
+    preview = run_scanner("preview", "- <!-- new --> hidden note\n- Real bullet\n")
+    assert preview_leads(preview) == ["Real bullet"]
+    # The parser agrees too: the item keeps its column, so a heading inside is nested, not indexed.
+    text = "## 1.0\n\n- <!-- hidden\n\n  ## 2.0\n"
+    assert [e.version for e in changelog_module.parse_changelog(text)] == ["1.0"]
diff --git a/tests/test_mapper_no_duplicate_keys.py b/tests/test_mapper_no_duplicate_keys.py
new file mode 100644
index 0000000000..42e49415fd
--- /dev/null
+++ b/tests/test_mapper_no_duplicate_keys.py
@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Guard against duplicate keys in the ``__INT_TO_FLOAT_MAPPER`` registry.
+
+Duplicate keys in the dict literal silently overwrite earlier entries.
+We inspect the source with ``ast`` to ensure there are no duplicates.
+"""
+
+import ast
+import os
+
+MAPPER_PATH = os.path.join(os.path.dirname(__file__), os.pardir, "unsloth", "models", "mapper.py")
+
+
+def _duplicate_int_to_float_keys():
+    with open(MAPPER_PATH, encoding = "utf-8") as f:
+        tree = ast.parse(f.read(), MAPPER_PATH)
+
+    for node in ast.walk(tree):
+        if not isinstance(node, ast.Assign):
+            continue
+        for target in node.targets:
+            # Private names are mangled at code generation, which ``ast.parse``
+            # never reaches, so the identifier reads exactly as written.
+            if isinstance(target, ast.Name) and target.id == "__INT_TO_FLOAT_MAPPER":
+                if not isinstance(node.value, ast.Dict):
+                    continue
+                # mapper.py reads the nested per-precision dicts directly, so
+                # check every dict. Count per dict: "16" and "8" legitimately
+                # repeat across sibling entries.
+                duplicates = {}
+                for mapping in ast.walk(node.value):
+                    if not isinstance(mapping, ast.Dict):
+                        continue
+                    seen = set()
+                    for k in mapping.keys:
+                        if not (isinstance(k, ast.Constant) and isinstance(k.value, str)):
+                            continue
+                        if k.value in seen:
+                            duplicates.setdefault(k.value, []).append(k.lineno)
+                        seen.add(k.value)
+                return duplicates
+    raise AssertionError("Could not find the __INT_TO_FLOAT_MAPPER dict literal in mapper.py")
+
+
+def test_int_to_float_mapper_has_no_duplicate_keys():
+    duplicates = _duplicate_int_to_float_keys()
+    assert not duplicates, (
+        "Duplicate keys in __INT_TO_FLOAT_MAPPER silently overwrite earlier "
+        "entries and corrupt model resolution. Remove the redundant "
+        f"literal(s), key -> line number(s) in mapper.py: {duplicates}"
+    )
diff --git a/tests/test_profile_startup_gate.py b/tests/test_profile_startup_gate.py
new file mode 100644
index 0000000000..66e6e16a89
--- /dev/null
+++ b/tests/test_profile_startup_gate.py
@@ -0,0 +1,243 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""Regression coverage for the startup profiler's budget gate, teardown and triggers."""
+
+from __future__ import annotations
+
+import ast
+import fnmatch
+import importlib.util
+import re
+import subprocess
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+import yaml
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+SCRIPT = REPO_ROOT / "scripts" / "profile_startup.py"
+WORKFLOW = REPO_ROOT / ".github" / "workflows" / "startup-profile-ci.yml"
+PROCESS_RS = REPO_ROOT / "studio" / "src-tauri" / "src" / "process.rs"
+
+# Checkout files that build the venv the workflow profiles.
+INSTALLER_INPUTS = (
+    "studio/setup.sh",
+    "studio/setup.ps1",
+    "studio/install_python_stack.py",
+)
+# Checkout file that defines the argv the profiler reproduces.
+LAUNCH_INPUTS = ("studio/src-tauri/src/process.rs",)
+
+
+def _load():
+    spec = importlib.util.spec_from_file_location("profile_startup", SCRIPT)
+    mod = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(mod)
+    return mod
+
+
+def _no_subprocesses(mod, monkeypatch):
+    # Keep the gate tests off the real interpreter and CLI.
+    monkeypatch.setattr(mod, "find_bin", lambda: None)
+    monkeypatch.setattr(mod, "profile_imports", lambda python, top = 15: {"ok": False, "error": ""})
+    monkeypatch.setattr(mod, "python_version_of", lambda python: "3.13.0")
+
+
+class _Proc:
+    """Stand-in for a still-running Popen."""
+
+    def __init__(self):
+        self.pid = 4321
+        self.terminated = False
+
+    def poll(self):
+        return None
+
+    def terminate(self):
+        self.terminated = True
+
+
+def _nt(mod, monkeypatch, returncode):
+    calls: list[list[str]] = []
+
+    def _run(argv, **kwargs):
+        calls.append(argv)
+        return subprocess.CompletedProcess(argv, returncode, "", "")
+
+    # Patch the module's own references, not the real os/subprocess the session shares.
+    monkeypatch.setattr(mod, "os", SimpleNamespace(name = "nt"))
+    monkeypatch.setattr(mod, "subprocess", SimpleNamespace(run = _run))
+    return calls
+
+
+def test_budget_fails_when_no_launch_was_measured(capsys, monkeypatch):
+    """A requested budget must not pass just because the CLI was never found."""
+    mod = _load()
+    _no_subprocesses(mod, monkeypatch)
+    rc = mod.main(["--max-healthz-seconds", "30"])
+    out = capsys.readouterr().out
+    assert rc == 1
+    assert "::error::" in out and "no healthz measurement" in out
+    assert "no unsloth CLI found" in out
+
+
+def _healthy_launch(
+    mod,
+    monkeypatch,
+    healthz = 1.5,
+):
+    monkeypatch.setattr(mod, "find_bin", lambda: "unsloth")
+    monkeypatch.setattr(
+        mod,
+        "profile_launch",
+        lambda bin_path, port, **kw: {
+            "spawn_seconds": 0.1,
+            "healthz_seconds": healthz,
+            "lifespan_ms": 100.0,
+            "reached_healthz": True,
+            "log_tail": [],
+        },
+    )
+
+
+def test_budget_still_passes_when_a_launch_was_measured(monkeypatch):
+    """The fail-closed branch must not swallow a genuinely healthy run."""
+    mod = _load()
+    _no_subprocesses(mod, monkeypatch)
+    _healthy_launch(mod, monkeypatch)
+    assert mod.main(["--max-healthz-seconds", "30"]) == 0
+    assert mod.main(["--max-healthz-seconds", "1"]) == 1
+
+
+# "=" form for -inf: a bare "-inf" is an option token to argparse, not a value.
+@pytest.mark.parametrize(
+    "bad", ["--max-healthz-seconds=nan", "--max-healthz-seconds=inf", "--max-healthz-seconds=-inf"]
+)
+def test_budget_rejects_non_finite_values(bad, capsys, monkeypatch):
+    """`med > nan` and `med > inf` are always False, so the gate would never bind."""
+    mod = _load()
+    _no_subprocesses(mod, monkeypatch)
+    _healthy_launch(mod, monkeypatch)
+    with pytest.raises(SystemExit) as exc:
+        mod.main([bad])
+    assert exc.value.code == 2
+    assert "finite" in capsys.readouterr().err
+
+
+def test_budget_rejects_import_only(capsys):
+    """--import-only launches nothing, so a budget on it could only ever pass."""
+    mod = _load()
+    with pytest.raises(SystemExit) as exc:
+        mod.main(["--import-only", "--max-healthz-seconds", "30"])
+    assert exc.value.code == 2
+    assert "--import-only" in capsys.readouterr().err
+
+
+def test_terminate_tree_falls_back_when_taskkill_fails(monkeypatch):
+    """A nonzero taskkill must still reach terminate(), not return silently."""
+    mod = _load()
+    calls = _nt(mod, monkeypatch, returncode = 1)
+    proc = _Proc()
+    mod._terminate_tree(proc)
+    assert calls == [["taskkill", "/PID", "4321", "/T", "/F"]]
+    assert proc.terminated
+
+
+def test_terminate_tree_falls_back_when_taskkill_raises(monkeypatch):
+    """A missing or hung taskkill must reach terminate() too."""
+    mod = _load()
+    monkeypatch.setattr(mod, "os", SimpleNamespace(name = "nt"))
+
+    def _boom(argv, **kwargs):
+        raise FileNotFoundError(argv)
+
+    monkeypatch.setattr(mod, "subprocess", SimpleNamespace(run = _boom))
+    proc = _Proc()
+    mod._terminate_tree(proc)
+    assert proc.terminated
+
+
+def test_terminate_tree_returns_on_successful_taskkill(monkeypatch):
+    mod = _load()
+    _nt(mod, monkeypatch, returncode = 0)
+    proc = _Proc()
+    mod._terminate_tree(proc)
+    assert not proc.terminated
+
+
+def test_terminate_tree_skips_an_exited_process(monkeypatch):
+    mod = _load()
+    calls = _nt(mod, monkeypatch, returncode = 0)
+    proc = _Proc()
+    proc.poll = lambda: 0
+    mod._terminate_tree(proc)
+    assert calls == [] and not proc.terminated
+
+
+def _trigger_paths():
+    wf = yaml.safe_load(WORKFLOW.read_text(encoding = "utf-8"))
+    # YAML 1.1 turns the bare `on:` key into True.
+    on = wf.get("on") or wf[True]
+    return [p for p in on["pull_request"]["paths"] if not p.startswith("!")]
+
+
+@pytest.mark.parametrize("rel", INSTALLER_INPUTS)
+def test_workflow_triggers_on_studio_installer_inputs(rel):
+    """A setup script that changes the profiled venv must schedule a measurement."""
+    assert (REPO_ROOT / rel).is_file(), f"{rel} moved; revisit the trigger list"
+    paths = _trigger_paths()
+    assert any(fnmatch.fnmatch(rel, p) for p in paths), f"{rel} not covered by {paths}"
+
+
+def test_studio_installer_inputs_are_on_the_local_install_path():
+    """Anchor the list above: these files are what --local actually executes."""
+    # install.ps1 reaches setup.ps1 through the editable install, not by name.
+    assert "studio/setup.sh" in (REPO_ROOT / "install.sh").read_text(encoding = "utf-8")
+    for setup in ("studio/setup.sh", "studio/setup.ps1"):
+        text = (REPO_ROOT / setup).read_text(encoding = "utf-8", errors = "replace")
+        assert "install_python_stack.py" in text
+
+
+@pytest.mark.parametrize("rel", LAUNCH_INPUTS)
+def test_workflow_triggers_on_the_desktop_launch_command(rel):
+    """The profiler copies process.rs's argv, so a change there must be measured."""
+    assert (REPO_ROOT / rel).is_file(), f"{rel} moved; revisit the trigger list"
+    paths = _trigger_paths()
+    assert any(fnmatch.fnmatch(rel, p) for p in paths), f"{rel} not covered by {paths}"
+
+
+def _desktop_backend_argv():
+    body = re.search(
+        r"fn backend_args\(port: u16\) -> Vec<String> \{(.*?)\n\}",
+        PROCESS_RS.read_text(encoding = "utf-8"),
+        re.S,
+    )
+    assert body, "backend_args moved; revisit the trigger list"
+    return re.findall(r'"([^"]+)"', body.group(1))
+
+
+def _profiler_argv():
+    tree = ast.parse(SCRIPT.read_text(encoding = "utf-8"))
+    fn = next(
+        n for n in ast.walk(tree) if isinstance(n, ast.FunctionDef) and n.name == "profile_launch"
+    )
+    call = next(
+        n for n in ast.walk(fn) if isinstance(n, ast.Call) and ast.unparse(n.func).endswith("Popen")
+    )
+    return [e.value for e in call.args[0].elts if isinstance(e, ast.Constant)]
+
+
+def test_profiler_spawns_the_desktop_backend_argv():
+    """Anchor the trigger above: these two argv lists must stay identical."""
+    assert _profiler_argv() == _desktop_backend_argv()
+
+
+@pytest.mark.skipif(sys.platform == "win32", reason = "posix branch")
+def test_terminate_tree_posix_uses_terminate():
+    mod = _load()
+    proc = _Proc()
+    mod._terminate_tree(proc)
+    assert proc.terminated
diff --git a/unsloth/_gpu_init.py b/unsloth/_gpu_init.py
index 682f3ae6c6..7e8f9ced46 100644
--- a/unsloth/_gpu_init.py
+++ b/unsloth/_gpu_init.py
@@ -303,13 +303,19 @@ if DEVICE_TYPE == "cuda":
     # Try loading bitsandbytes and triton
     try:
         import bitsandbytes as bnb
+
+        # Bind the submodule by name: a half-imported bitsandbytes leaves the parent
+        # without a `functional` attribute, which would otherwise be misreported below
+        # as a CUDA linking failure. See unsloth/kernels/utils.py.
+        import bitsandbytes.functional as bnb_functional
     except:
         print(
             "Unsloth: `bitsandbytes` is not installed - 4bit QLoRA unallowed, but 16bit and full finetuning works!"
         )
         bnb = None
+        bnb_functional = None
     try:
-        cdequantize_blockwise_fp32 = bnb.functional.lib.cdequantize_blockwise_fp32
+        cdequantize_blockwise_fp32 = bnb_functional.lib.cdequantize_blockwise_fp32
         libcuda_dirs()
     except:
         if hasattr(os, "geteuid") and os.geteuid() == 0:
@@ -351,7 +357,7 @@ if DEVICE_TYPE == "cuda":
                         pass
                 else:
                     from triton.common.build import libcuda_dirs
-                cdequantize_blockwise_fp32 = bnb.functional.lib.cdequantize_blockwise_fp32
+                cdequantize_blockwise_fp32 = bnb_functional.lib.cdequantize_blockwise_fp32
                 libcuda_dirs()
             except:
                 warnings.warn(
diff --git a/unsloth/bnb_availability.py b/unsloth/bnb_availability.py
new file mode 100644
index 0000000000..9d14bbb0f3
--- /dev/null
+++ b/unsloth/bnb_availability.py
@@ -0,0 +1,96 @@
+# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Can bitsandbytes actually run a 4bit kernel here? A successful import does not say.
+
+From 0.46 a wheel whose native library never loaded still imports and hands back a
+`throw_on_call` closure for every symbol, so attribute reads alone see a healthy wheel,
+`ALLOW_BITSANDBYTES` stays true and 4bit dies inside a kernel instead of falling back to
+16bit up front. A real handle is a ctypes function pointer and carries `restype`; a
+deferred failure is a plain Python function and does not. That is the whole test, applied
+to every probed handle: the same verdict gates the module-scope binds in kernels/utils.py,
+where one bad symbol is the crash this exists to prevent.
+
+Decides the capability flags only, never importability - a CPU-only install is exactly
+this shape and its Python side works. A leaf module: imports nothing from unsloth
+(device_type.py imports it very early, so anything else is a cycle) and takes the
+device type as an argument.
+"""
+
+__all__ = [
+    "bitsandbytes_symbols",
+    "check_native_kernels",
+    "native_kernels_ready",
+]
+
+# The ctypes handles kernels/utils.py binds at module scope; a test asserts they match.
+_C_SYMBOLS = (
+    "cdequantize_blockwise_fp32",
+    "cdequantize_blockwise_fp16_nf4",
+    "cdequantize_blockwise_bf16_nf4",
+)
+# 4bit inference is a gemv on xpu and a naive gemm elsewhere; probing the wrong pair
+# would write off a perfectly good wheel.
+_C_SYMBOLS_XPU = (
+    "cgemv_4bit_inference_fp16",
+    "cgemv_4bit_inference_bf16",
+)
+_C_SYMBOLS_GEMM = (
+    "cgemm_4bit_inference_naive_fp16",
+    "cgemm_4bit_inference_naive_bf16",
+)
+
+
+def bitsandbytes_symbols(device_type):
+    """Names kernels/utils.py reads off `bitsandbytes.functional.lib`."""
+    tail = _C_SYMBOLS_XPU if device_type == "xpu" else _C_SYMBOLS_GEMM
+    return _C_SYMBOLS + tail
+
+
+def check_native_kernels(bnb, device_type):
+    """Raise unless every handle kernels/utils.py is about to bind is a real kernel.
+
+    All of them: one that resolves here but not at the bind gives back the AttributeError
+    this prevents. Partial export costs 8bit too (`ALLOW_BITSANDBYTES` gates both), but a
+    wheel missing a symbol is a shape no flag makes safe. Safe to repeat - ctypes caches
+    each handle on first lookup, so these are the ones bound later.
+    """
+    if bnb is None:
+        raise ImportError("Unsloth: `bitsandbytes` is not installed.")
+    functional = getattr(bnb, "functional", None)
+    if functional is None:
+        # A part-initialised bitsandbytes leaves the parent without the attribute while
+        # the submodule stays in sys.modules, which `import x.y as z` reads directly.
+        import bitsandbytes.functional as functional
+
+    lib = functional.lib
+    if lib is None:
+        # 0.45.5, the floor in pyproject.toml, on a native-load failure.
+        raise AttributeError("Unsloth: `bitsandbytes.functional.lib` is None.")
+    for symbol in bitsandbytes_symbols(device_type):
+        handle = getattr(lib, symbol)  # AttributeError here is itself a failed check
+        if not hasattr(handle, "restype"):
+            raise AttributeError(
+                f"Unsloth: `bitsandbytes.functional.lib.{symbol}` is not a native "
+                "function pointer - the bitsandbytes native library did not load."
+            )
+
+
+def native_kernels_ready(bnb, device_type):
+    """Is the bitsandbytes native library alive? Gates the flags, never the import."""
+    try:
+        check_native_kernels(bnb, device_type)
+    except Exception:
+        return False
+    return True
diff --git a/unsloth/device_type.py b/unsloth/device_type.py
index 058e166b08..968062c7c1 100644
--- a/unsloth/device_type.py
+++ b/unsloth/device_type.py
@@ -27,6 +27,7 @@ import functools
 import inspect
 import os
 from unsloth_zoo.utils import Version
+from .bnb_availability import native_kernels_ready
 
 
 def is_mlx_available():
@@ -117,17 +118,20 @@ DEVICE_COUNT: int = get_device_count()
 ALLOW_PREQUANTIZED_MODELS: bool = True
 # HSA_STATUS_ERROR_EXCEPTION checks - sometimes AMD fails for BnB
 ALLOW_BITSANDBYTES: bool = True
-# Unusable bitsandbytes on any backend, not just hip: clear the flags the loader
-# reads before it selects a 4bit checkpoint. Same guarded import the fallbacks in
-# _gpu_init.py and kernels/utils.py use rather than a find_spec probe, so an
-# installed-but-broken wheel (missing .so, wrong ROCm/CUDA build) is treated as
-# unavailable by all three, not only by the ones that import it.
+# Unusable bitsandbytes on any backend, not just hip: clear the flags the loader reads
+# before it picks a 4bit checkpoint. A guarded import, not find_spec, since importable
+# is not usable - from 0.46 a dead native library still resolves every ctypes handle to
+# a closure that raises only when called, so 4bit would die mid-run, not fall back here.
 try:
     import bitsandbytes as _bnb_probe
-    del _bnb_probe
 except Exception:
     ALLOW_PREQUANTIZED_MODELS = False
     ALLOW_BITSANDBYTES = False
+else:
+    if not native_kernels_ready(_bnb_probe, DEVICE_TYPE):
+        ALLOW_PREQUANTIZED_MODELS = False
+        ALLOW_BITSANDBYTES = False
+    del _bnb_probe
 # gfx906 (MI50 / Radeon VII / Vega 20): Dynamo/Inductor codegen is broken on this
 # legacy GCN arch (ROCm dropped it after 6.3) - compiled graphs crash or miscompile
 # while the eager path trains fine. Default compile off; setdefault so a user
diff --git a/unsloth/kernels/utils.py b/unsloth/kernels/utils.py
index 2118e65aef..839eb9db84 100644
--- a/unsloth/kernels/utils.py
+++ b/unsloth/kernels/utils.py
@@ -29,6 +29,7 @@ from ..device_type import (
     DEVICE_COUNT,
     ALLOW_PREQUANTIZED_MODELS,
 )
+from ..bnb_availability import native_kernels_ready
 from .fp8 import weight_dequant, fp8_linear
 import functools
 
@@ -135,11 +136,18 @@ def calculate_settings(
 HAS_CUDA_STREAM = False
 try:
     import bitsandbytes as bnb
+
+    # If an earlier `import bitsandbytes` died inside __init__, CPython evicts only
+    # the parent from sys.modules and keeps its submodules, so this retry re-executes
+    # __init__ without rebinding `bnb.functional`. `import x.y as z` reads sys.modules
+    # directly and survives that, plain attribute access does not.
+    import bitsandbytes.functional as bnb_functional
 except Exception:
     # device_type.py already degrades to 16bit/full finetuning when bnb is missing
     # (e.g. gfx906, whose generic wheel has no kernels). Keep the import working and
     # fail only if a 4bit path is actually entered.
     bnb = None
+    bnb_functional = None
 
 
 def _bnb_required(*args, **kwargs):
@@ -152,7 +160,7 @@ def _bnb_required(*args, **kwargs):
 if bnb is not None:
     # https://github.com/bitsandbytes-foundation/bitsandbytes/pull/1330/files
     HAS_CUDA_STREAM = Version(bnb.__version__) > Version("0.43.3")
-    get_ptr = bnb.functional.get_ptr
+    get_ptr = bnb_functional.get_ptr
 else:
     get_ptr = _bnb_required
 
@@ -252,25 +260,28 @@ else:
 # Bitsandbytes operations
 ctypes_c_int = ctypes.c_int
 ctypes_c_int32 = ctypes.c_int32
-if bnb is None:
+# Same verdict device_type.py used to clear ALLOW_BITSANDBYTES, applied to the binds
+# themselves. 0.45.5 leaves `functional.lib = None` when the native library fails to
+# load, so these lookups would kill `import unsloth` instead of degrading to 16bit.
+if bnb is None or not native_kernels_ready(bnb, DEVICE_TYPE):
     cdequantize_blockwise_fp32 = _bnb_required
     cdequantize_blockwise_fp16_nf4 = _bnb_required
     cdequantize_blockwise_bf16_nf4 = _bnb_required
     cgemm_4bit_inference_naive_fp16 = _bnb_required
     cgemm_4bit_inference_naive_bf16 = _bnb_required
 else:
-    cdequantize_blockwise_fp32 = bnb.functional.lib.cdequantize_blockwise_fp32
-    cdequantize_blockwise_fp16_nf4 = bnb.functional.lib.cdequantize_blockwise_fp16_nf4
-    cdequantize_blockwise_bf16_nf4 = bnb.functional.lib.cdequantize_blockwise_bf16_nf4
+    cdequantize_blockwise_fp32 = bnb_functional.lib.cdequantize_blockwise_fp32
+    cdequantize_blockwise_fp16_nf4 = bnb_functional.lib.cdequantize_blockwise_fp16_nf4
+    cdequantize_blockwise_bf16_nf4 = bnb_functional.lib.cdequantize_blockwise_bf16_nf4
 
     if DEVICE_TYPE == "xpu":
         # https://github.com/bitsandbytes-foundation/bitsandbytes/blob/c3b8de268fdb55a88f92feada23fc811a1e6877a/bitsandbytes/backends/xpu/ops.py#L115
         # for xpu, inference gemv using above link
-        cgemm_4bit_inference_naive_fp16 = bnb.functional.lib.cgemv_4bit_inference_fp16
-        cgemm_4bit_inference_naive_bf16 = bnb.functional.lib.cgemv_4bit_inference_bf16
+        cgemm_4bit_inference_naive_fp16 = bnb_functional.lib.cgemv_4bit_inference_fp16
+        cgemm_4bit_inference_naive_bf16 = bnb_functional.lib.cgemv_4bit_inference_bf16
     else:
-        cgemm_4bit_inference_naive_fp16 = bnb.functional.lib.cgemm_4bit_inference_naive_fp16
-        cgemm_4bit_inference_naive_bf16 = bnb.functional.lib.cgemm_4bit_inference_naive_bf16
+        cgemm_4bit_inference_naive_fp16 = bnb_functional.lib.cgemm_4bit_inference_naive_fp16
+        cgemm_4bit_inference_naive_bf16 = bnb_functional.lib.cgemm_4bit_inference_naive_bf16
 
 
 torch_device_stream = (
diff --git a/unsloth/models/mapper.py b/unsloth/models/mapper.py
index 4558bb0f28..747b3bb986 100644
--- a/unsloth/models/mapper.py
+++ b/unsloth/models/mapper.py
@@ -94,10 +94,6 @@ __INT_TO_FLOAT_MAPPER = \
         "unsloth/llama-2-7b-chat",
         "meta-llama/Llama-2-7b-chat-hf",
     ),
-    "unsloth/llama-2-7b-chat-bnb-4bit" : (
-        "unsloth/llama-2-7b-chat",
-        "meta-llama/Llama-2-7b-chat-hf",
-    ),
     "unsloth/Mixtral-8x7B-v0.1-unsloth-bnb-4bit" : (
         "unsloth/Mixtral-8x7B-v0.1",
         "mistralai/Mixtral-8x7B-v0.1",
diff --git a/unsloth_cli/commands/studio.py b/unsloth_cli/commands/studio.py
index 9fd264ddf5..560df4aea2 100644
--- a/unsloth_cli/commands/studio.py
+++ b/unsloth_cli/commands/studio.py
@@ -19,7 +19,7 @@ import urllib.error
 import urllib.request
 from datetime import datetime, timezone
 from pathlib import Path
-from typing import List, Literal, Optional
+from typing import List, Literal, Optional, Sequence
 import typer
 
 from unsloth_cli import _studio_deps
@@ -483,9 +483,12 @@ def _write_auth_secret(path: Path, secret: str) -> None:
             os.chmod(tmp_path, 0o600)
         except OSError:
             pass
-        with os.fdopen(fd, "w", encoding = "utf-8") as f:
+        # newline pins LF: text mode writes CRLF on Windows, and `$(cat ...)`
+        # strips the LF but leaves the CR glued to the credential.
+        with os.fdopen(fd, "w", encoding = "utf-8", newline = "\n") as f:
             fd = -1
-            f.write(secret)
+            # Newline so `cat` doesn't run it into the shell prompt; readers strip.
+            f.write(secret + "\n")
         os.replace(tmp_path, path)
     except Exception:
         if fd >= 0:
@@ -502,6 +505,8 @@ def _connect_auth_db() -> sqlite3.Connection:
     auth_dir = STUDIO_HOME / "auth"
     auth_dir.mkdir(parents = True, exist_ok = True)
     conn = sqlite3.connect(auth_dir / "auth.db")
+    # A live server writes this DB while the CLI runs; the default lock wait is zero.
+    conn.execute("PRAGMA busy_timeout=5000")
     # Mirror backend storage.get_connection: this path can create auth/ and
     # auth.db (the pre-exposure gate writes here first), and sqlite3.connect
     # makes the DB 0644 under a 022 umask. Keep both private.
@@ -529,7 +534,8 @@ def _connect_auth_db() -> sqlite3.Connection:
             token_hash TEXT NOT NULL,
             username TEXT NOT NULL,
             expires_at TEXT NOT NULL,
-            is_desktop INTEGER NOT NULL DEFAULT 0
+            is_desktop INTEGER NOT NULL DEFAULT 0,
+            secret_gen TEXT
         );
         """
     )
@@ -564,6 +570,8 @@ def _connect_auth_db() -> sqlite3.Connection:
     refresh_columns = {row[1] for row in conn.execute("PRAGMA table_info(refresh_tokens)")}
     if "is_desktop" not in refresh_columns:
         conn.execute("ALTER TABLE refresh_tokens ADD COLUMN is_desktop INTEGER NOT NULL DEFAULT 0")
+    if "secret_gen" not in refresh_columns:
+        conn.execute("ALTER TABLE refresh_tokens ADD COLUMN secret_gen TEXT")
     conn.commit()
     return conn
 
@@ -697,12 +705,30 @@ def _bootstrap_deadline_active() -> bool:
         return True
 
 
-def _cli_update_password(conn: sqlite3.Connection, username: str, new_password: str) -> None:
+def _generate_reset_password() -> str:
+    """Readable 4-word passphrase; the user has to type this one back in."""
+    try:
+        import diceware
+        return diceware.get_passphrase(
+            options = diceware.handle_options(args = ["-n", "4", "-d", "", "-c"])
+        )
+    except Exception:
+        return secrets.token_urlsafe(24)
+
+
+def _cli_update_password(
+    conn: sqlite3.Connection,
+    username: str,
+    new_password: str,
+    *,
+    revoke_api_keys: bool = False,
+) -> None:
     """CLI mirror of backend update_password + change-password route effects.
 
     One transaction: rehash, rotate the JWT secret, clear must_change_password,
-    revoke refresh tokens (PR #6651 finding), and drop the desktop secret. File
-    cleanup happens after commit; a failed unlink must not roll the change back.
+    revoke refresh tokens (PR #6651 finding), drop the desktop secret, and (for a
+    reset) the API keys the old credential could have minted. File cleanup happens
+    after commit; a failed unlink must not roll the change back.
     """
     password_salt, password_hash = _hash_password(new_password)
     with conn:
@@ -719,6 +745,8 @@ def _cli_update_password(conn: sqlite3.Connection, username: str, new_password:
             "DELETE FROM app_secrets WHERE key IN (?, ?)",
             (DESKTOP_SECRET_HASH_KEY, DESKTOP_SECRET_CREATED_AT_KEY),
         )
+        if revoke_api_keys:
+            conn.execute("DELETE FROM api_keys")
     for stale in (BOOTSTRAP_PASSWORD_FILE, DESKTOP_SECRET_FILE):
         stale_path = STUDIO_HOME / "auth" / stale
         try:
@@ -728,8 +756,8 @@ def _cli_update_password(conn: sqlite3.Connection, username: str, new_password:
             # change back. But a locked-yet-writable file (Windows AV, read-only
             # auth dir) must be truncated: otherwise its stale plaintext survives
             # and generate_bootstrap_password() would re-validate this revoked
-            # credential after a later reset-password deletes auth.db. Mirrors
-            # backend clear_bootstrap_password().
+            # credential if auth.db is ever recreated. Mirrors backend
+            # clear_bootstrap_password().
             try:
                 stale_path.write_text("", encoding = "utf-8")
                 cleared = True
@@ -787,8 +815,8 @@ def _apply_supplied_password_before_launch(supplied_password: "str | None") -> N
         if not row[2]:
             typer.echo(
                 "Error: an Unsloth admin password is already set; --password only sets "
-                "the initial password. Run `unsloth studio reset-password` first "
-                "(or change it in the UI).",
+                "the initial password. Change it in the UI, or run `unsloth studio "
+                "reset-password` for a new one.",
                 err = True,
             )
             raise typer.Exit(1)
@@ -2237,6 +2265,9 @@ def run(
         # Headless serving prints its own URL/API-key banner; the Tauri-only
         # TAURI_PORT line would corrupt that machine-parseable output.
         emit_tauri_port = False,
+        # We read the bound port back below, so a fallback past another Studio is
+        # safe here and keeps side-by-side model runs working.
+        abort_if_own_studio = False,
     )
     # Forward the frontend validated before the gate (in-venv path).
     if resolved_frontend is not None:
@@ -2396,6 +2427,7 @@ def run(
 # ── unsloth studio stop ───────────────────────────────────────────────
 
 _PID_FILE = STUDIO_HOME / "studio.pid"
+PID_FILE_GLOB = "studio-*.pid"
 
 
 def _pid_alive(pid: int) -> bool:
@@ -2425,58 +2457,210 @@ def _pid_alive(pid: int) -> bool:
     return True
 
 
-@studio_app.command()
-def stop():
-    """Stop a running Unsloth Studio server.
+def _parse_pid_record(text: str) -> "tuple[int, float | None] | None":
+    """Parse ``pid`` / optional ``create_time`` from PID file contents."""
+    lines = text.splitlines()
+    if not lines or not lines[0].strip().isdigit():
+        return None
+    try:
+        # isdigit() is not enough: "²".isdigit() is True but int() rejects it.
+        pid = int(lines[0].strip())
+    except ValueError:
+        return None
+    # kill(0) signals our whole process group; kill(1) is init. Never either.
+    if pid < 2:
+        return None
+    created = None
+    if len(lines) > 1:
+        try:
+            created = float(lines[1].strip())
+        except ValueError:
+            created = None
+    return pid, created
 
-    Reads the PID from ~/.unsloth/studio/studio.pid and sends SIGTERM
-    (or TerminateProcess on Windows) to shut it down gracefully.
+
+def _read_pid_record(path: Path) -> "tuple[int, float | None] | None":
+    """Parse ``pid`` / optional ``create_time`` from a PID file."""
+    try:
+        text = path.read_text(encoding = "utf-8")
+    except (OSError, UnicodeDecodeError):
+        return None
+    return _parse_pid_record(text)
+
+
+def _unlink_quietly(path: Path) -> None:
+    """Drop a record without letting one bad file end the loop.
+
+    An undeletable record must not stop us reaching the other servers -- that is
+    the orphan this command exists to prevent.
     """
+    try:
+        path.unlink(missing_ok = True)
+    except OSError as e:
+        typer.echo(f"Could not remove PID file {path.name}: {e}", err = True)
+
+
+def _report_unreadable(paths: "list[Path]") -> None:
+    """Say which servers we could not reach, since `stop` is about to exit 1."""
+    names = ", ".join(sorted(p.name for p in paths))
+    typer.echo(
+        f"Could not read {len(paths)} PID file(s): {names}. A server recorded "
+        f"there may still be running; re-run with permission to read "
+        f"{STUDIO_HOME} to stop it.",
+        err = True,
+    )
+
+
+def _pid_file_entries(
+    unreadable: "list[Path] | None" = None,
+) -> "list[tuple[int, list[float | None], list[Path]]]":
+    """(pid, create_times, files) per recorded server, including the legacy studio.pid.
+
+    Paths that could not be read are appended to `unreadable` when given, so the
+    caller can tell "nothing is running" apart from "something is running and we
+    could not see it".
+
+    Grouped by PID: a server writes both its per-port file and studio.pid, and
+    signalling twice would hit the SIG_DFL the first SIGTERM installs, hard-killing
+    it mid-shutdown. Every recorded time is kept -- a stale file and a live server
+    can share a PID, and the stale one must not veto the live one.
+    """
+    by_pid: "dict[int, tuple[list[float | None], list[Path]]]" = {}
+    try:
+        paths = sorted(STUDIO_HOME.glob(PID_FILE_GLOB)) + [_PID_FILE]
+    except OSError:
+        paths = [_PID_FILE]
+    seen = set()
+    for path in paths:
+        if path in seen or not path.is_file():
+            continue
+        seen.add(path)
+        try:
+            text = path.read_text(encoding = "utf-8")
+        except (OSError, UnicodeDecodeError) as e:
+            # Unreadable is not the same as invalid. A root-owned record, or one
+            # caught mid-write, still belongs to a live server, and deleting it
+            # strands that server -- the bug this command exists to fix.
+            typer.echo(f"Cannot read PID file {path.name}: {e}", err = True)
+            if unreadable is not None:
+                unreadable.append(path)
+            continue
+        record = _parse_pid_record(text)
+        if record is None:
+            typer.echo(f"Ignoring invalid PID file {path.name}")
+            _unlink_quietly(path)
+            continue
+        pid, created = record
+        created_times, files = by_pid.setdefault(pid, ([], []))
+        created_times.append(created)
+        files.append(path)
+    return [(pid, times, files) for pid, (times, files) in by_pid.items()]
+
+
+def _pid_is_studio_server(pid: int, created_times: "Sequence[float | None]" = ()) -> bool:
+    """False only when a recorded start time proves this PID is a different process.
+
+    Any recorded time matching is enough -- a stale record must not veto a live
+    server that reused the PID. Records with no time at all (a legacy studio.pid,
+    or a server started without psutil) cannot be checked, so they are trusted:
+    the old `stop` signalled with no checks at all, and skipping a live server is
+    the orphan bug this exists to fix.
+
+    An untimed record sitting *alongside* a timed one carries no information, so
+    it must not cancel the timed one either. Every current server writes both a
+    timed per-port record and an untimed studio.pid, so letting the untimed half
+    win made this check inert exactly where it matters and let `stop` SIGTERM an
+    unrelated process that had inherited the PID.
+    """
+    known = [c for c in created_times if c is not None]
+    if not known:
+        return True
+    try:
+        import psutil
+        actual = psutil.Process(pid).create_time()
+    except Exception:
+        return True
+    return any(abs(actual - c) < 1.0 for c in known)
+
+
+def _signal_stop(pid: int) -> "str | None":
+    """SIGTERM (or taskkill) the pid. Returns an error string, or None on success."""
     import signal as _signal
 
-    if not _PID_FILE.is_file():
-        typer.echo("No running Unsloth server found (no PID file).")
-        raise typer.Exit(0)
-
-    pid_text = _PID_FILE.read_text(encoding = "utf-8").strip()
-    if not pid_text.isdigit():
-        typer.echo(f"Invalid PID file contents: {pid_text}")
-        _PID_FILE.unlink(missing_ok = True)
-        raise typer.Exit(1)
-
-    pid = int(pid_text)
-
-    # Check if still alive (os.kill(pid, 0) is invalid on Windows -- see _pid_alive).
-    if not _pid_alive(pid):
-        typer.echo(f"Unsloth server (PID {pid}) is not running. Cleaning up stale PID file.")
-        _PID_FILE.unlink(missing_ok = True)
-        raise typer.Exit(0)
-
-    # Send SIGTERM (graceful shutdown) or TerminateProcess on Windows
+    if pid < 2:
+        return f"refusing to signal PID {pid}"
     try:
         if sys.platform == "win32":
             # /T also stops llama-server children, which otherwise keep GPU and port.
             subprocess.run(["taskkill", "/PID", str(pid), "/T", "/F"], check = True)
         else:
             os.kill(pid, _signal.SIGTERM)
-        typer.echo(f"Sent shutdown signal to Unsloth server (PID {pid}).")
     except ProcessLookupError:
-        typer.echo(f"Unsloth server (PID {pid}) already exited.")
-        _PID_FILE.unlink(missing_ok = True)
-        raise typer.Exit(0)
+        return None
     except Exception as e:
-        typer.echo(f"Failed to stop Unsloth server (PID {pid}): {e}", err = True)
-        raise typer.Exit(1)
+        return str(e)
+    return None
 
-    # Wait briefly for the process to exit and clean up.
+
+@studio_app.command()
+def stop():
+    """Stop every running Unsloth Studio server for this STUDIO_HOME.
+
+    The port fallback can leave more than one running, so stop them all.
+    """
+    unreadable: "list[Path]" = []
+    entries = _pid_file_entries(unreadable)
+    if not entries:
+        if unreadable:
+            # Reporting success here would be a lie: the records we could not
+            # read are kept, and the servers behind them are still serving.
+            _report_unreadable(unreadable)
+            raise typer.Exit(1)
+        typer.echo("No running Unsloth server found (no PID file).")
+        raise typer.Exit(0)
+
+    signalled, failed = [], []
+    for pid, created_times, paths in entries:
+        if not _pid_alive(pid) or not _pid_is_studio_server(pid, created_times):
+            for path in paths:
+                _unlink_quietly(path)
+            continue
+        error = _signal_stop(pid)
+        if error is not None:
+            failed.append((pid, error))
+            typer.echo(f"Failed to stop Unsloth server (PID {pid}): {error}", err = True)
+            continue
+        typer.echo(f"Sent shutdown signal to Unsloth server (PID {pid}).")
+        signalled.append((pid, paths))
+
+    if not signalled and not failed:
+        if unreadable:
+            _report_unreadable(unreadable)
+            raise typer.Exit(1)
+        typer.echo("No running Unsloth server found (cleaned up stale PID files).")
+        raise typer.Exit(0)
+
+    pending = list(signalled)
     for _ in range(10):
+        if not pending:
+            break
         time.sleep(0.5)
-        if not _pid_alive(pid):
-            _PID_FILE.unlink(missing_ok = True)
-            typer.echo("Unsloth server stopped.")
-            raise typer.Exit(0)
+        for entry in list(pending):
+            pid, paths = entry
+            if not _pid_alive(pid):
+                for path in paths:
+                    _unlink_quietly(path)
+                pending.remove(entry)
 
-    typer.echo("Unsloth server is shutting down (may take a few seconds).")
+    stopped = len(signalled) - len(pending)
+    if stopped:
+        typer.echo(f"Unsloth server{'s' if stopped > 1 else ''} stopped ({stopped}).")
+    for pid, _paths in pending:
+        typer.echo(f"Unsloth server (PID {pid}) is shutting down (may take a few seconds).")
+    if unreadable:
+        _report_unreadable(unreadable)
+    if failed or unreadable:
+        raise typer.Exit(1)
 
 
 # ── unsloth studio setup / update ─────────────────────────────────────
@@ -2890,59 +3074,33 @@ def provision_desktop_auth():
 def reset_password():
     """Reset the Unsloth admin password.
 
-    Deletes the auth database so that a fresh admin account with a new
-    random password is created on the next server start.  The Unsloth
-    server must be restarted after running this command.
+    Rotates the credential in place: a running Unsloth accepts the new password on
+    its next request, so there is nothing to restart. Shared /p preview links are
+    not revoked -- rotate those in Settings if the old password leaked.
     """
-    auth_dir = STUDIO_HOME / "auth"
-    db_file = auth_dir / "auth.db"
-    stale_files = [
-        auth_dir / BOOTSTRAP_PASSWORD_FILE,
-        auth_dir / DESKTOP_SECRET_FILE,
-    ]
-    had_db = db_file.exists()
-
-    # Delete auth.db FIRST and prove it is gone before touching the seeded
-    # credential files. If it cannot be removed (a running Unsloth or Windows
-    # holds it open, or a read-only auth dir), abort with the credential files
-    # untouched: deleting them while an un-resettable DB (must_change_password=1)
-    # survives would lock a forgotten-password reset out of any recovery
-    # credential. Failing here leaves a consistent, still-recoverable state.
+    new_password = _generate_reset_password()
     try:
-        db_file.unlink(missing_ok = True)
-    except OSError as exc:
+        conn = _connect_auth_db()
+    except (OSError, sqlite3.Error) as exc:
         typer.echo(
-            f"Error: could not delete the auth database ({exc}). Stop any running "
-            "Unsloth and retry; no credential files were changed.",
+            f"Error: could not open the auth database ({exc}). Check that "
+            f"{STUDIO_HOME / 'auth'} is writable; if auth.db itself is unreadable, stop "
+            "Unsloth, delete it, and start again to re-seed.",
             err = True,
         )
         raise typer.Exit(1)
 
-    # The DB is gone, so the next start re-seeds. Invalidate the seeded plaintext
-    # credential files so that re-seed generates a FRESH password instead of
-    # reusing a stale one: unlink only ignores FileNotFoundError, so a
-    # locked/undeletable file (Windows AV, read-only dir) would otherwise survive
-    # and generate_bootstrap_password() would read it back and re-validate the
-    # credential this reset revoked. Truncate on unlink failure; if a file can be
-    # neither removed nor truncated, fail closed -- the DB is already gone, so a
-    # surviving plaintext would be reused, and the user must remove it manually.
-    for path in stale_files:
-        try:
-            path.unlink(missing_ok = True)
-        except OSError:
-            try:
-                path.write_text("", encoding = "utf-8")
-            except OSError as exc:
-                typer.echo(
-                    f"Error: could not remove or clear {path.name} ({exc}); delete "
-                    "it manually before restarting Unsloth or the old password may "
-                    "be reused.",
-                    err = True,
-                )
-                raise typer.Exit(1)
+    try:
+        _ensure_cli_default_admin(conn)
+        _cli_update_password(conn, DEFAULT_ADMIN_USERNAME, new_password, revoke_api_keys = True)
+    except (OSError, sqlite3.Error) as exc:
+        typer.echo(f"Error: could not reset the password ({exc}).", err = True)
+        raise typer.Exit(1)
+    finally:
+        conn.close()
 
-    if not had_db:
-        typer.echo("No auth database found -- nothing to reset.")
-        raise typer.Exit(0)
-
-    typer.echo("Auth database deleted. Restart Unsloth Studio to get a new password.")
+    typer.echo(f"New password for '{DEFAULT_ADMIN_USERNAME}': {new_password}")
+    typer.echo(
+        "Sessions and API keys revoked. A running Unsloth takes it on the next request, "
+        "though repeated failed logins can hold the rate limit shut for up to a minute."
+    )
diff --git a/unsloth_cli/tests/conftest.py b/unsloth_cli/tests/conftest.py
new file mode 100644
index 0000000000..bb42914e69
--- /dev/null
+++ b/unsloth_cli/tests/conftest.py
@@ -0,0 +1,26 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Shared fixtures for the unsloth_cli tests."""
+
+import sys
+import types
+
+import pytest
+
+
+@pytest.fixture
+def stub_tool_policy_state(monkeypatch):
+    """Stub the backend's `state.tool_policy`, which run() imports in-venv.
+
+    It lives under studio/backend, so it only imports once something has put
+    that directory on sys.path. Tests that reach the in-venv branch of run()
+    used to get that for free from whichever file ran earlier and did it as a
+    side effect, which made them pass only in a full-directory run.
+    """
+    state_mod = types.ModuleType("state")
+    tp_mod = types.ModuleType("state.tool_policy")
+    tp_mod.set_tool_policy = lambda *a, **k: None
+    state_mod.tool_policy = tp_mod
+    monkeypatch.setitem(sys.modules, "state", state_mod)
+    monkeypatch.setitem(sys.modules, "state.tool_policy", tp_mod)
diff --git a/unsloth_cli/tests/test_studio_password_prompt.py b/unsloth_cli/tests/test_studio_password_prompt.py
index 6e9a2c1d52..753c22edc2 100644
--- a/unsloth_cli/tests/test_studio_password_prompt.py
+++ b/unsloth_cli/tests/test_studio_password_prompt.py
@@ -251,7 +251,7 @@ def test_studio_default_prompt_rejects_current_password(monkeypatch, tmp_path):
     studio_mod = _studio()
     events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
     _seed_auth(studio_mod)
-    bootstrap_pw = (tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE).read_text()
+    bootstrap_pw = (tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE).read_text().strip()
 
     _invoke_studio_default(monkeypatch, events, ["--secure"])
 
@@ -626,6 +626,12 @@ def test_studio_default_in_venv_broken_backend_exits_before_stripping_bootstrap(
 
     # Pretend we are already inside the studio venv, with a broken backend.
     monkeypatch.setattr(sys, "prefix", str(tmp_path / "unsloth_studio"))
+    # A built dist is not present in a fresh clone. The missing-frontend gate
+    # runs first and has its own test below; stub it so this one reaches the
+    # backend check it is actually about.
+    monkeypatch.setattr(
+        studio_mod, "_find_frontend_dist", lambda: Path("/fake/studio/frontend/dist")
+    )
 
     def _boom():
         raise ImportError("cannot import backend run.py")
@@ -957,7 +963,9 @@ def test_run_reexec_forwards_resolved_frontend_on_public_launch(monkeypatch, tmp
 
     exec_argv = [argv for kind, argv in events if kind == "exec"][0]
     assert "--frontend" in exec_argv, exec_argv
-    assert exec_argv[exec_argv.index("--frontend") + 1] == "/fake/studio/frontend/dist", exec_argv
+    # str(Path(...)), not the literal: Windows renders it with backslashes.
+    expected_dist = str(Path("/fake/studio/frontend/dist"))
+    assert exec_argv[exec_argv.index("--frontend") + 1] == expected_dist, exec_argv
 
 
 def test_run_non_tty_persists_seeded_admin_on_fresh_home(monkeypatch, tmp_path):
@@ -1032,50 +1040,173 @@ def test_bootstrap_deadline_active_mirrors_backend_parsing(monkeypatch, raw, exp
     assert studio_mod._bootstrap_deadline_active() is expected
 
 
-def test_reset_password_truncates_locked_bootstrap_after_db_delete(monkeypatch, tmp_path):
-    # reset-password deletes auth.db first, then invalidates the seeded credential
-    # files. A locked/undeletable .bootstrap_password must be truncated so its
-    # stale plaintext cannot be re-seeded (generate_bootstrap_password reuses a
-    # non-empty file), while the reset still succeeds.
-    import pathlib
-
-    studio_mod = _studio()
-    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
-    _seed_auth(studio_mod)
-    auth_dir = tmp_path / "auth"
-    bootstrap_file = auth_dir / studio_mod.BOOTSTRAP_PASSWORD_FILE
-    db_file = auth_dir / "auth.db"
-    assert bootstrap_file.exists() and db_file.exists()
-    assert bootstrap_file.read_text().strip()
-
-    _real_unlink = pathlib.Path.unlink
-
-    def _boom_unlink(self, *a, **k):
-        if self.name == studio_mod.BOOTSTRAP_PASSWORD_FILE:
-            raise OSError("locked")
-        return _real_unlink(self, *a, **k)
-
-    monkeypatch.setattr(pathlib.Path, "unlink", _boom_unlink)
-
+def _reset_password_cli(studio_mod):
     import typer as _typer
 
     app = _typer.Typer()
     app.command()(studio_mod.reset_password)
-    result = CliRunner().invoke(app, [], catch_exceptions = True)
+    return CliRunner().invoke(app, [], catch_exceptions = True)
+
+
+def _password_works(studio_mod, candidate):
+    conn = studio_mod._connect_auth_db()
+    try:
+        row = conn.execute(
+            "SELECT password_salt, password_hash FROM auth_user WHERE username = ?",
+            (studio_mod.DEFAULT_ADMIN_USERNAME,),
+        ).fetchone()
+    finally:
+        conn.close()
+    return studio_mod._pbkdf2_hex(candidate, row[0].encode("utf-8")) == row[1]
+
+
+def _printed_password(result):
+    line = next(l for l in result.output.splitlines() if l.startswith("New password for"))
+    return line.split(": ", 1)[1].strip()
+
+
+def test_reset_password_rotates_in_place_without_deleting_the_db(monkeypatch, tmp_path):
+    # The DB survives, so a running server keeps its admin row and the new password.
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    _seed_auth(studio_mod)
+    db_file = tmp_path / "auth" / "auth.db"
+    before = _auth_state(studio_mod)
+
+    result = _reset_password_cli(studio_mod)
 
     assert result.exit_code == 0, result.output
-    assert not db_file.exists()
-    # The locked file survives, but truncated -- no reusable plaintext.
-    assert bootstrap_file.exists()
-    assert bootstrap_file.read_text() == ""
+    assert db_file.exists()
+    after = _auth_state(studio_mod)
+    assert after["password_hash"] != before["password_hash"]
+    assert after["jwt_secret"] != before["jwt_secret"]
+    assert _password_works(studio_mod, _printed_password(result))
+
+
+def test_reset_password_waits_out_a_concurrent_writer(monkeypatch, tmp_path):
+    # The CLI now writes while the server does; without a busy_timeout this fails.
+    import threading
+    import time
+
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    _seed_auth(studio_mod)
+    released = threading.Event()
+
+    def hold_write_lock():
+        conn = sqlite3.connect(_auth_db(tmp_path))
+        conn.execute("BEGIN IMMEDIATE")
+        conn.execute(
+            "INSERT INTO refresh_tokens (token_hash, username, expires_at) "
+            "VALUES ('held', 'unsloth', '2099-01-01T00:00:00')"
+        )
+        time.sleep(0.5)
+        conn.rollback()
+        conn.close()
+        released.set()
+
+    holder = threading.Thread(target = hold_write_lock)
+    holder.start()
+    time.sleep(0.1)
+    result = _reset_password_cli(studio_mod)
+    holder.join()
+
+    assert released.is_set()
+    assert result.exit_code == 0, result.output
+    assert _password_works(studio_mod, _printed_password(result))
+
+
+def test_reset_password_revokes_sessions_and_api_keys(monkeypatch, tmp_path):
+    # Deleting auth.db used to drop these implicitly.
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    _seed_auth(studio_mod)
+    conn = studio_mod._connect_auth_db()
+    conn.execute(
+        "INSERT INTO api_keys (username, key_prefix, key_hash, name, created_at) "
+        "VALUES (?, 'sk-x', 'hash', 'k', '2026-01-01T00:00:00')",
+        (studio_mod.DEFAULT_ADMIN_USERNAME,),
+    )
+    conn.commit()
+    conn.close()
+
+    assert _reset_password_cli(studio_mod).exit_code == 0
+
+    conn = studio_mod._connect_auth_db()
+    try:
+        assert conn.execute("SELECT COUNT(*) FROM api_keys").fetchone()[0] == 0
+        assert conn.execute("SELECT COUNT(*) FROM refresh_tokens").fetchone()[0] == 0
+    finally:
+        conn.close()
+
+
+def test_reset_password_leaves_the_account_ready_to_log_in(monkeypatch, tmp_path):
+    # must_change_password stays 0 on purpose: at 1 a running server injects its
+    # startup-cached (now wrong) bootstrap password into the login page.
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    _seed_auth(studio_mod)
+
+    assert _reset_password_cli(studio_mod).exit_code == 0
+
+    assert _auth_state(studio_mod)["must_change_password"] == 0
+    assert not (tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE).exists()
+
+
+def test_reset_password_seeds_the_admin_when_no_db_exists(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+
+    result = _reset_password_cli(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert _password_works(studio_mod, _printed_password(result))
+
+
+def test_reset_password_reports_an_unwritable_auth_dir(monkeypatch, tmp_path):
+    # _connect_auth_db creates auth/ before it opens SQLite, so a read-only Unsloth
+    # home raises OSError, not sqlite3.Error.
+    import pathlib
+
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+
+    def _boom_mkdir(self, *a, **k):
+        raise PermissionError("read-only")
+
+    monkeypatch.setattr(pathlib.Path, "mkdir", _boom_mkdir)
+
+    result = _reset_password_cli(studio_mod)
+
+    assert result.exit_code == 1, result.output
+    assert not isinstance(result.exception, OSError)
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "could not open the auth database" in combined.lower()
+
+
+def test_reset_password_reports_an_unreadable_db(monkeypatch, tmp_path):
+    # Deleting a corrupt DB here would revive the bug: a running server would be
+    # left with no admin row, rejecting the correct password until restarted.
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    auth_dir = tmp_path / "auth"
+    auth_dir.mkdir()
+    (auth_dir / "auth.db").write_text("not a database")
+
+    result = _reset_password_cli(studio_mod)
+
+    assert result.exit_code == 1, result.output
+    assert (auth_dir / "auth.db").exists()
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "could not open the auth database" in combined.lower()
 
 
 def test_cli_update_password_truncates_locked_bootstrap_after_change(monkeypatch, tmp_path):
     # After a CLI/interactive password change the seeded .bootstrap_password is
     # deleted. If it cannot be unlinked but is still writable (locked file /
     # read-only dir), it must be TRUNCATED so its stale plaintext cannot be
-    # re-seeded by generate_bootstrap_password() after a later reset-password
-    # deletes auth.db. The change is already committed, so it must NOT roll back.
+    # re-seeded by generate_bootstrap_password() if auth.db is ever recreated. The
+    # change is already committed, so it must NOT roll back.
     import pathlib
 
     studio_mod = _studio()
@@ -1103,88 +1234,6 @@ def test_cli_update_password_truncates_locked_bootstrap_after_change(monkeypatch
     assert bootstrap_file.read_text() == ""
 
 
-def test_reset_password_fails_closed_when_db_cannot_be_deleted(monkeypatch, tmp_path):
-    # If auth.db cannot be removed (running Unsloth / Windows lock, read-only dir),
-    # reset must abort BEFORE touching the credential files -- deleting them while
-    # an un-resettable must_change_password=1 DB survives would lock a
-    # forgotten-password reset out with no recovery credential.
-    import pathlib
-
-    studio_mod = _studio()
-    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
-    _seed_auth(studio_mod)
-    auth_dir = tmp_path / "auth"
-    bootstrap_file = auth_dir / studio_mod.BOOTSTRAP_PASSWORD_FILE
-    db_file = auth_dir / "auth.db"
-    assert bootstrap_file.exists() and db_file.exists()
-
-    _real_unlink = pathlib.Path.unlink
-
-    def _boom_unlink(self, *a, **k):
-        if self.name == "auth.db":
-            raise OSError("database is locked")
-        return _real_unlink(self, *a, **k)
-
-    monkeypatch.setattr(pathlib.Path, "unlink", _boom_unlink)
-
-    import typer as _typer
-
-    app = _typer.Typer()
-    app.command()(studio_mod.reset_password)
-    result = CliRunner().invoke(app, [], catch_exceptions = True)
-
-    assert result.exit_code == 1, result.output
-    # DB still there; credential files untouched (no lockout, no half-done reset).
-    assert db_file.exists()
-    assert bootstrap_file.exists()
-    assert bootstrap_file.read_text().strip()
-    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
-    assert "could not delete the auth database" in combined.lower()
-
-
-def test_reset_password_fails_closed_when_credential_cannot_be_invalidated(monkeypatch, tmp_path):
-    # If a seeded credential file can be neither unlinked nor truncated, reset must
-    # fail closed: auth.db is already gone, so a surviving plaintext would be
-    # re-seeded and re-validate the revoked password.
-    import pathlib
-
-    studio_mod = _studio()
-    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
-    _seed_auth(studio_mod)
-    auth_dir = tmp_path / "auth"
-    bootstrap_file = auth_dir / studio_mod.BOOTSTRAP_PASSWORD_FILE
-    db_file = auth_dir / "auth.db"
-    assert bootstrap_file.exists() and db_file.exists()
-
-    _real_unlink = pathlib.Path.unlink
-    _real_write_text = pathlib.Path.write_text
-
-    def _boom_unlink(self, *a, **k):
-        if self.name == studio_mod.BOOTSTRAP_PASSWORD_FILE:
-            raise OSError("locked")
-        return _real_unlink(self, *a, **k)
-
-    def _boom_write_text(self, *a, **k):
-        if self.name == studio_mod.BOOTSTRAP_PASSWORD_FILE:
-            raise OSError("read-only")
-        return _real_write_text(self, *a, **k)
-
-    monkeypatch.setattr(pathlib.Path, "unlink", _boom_unlink)
-    monkeypatch.setattr(pathlib.Path, "write_text", _boom_write_text)
-
-    import typer as _typer
-
-    app = _typer.Typer()
-    app.command()(studio_mod.reset_password)
-    result = CliRunner().invoke(app, [], catch_exceptions = True)
-
-    assert result.exit_code == 1, result.output
-    # auth.db was deleted first; the un-invalidatable file is reported for manual removal.
-    assert not db_file.exists()
-    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
-    assert "delete it manually" in combined.lower()
-
-
 def test_connect_auth_db_creates_private_files(monkeypatch, tmp_path):
     # Fresh install: the CLI gate writes the password hash + JWT secret before
     # the backend ever runs, so this path must apply the same 0700/0600 modes
@@ -1204,6 +1253,37 @@ def test_connect_auth_db_creates_private_files(monkeypatch, tmp_path):
     assert stat.S_IMODE((auth_dir / "auth.db").stat().st_mode) == 0o600
 
 
+def test_write_auth_secret_terminates_the_file_with_a_newline(monkeypatch, tmp_path):
+    # Shared by .bootstrap_password and .desktop_secret; every reader strips.
+    studio_mod = _studio()
+    path = tmp_path / ".desktop_secret"
+
+    studio_mod._write_auth_secret(path, "desktop-abc123")
+
+    # Bytes: read_text would decode CRLF back to "\n" and hide a CR.
+    assert path.read_bytes() == b"desktop-abc123\n"
+
+
+def test_seeded_bootstrap_file_ends_with_a_newline(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    _seed_auth(studio_mod)
+
+    raw = (tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE).read_bytes()
+
+    assert raw.endswith(b"\n") and not raw.endswith(b"\r\n")
+
+    conn = sqlite3.connect(_auth_db(tmp_path))
+    try:
+        salt, pwd_hash = conn.execute(
+            "SELECT password_salt, password_hash FROM auth_user WHERE username = ?",
+            (studio_mod.DEFAULT_ADMIN_USERNAME,),
+        ).fetchone()
+    finally:
+        conn.close()
+    assert studio_mod._pbkdf2_hex(raw.decode("utf-8").strip(), salt.encode("utf-8")) == pwd_hash
+
+
 # ── non-interactive --password / UNSLOTH_STUDIO_PASSWORD / stdin ──────
 
 
@@ -1284,7 +1364,7 @@ def test_studio_default_password_must_differ_fails_closed(monkeypatch, tmp_path)
     studio_mod = _studio()
     events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
     _seed_auth(studio_mod)
-    bootstrap_pw = (tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE).read_text()
+    bootstrap_pw = (tmp_path / "auth" / studio_mod.BOOTSTRAP_PASSWORD_FILE).read_text().strip()
 
     result = _invoke_studio_default(monkeypatch, events, ["--secure", "--password", bootstrap_pw])
 
@@ -1368,30 +1448,3 @@ def test_studio_default_password_applies_on_headless_wildcard_no_tunnel(monkeypa
     assert after["must_change_password"] == 0
     assert after["password_hash"] != before["password_hash"]
     assert "--password" not in _exec_argv(events)
-
-
-def test_reset_password_then_password_roundtrip(monkeypatch, tmp_path):
-    # After reset-password wipes the DB, the next start re-seeds a fresh admin
-    # that again requires a change, so --password can set a new initial password.
-    import typer
-
-    studio_mod = _studio()
-    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
-    _seed_auth(studio_mod)
-    conn = studio_mod._connect_auth_db()
-    studio_mod._cli_update_password(conn, studio_mod.DEFAULT_ADMIN_USERNAME, "first-password-1")
-    conn.close()
-    assert _auth_state(studio_mod)["must_change_password"] == 0
-
-    # reset-password deletes the auth DB + seeded credential files.
-    try:
-        studio_mod.reset_password()
-    except typer.Exit:
-        pass
-    assert not (tmp_path / "auth" / "auth.db").exists()
-
-    # A restart re-seeds (ensure_default_admin, must_change=1); --password sets anew.
-    events = _install_prompt_env(monkeypatch, tmp_path, interactive = True)
-    _invoke_studio_default(monkeypatch, events, ["--secure", "--password", "second-password-2"])
-    assert [kind for kind, _ in events] == ["exec"], events
-    assert _auth_state(studio_mod)["must_change_password"] == 0
diff --git a/unsloth_cli/tests/test_studio_run_parallel_flag.py b/unsloth_cli/tests/test_studio_run_parallel_flag.py
index f1a4e69b81..9a3260d699 100644
--- a/unsloth_cli/tests/test_studio_run_parallel_flag.py
+++ b/unsloth_cli/tests/test_studio_run_parallel_flag.py
@@ -606,14 +606,14 @@ def test_studio_default_exposes_parallel_option():
     assert "--parallel" in decls
     assert "--n-parallel" in decls
     assert (
-        getattr(opt, "default", None) == 1
-    ), "studio_default --parallel must default to 1 (pre-PR); `run` is 4"
+        getattr(opt, "default", None) == studio_mod._PARALLEL_DEFAULT_PLAIN
+    ), "studio_default --parallel must use _PARALLEL_DEFAULT_PLAIN"
     assert getattr(opt, "min", None) == 1
     assert getattr(opt, "max", None) == 64
 
 
 @pytest.mark.parametrize("value", [1, 4, 8, 64])
-def test_in_venv_path_passes_parallel_to_run_server(monkeypatch, value):
+def test_in_venv_path_passes_parallel_to_run_server(monkeypatch, value, stub_tool_policy_state):
     """In-venv path must forward --parallel to
     run_server(llama_parallel_slots=N), not the old hardcoded 4."""
     studio_mod = _load_run_command()
@@ -679,7 +679,6 @@ def test_api_only_option_is_registered():
     "extra,present",
     [
         (["--api-only"], True),
-        (["--secure", "--api-only"], True),  # secure headless path
         ([], False),
     ],
 )
@@ -691,8 +690,25 @@ def test_reexec_forwards_api_only(monkeypatch, extra, present):
     assert ("--api-only" in argv) is present, argv
 
 
+def test_secure_api_only_is_refused_before_any_reexec(monkeypatch, tmp_path):
+    """`--secure --api-only` used to re-exec; the pre-exposure gate now refuses
+    it, because api-only has no login page and the bootstrap deadline does not
+    apply, so the seeded password could never be changed."""
+    studio_mod = _load_run_command()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+
+    result, captured = _invoke_run(monkeypatch, _BASE + ["--secure", "--api-only"])
+
+    assert captured == [], captured
+    assert result.exit_code != 0
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "default admin password was never changed" in combined.lower()
+
+
 @pytest.mark.parametrize("extra,expected", [(["--api-only"], True), ([], False)])
-def test_in_venv_path_passes_api_only_to_run_server(monkeypatch, extra, expected):
+def test_in_venv_path_passes_api_only_to_run_server(
+    monkeypatch, extra, expected, stub_tool_policy_state
+):
     """In-venv path must forward --api-only to run_server(api_only=...)."""
     studio_mod = _load_run_command()
 
diff --git a/unsloth_cli/tests/test_studio_secure_flag.py b/unsloth_cli/tests/test_studio_secure_flag.py
index 2a67aad95a..118f227949 100644
--- a/unsloth_cli/tests/test_studio_secure_flag.py
+++ b/unsloth_cli/tests/test_studio_secure_flag.py
@@ -244,7 +244,7 @@ class _RunServerCaptured(SystemExit):
         self.kwargs = dict(kwargs)
 
 
-def test_run_in_venv_passes_secure_and_forces_host(monkeypatch, tmp_path):
+def test_run_in_venv_passes_secure_and_forces_host(monkeypatch, tmp_path, stub_tool_policy_state):
     import types
 
     studio_mod = _studio()
@@ -261,6 +261,11 @@ def test_run_in_venv_passes_secure_and_forces_host(monkeypatch, tmp_path):
 
     fake_venv = tmp_path / "unsloth_studio"
     monkeypatch.setattr(sys, "prefix", str(fake_venv))
+    # A built dist is not present in a fresh clone, and without it the public
+    # launch gate exits before run_server is ever reached.
+    monkeypatch.setattr(
+        studio_mod, "_find_frontend_dist", lambda: Path("/fake/studio/frontend/dist")
+    )
 
     from unsloth_cli import _tool_policy as _tp_mod
 
diff --git a/unsloth_cli/tests/test_studio_stop.py b/unsloth_cli/tests/test_studio_stop.py
new file mode 100644
index 0000000000..74e34d4fa1
--- /dev/null
+++ b/unsloth_cli/tests/test_studio_stop.py
@@ -0,0 +1,530 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""`unsloth studio stop` must stop every server it started.
+
+With one PID file the second launch overwrote the first entry, so stop killed
+the newer server, claimed success, and left the older one serving.
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+from typer.testing import CliRunner
+
+
+_REPO_ROOT = Path(__file__).resolve().parents[2]
+if str(_REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(_REPO_ROOT))
+
+
+def _studio():
+    from unsloth_cli.commands import studio as _studio_mod
+    return _studio_mod
+
+
+# Captured before _install stubs it, for the tests that exercise it.
+_REAL_IS_STUDIO_SERVER = _studio()._pid_is_studio_server
+
+
+def _install(
+    monkeypatch,
+    tmp_path,
+    *,
+    alive,
+    killed = None,
+):
+    """Point the CLI at tmp_path and fake process liveness."""
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    monkeypatch.setattr(studio_mod, "_PID_FILE", tmp_path / "studio.pid")
+    monkeypatch.setattr(studio_mod.time, "sleep", lambda _s: None)
+
+    live = set(alive)
+    killed = killed if killed is not None else []
+
+    monkeypatch.setattr(studio_mod, "_pid_alive", lambda pid: pid in live)
+    monkeypatch.setattr(studio_mod, "_pid_is_studio_server", lambda pid, created_times = (): True)
+
+    def fake_kill(pid, _sig):
+        killed.append(pid)
+        live.discard(pid)
+
+    monkeypatch.setattr(studio_mod.os, "kill", fake_kill)
+    monkeypatch.setattr(sys, "platform", "linux")
+    return studio_mod, live, killed
+
+
+def _write_pid(tmp_path, name, pid):
+    (tmp_path / name).write_text(str(pid), encoding = "utf-8")
+
+
+def _run_stop(studio_mod):
+    import typer as _typer
+
+    app = _typer.Typer()
+    app.add_typer(studio_mod.studio_app, name = "studio")
+    return CliRunner().invoke(app, ["studio", "stop"])
+
+
+def test_stop_kills_every_recorded_server(monkeypatch, tmp_path):
+    studio_mod, _live, killed = _install(monkeypatch, tmp_path, alive = {8550, 8600})
+    _write_pid(tmp_path, "studio-8901-8550.pid", 8550)
+    _write_pid(tmp_path, "studio-8902-8600.pid", 8600)
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert sorted(killed) == [8550, 8600]
+    assert not list(tmp_path.glob("studio-*.pid"))
+
+
+def test_stop_does_not_leave_the_older_instance_running(monkeypatch, tmp_path):
+    # The reported symptom: stop claimed success while instance A kept serving.
+    studio_mod, live, _killed = _install(monkeypatch, tmp_path, alive = {8550, 8600})
+    _write_pid(tmp_path, "studio-8901-8550.pid", 8550)
+    _write_pid(tmp_path, "studio-8902-8600.pid", 8600)
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert live == set()
+
+
+def test_stop_signals_each_server_once(monkeypatch, tmp_path):
+    # A server writes its per-port file AND studio.pid. It stays alive while it
+    # shuts down gracefully, so a second SIGTERM would hit the SIG_DFL the first
+    # one installs and hard-kill it mid-cleanup.
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    monkeypatch.setattr(studio_mod, "_PID_FILE", tmp_path / "studio.pid")
+    monkeypatch.setattr(studio_mod.time, "sleep", lambda _s: None)
+    monkeypatch.setattr(studio_mod, "_pid_alive", lambda pid: True)
+    monkeypatch.setattr(studio_mod, "_pid_is_studio_server", lambda pid, created_times = (): True)
+    killed = []
+    monkeypatch.setattr(studio_mod.os, "kill", lambda pid, _sig: killed.append(pid))
+    monkeypatch.setattr(sys, "platform", "linux")
+    _write_pid(tmp_path, "studio-8901-8550.pid", 8550)
+    _write_pid(tmp_path, "studio.pid", 8550)
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert killed == [8550]
+    assert result.output.lower().count("sent shutdown signal") == 1
+
+
+def test_stop_removes_every_stale_file_for_one_pid(monkeypatch, tmp_path):
+    studio_mod, _live, killed = _install(monkeypatch, tmp_path, alive = set())
+    _write_pid(tmp_path, "studio-8901-8550.pid", 8550)
+    _write_pid(tmp_path, "studio.pid", 8550)
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert killed == []
+    assert not list(tmp_path.glob("*.pid"))
+
+
+def test_stop_does_not_signal_a_reused_pid(monkeypatch, tmp_path):
+    # Crash leaves a per-port file behind, the OS hands that PID to something
+    # else: stop must drop the record, not SIGTERM an unrelated process.
+    studio_mod, _live, killed = _install(monkeypatch, tmp_path, alive = {8550})
+    monkeypatch.setattr(studio_mod, "_pid_is_studio_server", lambda pid, created_times = (): False)
+    _write_pid(tmp_path, "studio-8901-8550.pid", 8550)
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert killed == []
+    assert not (tmp_path / "studio-8901-8550.pid").exists()
+
+
+def test_stop_signals_a_live_server_whose_pid_has_a_stale_record(monkeypatch, tmp_path):
+    # Crash leaves studio-8888-8550.pid, the OS reuses 8550 for a new server on
+    # another port. The stale timestamp must not veto the live one.
+    studio_mod, _live, killed = _install(monkeypatch, tmp_path, alive = {8550})
+    monkeypatch.setattr(studio_mod, "_pid_is_studio_server", _REAL_IS_STUDIO_SERVER)
+
+    class _FakeProcess:
+        def __init__(self, pid):
+            self.pid = pid
+
+        def create_time(self):
+            return 999.0
+
+    monkeypatch.setitem(sys.modules, "psutil", SimpleNamespace(Process = _FakeProcess))
+    (tmp_path / "studio-8888-8550.pid").write_text("8550\n111.5", encoding = "utf-8")
+    (tmp_path / "studio-9000-8550.pid").write_text("8550\n999.0", encoding = "utf-8")
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert killed == [8550]
+    assert not list(tmp_path.glob("studio-*.pid"))
+
+
+def test_a_bare_run_py_command_line_is_not_rejected(monkeypatch):
+    # `cd studio/backend && python run.py --port 8901` has no "studio" or "unsloth"
+    # in argv. Guessing from the command line deleted its record without stopping it.
+    studio_mod = _studio()
+
+    class _FakeProcess:
+        def __init__(self, pid):
+            self.pid = pid
+
+        def cmdline(self):
+            return ["python", "run.py", "--port", "8901"]
+
+        def create_time(self):
+            return 111.5
+
+    monkeypatch.setitem(sys.modules, "psutil", SimpleNamespace(Process = _FakeProcess))
+
+    assert studio_mod._pid_is_studio_server(8550) is True
+
+
+def test_an_untimed_record_is_trusted(monkeypatch):
+    # A legacy `python run.py --port 8901` has no telltale argv, and the in-venv
+    # path runs in-process. Guessing from the command line rejected real servers.
+    studio_mod = _studio()
+
+    assert studio_mod._pid_is_studio_server(8550) is True
+    assert studio_mod._pid_is_studio_server(8550, [None]) is True
+
+
+def test_an_unverifiable_record_is_still_stopped(monkeypatch):
+    # psutil is not a base CLI dependency, so the CLI meets timestamped records it
+    # cannot check. The old `stop` signalled with no checks at all -- skipping one
+    # would leave a live server running, the orphan bug this exists to fix.
+    studio_mod = _studio()
+    monkeypatch.setitem(sys.modules, "psutil", None)
+
+    assert studio_mod._pid_is_studio_server(8550, [111.5]) is True
+    assert studio_mod._pid_is_studio_server(8550, [None]) is True
+
+
+def test_stop_signals_a_timestamped_record_without_psutil(monkeypatch, tmp_path):
+    # Multiple servers on different ports: only the newest is also in studio.pid,
+    # so the earlier ones are timestamp-only and must still be stopped.
+    studio_mod, _live, killed = _install(monkeypatch, tmp_path, alive = {8550})
+    monkeypatch.setattr(studio_mod, "_pid_is_studio_server", _REAL_IS_STUDIO_SERVER)
+    monkeypatch.setitem(sys.modules, "psutil", None)
+    (tmp_path / "studio-8901-8550.pid").write_text("8550\n111.5", encoding = "utf-8")
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert killed == [8550]
+    assert not (tmp_path / "studio-8901-8550.pid").exists()
+
+
+def test_the_untimed_legacy_record_does_not_cancel_a_timed_one(monkeypatch):
+    # Every current server writes BOTH a timed per-port record and an untimed
+    # studio.pid, so letting the untimed half win made this check inert exactly
+    # where it matters: after a crash and a PID reuse, `stop` SIGTERMed whatever
+    # unrelated process had inherited the PID. An untimed record carries no
+    # information, so it must not overrule a start time that says "not ours".
+    studio_mod = _studio()
+
+    class _FakeProcess:
+        def __init__(self, pid):
+            self.pid = pid
+
+        def create_time(self):
+            return 999.0
+
+    monkeypatch.setitem(sys.modules, "psutil", SimpleNamespace(Process = _FakeProcess))
+
+    assert studio_mod._pid_is_studio_server(8550, [111.5, None]) is False
+    assert studio_mod._pid_is_studio_server(8550, [111.5]) is False
+    # A matching time still wins over a stale sibling record.
+    assert studio_mod._pid_is_studio_server(8550, [111.5, 999.0]) is True
+    assert studio_mod._pid_is_studio_server(8550, [None, None]) is True
+
+
+def test_stop_does_not_signal_a_reused_pid_recorded_in_both_files(monkeypatch, tmp_path):
+    # End to end for the case above: a crashed server left studio-8901-8550.pid
+    # and studio.pid, and 8550 now belongs to something else entirely.
+    studio_mod, _live, killed = _install(monkeypatch, tmp_path, alive = {8550})
+    monkeypatch.setattr(studio_mod, "_pid_is_studio_server", _REAL_IS_STUDIO_SERVER)
+
+    class _FakeProcess:
+        def __init__(self, pid):
+            self.pid = pid
+
+        def create_time(self):
+            return 999.0
+
+    monkeypatch.setitem(sys.modules, "psutil", SimpleNamespace(Process = _FakeProcess))
+    (tmp_path / "studio-8901-8550.pid").write_text("8550\n111.5\n127.0.0.1", encoding = "utf-8")
+    (tmp_path / "studio.pid").write_text("8550", encoding = "utf-8")
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert killed == []
+    assert not list(tmp_path.glob("*.pid"))
+
+
+def test_pid_identity_check_trusts_the_record_without_psutil(monkeypatch):
+    # No psutil: fall back to trusting the record rather than never stopping.
+    studio_mod = _studio()
+    monkeypatch.setitem(sys.modules, "psutil", None)
+
+    assert studio_mod._pid_is_studio_server(8550) is True
+
+
+def test_pid_identity_check_uses_the_recorded_start_time(monkeypatch):
+    studio_mod = _studio()
+
+    class _FakeProcess:
+        def __init__(self, pid):
+            self.pid = pid
+
+        def create_time(self):
+            return 111.5
+
+    monkeypatch.setitem(sys.modules, "psutil", SimpleNamespace(Process = _FakeProcess))
+
+    assert studio_mod._pid_is_studio_server(8550, [111.5]) is True
+    assert studio_mod._pid_is_studio_server(8550, [999.0]) is False
+
+
+def test_stop_drops_a_record_whose_start_time_no_longer_matches(monkeypatch, tmp_path):
+    # The PID was reused: same number, different process.
+    studio_mod, _live, killed = _install(monkeypatch, tmp_path, alive = {8550})
+    monkeypatch.setattr(studio_mod, "_pid_is_studio_server", _REAL_IS_STUDIO_SERVER)
+
+    class _FakeProcess:
+        def __init__(self, pid):
+            self.pid = pid
+
+        def create_time(self):
+            return 999.0
+
+    monkeypatch.setitem(sys.modules, "psutil", SimpleNamespace(Process = _FakeProcess))
+    (tmp_path / "studio-8901-8550.pid").write_text("8550\n111.5", encoding = "utf-8")
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert killed == []
+    assert not (tmp_path / "studio-8901-8550.pid").exists()
+    # Dropped for the start-time mismatch, not because the record looked corrupt.
+    assert "invalid pid file" not in result.output.lower()
+
+
+def test_stop_reads_the_legacy_single_pid_file(monkeypatch, tmp_path):
+    studio_mod, _live, killed = _install(monkeypatch, tmp_path, alive = {4242})
+    _write_pid(tmp_path, "studio.pid", 4242)
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert killed == [4242]
+    assert not (tmp_path / "studio.pid").exists()
+
+
+def test_stop_reports_nothing_running_without_pid_files(monkeypatch, tmp_path):
+    studio_mod, _live, _killed = _install(monkeypatch, tmp_path, alive = set())
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert "no running unsloth server" in result.output.lower()
+
+
+def test_stop_cleans_stale_pid_files_without_claiming_a_stop(monkeypatch, tmp_path):
+    studio_mod, _live, killed = _install(monkeypatch, tmp_path, alive = set())
+    _write_pid(tmp_path, "studio-8901-8550.pid", 8550)
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert killed == []
+    assert not (tmp_path / "studio-8901-8550.pid").exists()
+    assert "stopped" not in result.output.lower()
+
+
+def test_stop_does_not_claim_a_stop_while_a_server_is_still_alive(monkeypatch, tmp_path):
+    # SIGTERM delivered but it never exits: don't claim a stop, keep the file.
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    monkeypatch.setattr(studio_mod, "_PID_FILE", tmp_path / "studio.pid")
+    monkeypatch.setattr(studio_mod.time, "sleep", lambda _s: None)
+    monkeypatch.setattr(studio_mod, "_pid_alive", lambda pid: True)
+    monkeypatch.setattr(studio_mod, "_pid_is_studio_server", lambda pid, created_times = (): True)
+    monkeypatch.setattr(studio_mod.os, "kill", lambda pid, sig: None)
+    monkeypatch.setattr(sys, "platform", "linux")
+    _write_pid(tmp_path, "studio-8901-8550.pid", 8550)
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert "shutting down" in result.output.lower()
+    assert "stopped" not in result.output.lower()
+    assert (tmp_path / "studio-8901-8550.pid").exists()
+
+
+def test_stop_continues_after_one_server_fails_to_stop(monkeypatch, tmp_path):
+    studio_mod = _studio()
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", tmp_path)
+    monkeypatch.setattr(studio_mod, "_PID_FILE", tmp_path / "studio.pid")
+    monkeypatch.setattr(studio_mod.time, "sleep", lambda _s: None)
+    live = {8550, 8600}
+    monkeypatch.setattr(studio_mod, "_pid_alive", lambda pid: pid in live)
+    monkeypatch.setattr(studio_mod, "_pid_is_studio_server", lambda pid, created_times = (): True)
+
+    def fake_kill(pid, _sig):
+        if pid == 8550:
+            raise PermissionError("not permitted")
+        live.discard(pid)
+
+    monkeypatch.setattr(studio_mod.os, "kill", fake_kill)
+    monkeypatch.setattr(sys, "platform", "linux")
+    _write_pid(tmp_path, "studio-8901-8550.pid", 8550)
+    _write_pid(tmp_path, "studio-8902-8600.pid", 8600)
+
+    result = _run_stop(studio_mod)
+
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert result.exit_code == 1, combined
+    assert 8600 not in live
+    assert "8550" in combined
+
+
+def test_stop_never_signals_pid_zero_or_init(monkeypatch, tmp_path):
+    # os.kill(0, SIGTERM) hits our whole process group -- the shell and its jobs.
+    studio_mod, _live, killed = _install(monkeypatch, tmp_path, alive = {0, 1})
+    _write_pid(tmp_path, "studio-8901-0.pid", 0)
+    _write_pid(tmp_path, "studio-8902-1.pid", 1)
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert killed == []
+    assert not list(tmp_path.glob("*.pid"))
+
+
+def test_signal_stop_refuses_pid_zero_or_init(monkeypatch, tmp_path):
+    studio_mod, _live, killed = _install(monkeypatch, tmp_path, alive = {0, 1})
+
+    assert studio_mod._signal_stop(0) is not None
+    assert studio_mod._signal_stop(1) is not None
+    assert killed == []
+
+
+def test_stop_discards_a_corrupt_pid_file(monkeypatch, tmp_path):
+    studio_mod, _live, _killed = _install(monkeypatch, tmp_path, alive = set())
+    (tmp_path / "studio-8901-8550.pid").write_text("not-a-pid", encoding = "utf-8")
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert not (tmp_path / "studio-8901-8550.pid").exists()
+
+
+def test_stop_keeps_a_record_it_cannot_read(monkeypatch, tmp_path):
+    # A root-owned record, or one caught mid-write, still belongs to a live
+    # server. Deleting it is `stop` manufacturing the orphan it exists to fix.
+    studio_mod, _live, killed = _install(monkeypatch, tmp_path, alive = {8550})
+    path = tmp_path / "studio-8901-8550.pid"
+    path.write_text("8550", encoding = "utf-8")
+    real_read_text = Path.read_text
+
+    def deny(self, *args, **kwargs):
+        if self == path:
+            raise PermissionError(13, "Permission denied")
+        return real_read_text(self, *args, **kwargs)
+
+    monkeypatch.setattr(Path, "read_text", deny)
+
+    result = _run_stop(studio_mod)
+
+    assert path.exists(), "an unreadable record must not be deleted"
+    assert "cannot read" in (result.output + (result.stderr or "")).lower()
+
+
+def test_stop_does_not_claim_success_when_the_only_record_is_unreadable(monkeypatch, tmp_path):
+    # A server started under sudo leaves a record we cannot read. Printing "no
+    # running server" and exiting 0 tells the user the opposite of the truth.
+    studio_mod, _live, killed = _install(monkeypatch, tmp_path, alive = {8550})
+    path = tmp_path / "studio-8901-8550.pid"
+    path.write_text("8550", encoding = "utf-8")
+    real_read_text = Path.read_text
+
+    def deny(self, *args, **kwargs):
+        if self == path:
+            raise PermissionError(13, "Permission denied")
+        return real_read_text(self, *args, **kwargs)
+
+    monkeypatch.setattr(Path, "read_text", deny)
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 1, "an unreachable server is not a successful stop"
+    output = result.output + (result.stderr or "")
+    assert "no running unsloth server" not in output.lower()
+    assert killed == []
+
+
+def test_stop_reports_failure_when_one_record_is_unreadable_but_another_stops(
+    monkeypatch, tmp_path
+):
+    # Stopping the servers we can see is still a partial result, and exiting 0
+    # would hide the one we could not.
+    studio_mod, _live, killed = _install(monkeypatch, tmp_path, alive = {8550, 8600})
+    _write_pid(tmp_path, "studio-8901-8550.pid", 8550)
+    hidden = tmp_path / "studio-8902-8600.pid"
+    hidden.write_text("8600", encoding = "utf-8")
+    real_read_text = Path.read_text
+
+    def deny(self, *args, **kwargs):
+        if self == hidden:
+            raise PermissionError(13, "Permission denied")
+        return real_read_text(self, *args, **kwargs)
+
+    monkeypatch.setattr(Path, "read_text", deny)
+
+    result = _run_stop(studio_mod)
+
+    assert killed == [8550], "the readable server must still be stopped"
+    assert result.exit_code == 1
+    assert hidden.exists()
+
+
+def test_stop_reaches_every_server_when_one_record_cannot_be_removed(monkeypatch, tmp_path):
+    # One undeletable stale record must not end the loop before the live servers.
+    studio_mod, _live, killed = _install(monkeypatch, tmp_path, alive = {8600})
+    _write_pid(tmp_path, "studio-8901-8550.pid", 8550)  # dead -> stop prunes it
+    _write_pid(tmp_path, "studio-8902-8600.pid", 8600)  # live -> stop signals it
+    real_unlink = Path.unlink
+
+    def deny(self, *args, **kwargs):
+        if self.name == "studio-8901-8550.pid":
+            raise PermissionError(13, "Permission denied")
+        return real_unlink(self, *args, **kwargs)
+
+    monkeypatch.setattr(Path, "unlink", deny)
+
+    result = _run_stop(studio_mod)
+
+    assert killed == [8600], "the live server must still be signalled"
+    assert result.exit_code == 0, result.output
+
+
+def test_a_record_whose_pid_is_not_ascii_digits_is_discarded(monkeypatch, tmp_path):
+    # A superscript two passes isdigit() but int() rejects it, so that gate alone
+    # let a ValueError escape _read_pid_record and abort the whole command.
+    studio_mod, _live, _killed = _install(monkeypatch, tmp_path, alive = set())
+    (tmp_path / "studio-8901-1.pid").write_text("²", encoding = "utf-8")
+
+    result = _run_stop(studio_mod)
+
+    assert result.exit_code == 0, result.output
+    assert not (tmp_path / "studio-8901-1.pid").exists()