Merge remote-tracking branch 'origin/main' into feature/rag

# Conflicts:
#	studio/backend/core/inference/tools.py
#	studio/frontend/src/components/assistant-ui/sources.tsx
#	studio/frontend/src/components/assistant-ui/thread.tsx
#	studio/frontend/src/features/chat/api/chat-adapter.ts
#	studio/frontend/src/features/chat/chat-settings-sheet.tsx
#	studio/frontend/src/features/chat/shared-composer.tsx
#	studio/frontend/src/features/chat/stores/chat-runtime-store.ts

.github/workflows/security-audit.yml

@@ -72,6 +72,31 @@ concurrency:
 permissions:
   contents: read
 
+# ──────────────────────────────────────────────────────────────────────
+# Network-resilience knobs, applied to every job/step. These add retries
+# and backoff ONLY; they do not relax a single integrity check. cargo
+# still resolves against Cargo.lock (--locked), pip still verifies the
+# wheels it downloads, npm still enforces package-lock integrity, the
+# harden-runner egress allowlists below are unchanged, and every action
+# stays SHA-pinned. The advisory-audit run on 2026-05-29 red-failed when
+# one crates.io tarball fetch hit "Recv failure: Connection reset by
+# peer" (curl 56); cargo's default of 3 retries over an HTTP/2-multiplexed
+# connection did not recover. The settings below make that class of
+# transient fault self-heal instead of failing the whole run.
+env:
+  # pip: raise the built-in retry count and per-connection timeout.
+  PIP_RETRIES: "10"
+  PIP_DEFAULT_TIMEOUT: "60"
+  # cargo: retry network ops and disable HTTP/2 multiplexing -- the
+  # documented mitigation for the curl-56 connection resets above.
+  CARGO_NET_RETRY: "10"
+  CARGO_HTTP_MULTIPLEXING: "false"
+  CARGO_NET_GIT_FETCH_WITH_CLI: "true"
+  # npm: retry registry fetches with capped exponential backoff.
+  NPM_CONFIG_FETCH_RETRIES: "5"
+  NPM_CONFIG_FETCH_RETRY_MINTIMEOUT: "2000"
+  NPM_CONFIG_FETCH_RETRY_MAXTIMEOUT: "60000"
+
 jobs:
   # ─────────────────────────────────────────────────────────────────────
   # Combined advisory-DB audit: pip-audit + npm audit + cargo audit
@@ -140,7 +165,7 @@ jobs:
 
       - uses: dtolnay/rust-toolchain@29eef336d9b2848a0b548edc03f92a220660cdb8  # stable @ 2026-03-27
 
-      - uses: swatinem/rust-cache@e18b497796c12c097a38f9edb9d0641fb99eee32  # v2.9.1
+      - uses: swatinem/rust-cache@c19371144df3bb44fab255c43d04cbc2ab54d1c4  # v2.9.1
         with:
           workspaces: studio/src-tauri -> target
 
@@ -153,8 +178,23 @@ jobs:
         # crashes with a TOML parse error on that file.
         # npm audit is bundled with the node toolchain, no install.
         run: |
-          python -m pip install --upgrade pip 'pip-audit>=2.7'
-          cargo install --locked --version '^0.22' cargo-audit
+          retry() {  # retry <max-attempts> <command...> with exponential backoff
+            local max="$1"; shift
+            local n=1 delay=5
+            until "$@"; do
+              if [ "$n" -ge "$max" ]; then
+                echo "::error::command failed after ${n} attempts: $*" >&2
+                return 1
+              fi
+              echo "attempt ${n}/${max} failed; retrying in ${delay}s: $*" >&2
+              sleep "$delay"; n=$((n + 1)); delay=$((delay * 2))
+            done
+          }
+          retry 5 python -m pip install --upgrade pip 'pip-audit>=2.7'
+          # --locked keeps the resolved tree identical to Cargo.lock; the
+          # CARGO_NET_* env above plus this outer loop survive transient
+          # crates.io connection resets without weakening that guarantee.
+          retry 5 cargo install --locked --version '^0.22' cargo-audit
 
       # ─────────────────────────────────────────────────────────────
       # Python: pip-audit
@@ -330,32 +370,60 @@ jobs:
       # ─────────────────────────────────────────────────────────────
       # OSV-Scanner: cross-ecosystem advisory DB (PyPI + npm + cargo)
       # ─────────────────────────────────────────────────────────────
+      - name: Download + verify OSV-Scanner
+        # Split out from the scan below so binary integrity is a HARD gate:
+        # a checksum mismatch (swapped release asset, the Trivy-style pivot
+        # this workflow refuses) fails the job instead of being swallowed by
+        # the scan step's continue-on-error. A download still failing after
+        # retries is transient, so we skip the scan rather than red-fail.
+        # SHA-256 verified BEFORE chmod +x / exec. Bump OSV_SHA256 in lockstep
+        # with OSV_VERSION (value from the release's osv-scanner_SHA256SUMS).
+        run: |
+          set -euo pipefail
+          OSV_VERSION="v2.0.2"
+          OSV_SHA256="3abcfd7126c453a00421487e721b296e0cb68085bd431d6cef60872774170fc8"
+          if ! curl --proto '=https' --tlsv1.2 -fsSL \
+            --retry 5 --retry-delay 3 --retry-connrefused --retry-all-errors \
+            -o /tmp/osv-scanner \
+            "https://github.com/google/osv-scanner/releases/download/${OSV_VERSION}/osv-scanner_linux_amd64"; then
+            echo "::warning::osv-scanner download failed after retries; skipping scan" >&2
+            rm -f /tmp/osv-scanner
+            exit 0   # transient availability: do not red-fail the job
+          fi
+          if ! echo "${OSV_SHA256}  /tmp/osv-scanner" | sha256sum -c -; then
+            echo "::error::osv-scanner checksum mismatch; refusing to execute" >&2
+            rm -f /tmp/osv-scanner
+            exit 1   # integrity failure: hard-fail
+          fi
+          chmod +x /tmp/osv-scanner
+          /tmp/osv-scanner --version
+
       - name: OSV-Scanner (PyPI + npm + cargo, cross-ecosystem advisories)
         # OSV's advisory feed is a superset of GitHub-Advisory + RustSec
         # + npm advisories; running it alongside the per-ecosystem audit
         # tools catches CVEs that haven't propagated to the per-ecosystem
         # DBs yet (e.g. langchain-core CVE-2025-68664 was on OSV before
         # GitHub Advisory). Single binary, one transitive resolver, all
-        # three lockfile types in one pass. Non-blocking until baselines
-        # close.
+        # three lockfile types in one pass. Binary is checksum-verified in
+        # the step above; only the advisory scan stays non-blocking until
+        # baselines close.
         continue-on-error: true
         run: |
           set +e
-          # OSV-Scanner ships a raw binary (no tarball) in v2.x.
-          curl -fsSL -o /tmp/osv-scanner \
-            https://github.com/google/osv-scanner/releases/download/v2.0.2/osv-scanner_linux_amd64
-          chmod +x /tmp/osv-scanner
-          /tmp/osv-scanner --version
-          /tmp/osv-scanner scan source \
-            --lockfile=studio/frontend/package-lock.json \
-            --lockfile=studio/src-tauri/Cargo.lock \
-            --lockfile=requirements.txt:audit-reqs/unsloth-deps.txt \
-            --lockfile=requirements.txt:audit-reqs/studio.txt \
-            --lockfile=requirements.txt:audit-reqs/no-torch-runtime.txt \
-            --lockfile=requirements.txt:audit-reqs/overrides.txt \
-            --lockfile=requirements.txt:audit-reqs/extras.txt \
-            --lockfile=requirements.txt:audit-reqs/extras-no-deps.txt \
-            --format=table 2>&1 | tee logs-osv-scanner.txt
+          if [ ! -x /tmp/osv-scanner ]; then
+            echo "osv-scanner unavailable this run; skipping scan" | tee logs-osv-scanner.txt
+          else
+            /tmp/osv-scanner scan source \
+              --lockfile=studio/frontend/package-lock.json \
+              --lockfile=studio/src-tauri/Cargo.lock \
+              --lockfile=requirements.txt:audit-reqs/unsloth-deps.txt \
+              --lockfile=requirements.txt:audit-reqs/studio.txt \
+              --lockfile=requirements.txt:audit-reqs/no-torch-runtime.txt \
+              --lockfile=requirements.txt:audit-reqs/overrides.txt \
+              --lockfile=requirements.txt:audit-reqs/extras.txt \
+              --lockfile=requirements.txt:audit-reqs/extras-no-deps.txt \
+              --format=table 2>&1 | tee logs-osv-scanner.txt
+          fi
           {
             echo "## OSV-Scanner (cross-ecosystem)"
             echo
@@ -1075,7 +1143,23 @@ jobs:
         # new-install-script gate below protects against, and we must
         # not run any third-party hook to set up the audit.
         working-directory: studio/frontend
-        run: npm ci --ignore-scripts
+        run: |
+          retry() {  # retry <max-attempts> <command...> with exponential backoff
+            local max="$1"; shift
+            local n=1 delay=5
+            until "$@"; do
+              if [ "$n" -ge "$max" ]; then
+                echo "::error::command failed after ${n} attempts: $*" >&2
+                return 1
+              fi
+              echo "attempt ${n}/${max} failed; retrying in ${delay}s: $*" >&2
+              sleep "$delay"; n=$((n + 1)); delay=$((delay * 2))
+            done
+          }
+          # --ignore-scripts is mandatory here (no third-party hook runs);
+          # the retry only re-attempts the registry fetch, it never relaxes
+          # that flag or the package-lock integrity check npm ci enforces.
+          retry 5 npm ci --ignore-scripts
 
       - name: npm audit signatures (informational)
         # Surfaces unsigned / mis-signed packages from the npm

.github/workflows/studio-inference-smoke.yml

@@ -296,6 +296,8 @@ jobs:
       - name: Upload logs
         # Always upload so green runs are still reviewable.
         if: always()
+        # Diagnostic only: a transient artifact-service drop must not fail a green job.
+        continue-on-error: true
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
         with:
           name: openai-anthropic-log
@@ -771,6 +773,8 @@ jobs:
       - name: Upload logs
         # Always upload so green runs are still reviewable.
         if: always()
+        # Diagnostic only: a transient artifact-service drop must not fail a green job.
+        continue-on-error: true
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
         with:
           name: tool-calling-log
@@ -1043,6 +1047,8 @@ jobs:
       - name: Upload logs
         # Always upload so green runs are still reviewable.
         if: always()
+        # Diagnostic only: a transient artifact-service drop must not fail a green job.
+        continue-on-error: true
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
         with:
           name: json-images-log

.github/workflows/studio-mac-inference-smoke.yml

@@ -289,6 +289,8 @@ jobs:
       - name: Upload logs
         # Always upload so green runs are still reviewable.
         if: always()
+        # Diagnostic only: a transient artifact-service drop must not fail a green job.
+        continue-on-error: true
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
         with:
           name: openai-anthropic-log
@@ -649,6 +651,8 @@ jobs:
       - name: Upload logs
         # Always upload so green runs are still reviewable.
         if: always()
+        # Diagnostic only: a transient artifact-service drop must not fail a green job.
+        continue-on-error: true
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
         with:
           name: tool-calling-log
@@ -1025,6 +1029,8 @@ jobs:
       - name: Upload logs
         # Always upload so green runs are still reviewable.
         if: always()
+        # Diagnostic only: a transient artifact-service drop must not fail a green job.
+        continue-on-error: true
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
         with:
           name: json-images-log

.github/workflows/studio-windows-inference-smoke.yml

@@ -65,6 +65,18 @@ jobs:
         with:
           persist-credentials: false
 
+      # Fast GPU-free gate: parse setup.ps1 and run the Resolve-CudaToolkit unit
+      # test (deferred Windows CUDA Toolkit check) before the heavy GGUF smoke.
+      - name: setup.ps1 unit test (Resolve-CudaToolkit)
+        shell: pwsh
+        run: |
+          $errs = $null
+          [void][System.Management.Automation.Language.Parser]::ParseFile(
+            (Resolve-Path studio/setup.ps1).Path, [ref]$null, [ref]$errs)
+          if ($errs) { $errs | ForEach-Object { $_.ToString() }; exit 1 }
+          Write-Host "setup.ps1 parsed with no errors"
+          pwsh -NoProfile -File tests/studio/test_resolve_cuda_toolkit.ps1
+
       - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
         with:
           node-version: '22'
@@ -360,6 +372,9 @@ jobs:
 
       - name: Collect llama-server logs
         if: always()
+        # A transient Windows DLL-init crash (0xC0000142) in this diagnostic
+        # copy must not fail an otherwise-green job.
+        continue-on-error: true
         shell: bash
         # Copy llama-server's own stdout/stderr (teed by Studio under
         # ~/.unsloth/studio/logs/llama-server/) into the workspace so
@@ -373,6 +388,8 @@ jobs:
 
       - name: Upload logs
         if: always()
+        # Diagnostic only: a transient artifact-service drop must not fail a green job.
+        continue-on-error: true
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
         with:
           name: windows-openai-anthropic-log
@@ -788,6 +805,9 @@ jobs:
 
       - name: Collect llama-server logs
         if: always()
+        # A transient Windows DLL-init crash (0xC0000142) in this diagnostic
+        # copy must not fail an otherwise-green job.
+        continue-on-error: true
         shell: bash
         # Copy llama-server's own stdout/stderr (teed by Studio under
         # ~/.unsloth/studio/logs/llama-server/) into the workspace so
@@ -801,6 +821,8 @@ jobs:
 
       - name: Upload logs
         if: always()
+        # Diagnostic only: a transient artifact-service drop must not fail a green job.
+        continue-on-error: true
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
         with:
           name: windows-tool-calling-log
@@ -1186,6 +1208,9 @@ jobs:
 
       - name: Collect llama-server logs
         if: always()
+        # A transient Windows DLL-init crash (0xC0000142) in this diagnostic
+        # copy must not fail an otherwise-green job.
+        continue-on-error: true
         shell: bash
         # Copy llama-server's own stdout/stderr (teed by Studio under
         # ~/.unsloth/studio/logs/llama-server/) into the workspace so
@@ -1199,6 +1224,8 @@ jobs:
 
       - name: Upload logs
         if: always()
+        # Diagnostic only: a transient artifact-service drop must not fail a green job.
+        continue-on-error: true
         uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a  # v7.0.1
         with:
           name: windows-json-images-log

README.md

@@ -72,10 +72,13 @@ Unsloth Studio (Beta) works on **Windows, Linux, WSL** and **macOS**.
 ```bash
 curl -fsSL https://unsloth.ai/install.sh | sh
 ```
+Use the same command to update.
+
 #### Windows:
 ```powershell
 irm https://unsloth.ai/install.ps1 | iex
 ```
+Use the same command to update.
 
 #### Launch
 ```bash
@@ -83,9 +86,6 @@ unsloth studio -p 8888
 ```
 For cloud or global access, add `-H 0.0.0.0`. By default, Unsloth is accessible only locally.
 
-#### Update
-To update, use the same install commands above or use `unsloth studio update`.
-
 #### Docker
 Use our [Docker image](https://hub.docker.com/r/unsloth/unsloth) ```unsloth/unsloth``` container. Run:
 ```bash
@@ -171,7 +171,9 @@ unsloth studio -p 8888
 ```
 Then to update :
 ```bash
-unsloth studio update
+cd unsloth && git pull
+./install.sh --local
+unsloth studio -p 8888
 ```
 
 #### Developer installs: Windows PowerShell:
@@ -184,7 +186,9 @@ unsloth studio -p 8888
 ```
 Then to update :
 ```bash
-unsloth studio update
+cd unsloth && git pull
+./install.sh --local
+unsloth studio -p 8888
 ```
 
 #### Nightly: MacOS, Linux, WSL:

install.ps1

@@ -1566,7 +1566,7 @@ shell.Run cmd, 0, False
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.5.8" unsloth-zoo }
+            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.5.10" unsloth-zoo }
             if ($baseInstallExit -eq 0) {
                 # Resolve pydantic WITH deps so pip pins pydantic-core
                 # to the matching version (no-torch-runtime.txt below
@@ -1580,7 +1580,7 @@ shell.Run cmd, 0, False
                 }
             }
         } else {
-            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.5.8" unsloth-zoo }
+            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.5.10" unsloth-zoo }
         }
         if ($baseInstallExit -ne 0) {
             Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
@@ -1627,7 +1627,7 @@ shell.Run cmd, 0, False
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.5.8" unsloth-zoo }
+            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.5.10" unsloth-zoo }
             if ($baseInstallExit -eq 0) {
                 # Same pydantic-with-deps trick as the migrated branch.
                 $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython pydantic }
@@ -1639,7 +1639,7 @@ shell.Run cmd, 0, False
                 }
             }
         } elseif ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.5.8" unsloth-zoo }
+            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.5.10" unsloth-zoo }
         } else {
             $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --upgrade-package unsloth -- "$PackageName" }
         }
@@ -1667,7 +1667,7 @@ shell.Run cmd, 0, False
         Write-TauriLog "STEP" "Installing unsloth"
         substep "installing unsloth (this may take a few minutes)..."
         if ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython unsloth-zoo "unsloth>=2026.5.8" --torch-backend=auto }
+            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython unsloth-zoo "unsloth>=2026.5.10" --torch-backend=auto }
             if ($baseInstallExit -ne 0) {
                 Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
                 return (Exit-InstallFailure "Failed to install unsloth (exit code $baseInstallExit)" $baseInstallExit)

install.sh

@@ -1530,17 +1530,51 @@ if [ -x "$VENV_DIR/bin/python" ]; then
     : > "$VENV_DIR/.unsloth-studio-owned" 2>/dev/null || true
 fi
 
-# Guard against Python 3.13.8 torch import bug on Apple Silicon
-# (skip when the user explicitly chose a version via --python)
+# Guard against two independent Apple Silicon venv problems, in order:
+#   1. uv may create the venv from a cached x86_64 (Rosetta) Python when a
+#      same-version x86_64 build is already cached (often because uv itself
+#      is an x86_64 build). That venv reports x86_64 to wheel resolvers, and
+#      PyTorch ships no macOS wheels on the CPU index for any architecture,
+#      so the torch install can never resolve. Recreate it with an
+#      arch-explicit arm64 CPython.
+#   2. Python 3.13.8 has a known torch import bug.
+# The two are independent: a venv may be x86_64 and, once recreated, still
+# land on 3.13.8. So we re-inspect the interpreter between the checks instead
+# of chaining them with elif, guaranteeing both invariants hold on whatever
+# venv we end up with. Skip both when the user explicitly chose an interpreter
+# via --python.
 if [ -z "$_USER_PYTHON" ] && [ "$OS" = "macos" ] && [ "$_ARCH" = "arm64" ]; then
-    _PY_VER=$("$VENV_DIR/bin/python" -c \
-        "import sys; print('{}.{}.{}'.format(*sys.version_info[:3]))" 2>/dev/null || echo "")
+    _inspect_venv() {
+        "$VENV_DIR/bin/python" -c \
+            "import platform, sys; print(platform.machine(), '{}.{}.{}'.format(*sys.version_info[:3]))" \
+            2>/dev/null || echo " "
+    }
+    _info=$(_inspect_venv)
+    _VENV_ARCH=${_info%% *}
+    _PY_VER=${_info##* }
+
+    if [ "$_VENV_ARCH" = "x86_64" ]; then
+        echo "  WARNING: venv was created with an x86_64 (Rosetta) Python on Apple Silicon."
+        echo "  Recreating venv with native arm64 Python ${PYTHON_VERSION}..."
+        rm -rf "$VENV_DIR"
+        run_install_cmd "recreate venv (arm64)" uv venv "$VENV_DIR" \
+            --python "cpython-${PYTHON_VERSION}-macos-aarch64-none"
+        if [ -x "$VENV_DIR/bin/python" ]; then
+            : > "$VENV_DIR/.unsloth-studio-owned" 2>/dev/null || true
+        fi
+        # Re-inspect: the recreated arm64 venv may still be 3.13.8.
+        _info=$(_inspect_venv)
+        _VENV_ARCH=${_info%% *}
+        _PY_VER=${_info##* }
+    fi
+
     if [ "$_PY_VER" = "3.13.8" ]; then
         echo "  WARNING: Python 3.13.8 has a known torch import bug."
         echo "  Recreating venv with Python 3.12..."
         rm -rf "$VENV_DIR"
         PYTHON_VERSION="3.12"
-        run_install_cmd "recreate venv" uv venv "$VENV_DIR" --python "$PYTHON_VERSION"
+        run_install_cmd "recreate venv" uv venv "$VENV_DIR" \
+            --python "cpython-${PYTHON_VERSION}-macos-aarch64-none"
         if [ -x "$VENV_DIR/bin/python" ]; then
             : > "$VENV_DIR/.unsloth-studio-owned" 2>/dev/null || true
         fi
@@ -2049,7 +2083,7 @@ if [ "$_MIGRATED" = true ]; then
         # to prevent transitive torch resolution.
         run_install_cmd "install unsloth (migrated no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.5.8" unsloth-zoo
+            "unsloth>=2026.5.10" unsloth-zoo
         # Resolve pydantic WITH deps so pip pins pydantic-core to the
         # matching version (no-torch-runtime.txt below is --no-deps).
         # All transitive deps are torch-free.
@@ -2062,7 +2096,7 @@ if [ "$_MIGRATED" = true ]; then
     else
         run_install_cmd "install unsloth (migrated)" uv pip install --python "$_VENV_PY" \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.5.8" unsloth-zoo
+            "unsloth>=2026.5.10" unsloth-zoo
     fi
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         substep "overlaying local repo (editable)..."
@@ -2266,7 +2300,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
         run_install_cmd "install unsloth (no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --upgrade-package unsloth --upgrade-package unsloth-zoo \
-            "unsloth>=2026.5.8" unsloth-zoo
+            "unsloth>=2026.5.10" unsloth-zoo
         # Same pydantic-with-deps trick as the migrated branch.
         run_install_cmd "install pydantic (with deps for compatible core)" \
             uv pip install --python "$_VENV_PY" pydantic
@@ -2284,7 +2318,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         fi
     elif [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         run_install_cmd "install unsloth (local)" uv pip install --python "$_VENV_PY" \
-            --upgrade-package unsloth "unsloth>=2026.5.8" unsloth-zoo
+            --upgrade-package unsloth "unsloth>=2026.5.10" unsloth-zoo
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."
@@ -2316,7 +2350,7 @@ else
     tauri_log "STEP" "Installing Unsloth"
     substep "installing unsloth (this may take a few minutes)..."
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
-        run_install_cmd "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" unsloth-zoo "unsloth>=2026.5.8" --torch-backend=auto
+        run_install_cmd "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" unsloth-zoo "unsloth>=2026.5.10" --torch-backend=auto
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."

pyproject.toml

@@ -69,7 +69,7 @@ triton = [
 ]
 
 huggingfacenotorch = [
-    "unsloth_zoo>=2026.5.4",
+    "unsloth_zoo>=2026.5.5",
     "wheel>=0.42.0",
     "packaging",
     "numpy",
@@ -90,7 +90,7 @@ huggingfacenotorch = [
 ]
 huggingface = [
     "unsloth[huggingfacenotorch]",
-    "unsloth_zoo>=2026.5.4",
+    "unsloth_zoo>=2026.5.5",
     "torchvision",
     "unsloth[triton]",
 ]
@@ -580,7 +580,7 @@ colab-ampere-torch220 = [
     "flash-attn>=2.6.3 ; ('linux' in sys_platform)",
 ]
 colab-new = [
-    "unsloth_zoo>=2026.5.4",
+    "unsloth_zoo>=2026.5.5",
     "packaging",
     "tyro",
     "transformers>=4.51.3,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0,!=4.54.0,!=4.55.0,!=4.55.1,!=4.57.0,!=4.57.4,!=4.57.5,!=5.0.0,!=5.1.0,<=5.5.0",

studio/backend/core/data_recipe/service.py

@@ -176,12 +176,21 @@ def build_mcp_providers(
 ) -> list:
     from data_designer.config.mcp import LocalStdioMCPProvider, MCPProvider  # pyright: ignore[reportMissingImports]
 
+    # Same gate as the chat MCP path: stdio providers spawn a local subprocess,
+    # so only build them when this host allows it (desktop / explicit opt-in).
+    # Skip them otherwise so a recipe carried onto a hosted host cannot spawn.
+    from core.inference.mcp_client import stdio_mcp_enabled
+
+    stdio_allowed = stdio_mcp_enabled()
+
     providers: list[MCPProvider | LocalStdioMCPProvider] = []
     for provider in recipe.get("mcp_providers", []):
         if not isinstance(provider, dict):
             continue
         provider_type = provider.get("provider_type")
         if provider_type == "stdio":
+            if not stdio_allowed:
+                continue
             env = provider.get("env")
             if not isinstance(env, dict):
                 env = {}

studio/backend/core/inference/anthropic_compat.py

@@ -218,6 +218,8 @@ class AnthropicStreamEmitter:
     def __init__(self) -> None:
         self.block_index: int = 0
         self._text_block_open: bool = False
+        self._open_tool_call_id: Optional[str] = None
+        self._open_tool_args_sent: bool = False
         self._prev_text: str = ""
         self._usage: dict = {}
 
@@ -263,8 +265,10 @@ class AnthropicStreamEmitter:
     def finish(self, stop_reason: str = "end_turn") -> list[str]:
         """Close any open block and emit message_delta + message_stop."""
         events = []
-        if self._text_block_open:
+        if self._text_block_open or self._open_tool_call_id is not None:
             events.append(self._close_block())
+            self._open_tool_call_id = None
+            self._open_tool_args_sent = False
         events.append(
             build_anthropic_sse_event(
                 "message_delta",
@@ -310,12 +314,26 @@ class AnthropicStreamEmitter:
         return events
 
     def _handle_tool_start(self, event: dict) -> list[str]:
+        tool_call_id = event.get("tool_call_id", "")
+        args = event.get("arguments", {})
+        if tool_call_id and self._open_tool_call_id == tool_call_id:
+            return self._tool_arguments_delta(args)
+
         events = []
-        # Close current text block if open
+        # Close current text block if open.
         if self._text_block_open:
             events.append(self._close_block())
-        # Open a tool_use block
+        # Defensive: if a replacement/different tool_start arrives while a
+        # tool_use block is open, close the stale block before starting another.
+        elif self._open_tool_call_id is not None:
+            events.append(self._close_block())
+            self._open_tool_call_id = None
+            self._open_tool_args_sent = False
+
+        # Open a tool_use block.
         self.block_index += 1
+        self._open_tool_call_id = tool_call_id
+        self._open_tool_args_sent = False
         events.append(
             build_anthropic_sse_event(
                 "content_block_start",
@@ -324,35 +342,43 @@ class AnthropicStreamEmitter:
                     "index": self.block_index,
                     "content_block": {
                         "type": "tool_use",
-                        "id": event.get("tool_call_id", ""),
+                        "id": tool_call_id,
                         "name": event.get("tool_name", ""),
                         "input": {},
                     },
                 },
             )
         )
-        # Emit the arguments as input_json_delta
-        args = event.get("arguments", {})
-        if args:
-            events.append(
-                build_anthropic_sse_event(
-                    "content_block_delta",
-                    {
-                        "type": "content_block_delta",
-                        "index": self.block_index,
-                        "delta": {
-                            "type": "input_json_delta",
-                            "partial_json": json.dumps(args),
-                        },
-                    },
-                )
-            )
+        events.extend(self._tool_arguments_delta(args))
         return events
 
+    def _tool_arguments_delta(self, args: dict) -> list[str]:
+        if not args:
+            return []
+        if self._open_tool_args_sent:
+            return []
+        self._open_tool_args_sent = True
+        return [
+            build_anthropic_sse_event(
+                "content_block_delta",
+                {
+                    "type": "content_block_delta",
+                    "index": self.block_index,
+                    "delta": {
+                        "type": "input_json_delta",
+                        "partial_json": json.dumps(args),
+                    },
+                },
+            )
+        ]
+
     def _handle_tool_end(self, event: dict) -> list[str]:
         events = []
-        # Close the tool_use block
-        events.append(self._close_block())
+        # Close the tool_use block.
+        if self._open_tool_call_id is not None or self._text_block_open:
+            events.append(self._close_block())
+        self._open_tool_call_id = None
+        self._open_tool_args_sent = False
         # Emit custom tool_result event (non-standard, ignored by SDKs)
         events.append(
             build_anthropic_sse_event(

studio/backend/core/inference/llama_cpp.py

@@ -52,6 +52,7 @@ from utils.subprocess_compat import (
     windows_hidden_subprocess_kwargs as _windows_hidden_subprocess_kwargs,
 )
 from core.inference.tool_call_parser import (
+    RENDER_HTML_REPEAT_NUDGE,
     parse_tool_calls_from_text as _shared_parse_tool_calls_from_text,
 )
 
@@ -2604,6 +2605,105 @@ class LlamaCppBackend:
 
     # ── Lifecycle ─────────────────────────────────────────────────
 
+    # GGUF ``general.architecture`` values for diffusion / image models.
+    # llama.cpp proper has no such architectures, so loading one as a chat
+    # model dies with "unknown model architecture: '<arch>'". These match
+    # the patched stable-diffusion.cpp / ComfyUI-GGUF enums (LLM_ARCH_FLUX,
+    # LLM_ARCH_QWEN_IMAGE, ...). Unsloth publishes FLUX and Qwen-Image GGUFs
+    # under https://huggingface.co/collections/unsloth/unsloth-diffusion-ggufs.
+    # Matched exactly (not as a substring) so a chat arch merely containing a
+    # short token like "wan"/"sd1" (e.g. "taiwan") is not misrouted to Images.
+    _DIFFUSION_ARCHES = frozenset(
+        (
+            "qwen_image",
+            "flux",
+            "sd1",
+            "sdxl",
+            "sd3",
+            "aura",
+            "hidream",
+            "cosmos",
+            "ltxv",
+            "hyvid",
+            "wan",
+            "lumina2",
+        )
+    )
+
+    @staticmethod
+    def _classify_llama_start_failure(
+        output: str,
+        gguf_path: Optional[str],
+        model_identifier: Optional[str],
+    ) -> str:
+        """Explain *why* llama-server failed to start, from its output.
+
+        Several distinct failures all otherwise collapse into the same
+        opaque "invalid GGUF or out of memory" message. The worst case is
+        a diffusion / image GGUF (FLUX, Qwen-Image, ...) loaded as a chat
+        model: the file is perfectly valid and there is plenty of memory,
+        but llama.cpp has no such architecture, so the user is told to free
+        memory that was never the problem (issue #5842). Pick the most
+        specific message the captured output supports.
+        """
+        lowered = (output or "").lower()
+
+        # Detect Ollama source up front so the arch branch can keep the
+        # Ollama hint instead of the generic "unsupported arch" message.
+        gguf = gguf_path or ""
+        is_ollama = (
+            ".studio_links" in gguf
+            or os.sep + "ollama_links" + os.sep in gguf
+            or os.sep + ".cache" + os.sep + "ollama" + os.sep in gguf
+            or (model_identifier or "").startswith("ollama/")
+        )
+
+        # "unknown model architecture: '<arch>'": diffusion -> Images page,
+        # Ollama -> Ollama hint, else a precise "unsupported" message. Exact
+        # match so chat archs are never misrouted.
+        arch_match = re.search(r"unknown model architecture:\s*'([^']+)'", lowered)
+        if arch_match:
+            arch = arch_match.group(1)
+            if arch in LlamaCppBackend._DIFFUSION_ARCHES:
+                return (
+                    f"'{arch}' is a diffusion (image-generation) GGUF, which "
+                    "llama-server cannot run as a chat/completion model. Use "
+                    "Studio's Images page to generate with local diffusion "
+                    "GGUFs such as FLUX and Qwen-Image."
+                )
+            if is_ollama:
+                return (
+                    "Some Ollama models do not work with llama.cpp. Try a "
+                    "different model, or use this model directly through "
+                    "Ollama instead."
+                )
+            return (
+                f"llama.cpp does not support this GGUF's model architecture "
+                f"('{arch}'). The file is valid, but this model type cannot "
+                "be run with llama-server."
+            )
+
+        # Other Ollama compat failures that do not name an arch. Only when
+        # the output shows a GGUF compat issue, not OOM / missing binaries.
+        if is_ollama:
+            gguf_compat_hints = (
+                "key not found",
+                "unknown model architecture",
+                "failed to load model",
+            )
+            if any(h in lowered for h in gguf_compat_hints):
+                return (
+                    "Some Ollama models do not work with llama.cpp. Try a "
+                    "different model, or use this model directly through "
+                    "Ollama instead."
+                )
+
+        # Fallback: genuinely unknown failure (OOM, missing binary, ...).
+        return (
+            "llama-server failed to start. "
+            "Check that the GGUF file is valid and you have enough memory."
+        )
+
     def load_model(
         self,
         *,
@@ -3396,31 +3496,12 @@ class LlamaCppBackend:
                 # Wait for llama-server to become healthy
                 if not self._wait_for_health(timeout = 600.0):
                     self._kill_process()
-                    _gguf = gguf_path or ""
-                    _is_ollama = (
-                        ".studio_links" in _gguf
-                        or os.sep + "ollama_links" + os.sep in _gguf
-                        or os.sep + ".cache" + os.sep + "ollama" + os.sep in _gguf
-                        or (self._model_identifier or "").startswith("ollama/")
-                    )
-                    # Only show the Ollama-specific message when the server
-                    # output indicates a GGUF compatibility issue, not for
-                    # unrelated failures like OOM or missing binaries.
-                    if _is_ollama:
-                        _output = "\n".join(self._stdout_lines[-50:]).lower()
-                        _gguf_compat_hints = (
-                            "key not found",
-                            "unknown model architecture",
-                            "failed to load model",
-                        )
-                        if any(h in _output for h in _gguf_compat_hints):
-                            raise RuntimeError(
-                                "Some Ollama models do not work with llama.cpp. "
-                                "Try a different model, or use this model directly through Ollama instead."
-                            )
                     raise RuntimeError(
-                        "llama-server failed to start. "
-                        "Check that the GGUF file is valid and you have enough memory."
+                        self._classify_llama_start_failure(
+                            "\n".join(self._stdout_lines[-50:]),
+                            gguf_path,
+                            self._model_identifier,
+                        )
                     )
 
                 self._healthy = True
@@ -4553,6 +4634,7 @@ class LlamaCppBackend:
         # a transient failure are allowed (only block when the previous
         # identical call succeeded).
         _tool_call_history: list[tuple[str, bool]] = []  # (key, failed)
+        _render_html_succeeded = False
 
         # ── Re-prompt on plan-without-action ─────────────────
         # When the model describes what it intends to do (forward-looking
@@ -4627,6 +4709,7 @@ class LlamaCppBackend:
                 _iter_timings = None
                 _stream_done = False
                 _last_emitted = ""
+                provisional_render_html_tool_call_ids = set()
 
                 stream_timeout = httpx.Timeout(
                     connect = 10,
@@ -4736,6 +4819,33 @@ class LlamaCppBackend:
                                                 tool_calls_acc[idx]["function"][
                                                     "arguments"
                                                 ] += func["arguments"]
+                                            current_name = tool_calls_acc[idx][
+                                                "function"
+                                            ].get("name", "")
+                                            fallback_id = f"call_{idx}"
+                                            current_id = tool_calls_acc[idx].get(
+                                                "id", fallback_id
+                                            )
+                                            already_started = (
+                                                current_id
+                                                in provisional_render_html_tool_call_ids
+                                            )
+                                            has_real_id = current_id != fallback_id
+                                            if (
+                                                current_name == "render_html"
+                                                and not _render_html_succeeded
+                                                and not already_started
+                                                and has_real_id
+                                            ):
+                                                provisional_render_html_tool_call_ids.add(
+                                                    current_id
+                                                )
+                                                yield {
+                                                    "type": "tool_start",
+                                                    "tool_name": "render_html",
+                                                    "tool_call_id": current_id,
+                                                    "arguments": {},
+                                                }
                                         continue
 
                                     # ── Reasoning tokens ──
@@ -4917,13 +5027,25 @@ class LlamaCppBackend:
                                     "content": _stripped,
                                 }
                             )
+                            available_tool_names = [
+                                tool.get("function", {}).get("name")
+                                for tool in tools
+                                if isinstance(tool, dict)
+                                and isinstance(tool.get("function"), dict)
+                            ]
+                            available_tool_names = [
+                                name for name in available_tool_names if name
+                            ]
+                            tool_hint = (
+                                " or ".join(available_tool_names) or "an available tool"
+                            )
                             conversation.append(
                                 {
                                     "role": "user",
                                     "content": (
                                         "STOP. Do NOT write code or explain. "
                                         "You MUST call a tool NOW. "
-                                        "Call web_search or python immediately."
+                                        f"Call {tool_hint} immediately."
                                     ),
                                 }
                             )
@@ -5095,7 +5217,12 @@ class LlamaCppBackend:
                             arguments = json.loads(raw_args)
                         except (json.JSONDecodeError, ValueError):
                             if auto_heal_tool_calls:
-                                arguments = {"query": raw_args}
+                                heal_key = {
+                                    "python": "code",
+                                    "terminal": "command",
+                                    "render_html": "code",
+                                }.get(tool_name, "query")
+                                arguments = {heal_key: raw_args}
                             else:
                                 arguments = {"raw": raw_args}
                     else:
@@ -5132,14 +5259,18 @@ class LlamaCppBackend:
                         )
                     else:
                         status_text = f"Calling: {tool_name}"
-                    yield {"type": "status", "text": status_text}
+                    _repeat_render_html = (
+                        tool_name == "render_html" and _render_html_succeeded
+                    )
+                    if not _repeat_render_html:
+                        yield {"type": "status", "text": status_text}
 
-                    yield {
-                        "type": "tool_start",
-                        "tool_name": tool_name,
-                        "tool_call_id": tc.get("id", ""),
-                        "arguments": arguments,
-                    }
+                        yield {
+                            "type": "tool_start",
+                            "tool_name": tool_name,
+                            "tool_call_id": tc.get("id", ""),
+                            "arguments": arguments,
+                        }
 
                     # ── Duplicate call detection ──────────────
                     # str(dict) is stable here: arguments always comes from
@@ -5147,7 +5278,9 @@ class LlamaCppBackend:
                     # so insertion order is deterministic (Python 3.7+).
                     _tc_key = tool_name + str(arguments)
                     _prev = _tool_call_history[-1] if _tool_call_history else None
-                    if _prev and _prev[0] == _tc_key and not _prev[1]:
+                    if _repeat_render_html:
+                        result = RENDER_HTML_REPEAT_NUDGE
+                    elif _prev and _prev[0] == _tc_key and not _prev[1]:
                         result = (
                             "You already made this exact call. "
                             "Do not repeat the same tool call. "
@@ -5186,12 +5319,13 @@ class LlamaCppBackend:
                                 tool_context = tool_context,
                             )
 
-                    yield {
-                        "type": "tool_end",
-                        "tool_name": tool_name,
-                        "tool_call_id": tc.get("id", ""),
-                        "result": result,
-                    }
+                    if not _repeat_render_html:
+                        yield {
+                            "type": "tool_end",
+                            "tool_name": tool_name,
+                            "tool_call_id": tc.get("id", ""),
+                            "result": result,
+                        }
 
                     # Nudge model to try a different approach on errors
                     _error_prefixes = (
@@ -5207,6 +5341,8 @@ class LlamaCppBackend:
                     _is_error = isinstance(result, str) and result.lstrip().startswith(
                         _error_prefixes
                     )
+                    if tool_name == "render_html" and not _is_error:
+                        _render_html_succeeded = True
                     _tool_call_history.append((_tc_key, _is_error))
                     # Strip image sentinel before feeding result to the LLM
                     # (the full result with sentinel is still yielded via

studio/backend/core/inference/mcp_client.py

@@ -5,6 +5,9 @@ from __future__ import annotations
 
 import asyncio
 import json
+import os
+import shlex
+import sys
 from typing import Any, Optional
 
 from loggers import get_logger
@@ -16,7 +19,55 @@ MCP_TOOL_PREFIX = "mcp__"
 _oauth_token_store = None
 
 
+def is_stdio(address: str) -> bool:
+    """A non-HTTP address is a local stdio command, e.g.
+    'npx -y @modelcontextprotocol/server-filesystem /path'."""
+    return not address.strip().lower().startswith(("http://", "https://"))
+
+
+def parse_stdio_command(address: str) -> list[str]:
+    """Split a stdio command line into argv. Shared by route validation and the
+    transport so both agree on quoting (notably Windows backslash paths)."""
+    posix = sys.platform != "win32"
+    parts = shlex.split(address, posix = posix)
+    if not posix:
+        # posix=False keeps backslash paths intact but also keeps the surrounding
+        # quotes on a token. Strip a matched pair so the argv reaches the
+        # subprocess clean ('"C:\\Program Files\\node"' -> C:\\Program Files\\node).
+        parts = [
+            p[1:-1] if len(p) >= 2 and p[0] == p[-1] and p[0] in "\"'" else p
+            for p in parts
+        ]
+    return parts
+
+
+def stdio_mcp_enabled() -> bool:
+    """stdio MCP servers spawn local processes as the backend user (and bypass
+    the python/terminal sandbox), so they are only allowed when the backend
+    host is the user's own machine. The Tauri desktop app sets
+    UNSLOTH_STUDIO_ALLOW_STDIO_MCP=1 (see main.py); advanced localhost /
+    self-hosted users can opt in with the same variable. It stays off for
+    Colab and any network (0.0.0.0) bind."""
+    return os.environ.get("UNSLOTH_STUDIO_ALLOW_STDIO_MCP") == "1"
+
+
+# Probe timeouts for discovering a server's tool list. OAuth needs minutes for
+# first-connect/expired-token browser sign-in; stdio allows for first-run
+# package download (e.g. `npx -y ...`); HTTP fails fast.
+_HTTP_PROBE_TIMEOUT = 8.0
+_OAUTH_PROBE_TIMEOUT = 305.0
+_STDIO_PROBE_TIMEOUT = 60.0
+
+
+def probe_timeout(address: str, use_oauth: bool) -> float:
+    if use_oauth:
+        return _OAUTH_PROBE_TIMEOUT
+    return _STDIO_PROBE_TIMEOUT if is_stdio(address) else _HTTP_PROBE_TIMEOUT
+
+
 def parse_server_headers(server: dict) -> Optional[dict]:
+    """Parsed headers_json. For stdio servers this dict is the process
+    environment instead of HTTP headers (see _client)."""
     raw = server.get("headers_json")
     if not raw:
         return None
@@ -63,6 +114,28 @@ async def clear_oauth_tokens_async(url: str) -> None:
 
 def _client(url: str, headers: Optional[dict], use_oauth: bool = False):
     from fastmcp import Client
+
+    if is_stdio(url):
+        # Belt-and-suspenders: never spawn unless stdio is enabled on this host.
+        if not stdio_mcp_enabled():
+            raise PermissionError("stdio MCP servers are disabled on this host")
+        from fastmcp.client.transports import StdioTransport
+
+        parts = parse_stdio_command(url)
+        if not parts:
+            raise ValueError(f"Empty stdio command: {url!r}")
+        # env vars ride the headers field (merged over the SDK's safe default env).
+        # keep_alive=False tears the subprocess down on exit, so a one-shot
+        # probe/tool call never leaves an orphan process.
+        return Client(
+            StdioTransport(
+                command = parts[0],
+                args = parts[1:],
+                env = headers or None,
+                keep_alive = False,
+            )
+        )
+
     from fastmcp.client.transports import SSETransport, StreamableHttpTransport
     from fastmcp.mcp_config import infer_transport_type_from_url
 

studio/backend/core/inference/safetensors_agentic.py

@@ -18,6 +18,7 @@ cumulative text and dispatches them via ``core.inference.tools``.
 """
 
 import json
+import re
 import threading
 from typing import Callable, Generator, Optional
 from urllib.parse import urlparse
@@ -27,6 +28,7 @@ from loggers import get_logger
 from core.inference.tool_call_parser import (
     BUDGET_EXHAUSTED_NUDGE,
     DUPLICATE_CALL_NUDGE,
+    RENDER_HTML_REPEAT_NUDGE,
     TOOL_ERROR_NUDGE,
     TOOL_ERROR_PREFIXES,
     TOOL_XML_SIGNALS,
@@ -66,7 +68,34 @@ def _status_for_tool(tool_name: str, arguments: dict) -> str:
     return f"Calling: {tool_name}"
 
 
-_CANONICAL_HEAL_ARG = {"python": "code", "terminal": "command"}
+_CANONICAL_HEAL_ARG = {
+    "python": "code",
+    "terminal": "command",
+    "render_html": "code",
+}
+
+
+_FUNCTION_SIGNAL_RE = re.compile(r"<function=([\w-]+)>")
+_TOOL_CALL_NAME_RE = re.compile(r'"name"\s*:\s*"([\w-]+)"')
+
+
+def _detect_render_html_tool_start(content: str) -> bool:
+    """Return True when the first drained tool call is clearly render_html."""
+    function_match = _FUNCTION_SIGNAL_RE.search(content)
+    tool_call_index = content.find("<tool_call>")
+    if not function_match and tool_call_index < 0:
+        return False
+
+    if function_match and (
+        tool_call_index < 0 or function_match.start() < tool_call_index
+    ):
+        return function_match.group(1) == "render_html"
+
+    if tool_call_index >= 0:
+        name_match = _TOOL_CALL_NAME_RE.search(content[tool_call_index:])
+        return bool(name_match and name_match.group(1) == "render_html")
+
+    return False
 
 
 def _coerce_arguments(raw_args, *, heal: bool, tool_name: str = "") -> dict:
@@ -136,6 +165,7 @@ def run_safetensors_tool_loop(
     """
     conversation = list(messages)
     tool_call_history: list[tuple[str, bool]] = []
+    render_html_succeeded = False
     final_attempt_done = False
     allowed_tool_names = {
         (tool.get("function") or {}).get("name")
@@ -162,6 +192,8 @@ def run_safetensors_tool_loop(
         content_accum = ""
         cumulative_display = ""
         last_emitted = ""
+        provisional_render_html_started = False
+        provisional_render_html_id = f"call_{next_call_id}"
 
         gen = single_turn(conversation)
         prev_cumulative = ""
@@ -180,6 +212,18 @@ def run_safetensors_tool_loop(
             content_accum += delta
 
             if detect_state == _state_draining:
+                if (
+                    not render_html_succeeded
+                    and not provisional_render_html_started
+                    and _detect_render_html_tool_start(content_accum)
+                ):
+                    provisional_render_html_started = True
+                    yield {
+                        "type": "tool_start",
+                        "tool_name": "render_html",
+                        "tool_call_id": provisional_render_html_id,
+                        "arguments": {},
+                    }
                 continue
 
             if detect_state == _state_streaming:
@@ -197,6 +241,18 @@ def run_safetensors_tool_loop(
                         yield {"type": "content", "text": cleaned_before}
                     cumulative_display = candidate
                     detect_state = _state_draining
+                    if (
+                        not render_html_succeeded
+                        and not provisional_render_html_started
+                        and _detect_render_html_tool_start(content_accum)
+                    ):
+                        provisional_render_html_started = True
+                        yield {
+                            "type": "tool_start",
+                            "tool_name": "render_html",
+                            "tool_call_id": provisional_render_html_id,
+                            "arguments": {},
+                        }
                     continue
                 cumulative_display = candidate
                 cleaned = strip_tool_markup(cumulative_display)
@@ -223,6 +279,18 @@ def run_safetensors_tool_loop(
 
             if is_match:
                 detect_state = _state_draining
+                if (
+                    not render_html_succeeded
+                    and not provisional_render_html_started
+                    and _detect_render_html_tool_start(content_accum)
+                ):
+                    provisional_render_html_started = True
+                    yield {
+                        "type": "tool_start",
+                        "tool_name": "render_html",
+                        "tool_call_id": provisional_render_html_id,
+                        "arguments": {},
+                    }
             elif is_prefix and len(stripped) < _MAX_BUFFER_CHARS:
                 continue
             else:
@@ -283,6 +351,13 @@ def run_safetensors_tool_loop(
                 # literal "<tool_call>" prose is preserved.
                 if content_accum:
                     yield {"type": "content", "text": content_accum}
+                if provisional_render_html_started:
+                    yield {
+                        "type": "tool_end",
+                        "tool_name": "render_html",
+                        "tool_call_id": provisional_render_html_id,
+                        "result": "Error: render_html tool call could not be parsed.",
+                    }
                 yield {"type": "status", "text": ""}
                 return
             content_text = strip_tool_markup(content_accum, final = True)
@@ -309,16 +384,20 @@ def run_safetensors_tool_loop(
                 tool_name = tool_name,
             )
 
-            yield {"type": "status", "text": _status_for_tool(tool_name, arguments)}
-            yield {
-                "type": "tool_start",
-                "tool_name": tool_name,
-                "tool_call_id": tc.get("id", ""),
-                "arguments": arguments,
-            }
+            repeat_render_html = tool_name == "render_html" and render_html_succeeded
+            if not repeat_render_html:
+                yield {"type": "status", "text": _status_for_tool(tool_name, arguments)}
+                yield {
+                    "type": "tool_start",
+                    "tool_name": tool_name,
+                    "tool_call_id": tc.get("id", ""),
+                    "arguments": arguments,
+                }
 
             tc_key = tool_name + str(arguments)
-            if allowed_tool_names and tool_name not in allowed_tool_names:
+            if repeat_render_html:
+                result = RENDER_HTML_REPEAT_NUDGE
+            elif allowed_tool_names and tool_name not in allowed_tool_names:
                 result = (
                     f"Error: tool '{tool_name}' is not enabled for this "
                     "request. Use one of the enabled tools or provide a "
@@ -347,16 +426,19 @@ def run_safetensors_tool_loop(
                         logger.exception("Tool %s raised: %s", tool_name, exc)
                         result = f"Error: tool raised an exception: {exc}"
 
-            yield {
-                "type": "tool_end",
-                "tool_name": tool_name,
-                "tool_call_id": tc.get("id", ""),
-                "result": result,
-            }
+            if not repeat_render_html:
+                yield {
+                    "type": "tool_end",
+                    "tool_name": tool_name,
+                    "tool_call_id": tc.get("id", ""),
+                    "result": result,
+                }
 
             is_error = isinstance(result, str) and result.lstrip().startswith(
                 TOOL_ERROR_PREFIXES
             )
+            if tool_name == "render_html" and not is_error:
+                render_html_succeeded = True
             tool_call_history.append((tc_key, is_error))
 
             # Strip frontend image sentinel from the model's view.

studio/backend/core/inference/tool_call_parser.py

@@ -49,6 +49,12 @@ DUPLICATE_CALL_NUDGE = (
     "provide your final answer now."
 )
 
+RENDER_HTML_REPEAT_NUDGE = (
+    "Error: render_html was already called for this response. Do not call "
+    "render_html again in this response unless the user asks for changes. "
+    "Provide the final answer now."
+)
+
 TOOL_ERROR_NUDGE = (
     "\n\nThe tool call encountered an issue. Please try a different "
     "approach or rephrase your request."
@@ -70,6 +76,20 @@ _TC_FUNC_CLOSE_RE = re.compile(r"\s*</function>\s*$")
 # `issue-number`, `repo-name`); using `\w+` here dropped those keys.
 _TC_PARAM_START_RE = re.compile(r"<parameter=([\w-]+)>\s*")
 _TC_PARAM_CLOSE_RE = re.compile(r"\s*</parameter>\s*$")
+_PARAM_CLOSE_TAG = "</parameter>"
+_FUNC_CLOSE_TAG = "</function>"
+
+
+def _inside_open_parameter(content: str, pos: int) -> bool:
+    """Return True when ``pos`` falls inside an unclosed parameter value."""
+    last_param_start = -1
+    for match in _TC_PARAM_START_RE.finditer(content, 0, pos):
+        last_param_start = match.start()
+    if last_param_start < 0:
+        return False
+    last_param_close = content.rfind(_PARAM_CLOSE_TAG, 0, pos)
+    last_func_close = content.rfind(_FUNC_CLOSE_TAG, 0, pos)
+    return last_param_start > max(last_param_close, last_func_close)
 
 
 def strip_tool_markup(text: str, *, final: bool = False) -> str:
@@ -151,7 +171,11 @@ def parse_tool_calls_from_text(content: str, *, id_offset: int = 0) -> list[dict
     # optional; don't use </function> as body boundary because code
     # values can contain that literal.
     if not tool_calls:
-        func_starts = list(_TC_FUNC_START_RE.finditer(content))
+        func_starts = [
+            fm
+            for fm in _TC_FUNC_START_RE.finditer(content)
+            if not _inside_open_parameter(content, fm.start())
+        ]
         for idx, fm in enumerate(func_starts):
             func_name = fm.group(1)
             body_start = fm.end()

studio/backend/core/inference/tools.py

@@ -28,8 +28,11 @@ import urllib.request
 from core.inference.mcp_client import (
     MCP_TOOL_PREFIX,
     call_tool_sync,
+    is_stdio,
     list_tools_async,
     parse_server_headers,
+    probe_timeout,
+    stdio_mcp_enabled,
 )
 from storage import mcp_servers_db
 
@@ -418,6 +421,35 @@ _workdirs: dict[str, str] = {}
 
 # Non-matching session_ids collapse to ``_invalid`` to block cross-session escapes.
 _SESSION_ID_RE = re.compile(r"\A[A-Za-z0-9_\-]{1,64}\Z")
+_PROJECT_SESSION_PREFIX = "project-"
+
+
+def _get_project_workdir(session_id: str) -> str | None:
+    if not session_id.startswith(_PROJECT_SESSION_PREFIX):
+        return None
+    project_id = session_id[len(_PROJECT_SESSION_PREFIX) :]
+    if not project_id or not _SESSION_ID_RE.match(project_id):
+        return None
+    try:
+        from storage.studio_db import ensure_chat_project_workspace
+
+        project = ensure_chat_project_workspace(project_id)
+    except Exception:
+        logger.warning(
+            "Failed to resolve project sandbox for %s", session_id, exc_info = True
+        )
+        return None
+    if not project:
+        return None
+    root_path = project.get("rootPath")
+    sandbox_path = project.get("sandboxPath")
+    if not root_path or not sandbox_path:
+        return None
+    root_real = os.path.realpath(root_path)
+    sandbox_real = os.path.realpath(sandbox_path)
+    if sandbox_real != root_real and not sandbox_real.startswith(root_real + os.sep):
+        return None
+    return sandbox_real
 
 
 def _get_workdir(session_id: str | None = None) -> str:
@@ -427,7 +459,14 @@ def _get_workdir(session_id: str | None = None) -> str:
     if key not in _workdirs or not os.path.isdir(_workdirs[key]):
         home = os.path.expanduser("~")
         sandbox_root = os.path.join(home, "studio_sandbox")
-        if session_id and _SESSION_ID_RE.match(session_id):
+        project_workdir = (
+            _get_project_workdir(session_id)
+            if session_id and _SESSION_ID_RE.match(session_id)
+            else None
+        )
+        if project_workdir:
+            workdir = project_workdir
+        elif session_id and _SESSION_ID_RE.match(session_id):
             workdir = os.path.join(sandbox_root, session_id)
             if not os.path.realpath(workdir).startswith(
                 os.path.realpath(sandbox_root) + os.sep
@@ -450,6 +489,10 @@ def _get_workdir(session_id: str | None = None) -> str:
     return _workdirs[key]
 
 
+def get_sandbox_workdir(session_id: str | None = None) -> str:
+    return _get_workdir(session_id)
+
+
 WEB_SEARCH_TOOL = {
     "type": "function",
     "function": {
@@ -511,7 +554,6 @@ TERMINAL_TOOL = {
     },
 }
 
-
 # Lazy import: don't pull rag stack on inference paths that never see RAG.
 def _get_rag_tool_spec():
     from core.rag.tool import SEARCH_KNOWLEDGE_BASE_TOOL
@@ -519,7 +561,41 @@ def _get_rag_tool_spec():
     return SEARCH_KNOWLEDGE_BASE_TOOL
 
 
-ALL_TOOLS = [WEB_SEARCH_TOOL, PYTHON_TOOL, TERMINAL_TOOL, _get_rag_tool_spec()]
+RENDER_HTML_TOOL = {
+    "type": "function",
+    "function": {
+        "name": "render_html",
+        "description": (
+            "Render a self-contained HTML/CSS/JavaScript artifact for the user. "
+            "Call this at most once per assistant response unless the user "
+            "explicitly asks for changes in that response. Future user requests "
+            "for new artifacts may call render_html once. Put the entire document "
+            "in code, including any CSS in <style> tags and JavaScript in <script> tags."
+        ),
+        "parameters": {
+            "type": "object",
+            "properties": {
+                "code": {
+                    "type": "string",
+                    "description": "A complete self-contained HTML document.",
+                },
+                "title": {
+                    "type": "string",
+                    "description": "Short display title for the artifact.",
+                },
+            },
+            "required": ["code"],
+        },
+    },
+}
+
+ALL_TOOLS = [
+    WEB_SEARCH_TOOL,
+    PYTHON_TOOL,
+    TERMINAL_TOOL,
+    RENDER_HTML_TOOL,
+    _get_rag_tool_spec(),
+]
 
 
 # OpenAI's function.name regex: ^[a-zA-Z0-9_-]{1,64}$ -- enforced before
@@ -576,17 +652,19 @@ def _mcp_specs_for_server(server: dict, mcp_tools: list[dict]) -> list[dict]:
 
 async def get_enabled_mcp_tools() -> list[dict]:
     servers = [s for s in mcp_servers_db.list_servers() if s.get("is_enabled")]
+    # Never spawn stdio servers when stdio is disabled on this host (e.g. a DB
+    # carried over from a desktop install onto a Colab / network deployment).
+    if not stdio_mcp_enabled():
+        servers = [s for s in servers if not is_stdio(s["url"])]
     if not servers:
         return []
 
-    # OAuth probes need minutes for first-connect/expired-token browser
-    # sign-in; non-OAuth probes fail fast. Matches routes/mcp_servers.py.
     results = await asyncio.gather(
         *(
             list_tools_async(
                 url = s["url"],
                 headers = parse_server_headers(s),
-                timeout = 305.0 if s.get("use_oauth") else 8.0,
+                timeout = probe_timeout(s["url"], bool(s.get("use_oauth"))),
                 use_oauth = bool(s.get("use_oauth")),
             )
             for s in servers
@@ -611,6 +689,25 @@ async def get_enabled_mcp_tools() -> list[dict]:
 _TIMEOUT_UNSET = object()
 
 
+def _render_html_result(arguments: dict) -> str:
+    code = arguments.get("code")
+    if not isinstance(code, str) or not code.strip():
+        return "Error: render_html requires a non-empty code string."
+    title = arguments.get("title")
+    if isinstance(title, str) and title.strip():
+        safe_title = title.strip()[:120]
+        return (
+            f"Rendered HTML artifact: {safe_title}. Do not call render_html "
+            "again in this response unless the user asks for changes. For a later "
+            "user request for a new artifact, call render_html once."
+        )
+    return (
+        "Rendered HTML artifact. Do not call render_html again in this response "
+        "unless the user asks for changes. For a later user request for a new "
+        "artifact, call render_html once."
+    )
+
+
 def execute_tool(
     name: str,
     arguments: dict,
@@ -633,6 +730,8 @@ def execute_tool(
         f"execute_tool: name={name}, session_id={session_id}, timeout={timeout}"
     )
     effective_timeout = _EXEC_TIMEOUT if timeout is _TIMEOUT_UNSET else timeout
+    if name == "render_html":
+        return _render_html_result(arguments)
     if name.startswith(MCP_TOOL_PREFIX):
         try:
             _, server_id, tool_name = name.split("__", 2)
@@ -643,6 +742,8 @@ def execute_tool(
             return f"Error: MCP server '{server_id}' not found"
         if not server.get("is_enabled"):
             return f"Error: MCP server '{server_id}' is disabled"
+        if is_stdio(server["url"]) and not stdio_mcp_enabled():
+            return f"Error: stdio MCP server '{server_id}' is disabled on this host"
         return call_tool_sync(
             url = server["url"],
             headers = parse_server_headers(server),

studio/backend/main.py

@@ -298,6 +298,11 @@ def _load_desktop_owner() -> dict[str, str] | None:
 
 _DESKTOP_OWNER = _load_desktop_owner()
 
+# The Tauri desktop app runs the backend on the owner's own machine, so local
+# stdio MCP servers are safe there. setdefault lets an explicit "0" opt out.
+if _DESKTOP_OWNER:
+    os.environ.setdefault("UNSLOTH_STUDIO_ALLOW_STDIO_MCP", "1")
+
 
 def _desktop_owner() -> dict[str, str] | None:
     return _DESKTOP_OWNER
@@ -431,6 +436,7 @@ from starlette.requests import Request as _StarletteRequest  # noqa: E402
 
 
 _CSP_SCRIPT_NONCE_HEADER = "x-internal-script-nonce"
+_ARTIFACT_PREVIEW_FRAME_PATH = "/api/inference/artifact-preview-frame"
 
 
 # /content is Colab's working directory — more reliable than env vars which
@@ -484,6 +490,7 @@ def _build_csp(script_nonce: "str | None" = None) -> str:
         "style-src 'self' 'unsafe-inline'; "
         f"{script_src}; "
         "font-src 'self' data:; "
+        "frame-src 'self'; "
         f"frame-ancestors {frame_ancestors}; "
         "form-action 'self'; "
         "base-uri 'self'"
@@ -502,13 +509,13 @@ class SecurityHeadersMiddleware(BaseHTTPMiddleware):
         response.headers.setdefault("Content-Security-Policy", _build_csp(nonce))
         # Omit X-Frame-Options in Colab — CSP frame-ancestors handles it, and
         # DENY would block serve_kernel_port_as_iframe regardless of CSP.
-        if not _IS_COLAB:
+        if not _IS_COLAB and request.url.path != _ARTIFACT_PREVIEW_FRAME_PATH:
             response.headers.setdefault("X-Frame-Options", "DENY")
         response.headers.setdefault("X-Content-Type-Options", "nosniff")
         response.headers.setdefault("Referrer-Policy", "no-referrer")
         response.headers.setdefault(
             "Permissions-Policy",
-            "camera=(), microphone=(), geolocation=(), interest-cohort=()",
+            "camera=(), microphone=(), geolocation=()",
         )
         response.headers["server"] = "unsloth-studio"
         return response

studio/backend/models/inference.py

@@ -709,11 +709,12 @@ class ChatCompletionRequest(BaseModel):
     enabled_tools: Optional[list[str]] = Field(
         None,
         description = (
-            "[x-unsloth] List of enabled tool names. Local GGUF models accept "
-            "['web_search', 'python', 'terminal']. External providers accept "
-            "['web_search', 'web_fetch', 'code_execution'] for Anthropic and "
-            "['web_search', 'code_execution'] for OpenAI Responses. If None, "
-            "all local tools are enabled and no server-side tools are forwarded."
+            "[x-unsloth] List of enabled tool names. Local GGUF/safetensors models "
+            "accept ['web_search', 'python', 'terminal', 'render_html']. External "
+            "providers accept ['web_search', 'web_fetch', 'code_execution'] for "
+            "Anthropic and ['web_search', 'code_execution', 'image_generation'] for "
+            "OpenAI Responses. If None, all local tools are enabled and no "
+            "server-side tools are forwarded."
         ),
     )
     mcp_enabled: Optional[bool] = Field(

studio/backend/routes/chat_history.py

@@ -18,15 +18,21 @@ from storage.studio_db import (
     clear_chat_history,
     count_chat_threads,
     delete_chat_threads,
+    delete_chat_project,
+    ensure_chat_project_workspace,
+    get_chat_project,
     get_chat_thread,
     get_chat_message,
+    list_chat_projects,
     list_chat_legacy_imports,
     list_chat_settings,
     list_chat_messages,
     list_chat_messages_for_threads,
     list_chat_threads,
     sync_chat_messages,
+    update_chat_project,
     update_chat_thread,
+    upsert_chat_project,
     upsert_chat_legacy_imports,
     upsert_chat_message,
     upsert_chat_settings_merge,
@@ -42,6 +48,7 @@ class ChatThread(BaseModel):
     modelType: Literal["base", "lora", "model1", "model2"]
     modelId: str = ""
     pairId: Optional[str] = None
+    projectId: Optional[str] = None
     archived: bool = False
     createdAt: int
     openaiCodeExecContainerId: Optional[str] = None
@@ -53,6 +60,7 @@ class ChatThreadPatch(BaseModel):
     modelType: Optional[Literal["base", "lora", "model1", "model2"]] = None
     modelId: Optional[str] = None
     pairId: Optional[str] = None
+    projectId: Optional[str] = None
     archived: Optional[bool] = None
     createdAt: Optional[int] = None
     openaiCodeExecContainerId: Optional[str] = None
@@ -70,10 +78,33 @@ class ChatMessage(BaseModel):
     createdAt: int
 
 
+class ChatProject(BaseModel):
+    id: str
+    name: str
+    instructions: str = ""
+    rootPath: Optional[str] = None
+    sandboxPath: Optional[str] = None
+    archived: bool = False
+    createdAt: int
+    updatedAt: int
+
+
+class ChatProjectPatch(BaseModel):
+    name: Optional[str] = None
+    instructions: Optional[str] = None
+    archived: Optional[bool] = None
+    createdAt: Optional[int] = None
+    updatedAt: Optional[int] = None
+
+
 class ChatThreadListResponse(BaseModel):
     threads: list[ChatThread]
 
 
+class ChatProjectListResponse(BaseModel):
+    projects: list[ChatProject]
+
+
 class ChatMessageListResponse(BaseModel):
     messages: list[ChatMessage]
 
@@ -95,6 +126,7 @@ class ChatExportResponse(BaseModel):
     exportedAt: str
     version: int
     threadCount: int
+    projects: list[ChatProject] = Field(default_factory = list)
     threads: list[ChatThread]
     messages: list[ChatMessage]
 
@@ -135,6 +167,8 @@ class ChatSettingsPayload(BaseModel):
         Literal["none", "minimal", "low", "medium", "high", "max", "xhigh"]
     ] = None
     preserveThinking: Optional[bool] = None
+    collapseHtmlArtifacts: Optional[bool] = None
+    allowArtifactNetworkAccess: Optional[bool] = None
     autoHealToolCalls: Optional[bool] = None
     maxToolCallsPerMessage: Optional[int] = Field(default = None, ge = 1)
     toolCallTimeout: Optional[int] = Field(default = None, ge = 1)
@@ -176,12 +210,14 @@ class ChatImportLedgerRecordResponse(BaseModel):
 async def list_threads(
     model_type: Optional[str] = Query(None),
     pair_id: Optional[str] = Query(None),
+    project_id: Optional[str] = Query(None),
     include_archived: bool = Query(True),
     current_subject: str = Depends(get_current_subject),
 ):
     threads = list_chat_threads(
         model_type = model_type,
         pair_id = pair_id,
+        project_id = project_id,
         include_archived = include_archived,
     )
     return ChatThreadListResponse(threads = [ChatThread(**t) for t in threads])
@@ -192,6 +228,11 @@ async def save_thread(
     payload: ChatThread,
     current_subject: str = Depends(get_current_subject),
 ):
+    if payload.projectId and get_chat_project(payload.projectId) is None:
+        raise HTTPException(
+            status_code = 404,
+            detail = f"Project {payload.projectId} not found",
+        )
     return ChatThread(**upsert_chat_thread(payload.model_dump()))
 
 
@@ -216,6 +257,11 @@ async def patch_thread(
     for field in ("title", "modelType", "modelId", "archived", "createdAt"):
         if field in patch and patch[field] is None:
             raise HTTPException(status_code = 400, detail = f"{field} cannot be null")
+    if patch.get("projectId") and get_chat_project(patch["projectId"]) is None:
+        raise HTTPException(
+            status_code = 404,
+            detail = f"Project {patch['projectId']} not found",
+        )
     thread = update_chat_thread(
         thread_id,
         patch,
@@ -237,6 +283,77 @@ async def delete_threads(
     return {"status": "deleted"}
 
 
+@router.get("/projects", response_model = ChatProjectListResponse)
+async def list_projects(
+    include_archived: bool = Query(False),
+    current_subject: str = Depends(get_current_subject),
+):
+    return ChatProjectListResponse(
+        projects = [
+            ChatProject(**(ensure_chat_project_workspace(project["id"]) or project))
+            for project in list_chat_projects(include_archived = include_archived)
+        ]
+    )
+
+
+@router.post("/projects", response_model = ChatProject)
+async def save_project(
+    payload: ChatProject,
+    current_subject: str = Depends(get_current_subject),
+):
+    return ChatProject(**upsert_chat_project(payload.model_dump()))
+
+
+@router.get("/projects/{project_id}", response_model = ChatProject)
+async def get_project(
+    project_id: str,
+    current_subject: str = Depends(get_current_subject),
+):
+    project = ensure_chat_project_workspace(project_id)
+    if project is None:
+        raise HTTPException(
+            status_code = 404,
+            detail = f"Project {project_id} not found",
+        )
+    return ChatProject(**project)
+
+
+@router.patch("/projects/{project_id}", response_model = ChatProject)
+async def patch_project(
+    project_id: str,
+    payload: ChatProjectPatch,
+    current_subject: str = Depends(get_current_subject),
+):
+    patch = payload.model_dump(exclude_unset = True)
+    for field in ("name", "archived", "createdAt", "updatedAt"):
+        if field in patch and patch[field] is None:
+            raise HTTPException(status_code = 400, detail = f"{field} cannot be null")
+    project = update_chat_project(project_id, patch)
+    if project is not None:
+        project = ensure_chat_project_workspace(project_id)
+    if project is None:
+        raise HTTPException(
+            status_code = 404,
+            detail = f"Project {project_id} not found",
+        )
+    return ChatProject(**project)
+
+
+@router.delete("/projects/{project_id}", response_model = ChatProject)
+async def delete_project(
+    project_id: str,
+    delete_files: bool = Query(False),
+    current_subject: str = Depends(get_current_subject),
+):
+    project = delete_chat_project(project_id, delete_files = delete_files)
+    if project is None:
+        raise HTTPException(
+            status_code = 404,
+            detail = f"Project {project_id} not found",
+        )
+    return ChatProject(**project)
+
+
 @router.get("/threads/{thread_id}/messages", response_model = ChatMessageListResponse)
 async def get_thread_messages(
     thread_id: str,
@@ -392,11 +509,13 @@ async def export_history(current_subject: str = Depends(get_current_subject)):
     from datetime import datetime, timezone
 
     threads = list_chat_threads(include_archived = True)
+    projects = list_chat_projects(include_archived = True)
     messages = list_chat_messages_for_threads([thread["id"] for thread in threads])
     return ChatExportResponse(
         exportedAt = datetime.now(timezone.utc).isoformat(),
         version = 1,
         threadCount = len(threads),
+        projects = [ChatProject(**project) for project in projects],
         threads = [ChatThread(**thread) for thread in threads],
         messages = [ChatMessage(**message) for message in messages],
     )

studio/backend/routes/data_recipe/mcp.py

@@ -36,8 +36,18 @@ def list_mcp_tools(payload: McpToolsListRequest) -> McpToolsListResponse:
     providers: list[McpToolsProviderResult] = []
     tool_to_providers: dict[str, list[str]] = defaultdict(list)
 
+    from core.inference.mcp_client import stdio_mcp_enabled
+
     for provider_payload in payload.mcp_providers:
         provider_name = str(provider_payload.get("name", "")).strip()
+        if provider_payload.get("provider_type") == "stdio" and not stdio_mcp_enabled():
+            providers.append(
+                McpToolsProviderResult(
+                    name = provider_name,
+                    error = "Local (stdio) MCP servers are disabled on this host.",
+                )
+            )
+            continue
         built = build_mcp_providers({"mcp_providers": [provider_payload]})
         if len(built) != 1:
             providers.append(

studio/backend/routes/inference.py

@@ -239,6 +239,135 @@ router = APIRouter()
 studio_router = APIRouter()
 
 
+_ARTIFACT_PREVIEW_FRAME_ANCESTORS = "'self' tauri://localhost http://tauri.localhost"
+_ARTIFACT_PREVIEW_FRAME_STRICT_CSP = (
+    "default-src 'none'; "
+    "script-src 'unsafe-inline'; "
+    "style-src 'unsafe-inline'; "
+    "img-src data: blob:; "
+    "font-src data:; "
+    "media-src data: blob:; "
+    "connect-src 'none'; "
+    "object-src 'none'; "
+    "base-uri 'none'; "
+    "form-action 'none'; "
+    f"frame-ancestors {_ARTIFACT_PREVIEW_FRAME_ANCESTORS}; "
+    "sandbox allow-scripts"
+)
+_ARTIFACT_PREVIEW_FRAME_NETWORK_CSP = (
+    "default-src http: https: data: blob:; "
+    "script-src 'unsafe-inline' 'unsafe-eval' http: https: data: blob:; "
+    "script-src-elem 'unsafe-inline' http: https: data: blob:; "
+    "style-src 'unsafe-inline' http: https: data: blob:; "
+    "style-src-elem 'unsafe-inline' http: https: data: blob:; "
+    "img-src http: https: data: blob:; "
+    "font-src http: https: data: blob:; "
+    "media-src http: https: data: blob:; "
+    "connect-src http: https: ws: wss: data: blob:; "
+    "worker-src http: https: blob:; "
+    "object-src 'none'; "
+    "base-uri 'none'; "
+    "form-action 'none'; "
+    f"frame-ancestors {_ARTIFACT_PREVIEW_FRAME_ANCESTORS}; "
+    "sandbox allow-scripts"
+)
+_ARTIFACT_PREVIEW_FRAME_HTML = """<!doctype html>
+<html>
+  <head><meta charset=\"utf-8\" /></head>
+  <body>
+    <script>
+      (() => {
+        const createMemoryStorage = () => {
+          const data = new Map();
+          return {
+            get length() { return data.size; },
+            key: (index) => Array.from(data.keys())[index] ?? null,
+            getItem: (key) => data.has(String(key)) ? data.get(String(key)) : null,
+            setItem: (key, value) => data.set(String(key), String(value)),
+            removeItem: (key) => data.delete(String(key)),
+            clear: () => data.clear(),
+          };
+        };
+        const installStorageFallback = (name) => {
+          try {
+            void window[name];
+            return;
+          } catch {
+            // Opaque-origin sandboxed frames throw SecurityError for Web Storage.
+          }
+          try {
+            Object.defineProperty(window, name, {
+              value: createMemoryStorage(),
+              configurable: true,
+            });
+          } catch {
+            // Leave the sandbox failure contained in the artifact if the
+            // browser refuses to shadow the Web Storage accessor.
+          }
+        };
+        const installStorageFallbacks = () => {
+          installStorageFallback("localStorage");
+          installStorageFallback("sessionStorage");
+        };
+        const render = (html) => {
+          installStorageFallbacks();
+          document.open();
+          document.write(html);
+          document.close();
+        };
+        installStorageFallbacks();
+        window.addEventListener("message", (event) => {
+          const data = event.data;
+          if (!data || data.type !== "unsloth:artifact-html" || typeof data.html !== "string") return;
+          render(data.html);
+        });
+      })();
+    </script>
+  </body>
+</html>"""
+
+
+@studio_router.get("/artifact-preview-frame", include_in_schema = False)
+async def artifact_preview_frame(
+    request: Request,
+    allow_network: bool = False,
+    token: Optional[str] = None,
+):
+    """Serve the opaque sandbox shell used for client-side HTML artifacts."""
+
+    if allow_network:
+        auth_header = request.headers.get("authorization")
+        if auth_header and auth_header.lower().startswith("bearer "):
+            jwt_token = auth_header[7:]
+        elif token:
+            jwt_token = token
+        else:
+            raise HTTPException(
+                status_code = status.HTTP_401_UNAUTHORIZED,
+                detail = "Missing authentication token",
+            )
+        from fastapi.security import HTTPAuthorizationCredentials
+
+        creds = HTTPAuthorizationCredentials(scheme = "Bearer", credentials = jwt_token)
+        await get_current_subject(creds)
+
+    csp = (
+        _ARTIFACT_PREVIEW_FRAME_NETWORK_CSP
+        if allow_network
+        else _ARTIFACT_PREVIEW_FRAME_STRICT_CSP
+    )
+    return Response(
+        content = _ARTIFACT_PREVIEW_FRAME_HTML,
+        media_type = "text/html; charset=utf-8",
+        headers = {
+            "Cache-Control": "no-store",
+            "Content-Security-Policy": csp,
+            "Referrer-Policy": "no-referrer",
+            "X-Content-Type-Options": "nosniff",
+        },
+    )
+
+
 def _detect_safetensors_features(backend, chat_template: Optional[str]) -> dict:
     """Classify reasoning/tool capabilities via the GGUF classifier so
     flags match across backends. gpt-oss is overridden because Harmony
@@ -458,13 +587,28 @@ async def _await_cancel_then_close(cancel_event, resp) -> None:
         return
 
 
-# Appended to tool-use nudge to discourage plan-without-action
+# Appended to tool-use nudge to discourage plan-without-action.
+# Keep render_html guidance gated to turns where the artifact tool is actually
+# present in the tool schema; otherwise small local models can hallucinate a
+# missing tool call instead of following the fenced-HTML fallback prompt.
 _TOOL_ACTION_NUDGE = (
     " IMPORTANT: Always call tools directly -- never write code yourself."
     " Never describe what you plan to do -- just call the tool immediately."
-    " For any code request, call the python tool. For any factual question, call web_search."
-    " Do NOT output code blocks -- use the python tool instead."
+    " For non-artifact code requests, call the python tool when it is available."
+    " For factual questions that require current information, call web_search when it is available."
+    " Do NOT output raw code blocks when an enabled tool can satisfy the request."
 )
+_ARTIFACT_TOOL_ACTION_NUDGE = (
+    " For HTML, CSS, or JavaScript artifact requests, call render_html once when "
+    "it is available. After render_html succeeds, do not call it again in the "
+    "same response unless the user asks for changes. Future user requests for "
+    "new artifacts may call render_html once."
+)
+
+
+def _tool_action_nudge(has_artifact: bool) -> str:
+    return _TOOL_ACTION_NUDGE + (_ARTIFACT_TOOL_ACTION_NUDGE if has_artifact else "")
+
 
 # Strip tool-call XML the speculative buffer in core/inference/llama_cpp.py
 # split across the visible/DRAIN boundary. Four leak shapes:
@@ -2741,6 +2885,7 @@ async def openai_chat_completions(
             _tool_names = {t["function"]["name"] for t in tools_to_use}
             _has_web = "web_search" in _tool_names
             _has_code = "python" in _tool_names or "terminal" in _tool_names
+            _has_artifact = "render_html" in _tool_names
 
             _date_line = f"The current date is {_date.today().isoformat()}."
 
@@ -2762,34 +2907,34 @@ async def openai_chat_completions(
                 "Use code execution for math, calculations, data processing, "
                 "or to parse and analyze information from tool results."
             )
+            _artifact_tips = (
+                "Use render_html for HTML, CSS, or JavaScript artifact requests "
+                "with one complete self-contained HTML document in the code argument. "
+                "Call it once, then do not call it again in the same response unless "
+                "the user asks for changes. Future user requests for new artifacts may "
+                "call render_html once."
+            )
 
-            if _has_web and _has_code:
+            _tool_tip_parts = []
+            if _has_web:
+                _tool_tip_parts.append(_web_tips)
+            if _has_code:
+                _tool_tip_parts.append(_code_tips)
+            if _has_artifact:
+                _tool_tip_parts.append(_artifact_tips)
+
+            if _tool_tip_parts:
                 _nudge = (
                     _date_line + " "
                     "You have access to tools. When appropriate, prefer using "
                     "tools rather than answering from memory. "
-                    + _web_tips
-                    + " "
-                    + _code_tips
-                )
-            elif _has_code:
-                _nudge = (
-                    _date_line + " "
-                    "You have access to tools. When appropriate, prefer using "
-                    "code execution rather than answering from memory. " + _code_tips
-                )
-            elif _has_web:
-                _nudge = (
-                    _date_line + " "
-                    "You have access to tools. When appropriate, prefer using "
-                    "web search for up-to-date or uncertain factual "
-                    "information rather than answering from memory. " + _web_tips
+                    + " ".join(_tool_tip_parts)
                 )
             else:
                 _nudge = ""
 
             if _nudge:
-                _nudge += _TOOL_ACTION_NUDGE
+                _nudge += _tool_action_nudge(_has_artifact)
                 # Append nudge to system prompt (preserve user's prompt)
                 if system_prompt:
                     system_prompt = system_prompt.rstrip() + "\n\n" + _nudge
@@ -3271,6 +3416,7 @@ async def openai_chat_completions(
         _sf_tool_names = {t["function"]["name"] for t in _sf_tools_to_use}
         _sf_has_web = "web_search" in _sf_tool_names
         _sf_has_code = "python" in _sf_tool_names or "terminal" in _sf_tool_names
+        _sf_has_artifact = "render_html" in _sf_tool_names
 
         _sf_date_line = f"The current date is {_date.today().isoformat()}."
         _sf_model_size_b = _extract_model_size_b(model_name)
@@ -3289,35 +3435,35 @@ async def openai_chat_completions(
             "Use code execution for math, calculations, data processing, "
             "or to parse and analyze information from tool results."
         )
+        _sf_artifact_tips = (
+            "Use render_html for HTML, CSS, or JavaScript artifact requests "
+            "with one complete self-contained HTML document in the code argument. "
+            "Call it once, then do not call it again in the same response unless "
+            "the user asks for changes. Future user requests for new artifacts may "
+            "call render_html once."
+        )
 
-        if _sf_has_web and _sf_has_code:
+        _sf_tool_tip_parts = []
+        if _sf_has_web:
+            _sf_tool_tip_parts.append(_sf_web_tips)
+        if _sf_has_code:
+            _sf_tool_tip_parts.append(_sf_code_tips)
+        if _sf_has_artifact:
+            _sf_tool_tip_parts.append(_sf_artifact_tips)
+
+        if _sf_tool_tip_parts:
             _sf_nudge = (
                 _sf_date_line + " "
                 "You have access to tools. When appropriate, prefer using "
                 "tools rather than answering from memory. "
-                + _sf_web_tips
-                + " "
-                + _sf_code_tips
-            )
-        elif _sf_has_code:
-            _sf_nudge = (
-                _sf_date_line + " "
-                "You have access to tools. When appropriate, prefer using "
-                "code execution rather than answering from memory. " + _sf_code_tips
-            )
-        elif _sf_has_web:
-            _sf_nudge = (
-                _sf_date_line + " "
-                "You have access to tools. When appropriate, prefer using "
-                "web search for up-to-date or uncertain factual "
-                "information rather than answering from memory. " + _sf_web_tips
+                + " ".join(_sf_tool_tip_parts)
             )
         else:
             _sf_nudge = ""
 
         _sf_system_prompt = system_prompt
         if _sf_nudge:
-            _sf_nudge += _TOOL_ACTION_NUDGE
+            _sf_nudge += _tool_action_nudge(_sf_has_artifact)
             if _sf_system_prompt:
                 _sf_system_prompt = _sf_system_prompt.rstrip() + "\n\n" + _sf_nudge
             else:
@@ -3744,16 +3890,11 @@ async def serve_sandbox_file(
         )
 
     # ── Path containment check ──────────────────────────────────
-    home = os.path.expanduser("~")
-    sandbox_root = os.path.realpath(os.path.join(home, "studio_sandbox"))
-    safe_session = os.path.basename(session_id.replace("..", ""))
-    if not safe_session:
-        raise HTTPException(status_code = 404, detail = "Not found")
+    from core.inference.tools import get_sandbox_workdir
 
-    file_path = os.path.realpath(
-        os.path.join(sandbox_root, safe_session, safe_filename)
-    )
-    if not file_path.startswith(sandbox_root + os.sep):
+    sandbox_dir = os.path.realpath(get_sandbox_workdir(session_id))
+    file_path = os.path.realpath(os.path.join(sandbox_dir, safe_filename))
+    if file_path != sandbox_dir and not file_path.startswith(sandbox_dir + os.sep):
         raise HTTPException(
             status_code = status.HTTP_403_FORBIDDEN,
             detail = "Access denied",
@@ -4967,6 +5108,7 @@ async def anthropic_messages(
         _tool_names = {t["function"]["name"] for t in openai_tools}
         _has_web = "web_search" in _tool_names
         _has_code = "python" in _tool_names or "terminal" in _tool_names
+        _has_artifact = "render_html" in _tool_names
 
         _date_line = f"The current date is {_date.today().isoformat()}."
         _model_size_b = _extract_model_size_b(model_name)
@@ -4985,34 +5127,33 @@ async def anthropic_messages(
             "Use code execution for math, calculations, data processing, "
             "or to parse and analyze information from tool results."
         )
+        _artifact_tips = (
+            "Use render_html for HTML, CSS, or JavaScript artifact requests "
+            "with one complete self-contained HTML document in the code argument. "
+            "Call it once, then do not call it again in the same response unless "
+            "the user asks for changes. Future user requests for new artifacts may "
+            "call render_html once."
+        )
 
-        if _has_web and _has_code:
+        _tool_tip_parts = []
+        if _has_web:
+            _tool_tip_parts.append(_web_tips)
+        if _has_code:
+            _tool_tip_parts.append(_code_tips)
+        if _has_artifact:
+            _tool_tip_parts.append(_artifact_tips)
+
+        if _tool_tip_parts:
             _nudge = (
                 _date_line + " "
                 "You have access to tools. When appropriate, prefer using "
-                "tools rather than answering from memory. "
-                + _web_tips
-                + " "
-                + _code_tips
-            )
-        elif _has_code:
-            _nudge = (
-                _date_line + " "
-                "You have access to tools. When appropriate, prefer using "
-                "code execution rather than answering from memory. " + _code_tips
-            )
-        elif _has_web:
-            _nudge = (
-                _date_line + " "
-                "You have access to tools. When appropriate, prefer using "
-                "web search for up-to-date or uncertain factual "
-                "information rather than answering from memory. " + _web_tips
+                "tools rather than answering from memory. " + " ".join(_tool_tip_parts)
             )
         else:
             _nudge = ""
 
         if _nudge:
-            _nudge += _TOOL_ACTION_NUDGE
+            _nudge += _tool_action_nudge(_has_artifact)
             # Inject into system prompt
             if openai_messages and openai_messages[0].get("role") == "system":
                 openai_messages[0]["content"] = (
@@ -5200,6 +5341,7 @@ async def _anthropic_tool_non_streaming(run_gen, message_id, model_name):
     baseline, not against turn N's final length.
     """
     content_blocks: list = []
+    tool_blocks_by_id: dict[str, AnthropicResponseToolUseBlock] = {}
     usage = {}
     prev_text = ""
 
@@ -5218,13 +5360,25 @@ async def _anthropic_tool_non_streaming(run_gen, message_id, model_name):
                 else:
                     content_blocks.append(AnthropicResponseTextBlock(text = new))
         elif etype == "tool_start":
-            content_blocks.append(
-                AnthropicResponseToolUseBlock(
-                    id = event["tool_call_id"],
-                    name = event["tool_name"],
-                    input = event.get("arguments", {}),
-                )
+            tool_call_id = event["tool_call_id"]
+            arguments = event.get("arguments", {})
+            existing_tool_block = (
+                tool_blocks_by_id.get(tool_call_id) if tool_call_id else None
             )
+            if existing_tool_block is not None:
+                if arguments or not existing_tool_block.input:
+                    existing_tool_block.input = arguments
+                if event.get("tool_name") and not existing_tool_block.name:
+                    existing_tool_block.name = event["tool_name"]
+            else:
+                tool_block = AnthropicResponseToolUseBlock(
+                    id = tool_call_id,
+                    name = event["tool_name"],
+                    input = arguments,
+                )
+                if tool_call_id:
+                    tool_blocks_by_id[tool_call_id] = tool_block
+                content_blocks.append(tool_block)
         elif etype == "tool_end":
             prev_text = ""
         elif etype == "metadata":

studio/backend/routes/mcp_servers.py

@@ -11,8 +11,12 @@ from fastapi import APIRouter, Depends, HTTPException
 from auth.authentication import get_current_subject
 from core.inference.mcp_client import (
     clear_oauth_tokens_async,
+    is_stdio,
     list_tools_async,
     parse_server_headers,
+    parse_stdio_command,
+    probe_timeout,
+    stdio_mcp_enabled,
 )
 from models.mcp_servers import (
     McpServerCreate,
@@ -28,22 +32,48 @@ logger = structlog.get_logger(__name__)
 router = APIRouter()
 
 
-_PROBE_TIMEOUT_SECONDS = 8.0
-# When OAuth probes need to open a browser, wait long enough for the user to
-# sign in. Matches fastmcp's default OAuth callback_timeout (300 s) + slack.
-_OAUTH_PROBE_TIMEOUT_SECONDS = 305.0
+def _looks_like_command(value: str) -> bool:
+    """Whitespace is a one-way signal: a URL can't hold an unencoded space, so a
+    value with whitespace is definitely a command. No whitespace proves nothing
+    (a lone token may be a single-arg command or a scheme-less URL)."""
+    return any(ch.isspace() for ch in value)
 
 
 def _validate_url(url: str) -> str:
     trimmed = (url or "").strip()
     if not trimmed:
         raise HTTPException(status_code = 400, detail = "url must not be empty")
+    # When stdio is enabled on this host, a non-HTTP value is a local command.
+    # Reuse this field so stdio servers ride the existing CRUD/storage with no
+    # schema change. A lone token (example.com, /usr/bin/srv) is ambiguous, so we
+    # keep the existing behaviour and treat any non-HTTP value as a command here.
+    if stdio_mcp_enabled() and is_stdio(trimmed):
+        try:
+            parts = parse_stdio_command(trimmed)
+        except ValueError as exc:
+            raise HTTPException(status_code = 400, detail = f"Invalid command: {exc}")
+        if not parts or not parts[0].strip():
+            raise HTTPException(status_code = 400, detail = "command must not be empty")
+        if "://" in parts[0]:
+            # A URL-scheme first token is a mistyped URL, not a command. Reject
+            # it cleanly instead of exec-ing it (mirrors the frontend check).
+            raise HTTPException(
+                status_code = 400,
+                detail = "Enter an http(s):// URL, or a local command whose "
+                "first token is an executable (not a URL).",
+            )
+        return trimmed
     parsed = urlparse(trimmed)
     if parsed.scheme not in ("http", "https"):
-        raise HTTPException(
-            status_code = 400,
-            detail = "url must start with http:// or https://",
+        detail = (
+            "MCP server address must start with http:// or https:// "
+            "(for example https://example.com/mcp)."
         )
+        # Host-scoped wording ("this server"), not "desktop only": self-hosted
+        # hosts can opt in via the env var.
+        if _looks_like_command(trimmed):
+            detail += " Running a local command is not enabled on this server."
+        raise HTTPException(status_code = 400, detail = detail)
     if not parsed.netloc:
         raise HTTPException(status_code = 400, detail = "url is missing a host")
     return trimmed
@@ -91,6 +121,9 @@ async def create_mcp_server(
         raise HTTPException(status_code = 400, detail = "display_name must not be empty")
     url = _validate_url(payload.url)
     headers = _normalize_headers(payload.headers)
+    # OAuth is HTTP-only; force it off for stdio commands so a stale flag can't
+    # push the probe onto the 305s OAuth timeout. Backend is the enforcer.
+    use_oauth = payload.use_oauth and not is_stdio(url)
 
     server_id = uuid.uuid4().hex[:16]
     mcp_servers_db.create_server(
@@ -99,7 +132,7 @@ async def create_mcp_server(
         url = url,
         headers_json = json.dumps(headers) if headers else None,
         is_enabled = payload.is_enabled,
-        use_oauth = payload.use_oauth,
+        use_oauth = use_oauth,
     )
     return _row_to_response(mcp_servers_db.get_server(server_id))
 
@@ -132,6 +165,9 @@ def _changes_from_payload(payload: McpServerUpdate) -> dict:
                 status_code = 400, detail = "use_oauth must be true or false"
             )
         changes["use_oauth"] = payload.use_oauth
+    # stdio is OAuth-less: drop a stale OAuth flag when switching to a command.
+    if "url" in changes and is_stdio(changes["url"]):
+        changes["use_oauth"] = False
     return changes
 
 
@@ -147,6 +183,15 @@ async def update_mcp_server(
     changes = _changes_from_payload(payload)
     if not changes:
         raise HTTPException(status_code = 400, detail = "No fields to update")
+    # headers == HTTP headers (remote) or env vars (stdio). On a transport-type
+    # switch with no new headers, drop the old ones so env secrets are not
+    # re-sent as HTTP headers (or vice versa).
+    if (
+        "url" in changes
+        and is_stdio(changes["url"]) != is_stdio(old["url"])
+        and "headers_json" not in changes
+    ):
+        changes["headers_json"] = None
     # Clear persisted OAuth tokens when the URL changes or OAuth is
     # disabled; fastmcp keys tokens by URL and would otherwise let a
     # re-pointed server silently inherit the old account's credentials.
@@ -180,15 +225,19 @@ async def refresh_mcp_server_tools(
     server = mcp_servers_db.get_server(server_id)
     if not server:
         raise HTTPException(status_code = 404, detail = "MCP server not found")
+    # Refresh uses the stored address, so re-check the stdio gate here too: a
+    # stdio row from a desktop DB must not spawn on a hosted/network host.
+    if is_stdio(server["url"]) and not stdio_mcp_enabled():
+        raise HTTPException(
+            status_code = 400, detail = "stdio MCP servers are disabled on this host"
+        )
 
     use_oauth = bool(server.get("use_oauth"))
     try:
         tools = await list_tools_async(
             url = server["url"],
             headers = parse_server_headers(server),
-            timeout = _OAUTH_PROBE_TIMEOUT_SECONDS
-            if use_oauth
-            else _PROBE_TIMEOUT_SECONDS,
+            timeout = probe_timeout(server["url"], use_oauth),
             use_oauth = use_oauth,
         )
     except Exception as exc:  # noqa: BLE001 — surface transport+timeout errors to UI
@@ -212,9 +261,7 @@ async def test_mcp_server(
         tools = await list_tools_async(
             url = url,
             headers = headers,
-            timeout = _OAUTH_PROBE_TIMEOUT_SECONDS
-            if payload.use_oauth
-            else _PROBE_TIMEOUT_SECONDS,
+            timeout = probe_timeout(url, payload.use_oauth),
             use_oauth = payload.use_oauth,
         )
     except Exception as exc:  # noqa: BLE001

studio/backend/storage/studio_db.py

@@ -14,16 +14,19 @@ import json
 import logging
 import os
 import platform
+import re
+import shutil
 import sqlite3
 import threading
 from contextlib import contextmanager
 from datetime import datetime, timezone
+from pathlib import Path
 
 logger = logging.getLogger(__name__)
 from typing import Any, Iterable, Optional
 
 
-from utils.paths import studio_db_path, ensure_dir
+from utils.paths import project_workspaces_root, studio_db_path, ensure_dir
 
 
 def _denied_path_prefixes() -> list[str]:
@@ -56,6 +59,77 @@ def _denied_path_prefixes() -> list[str]:
 _schema_lock = threading.Lock()
 _schema_ready = False
 _SQLITE_IN_CHUNK_SIZE = 900
+_PROJECT_WORKSPACE_SUBDIRS = ("sandbox",)
+
+
+def _project_slug(name: str) -> str:
+    slug = re.sub(r"[^A-Za-z0-9._-]+", "-", name.strip()).strip(".-_")
+    return slug[:48] or "project"
+
+
+def _default_project_root(project: dict) -> str:
+    project_id = str(project["id"])
+    suffix = re.sub(r"[^A-Za-z0-9_-]+", "-", project_id)[:8].strip("-_") or "project"
+    folder_name = f"{_project_slug(str(project.get('name') or 'Project'))}-{suffix}"
+    return str(project_workspaces_root() / folder_name)
+
+
+def _ensure_project_workspace(root_path: str) -> str:
+    root = Path(root_path).expanduser()
+    root_resolved = ensure_dir(root).resolve()
+    for subdir in _PROJECT_WORKSPACE_SUBDIRS:
+        ensure_dir(root_resolved / subdir)
+    return str(root_resolved)
+
+
+def _delete_project_workspace(project: dict) -> None:
+    root_path = project.get("rootPath")
+    if not root_path:
+        return
+    root = Path(root_path).expanduser()
+    try:
+        root_resolved = root.resolve(strict = False)
+    except (OSError, RuntimeError, ValueError):
+        logger.warning(
+            "Skipping project workspace delete for invalid path %r", root_path
+        )
+        return
+
+    project_id = str(project["id"])
+    suffix = re.sub(r"[^A-Za-z0-9_-]+", "-", project_id)[:8].strip("-_") or "project"
+    if not root_resolved.name.endswith(f"-{suffix}"):
+        logger.warning(
+            "Skipping project workspace delete for unexpected project path %s",
+            root_resolved,
+        )
+        return
+    if root_resolved.parent == root_resolved or root_resolved == Path.home().resolve():
+        logger.warning(
+            "Skipping project workspace delete for unsafe project path %s",
+            root_resolved,
+        )
+        return
+    check = (
+        os.path.normcase(str(root_resolved))
+        if platform.system() == "Windows"
+        else str(root_resolved)
+    )
+    for prefix in _denied_path_prefixes():
+        if check == prefix or check.startswith(prefix + os.sep):
+            logger.warning(
+                "Skipping project workspace delete under denied path %s",
+                root_resolved,
+            )
+            return
+    if not root_resolved.exists():
+        return
+    if root_resolved.is_symlink() or not root_resolved.is_dir():
+        logger.warning(
+            "Skipping project workspace delete for non-directory path %s",
+            root_resolved,
+        )
+        return
+    shutil.rmtree(root_resolved)
 
 
 def _ensure_schema(conn: sqlite3.Connection) -> None:
@@ -120,6 +194,27 @@ def _ensure_schema(conn: sqlite3.Connection) -> None:
         )
         """
     )
+    conn.execute(
+        """
+        CREATE TABLE IF NOT EXISTS chat_projects (
+            id TEXT NOT NULL PRIMARY KEY,
+            name TEXT NOT NULL,
+            instructions TEXT,
+            root_path TEXT,
+            archived INTEGER NOT NULL DEFAULT 0,
+            created_at INTEGER NOT NULL,
+            updated_at INTEGER NOT NULL
+        )
+        """
+    )
+    chat_project_cols = {
+        row[1] for row in conn.execute("PRAGMA table_info(chat_projects)").fetchall()
+    }
+    if "root_path" not in chat_project_cols:
+        conn.execute("ALTER TABLE chat_projects ADD COLUMN root_path TEXT")
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS idx_chat_projects_archived_updated_at ON chat_projects(archived, updated_at)"
+    )
     conn.execute(
         """
         CREATE TABLE IF NOT EXISTS chat_threads (
@@ -128,16 +223,20 @@ def _ensure_schema(conn: sqlite3.Connection) -> None:
             model_type TEXT NOT NULL,
             model_id TEXT,
             pair_id TEXT,
+            project_id TEXT,
             archived INTEGER NOT NULL DEFAULT 0,
             created_at INTEGER NOT NULL,
             openai_code_exec_container_id TEXT,
-            anthropic_code_exec_container_id TEXT
+            anthropic_code_exec_container_id TEXT,
+            FOREIGN KEY(project_id) REFERENCES chat_projects(id) ON DELETE CASCADE
         )
         """
     )
     chat_thread_cols = {
         row[1] for row in conn.execute("PRAGMA table_info(chat_threads)").fetchall()
     }
+    if "project_id" not in chat_thread_cols:
+        conn.execute("ALTER TABLE chat_threads ADD COLUMN project_id TEXT")
     if "openai_code_exec_container_id" not in chat_thread_cols:
         conn.execute(
             "ALTER TABLE chat_threads ADD COLUMN openai_code_exec_container_id TEXT"
@@ -166,6 +265,9 @@ def _ensure_schema(conn: sqlite3.Connection) -> None:
     conn.execute(
         "CREATE INDEX IF NOT EXISTS idx_chat_threads_pair_id ON chat_threads(pair_id)"
     )
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS idx_chat_threads_project_id ON chat_threads(project_id)"
+    )
     conn.execute(
         "CREATE INDEX IF NOT EXISTS idx_chat_messages_thread_id_created_at ON chat_messages(thread_id, created_at)"
     )
@@ -855,6 +957,7 @@ def _chat_thread_from_row(row: sqlite3.Row) -> dict:
         "modelType": data["model_type"],
         "modelId": data.get("model_id") or "",
         "pairId": data.get("pair_id") or None,
+        "projectId": data.get("project_id") or None,
         "archived": bool(data["archived"]),
         "createdAt": data["created_at"],
         "openaiCodeExecContainerId": data.get("openai_code_exec_container_id"),
@@ -862,6 +965,21 @@ def _chat_thread_from_row(row: sqlite3.Row) -> dict:
     }
 
 
+def _chat_project_from_row(row: sqlite3.Row) -> dict:
+    data = dict(row)
+    root_path = data.get("root_path")
+    return {
+        "id": data["id"],
+        "name": data["name"],
+        "instructions": data.get("instructions") or "",
+        "rootPath": root_path or None,
+        "sandboxPath": os.path.join(root_path, "sandbox") if root_path else None,
+        "archived": bool(data["archived"]),
+        "createdAt": data["created_at"],
+        "updatedAt": data["updated_at"],
+    }
+
+
 def _chat_message_from_row(row: sqlite3.Row) -> dict:
     data = dict(row)
     message = {
@@ -887,13 +1005,14 @@ def upsert_chat_thread(thread: dict) -> dict:
         conn.execute(
             """
             INSERT INTO chat_threads
-                (id, title, model_type, model_id, pair_id, archived, created_at, openai_code_exec_container_id, anthropic_code_exec_container_id)
-            VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
+                (id, title, model_type, model_id, pair_id, project_id, archived, created_at, openai_code_exec_container_id, anthropic_code_exec_container_id)
+            VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
             ON CONFLICT(id) DO UPDATE SET
                 title = excluded.title,
                 model_type = excluded.model_type,
                 model_id = excluded.model_id,
                 pair_id = excluded.pair_id,
+                project_id = excluded.project_id,
                 archived = excluded.archived,
                 created_at = excluded.created_at,
                 openai_code_exec_container_id = excluded.openai_code_exec_container_id,
@@ -905,6 +1024,7 @@ def upsert_chat_thread(thread: dict) -> dict:
                 thread["modelType"],
                 thread.get("modelId") or "",
                 thread.get("pairId"),
+                thread.get("projectId"),
                 1 if thread.get("archived") else 0,
                 int(thread["createdAt"]),
                 thread.get("openaiCodeExecContainerId"),
@@ -923,6 +1043,7 @@ def update_chat_thread(id: str, patch: dict) -> Optional[dict]:
         "modelType": ("model_type", patch.get("modelType")),
         "modelId": ("model_id", patch.get("modelId")),
         "pairId": ("pair_id", patch.get("pairId")),
+        "projectId": ("project_id", patch.get("projectId")),
         "archived": ("archived", 1 if patch.get("archived") else 0),
         "createdAt": ("created_at", patch.get("createdAt")),
         "openaiCodeExecContainerId": (
@@ -968,6 +1089,7 @@ def get_chat_thread(id: str) -> Optional[dict]:
 def list_chat_threads(
     model_type: str | None = None,
     pair_id: str | None = None,
+    project_id: str | None = None,
     include_archived: bool = True,
 ) -> list[dict]:
     clauses = []
@@ -978,6 +1100,9 @@ def list_chat_threads(
     if pair_id is not None:
         clauses.append("pair_id = ?")
         values.append(pair_id)
+    if project_id is not None:
+        clauses.append("project_id = ?")
+        values.append(project_id)
     if not include_archived:
         clauses.append("archived = 0")
     where = f"WHERE {' AND '.join(clauses)}" if clauses else ""
@@ -1020,6 +1145,136 @@ def count_chat_threads() -> int:
         conn.close()
 
 
+def upsert_chat_project(project: dict) -> dict:
+    existing = get_chat_project(project["id"])
+    root_path = existing.get("rootPath") if existing else None
+    if not root_path:
+        root_path = _default_project_root(project)
+    root_path = _ensure_project_workspace(root_path)
+    conn = get_connection()
+    try:
+        conn.execute(
+            """
+            INSERT INTO chat_projects
+                (id, name, instructions, root_path, archived, created_at, updated_at)
+            VALUES (?, ?, ?, ?, ?, ?, ?)
+            ON CONFLICT(id) DO UPDATE SET
+                name = excluded.name,
+                instructions = excluded.instructions,
+                root_path = COALESCE(chat_projects.root_path, excluded.root_path),
+                archived = excluded.archived,
+                created_at = excluded.created_at,
+                updated_at = excluded.updated_at
+            """,
+            (
+                project["id"],
+                project["name"],
+                project.get("instructions") or "",
+                root_path,
+                1 if project.get("archived") else 0,
+                int(project["createdAt"]),
+                int(project["updatedAt"]),
+            ),
+        )
+        conn.commit()
+        return get_chat_project(project["id"]) or project
+    finally:
+        conn.close()
+
+
+def update_chat_project(id: str, patch: dict) -> Optional[dict]:
+    allowed = {
+        "name": ("name", patch.get("name")),
+        "instructions": ("instructions", patch.get("instructions")),
+        "archived": ("archived", 1 if patch.get("archived") else 0),
+        "createdAt": ("created_at", patch.get("createdAt")),
+        "updatedAt": ("updated_at", patch.get("updatedAt")),
+    }
+    assignments = []
+    values = []
+    for key, (column, value) in allowed.items():
+        if key in patch:
+            assignments.append(f"{column} = ?")
+            values.append(value)
+    if not assignments:
+        return get_chat_project(id)
+
+    conn = get_connection()
+    try:
+        conn.execute(
+            f"UPDATE chat_projects SET {', '.join(assignments)} WHERE id = ?",
+            (*values, id),
+        )
+        conn.commit()
+        row = conn.execute("SELECT * FROM chat_projects WHERE id = ?", (id,)).fetchone()
+        return _chat_project_from_row(row) if row is not None else None
+    finally:
+        conn.close()
+
+
+def ensure_chat_project_workspace(id: str) -> Optional[dict]:
+    project = get_chat_project(id)
+    if project is None:
+        return None
+    root_path = project.get("rootPath") or _default_project_root(project)
+    root_path = _ensure_project_workspace(root_path)
+    if project.get("rootPath") == root_path:
+        return project
+    conn = get_connection()
+    try:
+        conn.execute(
+            "UPDATE chat_projects SET root_path = ? WHERE id = ?",
+            (root_path, id),
+        )
+        conn.commit()
+    finally:
+        conn.close()
+    return get_chat_project(id)
+
+
+def get_chat_project(id: str) -> Optional[dict]:
+    conn = get_connection()
+    try:
+        row = conn.execute("SELECT * FROM chat_projects WHERE id = ?", (id,)).fetchone()
+        return _chat_project_from_row(row) if row is not None else None
+    finally:
+        conn.close()
+
+
+def list_chat_projects(include_archived: bool = False) -> list[dict]:
+    conn = get_connection()
+    try:
+        where = "" if include_archived else "WHERE archived = 0"
+        rows = conn.execute(
+            f"SELECT * FROM chat_projects {where} ORDER BY updated_at DESC"
+        ).fetchall()
+        return [_chat_project_from_row(row) for row in rows]
+    finally:
+        conn.close()
+
+
+def delete_chat_project(id: str, delete_files: bool = False) -> Optional[dict]:
+    conn = get_connection()
+    try:
+        conn.execute("BEGIN IMMEDIATE")
+        row = conn.execute("SELECT * FROM chat_projects WHERE id = ?", (id,)).fetchone()
+        if row is None:
+            conn.rollback()
+            return None
+        project = _chat_project_from_row(row)
+        conn.execute("DELETE FROM chat_threads WHERE project_id = ?", (id,))
+        conn.execute("DELETE FROM chat_projects WHERE id = ?", (id,))
+        conn.commit()
+        if delete_files:
+            _delete_project_workspace(project)
+        return project
+    except Exception:
+        conn.rollback()
+        raise
+    finally:
+        conn.close()
+
+
 class ChatMessageConflictError(RuntimeError):
     """Raised when a chat message id already belongs to another thread."""
 

studio/backend/tests/test_anthropic_messages.py

@@ -38,6 +38,7 @@ from routes.inference import (
     _normalize_anthropic_openai_images,
     _select_anthropic_server_tools,
     _anthropic_requested_studio_tools,
+    _anthropic_tool_non_streaming,
     anthropic_messages,
 )
 from state.tool_policy import reset_tool_policy, set_tool_policy
@@ -551,6 +552,54 @@ class TestAnthropicStreamEmitter:
         assert "tool_use" in events[1]
         assert "input_json_delta" in events[2]
 
+    def test_duplicate_tool_start_merges_into_open_tool_block(self):
+        e = AnthropicStreamEmitter()
+        e.start("msg_1", "m")
+        first_events = e.feed(
+            {
+                "type": "tool_start",
+                "tool_name": "render_html",
+                "tool_call_id": "call_0",
+                "arguments": {},
+            }
+        )
+        second_events = e.feed(
+            {
+                "type": "tool_start",
+                "tool_name": "render_html",
+                "tool_call_id": "call_0",
+                "arguments": {"code": "<!doctype html><html></html>"},
+            }
+        )
+
+        first_payloads = [
+            json.loads(event.split("data: ")[1]) for event in first_events
+        ]
+        second_payloads = [
+            json.loads(event.split("data: ")[1]) for event in second_events
+        ]
+
+        tool_starts = [
+            payload
+            for payload in first_payloads + second_payloads
+            if payload["type"] == "content_block_start"
+            and payload["content_block"]["type"] == "tool_use"
+        ]
+        assert len(tool_starts) == 1
+        assert tool_starts[0]["content_block"]["id"] == "call_0"
+        assert second_payloads == [
+            {
+                "type": "content_block_delta",
+                "index": tool_starts[0]["index"],
+                "delta": {
+                    "type": "input_json_delta",
+                    "partial_json": json.dumps(
+                        {"code": "<!doctype html><html></html>"}
+                    ),
+                },
+            }
+        ]
+
     def test_tool_end_closes_tool_opens_new_text_block(self):
         e = AnthropicStreamEmitter()
         e.start("msg_1", "m")
@@ -674,6 +723,49 @@ class TestAnthropicStreamEmitter:
         assert parsed["delta"]["text"] == "After tool"
 
 
+# =====================================================================
+# Non-streaming tool response tests
+# =====================================================================
+
+
+class TestAnthropicToolNonStreaming:
+    def test_duplicate_tool_start_replaces_provisional_tool_block(self):
+        def _run_gen():
+            yield {
+                "type": "tool_start",
+                "tool_name": "render_html",
+                "tool_call_id": "call_0",
+                "arguments": {},
+            }
+            yield {
+                "type": "tool_start",
+                "tool_name": "render_html",
+                "tool_call_id": "call_0",
+                "arguments": {"code": "<!doctype html><html></html>"},
+            }
+            yield {
+                "type": "tool_end",
+                "tool_name": "render_html",
+                "tool_call_id": "call_0",
+                "result": "Rendered HTML artifact.",
+            }
+
+        response = asyncio.run(_anthropic_tool_non_streaming(_run_gen, "msg_1", "m"))
+        body = json.loads(response.body)
+        tool_blocks = [
+            block for block in body["content"] if block["type"] == "tool_use"
+        ]
+
+        assert tool_blocks == [
+            {
+                "type": "tool_use",
+                "id": "call_0",
+                "name": "render_html",
+                "input": {"code": "<!doctype html><html></html>"},
+            }
+        ]
+
+
 # =====================================================================
 # Pass-through emitter tests (client-side tool execution path)
 # =====================================================================

studio/backend/tests/test_chat_history_storage.py

@@ -1,18 +1,49 @@
 # 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 os
+import platform
+import shutil
 import threading
+import uuid
+from pathlib import Path
 
 import pytest
 
 from storage import studio_db
 
 
-def _reset_studio_db(tmp_path, monkeypatch):
+def _reset_studio_db(tmp_path, monkeypatch, projects_home = None):
     monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setenv(
+        "UNSLOTH_STUDIO_PROJECTS_HOME",
+        str(projects_home if projects_home is not None else tmp_path / "Projects"),
+    )
     monkeypatch.setattr(studio_db, "_schema_ready", False)
 
 
+@pytest.fixture
+def workspace_projects_home(tmp_path):
+    """Projects root outside the platform delete denylist.
+
+    tmp_path resolves under /private/tmp on macOS, which the workspace
+    delete guard refuses by design. Linux/Windows tmp is not denied and is
+    used as-is; only the denied case falls back to a home subdir.
+    """
+    candidate = tmp_path / "Projects"
+    resolved = str(candidate.resolve())
+    check = os.path.normcase(resolved) if platform.system() == "Windows" else resolved
+    denied = studio_db._denied_path_prefixes()
+    if any(check == p or check.startswith(p + os.sep) for p in denied):
+        candidate = Path.home() / ".unsloth-studio-tests" / uuid.uuid4().hex
+    candidate.mkdir(parents = True, exist_ok = True)
+    try:
+        yield candidate
+    finally:
+        if ".unsloth-studio-tests" in candidate.parts:
+            shutil.rmtree(candidate, ignore_errors = True)
+
+
 def _thread(thread_id: str = "thread-1") -> dict:
     return {
         "id": thread_id,
@@ -41,6 +72,17 @@ def _message(
     }
 
 
+def _project(project_id: str = "project-1") -> dict:
+    return {
+        "id": project_id,
+        "name": "Research",
+        "instructions": "Use terse answers.",
+        "archived": False,
+        "createdAt": 1_700_000_000_000,
+        "updatedAt": 1_700_000_000_000,
+    }
+
+
 def test_sync_chat_messages_upserts_without_pruning(tmp_path, monkeypatch):
     _reset_studio_db(tmp_path, monkeypatch)
     studio_db.upsert_chat_thread(_thread())
@@ -63,6 +105,53 @@ def test_sync_chat_messages_upserts_without_pruning(tmp_path, monkeypatch):
     assert by_id["msg-2"]["content"] == [{"type": "text", "text": "updated text"}]
 
 
+def test_chat_projects_delete_cascades_threads_and_messages(
+    tmp_path,
+    monkeypatch,
+):
+    _reset_studio_db(tmp_path, monkeypatch)
+    project = studio_db.upsert_chat_project(_project())
+    assert project["rootPath"].startswith(str(tmp_path / "Projects"))
+    assert (tmp_path / "Projects" / "Research-project").exists()
+    assert (tmp_path / "Projects" / "Research-project" / "sandbox").is_dir()
+    assert not (tmp_path / "Projects" / "Research-project" / "chats").exists()
+    assert not (tmp_path / "Projects" / "Research-project" / "files").exists()
+    assert not (tmp_path / "Projects" / "Research-project" / "exports").exists()
+    studio_db.upsert_chat_thread({**_thread(), "projectId": "project-1"})
+    studio_db.upsert_chat_message(_message("msg-1", 1, "delete with project"))
+
+    [thread] = studio_db.list_chat_threads(project_id = "project-1")
+    assert thread["projectId"] == "project-1"
+
+    deleted = studio_db.delete_chat_project("project-1")
+
+    assert deleted is not None
+    assert deleted["id"] == "project-1"
+    assert studio_db.get_chat_project("project-1") is None
+    assert studio_db.list_chat_threads(project_id = "project-1") == []
+    assert studio_db.get_chat_thread("thread-1") is None
+    assert studio_db.list_chat_messages("thread-1") == []
+    assert (tmp_path / "Projects" / "Research-project").exists()
+
+
+def test_chat_project_delete_files_removes_workspace(
+    tmp_path, monkeypatch, workspace_projects_home
+):
+    _reset_studio_db(tmp_path, monkeypatch, projects_home = workspace_projects_home)
+    project = studio_db.upsert_chat_project(_project())
+    # Derive root from the created project so it tracks the projects home.
+    root = Path(project["rootPath"])
+    marker = root / "sandbox" / "marker.txt"
+    marker.write_text("created by code execution", encoding = "utf-8")
+
+    deleted = studio_db.delete_chat_project(project["id"], delete_files = True)
+
+    assert deleted is not None
+    assert deleted["rootPath"] == project["rootPath"]
+    assert not root.exists()
+    assert studio_db.get_chat_project(project["id"]) is None
+
+
 def test_sync_chat_messages_prunes_when_requested(tmp_path, monkeypatch):
     _reset_studio_db(tmp_path, monkeypatch)
     studio_db.upsert_chat_thread(_thread())

studio/backend/tests/test_llama_cpp_start_failure_classification.py

@@ -0,0 +1,144 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for LlamaCppBackend._classify_llama_start_failure.
+
+When llama-server exits before becoming healthy, load_model turns its
+captured stdout/stderr into a user-facing reason. A diffusion / image
+GGUF (FLUX, Qwen-Image, ...) is a valid file with plenty of memory, so
+the generic "invalid file or out of memory" message is actively
+misleading (issue #5842). These tests pin the classification.
+"""
+
+from __future__ import annotations
+
+import sys
+import types as _types
+from pathlib import Path
+
+import pytest
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+# Match the stubbing pattern in sibling tests so the module imports in a
+# lightweight env without fastapi.
+_loggers_stub = _types.ModuleType("loggers")
+_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+sys.modules.setdefault("loggers", _loggers_stub)
+# Give the structlog stub a real get_logger: a bare ModuleType poisons
+# sys.modules for later tests that call structlog.get_logger at import time.
+_structlog_stub = _types.ModuleType("structlog")
+_structlog_stub.get_logger = lambda *a, **k: __import__("logging").getLogger(
+    "structlog"
+)
+sys.modules.setdefault("structlog", _structlog_stub)
+if not hasattr(sys.modules["structlog"], "get_logger"):
+    sys.modules["structlog"].get_logger = _structlog_stub.get_logger
+
+from core.inference.llama_cpp import LlamaCppBackend  # noqa: E402
+
+_classify = LlamaCppBackend._classify_llama_start_failure
+
+# Real llama-server failure lines (lower-cased downstream anyway).
+_QWEN_IMAGE_OUT = (
+    "load_model: loading model 'qwen-image-edit-2511-Q4_K_M.gguf'\n"
+    "llama_model_load: error loading model: unknown model architecture: 'qwen_image'\n"
+    "llama_model_load_from_file_impl: failed to load model"
+)
+_OOM_OUT = (
+    "ggml_backend_cuda_buffer_type_alloc_buffer: allocating 12000.00 MiB on "
+    "device 0: cudaMalloc failed: out of memory"
+)
+
+
+class TestDiffusionArchitectures:
+    def test_qwen_image_routes_to_images_page(self):
+        msg = _classify(_QWEN_IMAGE_OUT, "/models/qwen-image.gguf", "local/qwen-image")
+        assert "diffusion" in msg.lower()
+        assert "Images page" in msg
+        assert "qwen_image" in msg
+        # Must NOT keep blaming memory / file validity.
+        assert "out of memory" not in msg.lower()
+        assert "enough memory" not in msg.lower()
+
+    # Parametrize over the production set so new arches are auto-covered.
+    @pytest.mark.parametrize("arch", sorted(LlamaCppBackend._DIFFUSION_ARCHES))
+    def test_every_diffusion_arch_is_recognised(self, arch):
+        out = f"error loading model: unknown model architecture: '{arch}'"
+        msg = _classify(out, f"/models/{arch}.gguf", f"local/{arch}")
+        assert "diffusion" in msg.lower()
+        assert "Images page" in msg
+        assert arch in msg
+
+
+class TestUnsupportedNonDiffusionArchitecture:
+    def test_unknown_llm_arch_says_unsupported_not_oom(self):
+        out = "error loading model: unknown model architecture: 'some_new_llm'"
+        msg = _classify(out, "/models/x.gguf", "local/x")
+        assert "some_new_llm" in msg
+        assert "architecture" in msg.lower()
+        # Specific, not the misleading memory message.
+        assert "enough memory" not in msg.lower()
+        assert "diffusion" not in msg.lower()
+
+    # Exact match: a chat arch merely containing a diffusion token (wan,
+    # sd1, flux, ...) must not be routed to the Images page.
+    @pytest.mark.parametrize(
+        "arch",
+        [
+            "taiwan",  # contains "wan"
+            "swan_llm",  # contains "wan"
+            "fluxion",  # contains "flux"
+            "sd1234",  # contains "sd1"
+            "sd3_chat",  # contains "sd3"
+            "aura2_text",  # contains "aura"
+            "cosmos_reason",  # contains "cosmos"
+            "qwen_image_text",  # contains "qwen_image"
+        ],
+    )
+    def test_arch_containing_diffusion_token_is_not_misrouted(self, arch):
+        out = f"error loading model: unknown model architecture: '{arch}'"
+        msg = _classify(out, f"/models/{arch}.gguf", f"local/{arch}")
+        assert arch in msg
+        assert "does not support" in msg.lower()
+        assert "diffusion" not in msg.lower()
+        assert "Images page" not in msg
+
+
+class TestOllamaAndFallback:
+    _OLLAMA_GGUF = (
+        f"/home/u/.ollama{__import__('os').sep}ollama_links"
+        f"{__import__('os').sep}m.gguf"
+    )
+
+    def test_ollama_compat_message_still_works(self):
+        out = "llama_model_load: error loading model: key not found"
+        msg = _classify(out, self._OLLAMA_GGUF, "ollama/llama3")
+        assert "Ollama" in msg
+
+    def test_ollama_unknown_arch_keeps_ollama_guidance(self):
+        # Ollama + non-diffusion unknown arch keeps the Ollama hint, not the
+        # generic llama.cpp "unsupported" message.
+        out = "error loading model: unknown model architecture: 'some_new_llm'"
+        msg = _classify(out, self._OLLAMA_GGUF, "ollama/some-new")
+        assert "Ollama" in msg
+        assert "directly through Ollama" in msg
+        assert "does not support" not in msg.lower()
+
+    def test_ollama_diffusion_arch_still_routes_to_images(self):
+        # Diffusion routing wins over the Ollama hint.
+        out = "error loading model: unknown model architecture: 'flux'"
+        msg = _classify(out, self._OLLAMA_GGUF, "ollama/flux")
+        assert "diffusion" in msg.lower()
+        assert "Images page" in msg
+
+    def test_generic_oom_keeps_memory_message(self):
+        msg = _classify(_OOM_OUT, "/models/big.gguf", "local/big")
+        assert "enough memory" in msg.lower()
+        assert "diffusion" not in msg.lower()
+
+    def test_empty_output_is_safe(self):
+        msg = _classify("", None, None)
+        assert "llama-server failed to start" in msg

studio/backend/tests/test_mcp_stdio_improvements.py

@@ -0,0 +1,236 @@
+"""Tests for the proposed PR #5863 improvements.
+
+Covers: _client() self-gating + keep_alive, OAuth normalised off for stdio
+(create + update), env/header dropped on a transport-type switch, and the
+backend rejecting a command whose first token is a URL scheme.
+
+Run from studio/backend:  python -m pytest tests/test_mcp_stdio_improvements.py -q
+"""
+
+import asyncio
+
+import pytest
+from fastapi import HTTPException
+
+from core.inference import mcp_client
+from storage import mcp_servers_db
+
+
+def _reset_db(tmp_path, monkeypatch):
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setattr(mcp_servers_db, "_schema_ready", False)
+
+
+def _enable(monkeypatch):
+    monkeypatch.setenv("UNSLOTH_STUDIO_ALLOW_STDIO_MCP", "1")
+
+
+def _disable(monkeypatch):
+    monkeypatch.delenv("UNSLOTH_STUDIO_ALLOW_STDIO_MCP", raising = False)
+
+
+# ── P1: _client() self-gates the stdio sink ─────────────────────────
+
+
+def test_client_refuses_stdio_when_disabled(monkeypatch):
+    _disable(monkeypatch)
+    with pytest.raises(PermissionError):
+        mcp_client._client("npx -y server /tmp", None)
+
+
+def test_client_builds_stdio_when_enabled_without_spawning(monkeypatch):
+    _enable(monkeypatch)
+    # Constructing the Client must not spawn the subprocess (spawn happens on
+    # __aenter__); we only assert it builds.
+    client = mcp_client._client("npx -y server /tmp", {"K": "v"})
+    assert client is not None
+
+
+def test_client_http_unaffected_by_gate(monkeypatch):
+    _disable(monkeypatch)
+    assert mcp_client._client("https://example.com/mcp", None) is not None
+
+
+# ── P3: OAuth normalised off for stdio (create + update) ────────────
+
+
+def test_create_forces_oauth_off_for_stdio(tmp_path, monkeypatch):
+    import routes.mcp_servers as routes_mcp
+    from models.mcp_servers import McpServerCreate
+
+    _reset_db(tmp_path, monkeypatch)
+    _enable(monkeypatch)
+    resp = asyncio.run(
+        routes_mcp.create_mcp_server(
+            McpServerCreate(
+                display_name = "FS", url = "npx -y server /tmp", use_oauth = True
+            ),
+            current_subject = "u",
+        )
+    )
+    assert resp.use_oauth is False
+    assert mcp_servers_db.get_server(resp.id)["use_oauth"] == 0
+
+
+def test_create_keeps_oauth_for_http(tmp_path, monkeypatch):
+    import routes.mcp_servers as routes_mcp
+    from models.mcp_servers import McpServerCreate
+
+    _reset_db(tmp_path, monkeypatch)
+    _enable(monkeypatch)
+    resp = asyncio.run(
+        routes_mcp.create_mcp_server(
+            McpServerCreate(display_name = "GH", url = "https://gh/mcp", use_oauth = True),
+            current_subject = "u",
+        )
+    )
+    assert resp.use_oauth is True
+
+
+def test_update_url_to_stdio_clears_oauth(tmp_path, monkeypatch):
+    import routes.mcp_servers as routes_mcp
+    from models.mcp_servers import McpServerUpdate
+
+    _reset_db(tmp_path, monkeypatch)
+    _enable(monkeypatch)
+    monkeypatch.setattr(mcp_client, "_oauth_token_store", None)
+    monkeypatch.setattr(
+        routes_mcp, "clear_oauth_tokens_async", lambda *a, **k: asyncio.sleep(0)
+    )
+    mcp_servers_db.create_server(
+        id = "s1", display_name = "A", url = "https://a/mcp", use_oauth = True
+    )
+    resp = asyncio.run(
+        routes_mcp.update_mcp_server(
+            "s1", McpServerUpdate(url = "npx -y server /tmp"), current_subject = "u"
+        )
+    )
+    assert resp.use_oauth is False
+
+
+# ── P4: env/headers dropped on a transport-type switch ──────────────
+
+
+def test_switch_stdio_to_http_drops_env(tmp_path, monkeypatch):
+    import routes.mcp_servers as routes_mcp
+    from models.mcp_servers import McpServerUpdate
+
+    _reset_db(tmp_path, monkeypatch)
+    _enable(monkeypatch)
+    mcp_servers_db.create_server(
+        id = "s1",
+        display_name = "A",
+        url = "npx server",
+        headers_json = '{"API_KEY": "secret"}',
+    )
+    resp = asyncio.run(
+        routes_mcp.update_mcp_server(
+            "s1", McpServerUpdate(url = "https://remote/mcp"), current_subject = "u"
+        )
+    )
+    # the stdio env must NOT survive as HTTP headers on the remote endpoint
+    assert resp.headers == {}
+    assert mcp_servers_db.get_server("s1")["headers_json"] is None
+
+
+def test_switch_keeps_explicitly_supplied_headers(tmp_path, monkeypatch):
+    import routes.mcp_servers as routes_mcp
+    from models.mcp_servers import McpServerUpdate
+
+    _reset_db(tmp_path, monkeypatch)
+    _enable(monkeypatch)
+    mcp_servers_db.create_server(
+        id = "s1",
+        display_name = "A",
+        url = "npx server",
+        headers_json = '{"API_KEY": "secret"}',
+    )
+    resp = asyncio.run(
+        routes_mcp.update_mcp_server(
+            "s1",
+            McpServerUpdate(
+                url = "https://remote/mcp", headers = {"Authorization": "Bearer new"}
+            ),
+            current_subject = "u",
+        )
+    )
+    assert resp.headers == {"Authorization": "Bearer new"}
+
+
+def test_same_transport_edit_keeps_headers(tmp_path, monkeypatch):
+    import routes.mcp_servers as routes_mcp
+    from models.mcp_servers import McpServerUpdate
+
+    _reset_db(tmp_path, monkeypatch)
+    _enable(monkeypatch)
+    mcp_servers_db.create_server(
+        id = "s1",
+        display_name = "A",
+        url = "npx server",
+        headers_json = '{"API_KEY": "secret"}',
+    )
+    # editing only the display name (still stdio) must not wipe env vars
+    resp = asyncio.run(
+        routes_mcp.update_mcp_server(
+            "s1", McpServerUpdate(display_name = "B"), current_subject = "u"
+        )
+    )
+    assert resp.headers == {"API_KEY": "secret"}
+
+
+# ── P5: reject a command whose first token is a URL scheme ───────────
+
+
+def test_validate_url_rejects_url_scheme_command_when_enabled(monkeypatch):
+    from routes.mcp_servers import _validate_url
+
+    _enable(monkeypatch)
+    for bad in ["ftp://host/x", "file:///etc/passwd", "ws://h/y"]:
+        with pytest.raises(HTTPException) as exc:
+            _validate_url(bad)
+        assert exc.value.status_code == 400
+
+
+def test_validate_url_allows_url_in_argument(monkeypatch):
+    from routes.mcp_servers import _validate_url
+
+    _enable(monkeypatch)
+    # :// inside an ARGUMENT (not the first token) is still a valid command
+    assert _validate_url("npx server --url https://x/mcp") == (
+        "npx server --url https://x/mcp"
+    )
+
+
+# ── P6: Data Recipe stdio path obeys the same host gate ─────────────
+# build_mcp_providers needs the data_designer plugin, which is only installed in
+# the Studio test job; skip there rather than fail the core matrix.
+
+_STDIO_RECIPE = {
+    "mcp_providers": [
+        {
+            "provider_type": "stdio",
+            "name": "fs",
+            "command": "npx",
+            "args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"],
+            "env": {},
+        }
+    ]
+}
+
+
+def test_data_recipe_skips_stdio_when_disabled(monkeypatch):
+    pytest.importorskip("data_designer")
+    _disable(monkeypatch)
+    from core.data_recipe.service import build_mcp_providers
+
+    # gate off -> the stdio provider is dropped (no subprocess can be spawned)
+    assert build_mcp_providers(_STDIO_RECIPE) == []
+
+
+def test_data_recipe_builds_stdio_when_enabled(monkeypatch):
+    pytest.importorskip("data_designer")
+    _enable(monkeypatch)
+    from core.data_recipe.service import build_mcp_providers
+
+    built = build_mcp_providers(_STDIO_RECIPE)
+    assert len(built) == 1  # constructed (not spawned) only when enabled

studio/backend/tests/test_mcp_stdio_pr5863.py

@@ -0,0 +1,401 @@
+"""Verification tests for PR #5863 (stdio MCP server support).
+
+Covers the pure helpers (is_stdio / parse_stdio_command / stdio_mcp_enabled /
+probe_timeout), the route-level _validate_url gate, and - most importantly -
+that the UNSLOTH_STUDIO_ALLOW_STDIO_MCP gate blocks the stdio transport at all
+five enforcement points (create, update, test, refresh, discovery, execute)
+when disabled, and reaches it when enabled. The transport (_client) is stubbed
+so no real subprocess is spawned; a recorder asserts whether it was reached.
+
+Run from studio/backend:  python -m pytest tests/test_mcp_stdio_pr5863.py -q
+"""
+
+import sys
+
+import pytest
+from fastapi import HTTPException
+
+from core.inference import mcp_client
+from storage import mcp_servers_db
+
+
+def _reset_db(tmp_path, monkeypatch):
+    monkeypatch.setenv("UNSLOTH_STUDIO_HOME", str(tmp_path))
+    monkeypatch.setattr(mcp_servers_db, "_schema_ready", False)
+
+
+def _enable(monkeypatch):
+    monkeypatch.setenv("UNSLOTH_STUDIO_ALLOW_STDIO_MCP", "1")
+
+
+def _disable(monkeypatch):
+    monkeypatch.delenv("UNSLOTH_STUDIO_ALLOW_STDIO_MCP", raising = False)
+
+
+# ── transport stub + recorder ───────────────────────────────────────
+
+
+class _FakeTool:
+    def __init__(self, name):
+        self._name = name
+
+    def model_dump(self, exclude_none = True):
+        return {"name": self._name, "description": f"{self._name} tool"}
+
+
+class _Block:
+    def __init__(self, text):
+        self.type = "text"
+        self.text = text
+
+
+class _FakeResult:
+    is_error = False
+
+    def __init__(self, text):
+        self.content = [_Block(text)]
+
+
+class _RecordingClient:
+    """Stands in for fastmcp.Client; records that the transport was opened."""
+
+    def __init__(self, url, headers, use_oauth, recorder):
+        recorder.append({"url": url, "headers": headers, "use_oauth": use_oauth})
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, *a):
+        return False
+
+    async def list_tools(self):
+        return [_FakeTool("list_directory"), _FakeTool("write_file")]
+
+    async def call_tool(self, name, args):
+        return _FakeResult(f"called {name}")
+
+
+@pytest.fixture
+def transport(monkeypatch):
+    """Patch mcp_client._client with a recorder. Returns the recorder list;
+    empty == the stdio transport was never reached."""
+    recorder = []
+    monkeypatch.setattr(
+        mcp_client,
+        "_client",
+        lambda url, headers, use_oauth = False: _RecordingClient(
+            url, headers, use_oauth, recorder
+        ),
+    )
+    return recorder
+
+
+# ── 1. is_stdio ─────────────────────────────────────────────────────
+
+
+@pytest.mark.parametrize(
+    "addr",
+    [
+        "http://localhost:8000/mcp",
+        "https://example.com/mcp",
+        "  https://example.com/mcp  ",
+        "HTTPS://EXAMPLE.COM/mcp",
+    ],
+)
+def test_is_stdio_false_for_http(addr):
+    assert mcp_client.is_stdio(addr) is False
+
+
+@pytest.mark.parametrize(
+    "addr",
+    [
+        "npx -y @modelcontextprotocol/server-filesystem /tmp",
+        "python -m some.module",
+        "uvx some-server --flag",
+        "/usr/local/bin/my-server",
+    ],
+)
+def test_is_stdio_true_for_commands(addr):
+    assert mcp_client.is_stdio(addr) is True
+
+
+# ── 2. parse_stdio_command ──────────────────────────────────────────
+
+
+def test_parse_basic_argv():
+    assert mcp_client.parse_stdio_command(
+        "npx -y @modelcontextprotocol/server-filesystem /tmp"
+    ) == ["npx", "-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
+
+
+def test_parse_keeps_url_argument_as_one_command():
+    # gemini "high": a :// inside an ARGUMENT must not break the command.
+    assert mcp_client.parse_stdio_command(
+        "npx server --endpoint https://example.com/mcp"
+    ) == ["npx", "server", "--endpoint", "https://example.com/mcp"]
+
+
+def test_parse_quoted_arg():
+    assert mcp_client.parse_stdio_command('python -m mod --name "a b"') == [
+        "python",
+        "-m",
+        "mod",
+        "--name",
+        "a b",
+    ]
+
+
+def test_parse_empty_returns_empty_list():
+    assert mcp_client.parse_stdio_command("   ") == []
+
+
+def test_parse_unclosed_quote_raises_valueerror():
+    with pytest.raises(ValueError):
+        mcp_client.parse_stdio_command('npx "unclosed')
+
+
+def test_parse_windows_strips_wrapping_quotes(monkeypatch):
+    # gemini "medium": posix=False keeps backslash paths but also the wrapping
+    # quotes; the PR strips a matched pair so argv[0] reaches the OS clean.
+    monkeypatch.setattr(sys, "platform", "win32")
+    parts = mcp_client.parse_stdio_command(
+        r'"C:\Program Files\node\node.exe" server.js'
+    )
+    assert parts[0] == r"C:\Program Files\node\node.exe"
+    assert parts[1] == "server.js"
+
+
+# ── 3. stdio_mcp_enabled ────────────────────────────────────────────
+
+
+@pytest.mark.parametrize("val", ["0", "false", "true", "", " 1 ", "yes", "2"])
+def test_stdio_disabled_for_non_exact_one(monkeypatch, val):
+    monkeypatch.setenv("UNSLOTH_STUDIO_ALLOW_STDIO_MCP", val)
+    assert mcp_client.stdio_mcp_enabled() is False
+
+
+def test_stdio_enabled_only_for_exact_one(monkeypatch):
+    _disable(monkeypatch)
+    assert mcp_client.stdio_mcp_enabled() is False
+    monkeypatch.setenv("UNSLOTH_STUDIO_ALLOW_STDIO_MCP", "1")
+    assert mcp_client.stdio_mcp_enabled() is True
+
+
+# ── 4. probe_timeout ────────────────────────────────────────────────
+
+
+def test_probe_timeout_matrix():
+    assert mcp_client.probe_timeout("https://x/mcp", False) == 8.0
+    assert mcp_client.probe_timeout("https://x/mcp", True) == 305.0
+    assert mcp_client.probe_timeout("npx server", False) == 60.0
+    # oauth wins regardless of address kind (documented behaviour)
+    assert mcp_client.probe_timeout("npx server", True) == 305.0
+
+
+# ── 5. _validate_url gate ───────────────────────────────────────────
+
+
+def test_validate_url_gate_off_rejects_stdio(monkeypatch):
+    _disable(monkeypatch)
+    from routes.mcp_servers import _validate_url
+
+    assert _validate_url("https://example.com/mcp") == "https://example.com/mcp"
+    # urlparse reads "localhost:8000" scheme as "localhost", so it lands here too.
+    for bad in [
+        "npx server",
+        "python -m mod",
+        "ftp://host",
+        "example.com",
+        "localhost:8000",
+        r"C:\node\node.exe server.js",
+    ]:
+        with pytest.raises(HTTPException) as exc:
+            _validate_url(bad)
+        assert exc.value.status_code == 400
+
+
+def test_validate_url_gate_off_message_depends_on_whitespace(monkeypatch):
+    # The message names a command only when the value has whitespace, and never
+    # says "desktop app only" (self-hosted hosts can opt in via the env var).
+    _disable(monkeypatch)
+    from routes.mcp_servers import _validate_url
+
+    with pytest.raises(HTTPException) as exc:
+        _validate_url("npx -y @modelcontextprotocol/server-filesystem /tmp")
+    cmd = exc.value.detail.lower()
+    assert "http://" in cmd and "https://" in cmd
+    assert "local command" in cmd
+    assert "desktop app" not in cmd
+
+    with pytest.raises(HTTPException) as exc:
+        _validate_url("example.com")
+    lone = exc.value.detail.lower()
+    assert "http://" in lone and "https://" in lone
+    assert "local command" not in lone
+
+
+def test_validate_url_gate_on_accepts_stdio(monkeypatch):
+    _enable(monkeypatch)
+    from routes.mcp_servers import _validate_url
+
+    assert _validate_url("npx -y server /tmp") == "npx -y server /tmp"
+    # http still works when stdio is on
+    assert _validate_url("https://x/mcp") == "https://x/mcp"
+    # url-bearing argument accepted as a command
+    assert _validate_url("npx server --url https://x/mcp") == (
+        "npx server --url https://x/mcp"
+    )
+    # A lone token is ambiguous; keep the prior behaviour and accept it as a
+    # command rather than guessing it's a URL (no regression for single binaries).
+    assert (
+        _validate_url("/usr/local/bin/my-mcp-server") == "/usr/local/bin/my-mcp-server"
+    )
+    assert _validate_url("mcp-server-sqlite") == "mcp-server-sqlite"
+    # empty / unparseable still rejected
+    for bad in ["   ", '"unclosed']:
+        with pytest.raises(HTTPException) as exc:
+            _validate_url(bad)
+        assert exc.value.status_code == 400
+
+
+# ── 6. gate enforcement at every spawn path (mocked transport) ──────
+
+
+def test_create_route_gate(tmp_path, monkeypatch, transport):
+    import asyncio
+
+    from models.mcp_servers import McpServerCreate
+    import routes.mcp_servers as routes_mcp
+
+    _reset_db(tmp_path, monkeypatch)
+    payload = McpServerCreate(display_name = "FS", url = "npx -y server /tmp")
+
+    _disable(monkeypatch)
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(routes_mcp.create_mcp_server(payload, current_subject = "u"))
+    assert exc.value.status_code == 400
+
+    _enable(monkeypatch)
+    resp = asyncio.run(routes_mcp.create_mcp_server(payload, current_subject = "u"))
+    assert resp.url == "npx -y server /tmp"
+
+
+def test_update_http_to_stdio_blocked_when_off(tmp_path, monkeypatch):
+    import asyncio
+
+    from models.mcp_servers import McpServerUpdate
+    import routes.mcp_servers as routes_mcp
+
+    _reset_db(tmp_path, monkeypatch)
+    _disable(monkeypatch)
+    mcp_servers_db.create_server(id = "s1", display_name = "A", url = "https://a/mcp")
+    # editing url -> stdio command must 400 (http->stdio edit bypass closed)
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(
+            routes_mcp.update_mcp_server(
+                "s1", McpServerUpdate(url = "npx server"), current_subject = "u"
+            )
+        )
+    assert exc.value.status_code == 400
+
+
+def test_test_route_gate(tmp_path, monkeypatch, transport):
+    import asyncio
+
+    from models.mcp_servers import McpServerTestRequest
+    import routes.mcp_servers as routes_mcp
+
+    _reset_db(tmp_path, monkeypatch)
+    req = McpServerTestRequest(url = "npx -y server /tmp")
+
+    _disable(monkeypatch)
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(routes_mcp.test_mcp_server(req, current_subject = "u"))
+    assert exc.value.status_code == 400
+    assert transport == []  # transport never opened
+
+    _enable(monkeypatch)
+    res = asyncio.run(routes_mcp.test_mcp_server(req, current_subject = "u"))
+    assert res.ok and res.tool_count == 2
+    assert len(transport) == 1
+
+
+def test_refresh_route_gate(tmp_path, monkeypatch, transport):
+    import asyncio
+
+    import routes.mcp_servers as routes_mcp
+
+    _reset_db(tmp_path, monkeypatch)
+    # a stdio row as if carried over from a desktop DB
+    mcp_servers_db.create_server(id = "stdio1", display_name = "FS", url = "npx server")
+
+    _disable(monkeypatch)
+    with pytest.raises(HTTPException) as exc:
+        asyncio.run(routes_mcp.refresh_mcp_server_tools("stdio1", current_subject = "u"))
+    assert exc.value.status_code == 400
+    assert transport == []
+
+    _enable(monkeypatch)
+    res = asyncio.run(
+        routes_mcp.refresh_mcp_server_tools("stdio1", current_subject = "u")
+    )
+    assert res.ok and res.tool_count == 2
+    assert len(transport) == 1
+
+
+def test_discovery_gate(tmp_path, monkeypatch, transport):
+    import asyncio
+
+    from core.inference.tools import get_enabled_mcp_tools
+
+    _reset_db(tmp_path, monkeypatch)
+    mcp_servers_db.create_server(
+        id = "stdio1", display_name = "FS", url = "npx server", is_enabled = True
+    )
+
+    _disable(monkeypatch)
+    assert asyncio.run(get_enabled_mcp_tools()) == []
+    assert transport == []  # filtered out before any probe
+
+    _enable(monkeypatch)
+    specs = asyncio.run(get_enabled_mcp_tools())
+    assert len(specs) == 2
+    assert len(transport) == 1
+
+
+def test_execute_gate(tmp_path, monkeypatch, transport):
+    from core.inference.tools import execute_tool
+
+    _reset_db(tmp_path, monkeypatch)
+    mcp_servers_db.create_server(
+        id = "stdio1", display_name = "FS", url = "npx server", is_enabled = True
+    )
+
+    _disable(monkeypatch)
+    out = execute_tool("mcp__stdio1__list_directory", {"path": "/tmp"})
+    assert "disabled on this host" in out
+    assert transport == []
+
+    _enable(monkeypatch)
+    out = execute_tool("mcp__stdio1__list_directory", {"path": "/tmp"})
+    assert out == "called list_directory"
+    assert len(transport) == 1
+
+
+# ── 7. env vars ride headers_json as the subprocess env ─────────────
+
+
+def test_stdio_env_passed_through(tmp_path, monkeypatch, transport):
+    from core.inference.tools import execute_tool
+
+    _reset_db(tmp_path, monkeypatch)
+    _enable(monkeypatch)
+    mcp_servers_db.create_server(
+        id = "stdio1",
+        display_name = "FS",
+        url = "npx server",
+        headers_json = '{"API_KEY": "sk-test"}',
+        is_enabled = True,
+    )
+    execute_tool("mcp__stdio1__list_directory", {})
+    assert transport[-1]["headers"] == {"API_KEY": "sk-test"}

studio/backend/tests/test_safetensors_tool_loop.py

@@ -28,12 +28,14 @@ Edge cases under coverage:
 """
 
 import threading
+from typing import cast
 
 import pytest
 
 from core.inference import safetensors_agentic
 from core.inference.safetensors_agentic import (
     _coerce_arguments,
+    _detect_render_html_tool_start,
     run_safetensors_tool_loop,
 )
 from core.inference.tool_call_parser import (
@@ -97,6 +99,17 @@ class TestParser:
         assert len(result) == 1
         assert "print('hi')" in result[0]["function"]["arguments"]
 
+    def test_function_signal_inside_parameter_is_literal(self):
+        text = (
+            "<function=python>"
+            "<parameter=code>print('<function=render_html>')</parameter>"
+            "</function>"
+        )
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "python"
+        assert "<function=render_html>" in result[0]["function"]["arguments"]
+
     def test_multiple_calls(self):
         text = (
             '<tool_call>{"name":"web_search","arguments":{"query":"a"}}</tool_call>'
@@ -118,6 +131,18 @@ class TestParser:
         assert has_tool_signal("hi <function=foo>...")
         assert not has_tool_signal("hello world")
 
+    def test_render_html_start_detector_uses_first_tool(self):
+        assert _detect_render_html_tool_start("<function=render_html>")
+        assert _detect_render_html_tool_start(
+            '<tool_call>{"name":"render_html","arguments":{"code":"<html>"}'
+        )
+        assert not _detect_render_html_tool_start(
+            "<function=python><parameter=code>'<function=render_html>'"
+        )
+        assert not _detect_render_html_tool_start(
+            '<tool_call>{"name":"python","arguments":{"code":"<function=render_html>"}}'
+        )
+
     def test_strip_markup_closed(self):
         text = "before <tool_call>{}</tool_call> after"
         assert strip_tool_markup(text) == "before  after"
@@ -281,6 +306,104 @@ class TestLoopBasic:
         contents = [e for e in events if e["type"] == "content"]
         assert "Result: 1" in contents[-1]["text"]
 
+    def test_render_html_emits_provisional_tool_start(self):
+        exec_fn = FakeExecuteTool(["Rendered HTML artifact."])
+        turn_iter = iter(
+            [
+                [
+                    "<function=render_html>",
+                    "<parameter=code><!doctype html><html>",
+                    "<body>Hi</body></html></parameter></function>",
+                ],
+                ["Done."],
+            ]
+        )
+
+        def _gen(_messages):
+            chunks = next(turn_iter)
+            acc = ""
+            for chunk in chunks:
+                acc += chunk
+                yield acc
+
+        loop = run_safetensors_tool_loop(
+            single_turn = _gen,
+            messages = [{"role": "user", "content": "make html"}],
+            tools = [{"type": "function", "function": {"name": "render_html"}}],
+            execute_tool = exec_fn,
+        )
+        events = _collect_events(loop)
+        tool_starts = [e for e in events if e["type"] == "tool_start"]
+
+        assert len(tool_starts) == 2
+        assert tool_starts[0]["tool_name"] == "render_html"
+        assert tool_starts[0]["arguments"] == {}
+        assert tool_starts[1]["tool_name"] == "render_html"
+        assert "<!doctype html>" in tool_starts[1]["arguments"]["code"]
+        assert exec_fn.calls[0][0] == "render_html"
+        assert "<!doctype html>" in exec_fn.calls[0][1]["code"]
+
+    def test_python_tool_containing_render_html_signal_does_not_emit_provisional_start(
+        self,
+    ):
+        loop, exec_fn = _make_loop(
+            turns = [
+                [
+                    "<function=python>",
+                    "<parameter=code>print('<function=render_html>')",
+                    "</parameter></function>",
+                ],
+                ["Done."],
+            ],
+            exec_results = ["ok"],
+        )
+        events = _collect_events(loop)
+        tool_starts = [e for e in events if e["type"] == "tool_start"]
+
+        assert len(tool_starts) == 1
+        assert tool_starts[0]["tool_name"] == "python"
+        assert exec_fn.calls == [
+            ("python", {"code": "print('<function=render_html>')"})
+        ]
+
+    def test_render_html_success_blocks_second_artifact_call(self):
+        exec_fn = FakeExecuteTool(["Rendered HTML artifact."])
+        turn_iter = iter(
+            [
+                [
+                    '<tool_call>{"name":"render_html",',
+                    '"arguments":{"code":"<html>one</html>"}}',
+                ],
+                [
+                    '<tool_call>{"name":"render_html",',
+                    '"arguments":{"code":"<html>two</html>"}}',
+                ],
+                ["Done."],
+            ]
+        )
+
+        def _gen(_messages):
+            chunks = next(turn_iter)
+            acc = ""
+            for chunk in chunks:
+                acc += chunk
+                yield acc
+
+        loop = run_safetensors_tool_loop(
+            single_turn = _gen,
+            messages = [{"role": "user", "content": "make html"}],
+            tools = [{"type": "function", "function": {"name": "render_html"}}],
+            execute_tool = exec_fn,
+        )
+        events = _collect_events(loop)
+        tool_starts = [e for e in events if e["type"] == "tool_start"]
+
+        assert exec_fn.calls == [("render_html", {"code": "<html>one</html>"})]
+        assert [e["arguments"] for e in tool_starts] == [
+            {},
+            {"code": "<html>one</html>"},
+        ]
+
     def test_truncated_unclosed_tool_call(self):
         loop, exec_fn = _make_loop(
             turns = [
@@ -596,6 +719,7 @@ class TestChatTemplateHelper:
         tok = self._Tok({"tools", "enable_thinking"})
         self.apply(tok, [], tools = [{}], enable_thinking = True)
         assert tok.call_count == 1
+        assert tok.last_kwargs is not None
         assert "tools" in tok.last_kwargs
         assert "enable_thinking" in tok.last_kwargs
 
@@ -782,7 +906,7 @@ class TestGptOssNameDetection:
 
     def test_empty_or_none_returns_false(self):
         assert is_gpt_oss_model_name("") is False
-        assert is_gpt_oss_model_name(None) is False
+        assert is_gpt_oss_model_name(cast(str, None)) is False
 
 
 if __name__ == "__main__":

studio/backend/utils/paths/__init__.py

@@ -25,6 +25,8 @@ from .storage_roots import (
     auth_root,
     auth_db_path,
     studio_db_path,
+    documents_root,
+    project_workspaces_root,
     tmp_root,
     seed_uploads_root,
     unstructured_seed_cache_root,
@@ -44,6 +46,10 @@ from .storage_roots import (
     resolve_dataset_path,
 )
 
+# Re-export shim: name-load the project-path helpers so the import-hoist
+# safety net sees them used here, not just listed in __all__ as strings.
+_REEXPORTED = (documents_root, project_workspaces_root)
+
 __all__ = [
     "normalize_path",
     "is_local_path",
@@ -62,6 +68,8 @@ __all__ = [
     "auth_root",
     "auth_db_path",
     "studio_db_path",
+    "documents_root",
+    "project_workspaces_root",
     "tmp_root",
     "seed_uploads_root",
     "unstructured_seed_cache_root",

studio/backend/utils/paths/storage_roots.py

@@ -96,6 +96,38 @@ def studio_db_path() -> Path:
     return studio_root() / "studio.db"
 
 
+def _xdg_user_dir(key: str) -> Path | None:
+    config = Path.home() / ".config" / "user-dirs.dirs"
+    try:
+        lines = config.read_text(encoding = "utf-8").splitlines()
+    except OSError:
+        return None
+    prefix = f"{key}="
+    for line in lines:
+        line = line.strip()
+        if not line.startswith(prefix):
+            continue
+        value = line[len(prefix) :].strip().strip('"')
+        if not value:
+            return None
+        return Path(value.replace("$HOME", str(Path.home()))).expanduser()
+    return None
+
+
+def documents_root() -> Path:
+    override = (os.environ.get("UNSLOTH_STUDIO_DOCUMENTS_HOME") or "").strip()
+    if override:
+        return Path(override).expanduser()
+    return _xdg_user_dir("XDG_DOCUMENTS_DIR") or (Path.home() / "Documents")
+
+
+def project_workspaces_root() -> Path:
+    override = (os.environ.get("UNSLOTH_STUDIO_PROJECTS_HOME") or "").strip()
+    if override:
+        return Path(override).expanduser()
+    return documents_root() / "Unsloth Studio" / "Projects"
+
+
 def tmp_root() -> Path:
     return Path(tempfile.gettempdir()) / "unsloth-studio"
 

studio/frontend/package-lock.json

@@ -14101,9 +14101,9 @@
       }
     },
     "node_modules/react-is": {
-      "version": "19.2.5",
-      "resolved": "https://registry.npmjs.org/react-is/-/react-is-19.2.5.tgz",
-      "integrity": "sha512-Dn0t8IQhCmeIT3wu+Apm1/YVsJXsGWi6k4sPdnBIdqMVtHtv0IGi6dcpNpNkNac0zB2uUAqNX3MHzN8c+z2rwQ==",
+      "version": "19.2.6",
+      "resolved": "https://registry.npmjs.org/react-is/-/react-is-19.2.6.tgz",
+      "integrity": "sha512-XjBR15BhXuylgWGuslhDKqlSayuqvqBX91BP8pauG8kd1zY8kotkNWbXksTCNRarse4kuGbe2kIY05ARtwNIvw==",
       "license": "MIT",
       "peer": true
     },

studio/frontend/src/app/router.tsx

@@ -14,6 +14,7 @@ import { Route as indexRoute } from "./routes/index";
 import { Route as knowledgeBasesRoute } from "./routes/knowledge-bases";
 import { Route as loginRoute } from "./routes/login";
 import { Route as onboardingRoute } from "./routes/onboarding";
+import { Route as projectsRoute } from "./routes/projects";
 import { Route as changePasswordRoute } from "./routes/change-password";
 import { Route as settingsRoute } from "./routes/settings";
 import { Route as studioRoute } from "./routes/studio";
@@ -28,6 +29,7 @@ const routeTree = rootRoute.addChildren([
   knowledgeBasesRoute,
   studioRoute,
   chatRoute,
+  projectsRoute,
   exportRoute,
   dataRecipesRoute,
   dataRecipeRoute,

studio/frontend/src/app/routes/__root.tsx

@@ -7,6 +7,7 @@ import { fetchDeviceType, usePlatformStore } from "@/config/env";
 import { SidebarInset, SidebarProvider } from "@/components/ui/sidebar";
 import { IngestionToastStack } from "@/features/rag/components/ingestion-toast-stack";
 import { SettingsDialog, useSettingsDialogStore } from "@/features/settings";
+import { useChatRuntimeStore } from "@/features/chat";
 import { useTrainingUnloadGuard } from "@/features/training";
 import { useSidebarPin } from "@/hooks/use-sidebar-pin";
 import { useT, type TranslationKey } from "@/i18n";
@@ -41,6 +42,7 @@ function RouteFallback() {
 const CHAT_ONLY_ALLOWED = new Set([
   "/",
   "/chat",
+  "/projects",
   "/login",
   "/signup",
   "/change-password",
@@ -111,6 +113,13 @@ function RootLayout() {
     return () => window.removeEventListener("keydown", handler);
   }, []);
 
+  useEffect(() => {
+    if (isChatRoute) return;
+    const chatRuntime = useChatRuntimeStore.getState();
+    chatRuntime.setActiveProjectId(null);
+    chatRuntime.setActiveThreadId(null);
+  }, [isChatRoute]);
+
   return (
     <AppProvider>
       <SettingsDialog />

studio/frontend/src/app/routes/chat.tsx

@@ -10,6 +10,7 @@ export type ChatSearch = {
   thread?: string;
   compare?: string;
   new?: string;
+  project?: string;
 };
 
 export const Route = createRoute({
@@ -21,6 +22,7 @@ export const Route = createRoute({
     thread: typeof search.thread === "string" ? search.thread : undefined,
     compare: typeof search.compare === "string" ? search.compare : undefined,
     new: typeof search.new === "string" ? search.new : undefined,
+    project: typeof search.project === "string" ? search.project : undefined,
   }),
   component: ChatPage,
 });

studio/frontend/src/app/routes/projects.tsx

@@ -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 { createRoute } from "@tanstack/react-router";
+import { lazy } from "react";
+import { requireAuth } from "../auth-guards";
+import { Route as rootRoute } from "./__root";
+
+const ProjectsPage = lazy(() =>
+  import("@/features/chat/projects-page").then((m) => ({
+    default: m.ProjectsPage,
+  })),
+);
+
+export const Route = createRoute({
+  getParentRoute: () => rootRoute,
+  path: "/projects",
+  staticData: { title: "Projects" },
+  beforeLoad: () => requireAuth(),
+  component: ProjectsPage,
+});

studio/frontend/src/components/app-sidebar.tsx

@@ -26,6 +26,9 @@ import {
   DropdownMenuItem,
   DropdownMenuSeparator,
   DropdownMenuShortcut,
+  DropdownMenuSub,
+  DropdownMenuSubContent,
+  DropdownMenuSubTrigger,
   DropdownMenuTrigger,
 } from "@/components/ui/dropdown-menu";
 import {
@@ -38,6 +41,7 @@ import {
 } from "@/components/ui/dialog";
 import { Input } from "@/components/ui/input";
 import { Button } from "@/components/ui/button";
+import { Switch } from "@/components/ui/switch";
 import { useAnimatedThemeToggle } from "@/components/ui/animated-theme-toggler";
 import { cn } from "@/lib/utils";
 import {
@@ -47,9 +51,12 @@ import {
   Delete02Icon,
   DownloadSquare01Icon,
   Edit03Icon,
+  FolderAddIcon,
+  Folder01Icon,
   Globe02Icon,
   HelpCircleIcon,
   Logout05Icon,
+  MoreVerticalIcon,
   Search01Icon,
   PowerIcon,
   PencilEdit02Icon,
@@ -68,11 +75,17 @@ import { ChevronDown, ChevronsUpDown, MoreHorizontalIcon, Moon, Sun } from "luci
 import { Link, useNavigate, useRouterState } from "@tanstack/react-router";
 import {
   ChatSearchDialog,
+  createChatProject,
+  deleteChatProject,
   deleteChatItem,
+  moveChatItemToProject,
   renameChatItem,
+  renameChatProject,
   useChatRuntimeStore,
+  useChatProjects,
   useChatSearchStore,
   useChatSidebarItems,
+  type ProjectRecord,
   type SidebarItem,
 } from "@/features/chat";
 import { useSettingsDialogStore } from "@/features/settings";
@@ -90,7 +103,7 @@ import {
   useTrainingRuntimeStore,
 } from "@/features/training";
 import type { TrainingRunSummary } from "@/features/training";
-import { useEffect, useRef, useState, type ReactNode } from "react";
+import { useEffect, useMemo, useRef, useState, type ReactNode } from "react";
 import { toast } from "@/lib/toast";
 import { ShutdownDialog } from "@/components/shutdown-dialog";
 import { translate, useT, type TranslationKey } from "@/i18n";
@@ -185,7 +198,7 @@ function NavItem({
   active: boolean;
   disabled?: boolean;
   onClick: () => void;
-  children?: React.ReactNode;
+  children?: ReactNode;
   dataTour?: string;
 }) {
   return (
@@ -197,7 +210,7 @@ function NavItem({
           onClick={onClick}
           isActive={active}
           data-tour={dataTour}
-          className="sidebar-nav-btn h-[35px] rounded-[10px] gap-[8.5px] px-2.5 font-medium group-data-[collapsible=icon]:!w-[32px] group-data-[collapsible=icon]:!rounded-[10px] group-data-[collapsible=icon]:mx-auto"
+          className="sidebar-nav-btn h-[33px] rounded-[10px] gap-[8.5px] px-2.5 font-medium group-data-[collapsible=icon]:!w-[32px] group-data-[collapsible=icon]:!rounded-[10px] group-data-[collapsible=icon]:mx-auto"
         >
           <HugeiconsIcon icon={icon} strokeWidth={1.75} className="size-icon! shrink-0 group-hover/menu-button:animate-icon-pop" />
           <span className="text-[14.5px] leading-[19px] tracking-nav">{label}</span>
@@ -231,22 +244,53 @@ export function AppSidebar() {
 
   const isChatRoute = pathname.startsWith("/chat");
   const isStudioRoute = pathname === "/studio" || pathname.startsWith("/studio/");
+  const [chatOpen, setChatOpen] = useState(true);
+  const [trainOpen, setTrainOpen] = useState(true);
+  const [runsOpen, setRunsOpen] = useState(true);
+
+  useEffect(() => {
+    if (!isChatRoute) return;
+    queueMicrotask(() => setChatOpen(true));
+  }, [isChatRoute]);
+  useEffect(() => {
+    if (!isStudioRoute) return;
+    queueMicrotask(() => setRunsOpen(true));
+  }, [isStudioRoute]);
 
   const scrollRef = useRef<HTMLDivElement | null>(null);
   const [scrolled, setScrolled] = useState(false);
-  useEffect(() => {
-    const el = scrollRef.current;
-    if (!el) return;
-    const handler = () => setScrolled(el.scrollTop > 0);
-    handler();
-    el.addEventListener("scroll", handler, { passive: true });
-    return () => el.removeEventListener("scroll", handler);
-  }, []);
+  // Bottom fade hides at the very bottom (and for short, non-scrolling lists)
+  // so the last row isn't washed out - Gemini-style.
+  const [canScrollDown, setCanScrollDown] = useState(false);
+  // Driven only from onScroll + a content-change effect below. Deliberately NO
+  // ResizeObserver: its callback-driven setState created a render loop (React
+  // #185). Both setters bail out when unchanged, so neither path can loop.
+  const syncScrollState = (el: HTMLDivElement) => {
+    const nextScrolled = el.scrollTop > 0;
+    setScrolled((prev) => (prev === nextScrolled ? prev : nextScrolled));
+    const nextCanScrollDown =
+      el.scrollHeight - el.scrollTop - el.clientHeight > 1;
+    setCanScrollDown((prev) =>
+      prev === nextCanScrollDown ? prev : nextCanScrollDown,
+    );
+  };
 
   const isRecipesRoute = pathname.startsWith("/data-recipes");
   const { displayTitle, avatarDataUrl } = useEffectiveProfile();
 
-  const { items: chatItems } = useChatSidebarItems();
+  const { projects } = useChatProjects();
+  const activeProjectId = isChatRoute
+    ? ((search.project as string | undefined) ?? null)
+    : null;
+  const { items: allChatItems } = useChatSidebarItems({
+    enabled: !isStudioRoute,
+    requireMessages: false,
+  });
+  const recentChatItems = useMemo(
+    () => allChatItems.filter((item) => !item.projectId),
+    [allChatItems],
+  );
+  const chatItems = allChatItems;
   const storeThreadId = useChatRuntimeStore((s) => s.activeThreadId);
   const setActiveThreadId = useChatRuntimeStore((s) => s.setActiveThreadId);
   const activeThreadId = isChatRoute
@@ -261,33 +305,87 @@ export function AppSidebar() {
     !chatOnly && isStudioRoute,
   );
   const activeJobId = useTrainingRuntimeStore((s) => s.jobId);
+  const currentRunViewActive = useTrainingRuntimeStore((s) => s.currentRunViewActive);
   const selectedHistoryRunId = useTrainingRuntimeStore((s) => s.selectedHistoryRunId);
   const setSelectedHistoryRunId = useTrainingRuntimeStore((s) => s.setSelectedHistoryRunId);
 
+  // Recompute the bottom-fade state on mount and whenever the list height can
+  // change (items load, sections collapse/expand, route switches the visible
+  // list) - onScroll never fires for short, non-scrolling lists. Guarded
+  // setState below means this can't loop even if a dep is a fresh reference.
+  useEffect(() => {
+    const el = scrollRef.current;
+    if (!el) return;
+    const next = el.scrollHeight - el.scrollTop - el.clientHeight > 1;
+    setCanScrollDown((prev) => (prev === next ? prev : next));
+  }, [
+    recentChatItems.length,
+    runItems.length,
+    projects.length,
+    chatOpen,
+    trainOpen,
+    runsOpen,
+    isStudioRoute,
+  ]);
+
   const chatDisabled = isTrainingRunning;
 
+  function chatSearchForProject(projectId: string | null) {
+    if (projectId) {
+      return { project: projectId };
+    }
+    return {
+      new: createNavigationNonce(),
+    };
+  }
+
+  function openNewChat(projectId = activeProjectId) {
+    if (chatDisabled) return;
+    setActiveThreadId(null);
+    useChatRuntimeStore.getState().setActiveProjectId(projectId);
+    navigate({ to: "/chat", search: chatSearchForProject(projectId) });
+    closeMobileIfOpen();
+  }
+
+  function openProject(projectId: string) {
+    if (chatDisabled) return;
+    setActiveThreadId(null);
+    useChatRuntimeStore.getState().setActiveProjectId(projectId);
+    navigate({ to: "/chat", search: { project: projectId } });
+    closeMobileIfOpen();
+  }
+
   async function handleDeleteThread(item: Parameters<typeof deleteChatItem>[0]) {
     await deleteChatItem(item, activeThreadId, (view) => {
       navigate({
         to: "/chat",
-        search: { new: view.newThreadNonce },
+        search: item.projectId
+          ? { project: item.projectId }
+          : { new: view.newThreadNonce },
       });
     });
   }
 
   type RenameTarget =
     | { kind: "chat"; item: SidebarItem; current: string }
+    | { kind: "project"; project: ProjectRecord; current: string }
     | { kind: "run"; run: TrainingRunSummary; current: string };
   const [renamingTarget, setRenamingTarget] = useState<RenameTarget | null>(
     null,
   );
   const [renameDraft, setRenameDraft] = useState("");
+  const [creatingProject, setCreatingProject] = useState(false);
+  const [projectNameDraft, setProjectNameDraft] = useState("");
+  const [projectCreateMoveTarget, setProjectCreateMoveTarget] =
+    useState<SidebarItem | null>(null);
   const renameTrimmed = renameDraft.trim();
   const nextRunDisplayName = renameTrimmed.length > 0 ? renameTrimmed : null;
   const renameDirty =
     renamingTarget !== null &&
     (renamingTarget.kind === "chat"
       ? renameTrimmed.length > 0 && renameTrimmed !== renamingTarget.current
+      : renamingTarget.kind === "project"
+        ? renameTrimmed.length > 0 && renameTrimmed !== renamingTarget.current
       : renameTrimmed.length > 0
         ? renameTrimmed !== renamingTarget.current
         : renamingTarget.run.display_name != null);
@@ -315,6 +413,16 @@ export function AppSidebar() {
       }
       return;
     }
+    if (target.kind === "project") {
+      try {
+        await renameChatProject(target.project.id, renameTrimmed);
+      } catch (err) {
+        toast.error("Failed to rename project", {
+          description: err instanceof Error ? err.message : undefined,
+        });
+      }
+      return;
+    }
     try {
       const updated = await renameTrainingRun(target.run.id, nextRunDisplayName);
       emitTrainingRunUpdated(updated);
@@ -327,13 +435,23 @@ export function AppSidebar() {
 
   type DeleteTarget =
     | { kind: "chat"; item: SidebarItem }
+    | { kind: "project"; project: ProjectRecord }
     | { kind: "run"; run: TrainingRunSummary };
   const [confirmingDelete, setConfirmingDelete] =
     useState<DeleteTarget | null>(null);
+  const [deleteProjectFiles, setDeleteProjectFiles] = useState(false);
+
+  useEffect(() => {
+    if (confirmingDelete?.kind !== "project") {
+      setDeleteProjectFiles(false);
+    }
+  }, [confirmingDelete]);
 
   async function commitDelete() {
     const target = confirmingDelete;
     if (!target) return;
+    const shouldDeleteProjectFiles =
+      target.kind === "project" && deleteProjectFiles;
     setConfirmingDelete(null);
     if (target.kind === "chat") {
       try {
@@ -345,6 +463,22 @@ export function AppSidebar() {
       }
       return;
     }
+    if (target.kind === "project") {
+      try {
+        await deleteChatProject(target.project.id, {
+          deleteFiles: shouldDeleteProjectFiles,
+        });
+        if (activeProjectId === target.project.id) {
+          useChatRuntimeStore.getState().setActiveProjectId(null);
+          navigate({ to: "/chat", search: { new: createNavigationNonce() } });
+        }
+      } catch (err) {
+        toast.error("Failed to delete project", {
+          description: err instanceof Error ? err.message : undefined,
+        });
+      }
+      return;
+    }
     if (target.run.status === "running") {
       toast.error(t("shell.toast.cannotDeleteRunningRun"));
       return;
@@ -362,6 +496,168 @@ export function AppSidebar() {
     }
   }
 
+  async function commitCreateProject() {
+    const name = projectNameDraft.trim();
+    if (!name) return;
+    const moveTarget = projectCreateMoveTarget;
+    try {
+      const project = await createChatProject(name);
+      if (moveTarget) {
+        await moveChatItemToProject(moveTarget, project.id);
+        if (activeThreadId === moveTarget.id) {
+          useChatRuntimeStore.getState().setActiveProjectId(project.id);
+        }
+      }
+      setCreatingProject(false);
+      setProjectNameDraft("");
+      setProjectCreateMoveTarget(null);
+      if (moveTarget) {
+        return;
+      } else {
+        openProject(project.id);
+      }
+    } catch (err) {
+      toast.error(moveTarget ? "Failed to create and move chat" : "Failed to create project", {
+        description: err instanceof Error ? err.message : undefined,
+      });
+    }
+  }
+
+  async function moveChatToProject(item: SidebarItem, projectId: string | null) {
+    if (item.projectId === projectId) return;
+    try {
+      await moveChatItemToProject(item, projectId);
+      if (activeThreadId === item.id) {
+        useChatRuntimeStore.getState().setActiveProjectId(projectId);
+      }
+    } catch (err) {
+      toast.error("Failed to move chat", {
+        description: err instanceof Error ? err.message : undefined,
+      });
+    }
+  }
+
+  function renderChatSidebarItem(
+    item: SidebarItem,
+    variant: "project" | "recent",
+  ) {
+    const itemClass =
+      variant === "project"
+        ? "group/project-chat-item relative"
+        : "group/recent-item relative";
+    const actionClass =
+      variant === "project"
+        ? "sidebar-row-action group-hover/project-chat-item:opacity-100 group-hover/project-chat-item:pointer-events-auto focus-visible:opacity-100 focus-visible:pointer-events-auto"
+        : "sidebar-row-action group-hover/recent-item:opacity-100 group-hover/recent-item:pointer-events-auto focus-visible:opacity-100 focus-visible:pointer-events-auto";
+    const buttonClass = cn(
+      "sidebar-nav-btn h-[33px] cursor-pointer rounded-[10px] pr-4 text-[14.5px] leading-[19px] tracking-nav font-medium",
+      variant === "project" ? "pl-[37px]" : "pl-2.5",
+      variant === "project"
+        ? "group-hover/project-chat-item:pr-8 group-has-[.sidebar-row-action[data-state=open]]/project-chat-item:pr-8"
+        : "group-hover/recent-item:pr-8 group-has-[.sidebar-row-action[data-state=open]]/recent-item:pr-8",
+    );
+
+    return (
+      <SidebarMenuItem key={item.id} className={itemClass}>
+        <SidebarMenuButton
+          data-testid="recent-thread"
+          data-thread-type={item.type}
+          data-thread-id={item.id}
+          isActive={activeThreadId === item.id}
+          className={buttonClass}
+          onClick={() => {
+            navigate({
+              to: "/chat",
+              search:
+                item.type === "single"
+                  ? {
+                      thread: item.id,
+                      ...(item.projectId ? { project: item.projectId } : {}),
+                    }
+                  : {
+                      compare: item.id,
+                      ...(item.projectId ? { project: item.projectId } : {}),
+                    },
+            });
+            closeMobileIfOpen();
+          }}
+        >
+          <span className="truncate">{item.title}</span>
+        </SidebarMenuButton>
+        <DropdownMenu>
+          <DropdownMenuTrigger asChild>
+            <button
+              type="button"
+              onClick={(e) => e.stopPropagation()}
+              aria-label="Chat options"
+              className={actionClass}
+            >
+              <span className="sidebar-row-action-glyph">
+                <HugeiconsIcon icon={MoreVerticalIcon} strokeWidth={1.75} className="size-icon" />
+              </span>
+            </button>
+          </DropdownMenuTrigger>
+          <DropdownMenuContent
+            side="bottom"
+            align="end"
+            sideOffset={4}
+            className="app-user-menu menu-soft-surface menu-flat-destructive ring-0 w-44 py-2 font-heading rounded-[14px] border-0"
+          >
+            <DropdownMenuItem onSelect={() => openRenameChat(item)}>
+              <HugeiconsIcon icon={Edit03Icon} strokeWidth={1.75} className="size-icon" />
+              <span>Rename</span>
+            </DropdownMenuItem>
+            <DropdownMenuSub>
+              <DropdownMenuSubTrigger>
+                <HugeiconsIcon icon={Folder01Icon} strokeWidth={1.75} className="size-icon" />
+                <span>Move to project</span>
+              </DropdownMenuSubTrigger>
+              <DropdownMenuSubContent
+                sideOffset={8}
+                alignOffset={-4}
+                className="app-user-menu menu-soft-surface menu-flat-destructive ring-0 w-56 py-2 font-heading rounded-[14px] border-0"
+              >
+                <DropdownMenuItem
+                  onSelect={() => {
+                    setProjectCreateMoveTarget(item);
+                    setProjectNameDraft("");
+                    setCreatingProject(true);
+                  }}
+                >
+                  <HugeiconsIcon icon={FolderAddIcon} strokeWidth={1.75} className="size-icon" />
+                  <span>New project</span>
+                </DropdownMenuItem>
+                <DropdownMenuItem
+                  disabled={!item.projectId}
+                  onSelect={() => void moveChatToProject(item, null)}
+                >
+                  <span>Recents</span>
+                </DropdownMenuItem>
+                {projects.map((project) => (
+                  <DropdownMenuItem
+                    key={project.id}
+                    disabled={item.projectId === project.id}
+                    onSelect={() => void moveChatToProject(item, project.id)}
+                  >
+                    <HugeiconsIcon icon={Folder01Icon} strokeWidth={1.75} className="size-icon" />
+                    <span className="truncate">{project.name}</span>
+                  </DropdownMenuItem>
+                ))}
+              </DropdownMenuSubContent>
+            </DropdownMenuSub>
+            <DropdownMenuItem
+              variant="destructive"
+              onSelect={() => setConfirmingDelete({ kind: "chat", item })}
+            >
+              <HugeiconsIcon icon={Delete02Icon} strokeWidth={1.75} className="size-icon" />
+              <span>Delete</span>
+            </DropdownMenuItem>
+          </DropdownMenuContent>
+        </DropdownMenu>
+      </SidebarMenuItem>
+    );
+  }
+
   return (
     <>
     <Sidebar
@@ -369,7 +665,7 @@ export function AppSidebar() {
       variant="sidebar"
       className="font-heading group-data-[collapsible=icon]:[&_[data-sidebar=sidebar]]:bg-white dark:group-data-[collapsible=icon]:[&_[data-sidebar=sidebar]]:bg-background"
     >
-      <SidebarHeader className="pl-[17px] pr-3 pt-[12px] pb-[8px] group-data-[collapsible=icon]:px-0">
+      <SidebarHeader className="pl-[17px] pr-3 pt-[14px] pb-[8px] group-data-[collapsible=icon]:px-0">
         {/* Expanded: compact logo + close toggle */}
         <div className="flex items-center justify-between gap-[8.5px] group-data-[collapsible=icon]:hidden">
           <Link
@@ -377,12 +673,7 @@ export function AppSidebar() {
             onClick={(event) => {
               event.preventDefault();
               if (chatDisabled) return;
-              setActiveThreadId(null);
-              closeMobileIfOpen();
-              void navigate({
-                to: "/chat",
-                search: { new: createNavigationNonce() },
-              });
+              openNewChat(null);
             }}
             className="flex items-center gap-[6px] select-none"
             aria-label={t("shell.aria.home")}
@@ -392,7 +683,7 @@ export function AppSidebar() {
               alt="Unsloth"
               className="h-[34px] w-[34px] rounded-full object-cover"
             />
-            <span className="font-heading text-[21px] font-semibold tracking-[-0.01em] dark:tracking-[0.02em] leading-none text-black dark:text-white">
+            <span className="font-heading text-[21px] font-semibold tracking-[0em] dark:tracking-[0.02em] leading-none text-black dark:text-white">
               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-[8px] 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)]">
@@ -405,7 +696,7 @@ export function AppSidebar() {
                 <button
                   type="button"
                   onClick={togglePinned}
-                  className="inline-flex h-[35px] w-[32px] 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-2 focus-visible:ring-ring"
+                  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-2 focus-visible:ring-ring"
                   aria-label={t("shell.aria.closeSidebar")}
                 >
                   <HugeiconsIcon icon={LayoutAlignLeftIcon} strokeWidth={1.75} className="size-icon" />
@@ -424,13 +715,13 @@ export function AppSidebar() {
 
         {/* Collapsed: panel icon doubles as expand trigger */}
         {!isMobile && (
-          <div className="hidden group-data-[collapsible=icon]:flex h-[35px] items-center justify-center w-full">
+          <div className="hidden group-data-[collapsible=icon]:flex h-[33px] items-center justify-center w-full">
             <Tooltip>
               <TooltipPrimitive.Trigger asChild>
                 <button
                   type="button"
                   onClick={togglePinned}
-                  className="inline-flex h-[35px] w-[32px] items-center justify-center rounded-[10px] text-nav-fg transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+                  className="inline-flex h-[33px] w-[32px] cursor-pointer items-center justify-center rounded-[10px] text-nav-fg transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
                   aria-label={t("shell.aria.openSidebar")}
                 >
                   <HugeiconsIcon icon={LayoutAlignLeftIcon} strokeWidth={1.75} className="size-icon" />
@@ -448,31 +739,37 @@ export function AppSidebar() {
         )}
       </SidebarHeader>
 
-      <SidebarGroup className="group-data-[collapsible=icon]:px-0 px-2 pt-[9px] pb-[8px] shrink-0">
+      <SidebarGroup className="group-data-[collapsible=icon]:px-0 px-2 pt-[9px] pb-px shrink-0">
         <SidebarGroupContent>
           <SidebarMenu>
             <NavItem
               icon={PencilEdit02Icon}
               label={t("shell.navigation.newChat")}
-              active={false}
+              active={
+                isChatRoute &&
+                !search.thread &&
+                !search.compare &&
+                !search.project
+              }
               disabled={chatDisabled}
-              onClick={() => {
-                if (chatDisabled) return;
-                setActiveThreadId(null);
-                navigate({ to: "/chat", search: { new: createNavigationNonce() } });
-                closeMobileIfOpen();
-              }}
+              onClick={() => openNewChat(null)}
             />
             <NavItem
               icon={ColumnInsertIcon}
               label={t("shell.navigation.compare")}
-              active={!!search.compare && !chatItems.some((i) => i.id === search.compare)}
+              active={
+                !!search.compare &&
+                !chatItems.some((i) => i.id === search.compare)
+              }
               disabled={chatDisabled}
               dataTour="chat-compare"
               onClick={() => {
                 if (chatDisabled) return;
                 setActiveThreadId(null);
-                navigate({ to: "/chat", search: { compare: createNavigationNonce() } });
+                navigate({
+                  to: "/chat",
+                  search: { compare: createNavigationNonce() },
+                });
                 closeMobileIfOpen();
               }}
             />
@@ -480,9 +777,10 @@ export function AppSidebar() {
               icon={Search01Icon}
               label={t("shell.navigation.search")}
               active={false}
-              disabled={chatDisabled}
               onClick={() => {
-                if (chatDisabled) return;
+                // Search is read-only over chat history and never runs
+                // inference, so it stays available while training (unlike
+                // New chat, which is gated on `chatDisabled`).
                 useChatSearchStore.getState().open();
                 closeMobileIfOpen();
               }}
@@ -491,140 +789,127 @@ export function AppSidebar() {
         </SidebarGroupContent>
       </SidebarGroup>
 
-      <SidebarGroup data-tour="navbar" className="group-data-[collapsible=icon]:px-0 px-2 pt-[9px] pb-[20px] shrink-0">
-        <SidebarGroupContent>
-          <SidebarMenu>
-            <NavItem
-              icon={TestTubeOutlineIcon}
-              label={t("shell.navigation.train")}
-              active={pathname === "/studio" || pathname.startsWith("/studio/")}
-              disabled={chatOnly}
-              onClick={() => {
-                if (chatOnly) return;
-                navigate({ to: "/studio" });
-                closeMobileIfOpen();
-              }}
-            />
+      <SidebarContent
+        ref={scrollRef}
+        onScroll={(e) => syncScrollState(e.currentTarget)}
+        className={cn(
+          // pb-2 keeps the last row's rounded highlight clear of the
+          // overflow clip edge so its bottom corners aren't shaved off.
+          "sidebar-scroll-fade gap-0 overflow-y-auto overscroll-contain min-h-0 pb-2",
+          scrolled && "is-scrolled",
+        )}
+      >
+        <SidebarGroup className="group-data-[collapsible=icon]:px-0 px-2 py-0 shrink-0">
+          <SidebarGroupContent>
+            <SidebarMenu>
+              <NavItem
+                icon={Folder01Icon}
+                label="Projects"
+                active={
+                  pathname === "/projects" || pathname.startsWith("/projects/")
+                }
+                onClick={() => {
+                  navigate({ to: "/projects" });
+                  closeMobileIfOpen();
+                }}
+              />
+            </SidebarMenu>
+          </SidebarGroupContent>
+        </SidebarGroup>
 
-            <NavItem
-              icon={ChefHatIcon}
-              label={t("shell.navigation.recipes")}
-              active={isRecipesRoute}
-              onClick={() => {
-                navigate({ to: "/data-recipes" });
-                closeMobileIfOpen();
-              }}
-            />
-
-            <NavItem
-              icon={DownloadSquare01Icon}
-              label={t("shell.navigation.export")}
-              active={pathname === "/export" || pathname.startsWith("/export/")}
-              disabled={chatOnly}
-              onClick={() => {
-                if (chatOnly) return;
-                navigate({ to: "/export" });
-                closeMobileIfOpen();
-              }}
-            />
-          </SidebarMenu>
-        </SidebarGroupContent>
-      </SidebarGroup>
-
-      <SidebarContent ref={scrollRef} className="gap-0 overflow-y-auto overscroll-contain min-h-0">
-        {!isStudioRoute && chatItems.length > 0 && (
-          <Collapsible
-            key={isChatRoute ? "chat-route" : "non-chat-route"}
-            defaultOpen
-            asChild
-          >
-          <SidebarGroup className="group-data-[collapsible=icon]:hidden px-0 py-0">
-            <SidebarGroupLabel className={cn("sidebar-sticky-label", scrolled && "is-scrolled")} asChild>
-              <CollapsibleTrigger className="cursor-pointer flex w-full items-center justify-between">
-                {t("shell.navigation.recents")}
-                <ChevronDown className="size-3.5 transition-transform duration-200 data-[state=open]:rotate-0 [[data-state=closed]_&]:rotate-[-90deg]" />
+        <Collapsible open={trainOpen} onOpenChange={setTrainOpen} asChild>
+          <SidebarGroup data-tour="navbar" className="group-data-[collapsible=icon]:hidden px-0 py-0">
+            <SidebarGroupLabel className={cn("sidebar-sticky-label sidebar-sticky-label-following", scrolled && "is-scrolled")} asChild>
+              <CollapsibleTrigger className="cursor-pointer flex w-full items-center gap-1 group/sb-collap">
+                {t("shell.navigation.train")}
+                <ChevronDown className="size-3.5 opacity-0 transition-[transform,opacity] duration-200 group-hover/sb-collap:opacity-100 group-focus-visible/sb-collap:opacity-100 data-[state=open]:rotate-0 [[data-state=closed]_&]:rotate-[-90deg]" />
               </CollapsibleTrigger>
             </SidebarGroupLabel>
             <CollapsibleContent>
-            <SidebarGroupContent className="px-2">
-              <SidebarMenu>
-                {chatItems.map((item) => (
-                  <SidebarMenuItem key={item.id} className="group/recent-item relative">
-                    <SidebarMenuButton
-                      data-testid="recent-thread"
-                      data-thread-type={item.type}
-                      data-thread-id={item.id}
-                      isActive={activeThreadId === item.id}
-                      className="sidebar-nav-btn h-[32px] rounded-[10px] pl-2.5 pr-2.5 group-hover/recent-item:pr-10 group-has-[.sidebar-row-action[data-state=open]]/recent-item:pr-10 text-[14.5px] leading-[19px] tracking-nav font-medium"
-                      onClick={() => {
-                        navigate({
-                          to: "/chat",
-                          search:
-                            item.type === "single"
-                              ? { thread: item.id }
-                              : { compare: item.id },
-                        });
-                        closeMobileIfOpen();
-                      }}
-                    >
-                      <span className="truncate">{item.title}</span>
-                    </SidebarMenuButton>
-                    <DropdownMenu>
-                      <DropdownMenuTrigger asChild>
-                        <button
-                          type="button"
-                          onClick={(e) => e.stopPropagation()}
-                          aria-label={t("shell.aria.chatOptions")}
-                          className="sidebar-row-action group-hover/recent-item:opacity-100 group-hover/recent-item:pointer-events-auto focus-visible:opacity-100 focus-visible:pointer-events-auto"
-                        >
-                          <span className="sidebar-row-action-glyph">
-                            <MoreHorizontalIcon strokeWidth={1.75} className="size-icon" />
-                          </span>
-                        </button>
-                      </DropdownMenuTrigger>
-                      <DropdownMenuContent
-                        side="bottom"
-                        align="end"
-                        sideOffset={4}
-                        className="app-user-menu menu-soft-surface menu-flat-destructive ring-0 w-44 py-2 font-heading rounded-[14px] border-0"
-                      >
-                        <DropdownMenuItem onSelect={() => openRenameChat(item)}>
-                          <HugeiconsIcon icon={Edit03Icon} strokeWidth={1.75} className="size-icon" />
-                          <span>{t("common.rename")}</span>
-                        </DropdownMenuItem>
-                        <DropdownMenuItem
-                          variant="destructive"
-                          onSelect={() => setConfirmingDelete({ kind: "chat", item })}
-                        >
-                          <HugeiconsIcon icon={Delete02Icon} strokeWidth={1.75} className="size-icon" />
-                          <span>{t("common.delete")}</span>
-                        </DropdownMenuItem>
-                      </DropdownMenuContent>
-                    </DropdownMenu>
-                  </SidebarMenuItem>
-                ))}
-              </SidebarMenu>
-            </SidebarGroupContent>
+              <SidebarGroupContent className="px-2">
+                <SidebarMenu>
+                  <NavItem
+                    icon={TestTubeOutlineIcon}
+                    label={t("shell.navigation.train")}
+                    active={pathname === "/studio" || pathname.startsWith("/studio/")}
+                    disabled={chatOnly}
+                    onClick={() => {
+                      if (chatOnly) return;
+                      navigate({ to: "/studio" });
+                      closeMobileIfOpen();
+                    }}
+                  />
+                  <NavItem
+                    icon={ChefHatIcon}
+                    label={t("shell.navigation.recipes")}
+                    active={isRecipesRoute}
+                    onClick={() => {
+                      navigate({ to: "/data-recipes" });
+                      closeMobileIfOpen();
+                    }}
+                  />
+                  <NavItem
+                    icon={DownloadSquare01Icon}
+                    label={t("shell.navigation.export")}
+                    active={pathname === "/export" || pathname.startsWith("/export/")}
+                    disabled={chatOnly}
+                    onClick={() => {
+                      if (chatOnly) return;
+                      navigate({ to: "/export" });
+                      closeMobileIfOpen();
+                    }}
+                  />
+                </SidebarMenu>
+              </SidebarGroupContent>
             </CollapsibleContent>
           </SidebarGroup>
+        </Collapsible>
+
+        {!isStudioRoute && (
+          <Collapsible open={chatOpen} onOpenChange={setChatOpen} asChild>
+            <SidebarGroup className="group-data-[collapsible=icon]:hidden px-0 py-0">
+              <SidebarGroupLabel className={cn("sidebar-sticky-label sidebar-sticky-label-following", scrolled && "is-scrolled")} asChild>
+                <CollapsibleTrigger className="cursor-pointer flex w-full items-center gap-1 group/sb-collap">
+                  {t("shell.navigation.recents")}
+                  <ChevronDown className="size-3.5 opacity-0 transition-[transform,opacity] duration-200 group-hover/sb-collap:opacity-100 group-focus-visible/sb-collap:opacity-100 data-[state=open]:rotate-0 [[data-state=closed]_&]:rotate-[-90deg]" />
+                </CollapsibleTrigger>
+              </SidebarGroupLabel>
+              <CollapsibleContent>
+                <SidebarGroupContent className="px-2">
+                  <SidebarMenu>
+                    {recentChatItems.map((item) =>
+                      renderChatSidebarItem(item, "recent"),
+                    )}
+                  </SidebarMenu>
+                </SidebarGroupContent>
+              </CollapsibleContent>
+            </SidebarGroup>
           </Collapsible>
         )}
 
         {isStudioRoute && runItems.length > 0 && !chatOnly && (
-          <Collapsible key="studio-runs-route" defaultOpen asChild>
+          <Collapsible open={runsOpen} onOpenChange={setRunsOpen} asChild>
           <SidebarGroup className="group-data-[collapsible=icon]:hidden px-0 py-0">
-            <SidebarGroupLabel className={cn("sidebar-sticky-label", scrolled && "is-scrolled")} asChild>
-              <CollapsibleTrigger className="cursor-pointer flex w-full items-center justify-between">
+            <SidebarGroupLabel className={cn("sidebar-sticky-label sidebar-sticky-label-following", scrolled && "is-scrolled")} asChild>
+              <CollapsibleTrigger className="cursor-pointer flex w-full items-center gap-1 group/sb-collap">
                 {t("shell.navigation.recents")}
-                <ChevronDown className="size-3.5 transition-transform duration-200 data-[state=open]:rotate-0 [[data-state=closed]_&]:rotate-[-90deg]" />
+                <ChevronDown className="size-3.5 opacity-0 transition-[transform,opacity] duration-200 group-hover/sb-collap:opacity-100 group-focus-visible/sb-collap:opacity-100 data-[state=open]:rotate-0 [[data-state=closed]_&]:rotate-[-90deg]" />
               </CollapsibleTrigger>
             </SidebarGroupLabel>
             <CollapsibleContent>
               <SidebarGroupContent className="px-2">
                 <SidebarMenu>
                   {runItems.map((run) => {
+                    // An explicit sidebar selection wins. Otherwise highlight
+                    // the active job only while the "Current Run" tab is the
+                    // view - that covers a live run (it auto-switches there) and
+                    // a just-finished/errored run you're still viewing, while
+                    // keeping the Configure tab unhighlighted even though
+                    // `activeJobId` stays pinned to the last job.
                     const isActiveRun =
-                      selectedHistoryRunId === run.id || activeJobId === run.id;
+                      selectedHistoryRunId != null
+                        ? run.id === selectedHistoryRunId
+                        : currentRunViewActive && run.id === activeJobId;
                     return (
                       <SidebarMenuItem
                         key={run.id}
@@ -703,7 +988,18 @@ export function AppSidebar() {
         )}
       </SidebarContent>
 
-      <SidebarFooter className="border-t border-sidebar-border group-data-[collapsible=icon]:border-transparent group-data-[collapsible=icon]:px-0">
+      <SidebarFooter className="relative group-data-[collapsible=icon]:px-0">
+        {/* Fade above the profile box, shown only while there's more list below
+            the fold; at the very bottom (or for short lists) it fades out so the
+            last row shows fully (Gemini-style). `right-2` keeps it clear of the
+            8px scrollbar gutter so the scrollbar isn't faded out. */}
+        <div
+          aria-hidden="true"
+          className={cn(
+            "pointer-events-none absolute left-0 right-2 bottom-full h-10 bg-gradient-to-t from-[var(--sidebar)] to-transparent transition-opacity duration-200",
+            canScrollDown ? "opacity-100" : "opacity-0",
+          )}
+        />
         <SidebarMenu>
           <SidebarMenuItem>
             <DropdownMenu>
@@ -711,14 +1007,14 @@ export function AppSidebar() {
                 <SidebarMenuButton
                   size="lg"
                   aria-label={t("shell.accountMenu", { name: displayTitle })}
-                  className="sidebar-nav-btn !h-[50px] gap-[8px] px-2 py-[9px] rounded-[10px]"
+                  className="sidebar-nav-btn !h-[40px] gap-[8px] px-2 py-[5px] rounded-[10px]"
                 >
                   <div className="shrink-0">
                     <UserAvatar
                       name={displayTitle}
                       imageUrl={avatarDataUrl}
                       size="sm"
-                      className="!size-8"
+                      className="!size-[30px]"
                     />
                   </div>
                   <div className="flex flex-col gap-0.5 leading-tight group-data-[collapsible=icon]:hidden">
@@ -818,7 +1114,10 @@ export function AppSidebar() {
     <Dialog
       open={confirmingDelete !== null}
       onOpenChange={(open) => {
-        if (!open) setConfirmingDelete(null);
+        if (!open) {
+          setConfirmingDelete(null);
+          setDeleteProjectFiles(false);
+        }
       }}
     >
       <DialogContent className="menu-flat-destructive corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-md">
@@ -826,7 +1125,9 @@ export function AppSidebar() {
           <DialogTitle>
             {confirmingDelete?.kind === "run"
               ? t("shell.dialog.deleteRun.title")
-              : t("shell.dialog.deleteChat.title")}
+              : confirmingDelete?.kind === "project"
+                ? "Delete project"
+                : t("shell.dialog.deleteChat.title")}
           </DialogTitle>
           <DialogDescription>
             {confirmingDelete?.kind === "run" ? (
@@ -842,9 +1143,37 @@ export function AppSidebar() {
                 "shell.dialog.deleteChat.description",
                 confirmingDelete.item.title,
               )
+            ) : confirmingDelete?.kind === "project" ? (
+              <>
+                Delete{" "}
+                <span className="font-medium text-foreground">
+                  &quot;{confirmingDelete.project.name}&quot;
+                </span>
+                ? Its chats will be permanently deleted.
+              </>
             ) : null}
           </DialogDescription>
         </DialogHeader>
+        {confirmingDelete?.kind === "project" ? (
+          <div className="flex items-start justify-between gap-4 rounded-md border border-border/60 bg-muted/35 px-3 py-2.5">
+            <label htmlFor="delete-project-files" className="min-w-0 space-y-1">
+              <span className="block text-sm font-medium text-foreground">
+                Delete files and sandbox folder
+              </span>
+              <span className="block break-words text-xs leading-5 text-muted-foreground">
+                {confirmingDelete.project.rootPath
+                  ? confirmingDelete.project.rootPath
+                  : "The project workspace folder will be removed from disk."}
+              </span>
+            </label>
+            <Switch
+              id="delete-project-files"
+              checked={deleteProjectFiles}
+              onCheckedChange={setDeleteProjectFiles}
+              aria-label="Delete project files and sandbox folder"
+            />
+          </div>
+        ) : null}
         <DialogFooter className="flex-wrap gap-2 sm:justify-end">
           <Button
             type="button"
@@ -858,7 +1187,9 @@ export function AppSidebar() {
             variant="destructive"
             onClick={() => void commitDelete()}
           >
-            {t("common.delete")}
+            {confirmingDelete?.kind === "project" && deleteProjectFiles
+              ? "Delete all"
+              : t("common.delete")}
           </Button>
         </DialogFooter>
       </DialogContent>
@@ -874,7 +1205,9 @@ export function AppSidebar() {
           <DialogTitle>
             {renamingTarget?.kind === "run"
               ? t("shell.dialog.renameRun.title")
-              : t("shell.dialog.renameChat.title")}
+              : renamingTarget?.kind === "project"
+                ? "Rename project"
+                : t("shell.dialog.renameChat.title")}
           </DialogTitle>
         </DialogHeader>
         <Input
@@ -891,12 +1224,16 @@ export function AppSidebar() {
           placeholder={
             renamingTarget?.kind === "run"
               ? t("shell.dialog.renameRun.placeholder")
-              : t("shell.dialog.renameChat.placeholder")
+              : renamingTarget?.kind === "project"
+                ? "Project name"
+                : t("shell.dialog.renameChat.placeholder")
           }
           aria-label={
             renamingTarget?.kind === "run"
               ? t("shell.dialog.renameRun.placeholder")
-              : t("shell.dialog.renameChat.placeholder")
+              : renamingTarget?.kind === "project"
+                ? "Project name"
+                : t("shell.dialog.renameChat.placeholder")
           }
           className="focus-visible:border-input focus-visible:ring-0"
         />
@@ -918,6 +1255,58 @@ export function AppSidebar() {
         </DialogFooter>
       </DialogContent>
     </Dialog>
+    <Dialog
+      open={creatingProject}
+      onOpenChange={(open) => {
+        setCreatingProject(open);
+        if (!open) {
+          setProjectNameDraft("");
+          setProjectCreateMoveTarget(null);
+        }
+      }}
+    >
+      <DialogContent className="corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-md">
+        <DialogHeader>
+          <DialogTitle>
+            {projectCreateMoveTarget ? "Move to new project" : "New project"}
+          </DialogTitle>
+        </DialogHeader>
+        <Input
+          value={projectNameDraft}
+          onChange={(event) => setProjectNameDraft(event.target.value)}
+          onKeyDown={(event) => {
+            if (event.key === "Enter") {
+              event.preventDefault();
+              void commitCreateProject();
+            }
+          }}
+          autoFocus
+          maxLength={120}
+          placeholder="Project name"
+          aria-label="Project name"
+          className="focus-visible:border-input focus-visible:ring-0"
+        />
+        <DialogFooter className="flex-wrap gap-2 sm:justify-end">
+          <Button
+            type="button"
+            variant="ghost"
+            onClick={() => {
+              setCreatingProject(false);
+              setProjectCreateMoveTarget(null);
+            }}
+          >
+            Cancel
+          </Button>
+          <Button
+            type="button"
+            onClick={() => void commitCreateProject()}
+            disabled={!projectNameDraft.trim()}
+          >
+            Create
+          </Button>
+        </DialogFooter>
+      </DialogContent>
+    </Dialog>
     </>
   );
 }

studio/frontend/src/components/assistant-ui/markdown-text.tsx

@@ -3,18 +3,19 @@
 
 "use client";
 
+import { ArtifactCard, useChatRuntimeStore } from "@/features/chat";
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { preprocessLaTeX } from "@/lib/latex";
 import { openLink } from "@/lib/open-link";
-import { INTERNAL, useMessagePartText } from "@assistant-ui/react";
+import { INTERNAL, useAuiState, useMessagePartText } from "@assistant-ui/react";
 import { Copy01Icon, Tick02Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { createCodePlugin } from "./code-plugin";
 import { createMathPlugin } from "@streamdown/math";
 import { mermaid } from "@streamdown/mermaid";
-import { DownloadIcon, Maximize2Icon, Minimize2Icon } from "lucide-react";
+import { DownloadIcon } from "lucide-react";
 import { useEffect, useMemo, useRef, useState } from "react";
 import { Block, type BlockProps, Streamdown } from "streamdown";
+import { createCodePlugin } from "./code-plugin";
 import "katex/dist/katex.min.css";
 import { AudioPlayer } from "./audio-player";
 import { unslothDarkTheme, unslothLightTheme } from "./code-themes";
@@ -26,11 +27,7 @@ const code = createCodePlugin({
 const { withSmoothContextProvider } = INTERNAL;
 
 const STREAMDOWN_COMPONENTS = {
-  a: ({
-    href,
-    children,
-    ...props
-  }: React.ComponentProps<"a">) => (
+  a: ({ href, children, ...props }: React.ComponentProps<"a">) => (
     <a
       href={href}
       rel="noopener noreferrer"
@@ -59,6 +56,34 @@ type CodeFence = {
   source: string;
 };
 
+type ToolCallPartLike = {
+  type?: string;
+  toolName?: string;
+  args?: unknown;
+  result?: unknown;
+};
+
+function isRenderableRenderHtmlToolPart(part: unknown): boolean {
+  const toolPart = part as ToolCallPartLike;
+  if (toolPart.type !== "tool-call" || toolPart.toolName !== "render_html") {
+    return false;
+  }
+  if (
+    typeof toolPart.result === "string" &&
+    toolPart.result.startsWith("Error:")
+  ) {
+    return false;
+  }
+  if (
+    typeof toolPart.result === "string" &&
+    toolPart.result.startsWith("Rendered HTML artifact")
+  ) {
+    return true;
+  }
+  const args = toolPart.args as { code?: unknown } | undefined;
+  return typeof args?.code === "string" && args.code.trim().length > 0;
+}
+
 function getMermaidSource(blockContent: string): string | null {
   const source = blockContent.match(MERMAID_SOURCE_RE)?.[1]?.trim();
   return source && source.length > 0 ? source : null;
@@ -123,7 +148,13 @@ function isHtmlFence(codeFence: CodeFence): boolean {
   return lang === "html" && !isSvgFence(codeFence);
 }
 
-const UNSAFE_SVG_RE = /<script[\s>]|on\w+\s*=|javascript:|<foreignObject[\s>]|<iframe[\s>]|<embed[\s>]|<object[\s>]/i;
+function isFullHtmlDocument(source: string): boolean {
+  const trimmed = source.trimStart();
+  return /^<!doctype\s+html\b/i.test(trimmed) || /^<html[\s>]/i.test(trimmed);
+}
+
+const UNSAFE_SVG_RE =
+  /<script[\s>]|on\w+\s*=|javascript:|<foreignObject[\s>]|<iframe[\s>]|<embed[\s>]|<object[\s>]/i;
 
 function sanitizeSvg(source: string): string | null {
   if (UNSAFE_SVG_RE.test(source)) return null;
@@ -145,96 +176,6 @@ function SvgPreview({ source }: { source: string }) {
   );
 }
 
-const HTML_PREVIEW_DEFAULT_HEIGHT = 400;
-const HTML_PREVIEW_MAX_HEIGHT = 800;
-
-function HtmlPreview({ source }: { source: string }) {
-  const iframeRef = useRef<HTMLIFrameElement>(null);
-  const [height, setHeight] = useState(HTML_PREVIEW_DEFAULT_HEIGHT);
-  const [enlarged, setEnlarged] = useState(false);
-
-  useEffect(() => {
-    const handler = (e: MessageEvent) => {
-      if (e.source !== iframeRef.current?.contentWindow) return;
-      if (typeof e.data?.htmlPreviewHeight === "number") {
-        setHeight(Math.min(Math.max(e.data.htmlPreviewHeight, 100), HTML_PREVIEW_MAX_HEIGHT));
-      }
-    };
-    window.addEventListener("message", handler);
-    return () => window.removeEventListener("message", handler);
-  }, []);
-
-  useEffect(() => {
-    if (!enlarged) return;
-    const handler = (e: KeyboardEvent) => {
-      if (e.key === "Escape") setEnlarged(false);
-    };
-    window.addEventListener("keydown", handler);
-    return () => window.removeEventListener("keydown", handler);
-  }, [enlarged]);
-
-  const resizeScript = `<script>new ResizeObserver(()=>{
-parent.postMessage({htmlPreviewHeight:document.documentElement.scrollHeight},"*");
-}).observe(document.documentElement);</script>`;
-
-  const srcDoc = source + resizeScript;
-
-  if (enlarged) {
-    return (
-      <>
-        <div className="mt-2 overflow-hidden rounded-lg border border-border" style={{ height }}>
-          {/* Placeholder keeps layout stable while overlay is shown */}
-        </div>
-        <div
-          className="fixed inset-0 z-50 flex flex-col bg-background/80 backdrop-blur-sm"
-          onClick={(e) => { if (e.target === e.currentTarget) setEnlarged(false); }}
-        >
-          <div className="flex items-center justify-end gap-2 px-4 py-2">
-            <button
-              type="button"
-              className="flex items-center gap-1.5 rounded-md border border-border bg-background px-3 py-1.5 text-sm text-muted-foreground transition-colors hover:bg-muted hover:text-foreground"
-              onClick={() => setEnlarged(false)}
-              title="Exit fullscreen (Esc)"
-            >
-              <Minimize2Icon className="size-4" />
-              Exit fullscreen
-            </button>
-          </div>
-          <div className="mx-4 mb-4 flex-1 overflow-hidden rounded-lg border border-border bg-background">
-            <iframe
-              ref={iframeRef}
-              srcDoc={srcDoc}
-              sandbox="allow-scripts"
-              style={{ width: "100%", height: "100%", border: "none", display: "block" }}
-              title="HTML preview"
-            />
-          </div>
-        </div>
-      </>
-    );
-  }
-
-  return (
-    <div className="group/html-preview relative mt-2 overflow-hidden rounded-lg border border-border">
-      <button
-        type="button"
-        className="absolute top-2 right-2 z-10 rounded-md border border-border bg-background/80 p-1.5 text-muted-foreground opacity-0 transition-all hover:bg-muted hover:text-foreground group-hover/html-preview:opacity-100 supports-[backdrop-filter]:backdrop-blur"
-        onClick={() => setEnlarged(true)}
-        title="Enlarge preview"
-      >
-        <Maximize2Icon className="size-4" />
-      </button>
-      <iframe
-        ref={iframeRef}
-        srcDoc={srcDoc}
-        sandbox="allow-scripts"
-        style={{ width: "100%", height, border: "none", display: "block" }}
-        title="HTML preview"
-      />
-    </div>
-  );
-}
-
 function downloadTextFile(filename: string, text: string): void {
   const blob = new Blob([text], { type: "text/plain;charset=utf-8" });
   const url = URL.createObjectURL(blob);
@@ -346,6 +287,12 @@ function CodeBlockActions({
 }
 
 function StreamdownBlock(props: BlockProps) {
+  const shouldCollapseHtmlArtifacts = useChatRuntimeStore(
+    (state) => state.artifactsEnabled || state.collapseHtmlArtifacts,
+  );
+  const messageHasRenderableRenderHtmlTool = useAuiState(({ message }) =>
+    message.parts.some(isRenderableRenderHtmlToolPart),
+  );
   const hasMermaidFence = props.content.includes("```mermaid");
   const mermaidSource = getMermaidSource(props.content);
   const codeFence = getCodeFence(props.content);
@@ -362,7 +309,9 @@ function StreamdownBlock(props: BlockProps) {
     return (
       <div className="relative isolate">
         <div className="my-4 rounded-xl border border-border bg-muted/30 p-4">
-          <div className="mb-2 text-xs font-medium text-muted-foreground">svg</div>
+          <div className="mb-2 text-xs font-medium text-muted-foreground">
+            svg
+          </div>
           <pre className="overflow-x-auto text-xs text-muted-foreground whitespace-pre-wrap break-all">
             <code>{codeFence.source}</code>
           </pre>
@@ -371,10 +320,17 @@ function StreamdownBlock(props: BlockProps) {
     );
   }
 
-  if (props.isIncomplete && codeFence && isHtmlFence(codeFence)) {
+  if (
+    shouldCollapseHtmlArtifacts &&
+    !messageHasRenderableRenderHtmlTool &&
+    props.isIncomplete &&
+    codeFence &&
+    isHtmlFence(codeFence) &&
+    isFullHtmlDocument(codeFence.source)
+  ) {
     return (
       <div className="my-4 flex h-48 items-center justify-center rounded-xl border border-border bg-muted/30 text-sm text-muted-foreground animate-pulse">
-        Loading preview...
+        Loading artifact preview...
       </div>
     );
   }
@@ -389,8 +345,24 @@ function StreamdownBlock(props: BlockProps) {
   }
 
   if (codeFence) {
-    const svgSource = !props.isIncomplete && isSvgFence(codeFence) ? sanitizeSvg(codeFence.source) : null;
-    const htmlSource = !props.isIncomplete && isHtmlFence(codeFence) ? codeFence.source : null;
+    const svgSource =
+      !props.isIncomplete && isSvgFence(codeFence)
+        ? sanitizeSvg(codeFence.source)
+        : null;
+    const htmlSource =
+      shouldCollapseHtmlArtifacts &&
+      !messageHasRenderableRenderHtmlTool &&
+      !props.isIncomplete &&
+      isHtmlFence(codeFence) &&
+      isFullHtmlDocument(codeFence.source)
+        ? codeFence.source
+        : null;
+    if (htmlSource) {
+      return (
+        <ArtifactCard code={htmlSource} title="HTML preview" source="fence" />
+      );
+    }
+
     return (
       <>
         <div className="relative isolate">
@@ -402,7 +374,6 @@ function StreamdownBlock(props: BlockProps) {
           />
         </div>
         {svgSource && <SvgPreview source={svgSource} />}
-        {htmlSource && <HtmlPreview source={htmlSource} />}
       </>
     );
   }

studio/frontend/src/components/assistant-ui/model-selector.tsx

@@ -17,7 +17,7 @@ import {
   CloudIcon,
   DashboardSquare01Icon,
   FolderSearchIcon,
-  Logout01Icon,
+  RemoveCircleIcon,
   Search01Icon,
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
@@ -326,7 +326,7 @@ function ModelSelectorContent({
             className="flex w-full items-center justify-center gap-1.5 rounded-md px-2 py-1.5 text-xs text-destructive transition-colors hover:bg-destructive/10"
             title="Eject model"
           >
-            <HugeiconsIcon icon={Logout01Icon} className="size-3.5" />
+            <HugeiconsIcon icon={RemoveCircleIcon} className="size-3.5" />
             Eject loaded model
           </button>
         </div>

studio/frontend/src/components/assistant-ui/sources.tsx

@@ -406,35 +406,49 @@ const SourcesGroup: FC = () => {
 
   return (
     <div className="relative mt-2 mb-3">
-      {/* Hidden measurement container — renders all badges to measure row positions */}
+      {/* Hidden measurement container. Renders all badges off-screen so we
+          can read each child's offsetTop and decide how many fit in two
+          rows. Wrapped in an absolute, h-0, overflow-hidden box so the
+          measurement pills do NOT contribute to the viewport's scrollable
+          overflow region. Without this clip, every hidden source row
+          adds ~30px to scrollHeight, producing a phantom empty scroll
+          area below the message: visible to users as unbounded blank
+          space below the assistant action bar. The inner div still
+          flex-wraps its children for measurement; offsetTop reads
+          correctly because the wrapper is positioned (absolute) and the
+          children's offsetTop is measured relative to it. */}
       <div
-        ref={containerRef}
-        aria-hidden={true}
-        className="flex w-full flex-wrap gap-1 invisible absolute pointer-events-none"
+        aria-hidden
+        className="absolute pointer-events-none overflow-hidden h-0 w-full left-0 top-0"
       >
-        {sources.map((source) => (
-          <span key={sourceKey(source)} className="inline-block">
-            {source.kind === "url" ? (
-              <Source href={source.url}>
-                <SourceIcon url={source.url} />
-                <SourceTitle>
-                  {source.title || extractDomain(source.url)}
-                </SourceTitle>
-              </Source>
-            ) : (
-              <Badge
-                variant="outline"
-                className="rounded-full inline-flex items-center gap-1.5"
-              >
-                <span className="font-mono text-[10px] font-semibold text-muted-foreground">
-                  [{source.chunkId}]
-                </span>
-                <FileTextIcon className="size-3 shrink-0 text-muted-foreground" />
-                <SourceTitle>{source.filename}</SourceTitle>
-              </Badge>
-            )}
-          </span>
-        ))}
+        <div
+          ref={containerRef}
+          className="flex w-full flex-wrap gap-1 invisible"
+        >
+          {sources.map((source) => (
+            <span key={sourceKey(source)} className="inline-block">
+              {source.kind === "url" ? (
+                <Source href={source.url}>
+                  <SourceIcon url={source.url} />
+                  <SourceTitle>
+                    {source.title || extractDomain(source.url)}
+                  </SourceTitle>
+                </Source>
+              ) : (
+                <Badge
+                  variant="outline"
+                  className="rounded-full inline-flex items-center gap-1.5"
+                >
+                  <span className="font-mono text-[10px] font-semibold text-muted-foreground">
+                    [{source.chunkId}]
+                  </span>
+                  <FileTextIcon className="size-3 shrink-0 text-muted-foreground" />
+                  <SourceTitle>{source.filename}</SourceTitle>
+                </Badge>
+              )}
+            </span>
+          ))}
+        </div>
       </div>
 
       {/* Visible container */}

studio/frontend/src/components/assistant-ui/thread.tsx

@@ -24,6 +24,7 @@ import { ToolFallback } from "@/components/assistant-ui/tool-fallback";
 import { ToolGroup } from "@/components/assistant-ui/tool-group";
 import { CodeExecutionToolUI } from "@/components/assistant-ui/tool-ui-code-execution";
 import { ImageGenerationToolUI } from "@/components/assistant-ui/tool-ui-image-generation";
+import { RenderHtmlToolUI } from "@/components/assistant-ui/tool-ui-render-html";
 import { PythonToolUI } from "@/components/assistant-ui/tool-ui-python";
 import { SearchKnowledgeBaseToolUI } from "@/components/assistant-ui/tool-ui-search-knowledge-base";
 import { TerminalToolUI } from "@/components/assistant-ui/tool-ui-terminal";
@@ -50,6 +51,7 @@ import {
 } from "@/components/ui/dropdown-menu";
 import { sentAudioNames } from "@/features/chat/api/chat-adapter";
 import { parseExternalModelId } from "@/features/chat/external-providers";
+import { McpComposerButton } from "@/features/chat/mcp-composer-button";
 import { getExternalReasoningCapabilities } from "@/features/chat/provider-capabilities";
 import { useChatRuntimeStore } from "@/features/chat/stores/chat-runtime-store";
 import { useExternalProvidersStore } from "@/features/chat/stores/external-providers-store";
@@ -82,6 +84,7 @@ import {
   ChevronRightIcon,
   DownloadIcon,
   FolderIcon,
+  FileTextIcon,
   GlobeIcon,
   HeadphonesIcon,
   ImageIcon,
@@ -382,7 +385,7 @@ const ThreadWelcome: FC<{
         <div className="aui-thread-welcome-message flex w-full flex-col justify-center gap-6 px-4">
           <div className="flex flex-col items-center gap-2 text-center">
             <img src={currentEmojiSrc} alt="Sloth mascot" className="size-20" />
-            <h1 className="aui-thread-welcome-message-inner fade-in slide-in-from-bottom-1 animate-in font-heading font-semibold text-2xl tracking-[-0.02em] duration-200">
+            <h1 className="aui-thread-welcome-message-inner fade-in slide-in-from-bottom-1 animate-in font-heading font-semibold text-2xl tracking-[0em] duration-200">
               Chat with your model
             </h1>
             <p className="aui-thread-welcome-message-inner fade-in slide-in-from-bottom-1 -mt-1 animate-in font-heading font-normal text-muted-foreground text-sm delay-75 duration-200">
@@ -396,14 +399,30 @@ const ThreadWelcome: FC<{
   );
 };
 
+export const ProjectComposer: FC<{
+  disabled?: boolean;
+  placeholder?: string;
+}> = ({ disabled, placeholder }) => {
+  return (
+    <GeneratedImageOverlayProvider>
+      <ComposerAnimated disabled={disabled} placeholder={placeholder} />
+    </GeneratedImageOverlayProvider>
+  );
+};
+
 const ComposerAnimated: FC<{
   disabled?: boolean;
+  placeholder?: string;
   threadId?: string | null;
-}> = ({ disabled, threadId }) => {
+}> = ({ disabled, placeholder, threadId }) => {
   return (
     <div className="relative mx-auto min-w-0 w-full max-w-(--thread-max-width)">
       <div className="relative z-10 w-full">
-        <Composer disabled={disabled} threadId={threadId} />
+        <Composer
+          disabled={disabled}
+          placeholder={placeholder}
+          threadId={threadId}
+        />
       </div>
     </div>
   );
@@ -435,8 +454,9 @@ const PendingAudioChip: FC = () => {
 
 const Composer: FC<{
   disabled?: boolean;
+  placeholder?: string;
   threadId?: string | null;
-}> = ({ disabled, threadId }) => {
+}> = ({ disabled, placeholder = "Send a message...", threadId }) => {
   const aui = useAui();
   const { overlay, closeOverlay } = useGeneratedImageOverlay();
   const setImageToolsEnabled = useChatRuntimeStore(
@@ -553,9 +573,7 @@ const Composer: FC<{
       <PendingDocChips docs={pendingDocs} onRemove={removeDoc} />
       <ToolStatusDisplay />
       <ComposerPrimitive.Input
-        placeholder={
-          overlay ? "Type your edits for your image" : "Send a message..."
-        }
+        placeholder={overlay ? "Type your edits for your image" : placeholder}
         className="aui-composer-input composer-input"
         minRows={1}
         maxRows={12}
@@ -996,8 +1014,8 @@ const PreserveThinkingToggle: FC = () => {
         disabled
           ? "cursor-not-allowed opacity-40"
           : preserveThinking
-            ? "text-primary hover:bg-primary/10 dark:hover:bg-white/[0.08]"
-            : "hover:bg-primary/10 dark:hover:bg-white/[0.08]",
+            ? "cursor-pointer text-primary hover:bg-primary/10 dark:hover:bg-white/[0.08]"
+            : "cursor-pointer hover:bg-primary/10 dark:hover:bg-white/[0.08]",
       )}
       aria-label={
         preserveThinking ? "Disable preserve think" : "Enable preserve think"
@@ -1185,6 +1203,29 @@ const RagToggle: FC = () => {
   );
 };
 
+const ArtifactsToggle: FC = () => {
+  const modelLoaded = useChatRuntimeStore(
+    (s) => !!s.params.checkpoint && !s.modelLoading,
+  );
+  const artifactsEnabled = useChatRuntimeStore((s) => s.artifactsEnabled);
+  const setArtifactsEnabled = useChatRuntimeStore((s) => s.setArtifactsEnabled);
+  const disabled = !modelLoaded;
+
+  return (
+    <button
+      type="button"
+      disabled={disabled}
+      onClick={() => setArtifactsEnabled(!artifactsEnabled)}
+      className="composer-pill-btn"
+      data-active={artifactsEnabled && !disabled ? "true" : "false"}
+      aria-label={artifactsEnabled ? "Disable artifacts" : "Enable artifacts"}
+    >
+      <FileTextIcon className="size-3.5" />
+      <span>Artifacts</span>
+    </button>
+  );
+};
+
 const ToolStatusDisplay: FC = () => {
   const toolStatus = useChatRuntimeStore((s) => s.toolStatus);
   const isThreadRunning = useAuiState(({ thread }) => thread.isRunning);
@@ -1331,6 +1372,8 @@ const ComposerAction: FC<{
         <CodeToolsToggle />
         <ImagesToggle />
         <RagToggle />
+        <ArtifactsToggle />
+        <McpComposerButton />
       </div>
       <div className="flex items-center gap-1">
         <ComposerPrimitive.If dictation={false}>
@@ -1459,6 +1502,7 @@ const AssistantMessage: FC = () => {
                 code_execution: CodeExecutionToolUI,
                 image_generation: ImageGenerationToolUI,
                 search_knowledge_base: SearchKnowledgeBaseToolUI,
+                render_html: RenderHtmlToolUI,
               },
               Fallback: ToolFallback,
             },

studio/frontend/src/components/assistant-ui/tool-group.tsx

@@ -8,6 +8,7 @@ import {
   type FC,
   type PropsWithChildren,
 } from "react";
+import { useAuiState } from "@assistant-ui/react";
 import { ChevronDownIcon, LoaderIcon } from "lucide-react";
 import { Wrench01Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
@@ -27,7 +28,8 @@ const toolGroupVariants = cva("aui-tool-group-root group/tool-group w-full", {
     variant: {
       outline: "corner-squircle rounded-lg border py-3",
       ghost: "rounded-lg bg-muted/10 py-2",
-      muted: "corner-squircle rounded-lg border border-muted-foreground/30 bg-muted/30 py-3",
+      muted:
+        "corner-squircle rounded-lg border border-muted-foreground/30 bg-muted/30 py-3",
     },
   },
   defaultVariants: { variant: "ghost" },
@@ -209,9 +211,17 @@ const ToolGroupImpl: FC<
   PropsWithChildren<{ startIndex: number; endIndex: number }>
 > = ({ children, startIndex, endIndex }) => {
   const toolCount = endIndex - startIndex + 1;
+  const containsArtifactTool = useAuiState(({ message }) =>
+    message.parts
+      .slice(startIndex, endIndex + 1)
+      .some(
+        (part) => part.type === "tool-call" && part.toolName === "render_html",
+      ),
+  );
 
-  // Single tool call — render directly without wrapper
-  if (toolCount <= 1) {
+  // Single tool calls and artifacts render directly so cards never hide inside
+  // a collapsed tool group.
+  if (toolCount <= 1 || containsArtifactTool) {
     return <>{children}</>;
   }
 

studio/frontend/src/components/assistant-ui/tool-ui-image-generation.tsx

@@ -10,7 +10,7 @@ import { DownloadIcon, ImageIcon, PencilIcon } from "lucide-react";
 import type { CSSProperties, MouseEvent } from "react";
 import { memo, useCallback, useEffect, useRef, useState } from "react";
 import { useGeneratedImageOverlay } from "./generated-image-overlay-context";
-import { Image, downloadImagePart } from "./image";
+import { downloadImagePart } from "./image";
 import {
   ToolFallbackContent,
   ToolFallbackRoot,
@@ -64,6 +64,8 @@ type GeneratedImagePart = {
 };
 
 const CAPTION_COLLAPSED_LINES = 4;
+const INLINE_IMAGE_MAX_WIDTH = 520;
+const INLINE_IMAGE_MAX_HEIGHT = 620;
 
 const extensionForMime = (mime: string): string => {
   switch (mime.toLowerCase()) {
@@ -100,6 +102,31 @@ const formatGeneratedImageLabel = (prompt: string): string => {
     : `Generated image: ${prompt}`;
 };
 
+const parseImageSize = (
+  size?: string,
+): { width: number; height: number } | null => {
+  const match = size?.match(/^(\d+)x(\d+)$/i);
+  if (!match) return null;
+  const width = Number(match[1]);
+  const height = Number(match[2]);
+  return width > 0 && height > 0 ? { width, height } : null;
+};
+
+const getInlineImageFrameWidth = ({
+  width,
+  height,
+}: {
+  width: number;
+  height: number;
+}): number =>
+  Math.round(
+    Math.min(
+      width,
+      INLINE_IMAGE_MAX_WIDTH,
+      (INLINE_IMAGE_MAX_HEIGHT * width) / height,
+    ),
+  );
+
 const loadingDots = Array.from({ length: 64 }, (_, index) => {
   const row = Math.floor(index / 8);
   const col = index % 8;
@@ -150,6 +177,16 @@ const ImageGenerationToolUIImpl: ToolCallMessagePartComponent = ({
     typeof result === "object" &&
     typeof (result as ImageGenerationResult).image_b64 === "string";
   const imageResult = isImageResult ? (result as ImageGenerationResult) : null;
+  const imageDimensions = parseImageSize(imageResult?.size);
+  const imageFrameStyle: CSSProperties = {
+    width: imageDimensions
+      ? getInlineImageFrameWidth(imageDimensions)
+      : INLINE_IMAGE_MAX_WIDTH,
+    maxWidth: "100%",
+  };
+  const imageBoxStyle: CSSProperties | undefined = imageDimensions
+    ? { aspectRatio: `${imageDimensions.width} / ${imageDimensions.height}` }
+    : undefined;
   const mime = imageResult?.image_mime || "image/png";
   const imageSrc = imageResult?.image_b64
     ? `data:${mime};base64,${imageResult.image_b64}`
@@ -202,8 +239,7 @@ const ImageGenerationToolUIImpl: ToolCallMessagePartComponent = ({
     const computedStyle = window.getComputedStyle(captionElement);
     const lineHeight = Number.parseFloat(computedStyle.lineHeight);
     const collapsedHeight =
-      (Number.isFinite(lineHeight) ? lineHeight : 20) *
-      CAPTION_COLLAPSED_LINES;
+      (Number.isFinite(lineHeight) ? lineHeight : 20) * CAPTION_COLLAPSED_LINES;
     const hasOverflow = captionElement.scrollHeight > collapsedHeight + 1;
     setPromptOverflow((current) =>
       current?.prompt === captionPrompt && current.canExpand === hasOverflow
@@ -289,26 +325,34 @@ const ImageGenerationToolUIImpl: ToolCallMessagePartComponent = ({
       />
       <ToolFallbackContent>
         {imagePart ? (
-          <figure className="m-0 flex flex-col gap-2">
-            <div className="group/generated-image relative aspect-square w-[480px] max-w-full overflow-hidden rounded-2xl bg-muted/25 shadow-lg shadow-foreground/5 dark:shadow-black/25">
-              <img
-                src={imagePart.image}
-                alt=""
-                aria-hidden={true}
-                className="pointer-events-none absolute inset-0 size-full scale-110 object-cover opacity-25 blur-2xl saturate-125"
-              />
-              <div className="pointer-events-none absolute inset-0 bg-background/45" />
+          <figure
+            className="m-0 flex max-w-full flex-col items-start gap-2 align-top"
+            style={imageFrameStyle}
+          >
+            <div
+              className="group/generated-image relative w-full overflow-hidden rounded-2xl align-top"
+              style={imageBoxStyle}
+            >
               <button
                 type="button"
-                className="relative z-10 block size-full cursor-zoom-in overflow-hidden rounded-2xl focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2 focus-visible:ring-offset-background"
+                className={cn(
+                  "block cursor-zoom-in rounded-2xl focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-ring",
+                  imageDimensions ? "size-full" : "max-w-full",
+                )}
                 onClick={showPreview}
                 aria-label="Open generated image preview"
               >
-                <Image.Preview
+                <img
                   src={imagePart.image}
                   alt={imageTitle}
-                  containerClassName="flex size-full min-h-0 items-center justify-center bg-transparent"
-                  className="size-full object-contain"
+                  width={imageDimensions?.width}
+                  height={imageDimensions?.height}
+                  className={cn(
+                    "block rounded-2xl object-contain",
+                    imageDimensions
+                      ? "size-full"
+                      : "h-auto max-h-[min(70vh,620px)] max-w-full",
+                  )}
                 />
               </button>
               <div className="pointer-events-none absolute inset-x-0 bottom-0 z-20 flex items-end justify-between gap-2 bg-gradient-to-t from-black/55 via-black/20 to-transparent p-3 opacity-100 transition-opacity sm:opacity-0 sm:group-hover/generated-image:opacity-100 sm:group-focus-within/generated-image:opacity-100">
@@ -335,7 +379,7 @@ const ImageGenerationToolUIImpl: ToolCallMessagePartComponent = ({
               </div>
             </div>
             {captionPrompt ? (
-              <figcaption className="max-w-[480px] text-xs leading-5 text-muted-foreground">
+              <figcaption className="w-full text-xs leading-5 text-muted-foreground">
                 <div
                   ref={captionRef}
                   className={cn(

studio/frontend/src/components/assistant-ui/tool-ui-render-html.tsx

@@ -0,0 +1,142 @@
+// 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 {
+  ArtifactCard,
+  useChatArtifactsStore,
+  useSelectedChatArtifact,
+} from "@/features/chat";
+import {
+  type ToolCallMessagePartComponent,
+  useAuiState,
+  useToolArgsStatus,
+} from "@assistant-ui/react";
+import { BrowserIcon } from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { memo, useEffect } from "react";
+
+// Context7 assistant-ui docs: tool UIs can read streaming args via
+// useToolArgsStatus, so render_html does not need to wait for tool completion.
+type RenderHtmlArgs = Record<string, unknown> & {
+  code?: string;
+  title?: string;
+};
+
+const RENDER_HTML_SESSION_STARTED_AT = Date.now();
+
+const RenderHtmlToolUIImpl: ToolCallMessagePartComponent = ({
+  args,
+  result,
+  status,
+  toolCallId,
+}) => {
+  const { propStatus } = useToolArgsStatus<RenderHtmlArgs>();
+  const parsedArgs = (args as RenderHtmlArgs) ?? {};
+  const code = typeof parsedArgs.code === "string" ? parsedArgs.code : "";
+  const hasCode = code.trim().length > 0;
+  const title =
+    typeof parsedArgs.title === "string" ? parsedArgs.title : "HTML artifact";
+  const isRunning = status?.type === "running";
+  const codeIsStreaming = propStatus.code === "streaming";
+
+  // Surface the backend error when the tool call completed with invalid
+  // args.  Backend success results start with "Rendered HTML artifact";
+  // error results start with "Error:".
+  const errorText =
+    status?.type === "complete" &&
+    typeof result === "string" &&
+    result.startsWith("Error:")
+      ? result
+      : null;
+  const messageId = useAuiState(({ message }) => message.id) ?? null;
+  const isMessageRunning = useAuiState(
+    ({ message }) => message.status?.type === "running",
+  );
+  const messageCreatedAtMs = useAuiState(({ message }) =>
+    message.createdAt instanceof Date ? message.createdAt.getTime() : null,
+  );
+  const isThreadRunning = useAuiState(({ thread }) => thread.isRunning);
+  const isLiveGeneratingArtifact =
+    isThreadRunning && isMessageRunning && (isRunning || codeIsStreaming);
+  const isStaleGeneratingArtifact =
+    !(isThreadRunning && isMessageRunning) && (isRunning || codeIsStreaming);
+  const messageCreatedThisSession =
+    messageCreatedAtMs != null &&
+    messageCreatedAtMs >= RENDER_HTML_SESSION_STARTED_AT - 1000;
+  const shouldAutoOpenArtifact =
+    (isLiveGeneratingArtifact && (hasCode || isRunning || codeIsStreaming)) ||
+    (hasCode && messageCreatedThisSession);
+  const selectedArtifact = useSelectedChatArtifact();
+  const closeArtifactSurface = useChatArtifactsStore(
+    (state) => state.closeArtifactSurface,
+  );
+
+  useEffect(() => {
+    if (!errorText) {
+      return;
+    }
+    if (!(messageId && toolCallId)) {
+      return;
+    }
+    if (selectedArtifact?.sourceToolCallId !== toolCallId) {
+      return;
+    }
+    if (selectedArtifact?.sourceMessageId !== messageId) {
+      return;
+    }
+    closeArtifactSurface();
+  }, [
+    closeArtifactSurface,
+    errorText,
+    messageId,
+    selectedArtifact,
+    toolCallId,
+  ]);
+
+  if (hasCode || (isLiveGeneratingArtifact && !errorText)) {
+    return (
+      <ArtifactCard
+        code={code}
+        title={title}
+        source="tool"
+        sourceToolCallId={toolCallId}
+        autoOpen={!errorText && shouldAutoOpenArtifact}
+        isStreaming={!errorText && isLiveGeneratingArtifact}
+      />
+    );
+  }
+
+  return (
+    <div className="relative my-2 flex min-h-[52px] w-full max-w-md items-center overflow-hidden rounded-lg border border-border/70 bg-muted/15 px-3 py-2 text-left dark:bg-muted/10">
+      <div className="relative z-10 flex min-w-0 flex-1 items-center gap-2.5">
+        <HugeiconsIcon
+          icon={BrowserIcon}
+          strokeWidth={1.75}
+          className="size-5 shrink-0 text-muted-foreground"
+        />
+        <span className="grid min-w-0 flex-1 gap-1">
+          <span className="truncate text-sm font-medium leading-tight text-foreground">
+            {errorText
+              ? "Artifact error"
+              : isStaleGeneratingArtifact
+                ? "Artifact interrupted"
+                : "Artifact unavailable"}
+          </span>
+          <span className="truncate text-[11px] leading-none text-muted-foreground">
+            {errorText ??
+              (isStaleGeneratingArtifact
+                ? "Refresh stopped this preview"
+                : "HTML artifact")}
+          </span>
+        </span>
+      </div>
+    </div>
+  );
+};
+
+export const RenderHtmlToolUI = memo(
+  RenderHtmlToolUIImpl,
+) as unknown as ToolCallMessagePartComponent;
+RenderHtmlToolUI.displayName = "RenderHtmlToolUI";

studio/frontend/src/components/ui/dropdown-menu.tsx

@@ -78,7 +78,7 @@ function DropdownMenuItem({
       data-inset={inset}
       data-variant={variant}
       className={cn(
-        "focus:bg-accent focus:text-accent-foreground data-[variant=destructive]:text-destructive data-[variant=destructive]:focus:bg-destructive/10 dark:data-[variant=destructive]:focus:bg-destructive/20 data-[variant=destructive]:focus:text-destructive data-[variant=destructive]:*:[svg]:text-destructive not-data-[variant=destructive]:focus:**:text-accent-foreground gap-2.5 rounded-lg px-3 py-2 text-sm [&_svg:not([class*='size-'])]:size-4 group/dropdown-menu-item relative flex cursor-default items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 data-[inset]:pl-8 [&_svg]:pointer-events-none [&_svg]:shrink-0",
+        "focus:bg-accent focus:text-accent-foreground data-[variant=destructive]:text-destructive data-[variant=destructive]:focus:bg-destructive/10 dark:data-[variant=destructive]:focus:bg-destructive/20 data-[variant=destructive]:focus:text-destructive data-[variant=destructive]:*:[svg]:text-destructive not-data-[variant=destructive]:focus:**:text-accent-foreground gap-2.5 rounded-lg px-3 py-2 text-sm [&_svg:not([class*='size-'])]:size-4 group/dropdown-menu-item relative flex cursor-pointer items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 data-[inset]:pl-8 [&_svg]:pointer-events-none [&_svg]:shrink-0",
         className,
       )}
       {...props}
@@ -96,7 +96,7 @@ function DropdownMenuCheckboxItem({
     <DropdownMenuPrimitive.CheckboxItem
       data-slot="dropdown-menu-checkbox-item"
       className={cn(
-        "focus:bg-accent focus:text-accent-foreground focus:**:text-accent-foreground gap-2.5 rounded-lg py-2 pr-8 pl-3 text-sm [&_svg:not([class*='size-'])]:size-4 relative flex cursor-default items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
+        "focus:bg-accent focus:text-accent-foreground focus:**:text-accent-foreground gap-2.5 rounded-lg py-2 pr-8 pl-3 text-sm [&_svg:not([class*='size-'])]:size-4 relative flex cursor-pointer items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
         className,
       )}
       checked={checked}
@@ -135,7 +135,7 @@ function DropdownMenuRadioItem({
     <DropdownMenuPrimitive.RadioItem
       data-slot="dropdown-menu-radio-item"
       className={cn(
-        "focus:bg-accent focus:text-accent-foreground focus:**:text-accent-foreground gap-2.5 rounded-lg py-2 pr-8 pl-3 text-sm [&_svg:not([class*='size-'])]:size-4 relative flex cursor-default items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
+        "focus:bg-accent focus:text-accent-foreground focus:**:text-accent-foreground gap-2.5 rounded-lg py-2 pr-8 pl-3 text-sm [&_svg:not([class*='size-'])]:size-4 relative flex cursor-pointer items-center outline-hidden select-none data-[disabled]:pointer-events-none data-[disabled]:opacity-50 [&_svg]:pointer-events-none [&_svg]:shrink-0",
         className,
       )}
       {...props}
@@ -221,7 +221,7 @@ function DropdownMenuSubTrigger({
       data-slot="dropdown-menu-sub-trigger"
       data-inset={inset}
       className={cn(
-        "focus:bg-accent focus:text-accent-foreground data-open:bg-accent data-open:text-accent-foreground not-data-[variant=destructive]:focus:**:text-accent-foreground gap-2 rounded-lg px-3 py-2 text-sm [&_svg:not([class*='size-'])]:size-4 flex cursor-default items-center outline-hidden select-none data-[inset]:pl-8 [&_svg]:pointer-events-none [&_svg]:shrink-0",
+        "focus:bg-accent focus:text-accent-foreground data-open:bg-accent data-open:text-accent-foreground not-data-[variant=destructive]:focus:**:text-accent-foreground gap-2 rounded-lg px-3 py-2 text-sm [&_svg:not([class*='size-'])]:size-4 flex cursor-pointer items-center outline-hidden select-none data-[inset]:pl-8 [&_svg]:pointer-events-none [&_svg]:shrink-0",
         className,
       )}
       {...props}

studio/frontend/src/components/ui/resizable.tsx

@@ -1,54 +1,54 @@
 // 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 type * as React from "react";
-import * as ResizablePrimitive from "react-resizable-panels";
-
-import { cn } from "@/lib/utils";
-
-function ResizablePanelGroup({
-  className,
-  ...props
-}: React.ComponentProps<typeof ResizablePrimitive.Group>): React.ReactElement {
-  return (
-    <ResizablePrimitive.Group
-      data-slot="resizable-panel-group"
-      className={cn(
-        "flex h-full w-full data-[panel-group-direction=vertical]:flex-col",
-        className,
-      )}
-      {...props}
-    />
-  );
-}
-
-function ResizablePanel({
-  ...props
-}: React.ComponentProps<typeof ResizablePrimitive.Panel>): React.ReactElement {
-  return <ResizablePrimitive.Panel data-slot="resizable-panel" {...props} />;
-}
-
-function ResizableHandle({
-  withHandle,
-  className,
-  ...props
-}: React.ComponentProps<typeof ResizablePrimitive.Separator> & {
-  withHandle?: boolean;
-}): React.ReactElement {
-  return (
-    <ResizablePrimitive.Separator
-      data-slot="resizable-handle"
-      className={cn(
-        "bg-border focus-visible:ring-ring relative flex w-px items-center justify-center after:absolute after:inset-y-0 after:left-1/2 after:w-1 after:-translate-x-1/2 focus-visible:ring-1 focus-visible:ring-offset-1 focus-visible:outline-hidden data-[panel-group-direction=vertical]:h-px data-[panel-group-direction=vertical]:w-full data-[panel-group-direction=vertical]:after:left-0 data-[panel-group-direction=vertical]:after:h-1 data-[panel-group-direction=vertical]:after:w-full data-[panel-group-direction=vertical]:after:translate-x-0 data-[panel-group-direction=vertical]:after:-translate-y-1/2 [&[data-panel-group-direction=vertical]>div]:rotate-90",
-        className,
-      )}
-      {...props}
-    >
-      {withHandle && (
-        <div className="bg-border h-6 w-1 rounded-lg z-10 flex shrink-0" />
-      )}
-    </ResizablePrimitive.Separator>
-  );
-}
-
-export { ResizablePanelGroup, ResizablePanel, ResizableHandle };
+import type * as React from "react";
+import * as ResizablePrimitive from "react-resizable-panels";
+
+import { cn } from "@/lib/utils";
+
+function ResizablePanelGroup({
+  className,
+  ...props
+}: React.ComponentProps<typeof ResizablePrimitive.Group>): React.ReactElement {
+  return (
+    <ResizablePrimitive.Group
+      data-slot="resizable-panel-group"
+      className={cn(
+        "flex h-full w-full data-[panel-group-direction=vertical]:flex-col",
+        className,
+      )}
+      {...props}
+    />
+  );
+}
+
+function ResizablePanel({
+  ...props
+}: React.ComponentProps<typeof ResizablePrimitive.Panel>): React.ReactElement {
+  return <ResizablePrimitive.Panel data-slot="resizable-panel" {...props} />;
+}
+
+function ResizableHandle({
+  withHandle,
+  className,
+  ...props
+}: React.ComponentProps<typeof ResizablePrimitive.Separator> & {
+  withHandle?: boolean;
+}): React.ReactElement {
+  return (
+    <ResizablePrimitive.Separator
+      data-slot="resizable-handle"
+      className={cn(
+        "group bg-border/80 relative z-10 flex w-px cursor-col-resize items-center justify-center transition-[background-color,box-shadow] duration-150 ease-out after:absolute after:inset-y-0 after:left-1/2 after:w-2 after:-translate-x-1/2 hover:bg-primary/80 hover:shadow-[0_0_16px_rgba(23,184,139,0.55)] active:bg-primary/90 active:shadow-[0_0_18px_rgba(23,184,139,0.7)] focus-visible:bg-primary/80 focus-visible:ring-1 focus-visible:ring-primary/50 focus-visible:ring-offset-1 focus-visible:outline-hidden data-[panel-group-direction=vertical]:h-px data-[panel-group-direction=vertical]:w-full data-[panel-group-direction=vertical]:cursor-row-resize data-[panel-group-direction=vertical]:after:left-0 data-[panel-group-direction=vertical]:after:h-2 data-[panel-group-direction=vertical]:after:w-full data-[panel-group-direction=vertical]:after:translate-x-0 data-[panel-group-direction=vertical]:after:-translate-y-1/2 [&[data-panel-group-direction=vertical]>div]:rotate-90",
+        className,
+      )}
+      {...props}
+    >
+      {withHandle && (
+        <div className="z-10 flex h-6 w-1 shrink-0 rounded-lg bg-border transition-[background-color,box-shadow,transform] duration-150 ease-out group-hover:scale-y-110 group-hover:bg-primary/80 group-hover:shadow-[0_0_12px_rgba(23,184,139,0.65)] group-active:bg-primary" />
+      )}
+    </ResizablePrimitive.Separator>
+  );
+}
+
+export { ResizablePanelGroup, ResizablePanel, ResizableHandle };

studio/frontend/src/components/ui/sidebar.tsx

@@ -30,7 +30,7 @@ import { LayoutAlignLeftIcon } from "@hugeicons/core-free-icons"
 
 const noop = () => {}
 
-const SIDEBAR_WIDTH = "16rem"
+const SIDEBAR_WIDTH = "17.5rem"
 const SIDEBAR_WIDTH_ICON = "3rem"
 const SIDEBAR_KEYBOARD_SHORTCUT = "b"
 
@@ -472,7 +472,7 @@ function SidebarGroupLabel({
       data-slot="sidebar-group-label"
       data-sidebar="group-label"
       className={cn(
-        "text-[#94a3b8] dark:text-[#666] ring-sidebar-ring h-auto pt-3 pb-2 px-4 rounded-md text-[10px] font-semibold uppercase tracking-[0.08em] group-data-[collapsible=icon]:-mt-8 group-data-[collapsible=icon]:opacity-0 focus-visible:ring-2 [&>svg]:size-3 flex shrink-0 items-center outline-hidden [&>svg]:shrink-0",
+        "text-[#94a3b8] dark:text-[#666] ring-sidebar-ring h-auto pt-3 pb-2 px-4 rounded-md text-[10px] font-semibold uppercase tracking-[0em] group-data-[collapsible=icon]:-mt-8 group-data-[collapsible=icon]:opacity-0 focus-visible:ring-2 [&>svg]:size-3 flex shrink-0 items-center outline-hidden [&>svg]:shrink-0",
         className
       )}
       {...props}

studio/frontend/src/components/ui/terminal.tsx

@@ -17,12 +17,20 @@ type TerminalProps = {
   className?: string
   sequence?: boolean
   startOnView?: boolean
+  /**
+   * Render every line in its final state immediately, skipping the typing /
+   * fade-in animations. Used when the terminal re-mounts for a run whose intro
+   * already played (e.g. navigating away from the training page and back), so
+   * the logs don't visually "restart" while the run itself keeps going.
+   */
+  instant?: boolean
 }
 
 type InternalLineProps = {
   __isActive?: boolean
   __onDone?: () => void
   __sequence?: boolean
+  __instant?: boolean
 }
 
 function useStartOnView(enabled: boolean): {
@@ -65,15 +73,18 @@ export function Terminal({
   className,
   sequence = true,
   startOnView = true,
+  instant = false,
 }: TerminalProps): ReactElement {
   const { ref, started } = useStartOnView(startOnView)
   const childElements = Children.toArray(children).filter(isValidElement)
   const [activeIndex, setActiveIndex] = useState(0)
-  const visibleIndex = sequence
-    ? started
-      ? activeIndex
-      : -1
-    : Number.MAX_SAFE_INTEGER
+  const visibleIndex = instant
+    ? Number.MAX_SAFE_INTEGER
+    : sequence
+      ? started
+        ? activeIndex
+        : -1
+      : Number.MAX_SAFE_INTEGER
 
   function handleLineDone(index: number): void {
     if (!sequence) {
@@ -99,7 +110,8 @@ export function Terminal({
       {childElements.map((child, index) =>
         cloneElement(child, {
           __sequence: sequence,
-          __isActive: !sequence || visibleIndex >= index,
+          __isActive: instant || !sequence || visibleIndex >= index,
+          __instant: instant,
           __onDone: () => handleLineDone(index),
           key: child.key ?? index,
         } as InternalLineProps)
@@ -122,11 +134,12 @@ export function AnimatedSpan({
   startOnView = false,
   __isActive,
   __sequence,
+  __instant,
   __onDone,
 }: AnimatedSpanProps): ReactElement {
   const { ref, started } = useStartOnView(startOnView)
-  const [visible, setVisible] = useState(false)
-  const doneRef = useRef(false)
+  const [visible, setVisible] = useState(Boolean(__instant))
+  const doneRef = useRef(Boolean(__instant))
   const onDoneRef = useRef(__onDone)
   const shouldStart = __sequence ? __isActive : started
 
@@ -180,11 +193,12 @@ export function TypingAnimation({
   startOnView = true,
   __isActive,
   __sequence,
+  __instant,
   __onDone,
 }: TypingAnimationProps): ReactElement {
   const { ref, started } = useStartOnView(startOnView)
-  const [typed, setTyped] = useState("")
-  const doneRef = useRef(false)
+  const [typed, setTyped] = useState(__instant ? children : "")
+  const doneRef = useRef(Boolean(__instant))
   const onDoneRef = useRef(__onDone)
   const shouldStart = __sequence ? __isActive : started
 

studio/frontend/src/features/chat/api/chat-adapter.ts

@@ -30,7 +30,6 @@ import {
 } from "../external-providers";
 import { pickFriendlyContainerName } from "../lib/friendly-names";
 import {
-  EXTERNAL_MAX_OUTPUT_TOKENS,
   clampReasoningEffortToLevels,
   getExternalMaxOutputTokens,
   getExternalMinOutputTokens,
@@ -59,6 +58,7 @@ import type {
 import type { ChatModelSummary } from "../types/runtime";
 import {
   getStoredChatThread,
+  getStoredChatProject,
   listStoredChatThreads,
   updateStoredChatThread,
 } from "../utils/chat-history-storage";
@@ -968,36 +968,6 @@ function toOpenAIMessages(message: RunMessage): SerializedMessage[] {
   return toolResults.length > 0 ? [base, ...toolResults] : [base];
 }
 
-// Thin singular wrapper: returns only the first serialized message
-// (without tool_calls or tool follow-ups) so the OpenAI image-edit
-// replay path can map a thread to flat OpenAI chat messages without
-// pulling in tool history.
-function toOpenAIMessage(message: RunMessage): {
-  role: "system" | "user" | "assistant";
-  content: OpenAIMessageContent;
-} | null {
-  const serialized = toOpenAIMessages(message);
-  if (serialized.length === 0) return null;
-  const first = serialized[0];
-  if (
-    first.role !== "system" &&
-    first.role !== "user" &&
-    first.role !== "assistant"
-  ) {
-    return null;
-  }
-  if (first.content === null || first.content === undefined) {
-    return null;
-  }
-  if (typeof first.content === "string" && !first.content) {
-    return null;
-  }
-  return {
-    role: first.role,
-    content: first.content as OpenAIMessageContent,
-  };
-}
-
 function extractImageBase64(input: string): string | undefined {
   if (!input) {
     return undefined;
@@ -1090,6 +1060,45 @@ async function resolveUseAdapter(
   }
 }
 
+async function resolveProjectInstructions(
+  threadId: string | undefined,
+): Promise<string> {
+  const projectId = await resolveProjectId(threadId);
+  if (!projectId) {
+    return "";
+  }
+
+  const project = await getStoredChatProject(projectId).catch(() => null);
+  if (!project || project.archived) {
+    return "";
+  }
+  return project.instructions?.trim() ?? "";
+}
+
+async function resolveProjectId(
+  threadId: string | undefined,
+): Promise<string | null> {
+  let projectId: string | null | undefined;
+  if (threadId) {
+    const thread = await getStoredChatThread(threadId).catch(() => null);
+    projectId = thread?.projectId ?? null;
+  }
+  if (!projectId) {
+    projectId = useChatRuntimeStore.getState().activeProjectId;
+  }
+  if (!projectId) {
+    return null;
+  }
+  return projectId;
+}
+
+async function resolveSandboxSessionId(
+  threadId: string | undefined,
+): Promise<string | undefined> {
+  const projectId = await resolveProjectId(threadId);
+  return projectId ? `project-${projectId}` : threadId;
+}
+
 /** Wait for an in-progress model load to finish (polls store every 500ms). */
 function waitForModelReady(abortSignal?: AbortSignal): Promise<void> {
   return new Promise((resolve, reject) => {
@@ -1426,6 +1435,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
       // the user switches chats while waiting for model load / auto-load.
       const resolvedThreadId =
         (unstable_threadId ?? runtime.activeThreadId) || undefined;
+      const sandboxSessionId = await resolveSandboxSessionId(resolvedThreadId);
       const resolvedThreadKey = resolvedThreadId ?? null;
       const pendingImageEditReferenceForRun = runtime.pendingImageEditReference;
       const selectedImageEditReference =
@@ -1496,6 +1506,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
         toolsEnabled,
         codeToolsEnabled,
         imageToolsEnabled,
+        artifactsEnabled,
         mcpEnabledForChat,
         webFetchToolsEnabled,
       } = runtime;
@@ -1701,7 +1712,14 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
 
       const safeSystemPrompt =
         typeof params.systemPrompt === "string" ? params.systemPrompt : "";
+      const projectInstructions =
+        await resolveProjectInstructions(resolvedThreadId);
       const systemPromptParts: string[] = [];
+      if (projectInstructions) {
+        systemPromptParts.push(
+          `<project_instructions>\n${projectInstructions}\n</project_instructions>`,
+        );
+      }
       if (safeSystemPrompt.trim()) {
         systemPromptParts.push(safeSystemPrompt.trim());
       }
@@ -1743,8 +1761,6 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
         outboundMessages.unshift({
           role: "system",
           content: systemPromptParts.join("\n\n"),
-        });
-      }
 
       if (ragPrefetchEnabled && ragToolEnabled && ragSource.kind !== "off") {
         const lastUser = [...outboundMessages]
@@ -1845,36 +1861,62 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
             "Do not return tool-call syntax inside your response.";
         }
       }
-      if (disabledToolGuard) {
-        const firstMessage = outboundMessages[0];
+      type OutboundMessage = (typeof outboundMessages)[number];
+      function addSystemInstruction(
+        targetMessages: OutboundMessage[],
+        text: string | null,
+      ): void {
+        if (!text) return;
+        const firstMessage = targetMessages[0];
         if (firstMessage?.role === "system") {
           if (typeof firstMessage.content === "string") {
-            outboundMessages[0] = {
+            targetMessages[0] = {
               ...firstMessage,
-              content: `${firstMessage.content}\n\n${disabledToolGuard}`,
+              content: `${firstMessage.content}\n\n${text}`,
             };
           } else {
-            outboundMessages[0] = {
+            targetMessages[0] = {
               ...firstMessage,
               content: [
                 ...(Array.isArray(firstMessage.content)
                   ? firstMessage.content
                   : []),
-                { type: "text", text: `\n\n${disabledToolGuard}` },
+                { type: "text", text: `\n\n${text}` },
               ],
             };
           }
-        } else {
-          outboundMessages.unshift({
-            role: "system",
-            content: disabledToolGuard,
-          });
+          return;
         }
+        targetMessages.unshift({ role: "system", content: text });
       }
+
       // Scan post-prune history so a refused user turn's image/audio
       // doesn't gate or mis-attribute the next non-refused turn.
       const imageBase64 = findLatestUserImageBase64(survivingMessages);
       const audioBase64 = findLatestUserAudioBase64(survivingMessages);
+      const hasOutboundImage = Boolean(imageBase64);
+
+      // Keep render_html local-only and mirror the backend image-turn gate.
+      // Artifacts are independent of Search/Code: if a local tool-capable
+      // model has Artifacts enabled, expose render_html even when no other
+      // tool pills are active.
+      const renderHtmlToolEnabledForThisTurn = Boolean(
+        !isExternalRequest &&
+          supportsTools &&
+          artifactsEnabled &&
+          !hasOutboundImage,
+      );
+      const artifactInstruction = artifactsEnabled
+        ? renderHtmlToolEnabledForThisTurn
+          ? "When the user asks for an HTML, CSS, or JavaScript artifact, call render_html once with one complete self-contained HTML document in the code argument. Embed CSS and JavaScript inside the document. After render_html succeeds, do not call it again in the same response unless the user asks for changes. Future user requests for new artifacts may call render_html once."
+          : "When the user asks for an HTML, CSS, or JavaScript artifact, return one complete self-contained fenced html code block. Embed CSS and JavaScript inside the document. Do not emit tool-call syntax."
+        : null;
+      const effectiveDisabledToolGuard =
+        disabledToolGuard && artifactsEnabled
+          ? `${disabledToolGuard} HTML, CSS, or JavaScript artifact requests can still be answered by following the artifact fallback instruction.`
+          : disabledToolGuard;
+      addSystemInstruction(outboundMessages, effectiveDisabledToolGuard);
+      addSystemInstruction(outboundMessages, artifactInstruction);
 
       // Block when ANY image is in the outbound payload (current or
       // prior turns) and the loaded model can't process images. Keeps
@@ -2074,7 +2116,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
       // /inference/cancel explicitly on abort.
       const onAbortCancel = () => {
         const body: Record<string, string> = { cancel_id: cancelId };
-        if (resolvedThreadId) body.session_id = resolvedThreadId;
+        if (sandboxSessionId) body.session_id = sandboxSessionId;
         // Plain fetch, not authFetch: authFetch redirects to login on
         // 401, which would kick the user out mid-stop.
         const token = getAuthToken();
@@ -2394,7 +2436,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
             image_base64: imageBase64,
             audio_base64: audioBase64,
             cancel_id: cancelId,
-            ...(resolvedThreadId ? { session_id: resolvedThreadId } : {}),
+            ...(sandboxSessionId ? { session_id: sandboxSessionId } : {}),
             ...(useAdapter === undefined ? {} : { use_adapter: useAdapter }),
             ...(supportsReasoning
               ? reasoningStyle === "reasoning_effort"
@@ -2410,6 +2452,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
             (toolsEnabled ||
               codeToolsEnabled ||
               ragToolPathTaken ||
+              renderHtmlToolEnabledForThisTurn ||
               mcpEnabledForChat)
               ? {
                   enable_tools: true,
@@ -2419,6 +2462,9 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                     ...(ragToolPathTaken ? ["search_knowledge_base"] : []),
                     ...(toolsEnabled ? ["web_search"] : []),
                     ...(codeToolsEnabled ? ["python", "terminal"] : []),
+                    ...(renderHtmlToolEnabledForThisTurn
+                      ? ["render_html"]
+                      : []),
                   ],
                   // Per-request scope for the LLM-invoked tool; tool path only.
                   ...(ragToolPathTaken
@@ -2550,13 +2596,25 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                     `${toolEvent.tool_name}_${Date.now()}`;
                   const toolArgs = (toolEvent.arguments ??
                     {}) as ToolCallMessagePart["args"];
-                  toolCallParts.push({
-                    type: "tool-call" as const,
-                    toolCallId: id,
-                    toolName: toolEvent.tool_name as string,
-                    argsText: JSON.stringify(toolArgs),
-                    args: toolArgs,
-                  });
+                  const idx = toolCallParts.findIndex(
+                    (p) => p.toolCallId === id,
+                  );
+                  if (idx !== -1) {
+                    toolCallParts[idx] = {
+                      ...toolCallParts[idx],
+                      toolName: toolEvent.tool_name as string,
+                      argsText: JSON.stringify(toolArgs),
+                      args: toolArgs,
+                    };
+                  } else {
+                    toolCallParts.push({
+                      type: "tool-call" as const,
+                      toolCallId: id,
+                      toolName: toolEvent.tool_name as string,
+                      argsText: JSON.stringify(toolArgs),
+                      args: toolArgs,
+                    });
+                  }
                 } else if (toolEvent.type === "tool_end") {
                   const id =
                     (toolEvent.tool_call_id as string) ||
@@ -2602,7 +2660,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                       const text = rawResult.slice(0, imgIdx);
                       // Fall back to "_default" to match the backend sandbox directory
                       // used when no session_id is provided (see tools.py _get_workdir).
-                      const sessionId = resolvedThreadId || "_default";
+                      const sessionId = sandboxSessionId || "_default";
                       try {
                         const images = JSON.parse(
                           rawResult.slice(imgIdx + imgMarker.length),

studio/frontend/src/features/chat/api/chat-api.ts

@@ -3,7 +3,12 @@
 
 import { authFetch } from "@/features/auth";
 import { formatFastApiDetail } from "@/lib/format-fastapi-error";
-import type { MessageRecord, ModelType, ThreadRecord } from "../types";
+import type {
+  MessageRecord,
+  ModelType,
+  ProjectRecord,
+  ThreadRecord,
+} from "../types";
 import type {
   AudioGenerationResponse,
   GgufVariantsResponse,
@@ -287,12 +292,14 @@ export async function listChatThreads(
   args: {
     modelType?: ModelType;
     pairId?: string;
+    projectId?: string | null;
     includeArchived?: boolean;
   } = {},
 ): Promise<ThreadRecord[]> {
   const params = new URLSearchParams();
   if (args.modelType) params.set("model_type", args.modelType);
   if (args.pairId) params.set("pair_id", args.pairId);
+  if (args.projectId) params.set("project_id", args.projectId);
   if (args.includeArchived !== undefined) {
     params.set("include_archived", String(args.includeArchived));
   }
@@ -353,6 +360,74 @@ export async function deleteChatThreads(threadIds: string[]): Promise<void> {
   notifyChatHistoryUpdated();
 }
 
+export async function listChatProjects(
+  args: { includeArchived?: boolean } = {},
+): Promise<ProjectRecord[]> {
+  const params = new URLSearchParams();
+  if (args.includeArchived !== undefined) {
+    params.set("include_archived", String(args.includeArchived));
+  }
+  const qs = params.toString();
+  const response = await authFetch(`/api/chat/projects${qs ? `?${qs}` : ""}`);
+  const data = await parseJsonOrThrow<{ projects: ProjectRecord[] }>(response);
+  return data.projects;
+}
+
+export async function getChatProject(
+  projectId: string,
+): Promise<ProjectRecord | null> {
+  const response = await authFetch(
+    `/api/chat/projects/${encodeURIComponent(projectId)}`,
+  );
+  if (response.status === 404) return null;
+  return parseJsonOrThrow<ProjectRecord>(response);
+}
+
+export async function saveChatProject(
+  project: ProjectRecord,
+): Promise<ProjectRecord> {
+  const response = await authFetch("/api/chat/projects", {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify(project),
+  });
+  const saved = await parseJsonOrThrow<ProjectRecord>(response);
+  notifyChatHistoryUpdated();
+  return saved;
+}
+
+export async function updateChatProject(
+  projectId: string,
+  patch: Partial<ProjectRecord>,
+): Promise<ProjectRecord> {
+  const response = await authFetch(
+    `/api/chat/projects/${encodeURIComponent(projectId)}`,
+    {
+      method: "PATCH",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify(patch),
+    },
+  );
+  const project = await parseJsonOrThrow<ProjectRecord>(response);
+  notifyChatHistoryUpdated();
+  return project;
+}
+
+export async function deleteChatProject(
+  projectId: string,
+  args: { deleteFiles?: boolean } = {},
+): Promise<void> {
+  const params = new URLSearchParams();
+  if (args.deleteFiles) params.set("delete_files", "true");
+  const qs = params.toString();
+  const response = await authFetch(
+    `/api/chat/projects/${encodeURIComponent(projectId)}${qs ? `?${qs}` : ""}`,
+    { method: "DELETE" },
+  );
+  await parseJsonOrThrow<ProjectRecord>(response);
+  notifyChatHistoryUpdated();
+}
+
 export async function listChatMessages(
   threadId: string,
 ): Promise<MessageRecord[]> {
@@ -464,6 +539,7 @@ export async function buildBackendChatExport(): Promise<{
   exportedAt: string;
   version: number;
   threadCount: number;
+  projects?: ProjectRecord[];
   threads: ThreadRecord[];
   messages: MessageRecord[];
 }> {

studio/frontend/src/features/chat/api/chat-settings-api.ts

@@ -30,6 +30,8 @@ export interface PersistedChatSettings {
   autoTitle?: boolean;
   reasoningEffort?: ReasoningEffort;
   preserveThinking?: boolean;
+  collapseHtmlArtifacts?: boolean;
+  allowArtifactNetworkAccess?: boolean;
   autoHealToolCalls?: boolean;
   maxToolCallsPerMessage?: number;
   toolCallTimeout?: number;

studio/frontend/src/features/chat/artifacts/artifact-card.tsx

@@ -0,0 +1,141 @@
+// 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 { cn } from "@/lib/utils";
+import { useAuiState } from "@assistant-ui/react";
+import { LayoutTwoColumnIcon as Layout2ColumnIcon } from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { useLayoutEffect, useMemo } from "react";
+import { useChatRuntimeStore } from "../stores/chat-runtime-store";
+import {
+  hasAutoOpenedArtifact,
+  rememberAutoOpenedArtifact,
+  useChatArtifactsStore,
+} from "./store";
+import {
+  type ChatArtifact,
+  type ChatArtifactSource,
+  createChatArtifact,
+} from "./types";
+
+export function ArtifactCard({
+  code,
+  title,
+  source,
+  sourceToolCallId,
+  sourceMessageId,
+  className,
+  autoOpen = false,
+  isStreaming = false,
+}: {
+  code: string;
+  title?: string | null;
+  source: ChatArtifactSource;
+  sourceToolCallId?: string | null;
+  sourceMessageId?: string | null;
+  className?: string;
+  autoOpen?: boolean;
+  isStreaming?: boolean;
+}) {
+  const activeThreadId = useChatRuntimeStore((state) => state.activeThreadId);
+  const messageIdFromContext = useAuiState(({ message }) => message.id);
+  const threadIdFromContext = useAuiState(
+    ({ threads }) => threads.mainThreadId,
+  );
+  const artifactThreadId = threadIdFromContext ?? activeThreadId ?? null;
+  const openArtifact = useChatArtifactsStore((state) => state.openArtifact);
+  const updateArtifact = useChatArtifactsStore((state) => state.updateArtifact);
+  const selectedArtifactId = useChatArtifactsStore(
+    (state) => state.selectedArtifactId,
+  );
+  const artifact = useMemo<ChatArtifact>(
+    () =>
+      createChatArtifact({
+        code,
+        title,
+        source,
+        sourceMessageId: sourceMessageId ?? messageIdFromContext ?? null,
+        sourceToolCallId: sourceToolCallId ?? null,
+        threadId: artifactThreadId,
+        isStreaming,
+      }),
+    [
+      artifactThreadId,
+      code,
+      isStreaming,
+      messageIdFromContext,
+      source,
+      sourceMessageId,
+      sourceToolCallId,
+      title,
+    ],
+  );
+  const surface = artifactThreadId ? "panel" : "overlay";
+
+  useLayoutEffect(() => {
+    if (selectedArtifactId === artifact.id) {
+      updateArtifact(artifact);
+    }
+
+    if (!autoOpen) {
+      return;
+    }
+    if (hasAutoOpenedArtifact(artifact.id)) {
+      return;
+    }
+
+    rememberAutoOpenedArtifact(artifact.id);
+    openArtifact(artifact, { surface });
+  }, [
+    artifact,
+    autoOpen,
+    openArtifact,
+    selectedArtifactId,
+    surface,
+    updateArtifact,
+  ]);
+
+  return (
+    <button
+      type="button"
+      className={cn(
+        "group/artifact-card relative my-2 flex min-h-[52px] w-full max-w-md cursor-pointer items-center overflow-hidden rounded-lg border border-border/70 bg-muted/15 px-3 py-2 text-left transition-colors hover:bg-muted/25 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2",
+        "dark:bg-muted/10 dark:hover:bg-muted/20",
+        isStreaming &&
+          "border-border/80 bg-muted/20 dark:border-border/70 dark:bg-muted/15",
+        className,
+      )}
+      onClick={() => openArtifact(artifact, { surface })}
+      aria-label={`Open ${artifact.title}`}
+    >
+      {isStreaming ? (
+        <span
+          aria-hidden={true}
+          className="artifact-card-shimmer pointer-events-none absolute inset-0 z-0 motion-reduce:hidden"
+        />
+      ) : null}
+      <div className="relative z-10 flex min-w-0 flex-1 items-center gap-2.5">
+        <HugeiconsIcon
+          icon={Layout2ColumnIcon}
+          strokeWidth={1.75}
+          className="size-5 shrink-0 text-muted-foreground"
+        />
+        <span className="grid min-w-0 flex-1 gap-1">
+          <span className="truncate text-sm font-medium leading-tight text-foreground">
+            {artifact.title}
+          </span>
+          <span className="truncate text-[11px] leading-none text-muted-foreground">
+            HTML artifact
+          </span>
+        </span>
+        {isStreaming ? (
+          <span className="shimmer shrink-0 rounded-full bg-primary/10 px-2 py-0.5 text-[10px] font-medium text-primary motion-reduce:animate-none">
+            Generating
+          </span>
+        ) : null}
+      </div>
+    </button>
+  );
+}

studio/frontend/src/features/chat/artifacts/artifact-surface.tsx

@@ -0,0 +1,362 @@
+// 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 { createCodePlugin } from "@/components/assistant-ui/code-plugin";
+import { CodeToggleIcon } from "@/components/assistant-ui/code-toggle-icon";
+import {
+  unslothDarkTheme,
+  unslothLightTheme,
+} from "@/components/assistant-ui/code-themes";
+import { Button } from "@/components/ui/button";
+import { copyToClipboard } from "@/lib/copy-to-clipboard";
+import { cn } from "@/lib/utils";
+import {
+  CheckIcon,
+  CopyIcon,
+  DownloadIcon,
+  EyeIcon,
+  Maximize2Icon,
+  XIcon,
+} from "lucide-react";
+import {
+  type KeyboardEvent,
+  useEffect,
+  useMemo,
+  useRef,
+  useState,
+} from "react";
+import { Streamdown } from "streamdown";
+import { ArtifactHtmlFrame, type ArtifactViewMode } from "./html-frame";
+import type { ChatArtifact } from "./types";
+import { getArtifactFilename } from "./types";
+
+const COPY_RESET_MS = 2000;
+const artifactSourceCodePlugin = createCodePlugin({
+  themes: [unslothLightTheme, unslothDarkTheme],
+});
+
+function buildHtmlFence(source: string): string {
+  const longestBacktickRun = Math.max(
+    2,
+    ...(source.match(/`+/g) ?? []).map((match) => match.length),
+  );
+  const fence = "`".repeat(longestBacktickRun + 1);
+  return `${fence}html\n${source}\n${fence}`;
+}
+// Sandboxed artifact iframes are intentionally excluded from the overlay focus
+// trap. Granting same-origin sandbox privileges would weaken isolation, so
+// keyboard users can reach Studio controls here while fully interactive artifact
+// content remains a known sandbox limitation.
+const FOCUSABLE_SELECTOR =
+  'a[href], button:not([disabled]), textarea:not([disabled]), input:not([disabled]), select:not([disabled]), [tabindex]:not([tabindex="-1"])';
+
+function getFocusableElements(container: HTMLElement): HTMLElement[] {
+  return Array.from(
+    container.querySelectorAll<HTMLElement>(FOCUSABLE_SELECTOR),
+  ).filter(
+    (element) =>
+      !element.hasAttribute("disabled") &&
+      element.getAttribute("aria-hidden") !== "true" &&
+      element.tabIndex !== -1,
+  );
+}
+
+function ArtifactLoadingLine() {
+  return (
+    <div className="absolute inset-x-0 bottom-0 h-[2.5px] overflow-hidden bg-border/45">
+      <span
+        aria-hidden={true}
+        className="artifact-loading-line block h-full rounded-full motion-reduce:hidden"
+      />
+    </div>
+  );
+}
+
+function ArtifactGeneratingPanel() {
+  return (
+    <div className="flex h-full min-h-0 flex-col items-center justify-center bg-muted/10 px-6 text-center">
+      <div className="max-w-[30ch] space-y-1.5">
+        <img
+          src="/Sloth%20emojis/sloth%20w%20pc%20transparent.png"
+          alt=""
+          aria-hidden={true}
+          className="mx-auto mb-3 size-20 object-contain"
+        />
+        <p className="text-sm font-medium text-foreground">
+          Building artifact preview…
+        </p>
+        <p className="text-xs leading-relaxed text-muted-foreground">
+          The preview will appear here when the HTML is ready.
+        </p>
+      </div>
+    </div>
+  );
+}
+
+function downloadTextFile(filename: string, text: string): void {
+  const blob = new Blob([text], { type: "text/html;charset=utf-8" });
+  const url = URL.createObjectURL(blob);
+  const anchor = document.createElement("a");
+  anchor.href = url;
+  anchor.download = filename;
+  document.body.appendChild(anchor);
+  anchor.click();
+  document.body.removeChild(anchor);
+  window.setTimeout(() => URL.revokeObjectURL(url), 0);
+}
+
+export function ArtifactSurface({
+  artifact,
+  variant,
+  onClose,
+  onOpenFullscreen,
+}: {
+  artifact: ChatArtifact;
+  variant: "panel" | "overlay";
+  onClose: () => void;
+  onOpenFullscreen?: () => void;
+}) {
+  const [viewMode, setViewMode] = useState<ArtifactViewMode>("preview");
+  const [copied, setCopied] = useState(false);
+  const copyResetRef = useRef<ReturnType<typeof setTimeout> | null>(null);
+  const surfaceRef = useRef<HTMLElement>(null);
+  const previousFocusRef = useRef<Element | null>(null);
+  const filename = getArtifactFilename(artifact);
+  const sourceMarkdown = useMemo(
+    () => buildHtmlFence(artifact.code),
+    [artifact.code],
+  );
+  const hasArtifactCode = artifact.code.trim().length > 0;
+  const isLoadingArtifact = Boolean(artifact.isStreaming);
+  const effectiveViewMode = isLoadingArtifact ? "preview" : viewMode;
+
+  useEffect(() => {
+    return () => {
+      if (copyResetRef.current) clearTimeout(copyResetRef.current);
+    };
+  }, []);
+
+  useEffect(() => {
+    if (variant !== "overlay") return;
+    previousFocusRef.current = document.activeElement;
+    const timeoutId = window.setTimeout(() => {
+      const surface = surfaceRef.current;
+      if (!surface) return;
+      const firstFocusable = getFocusableElements(surface)[0];
+      if (firstFocusable) {
+        firstFocusable.focus();
+      } else {
+        surface.focus();
+      }
+    }, 0);
+    return () => {
+      window.clearTimeout(timeoutId);
+      const previousFocus = previousFocusRef.current;
+      if (previousFocus instanceof HTMLElement) previousFocus.focus();
+    };
+  }, [variant]);
+
+  const handleCopy = async () => {
+    if (!(await copyToClipboard(artifact.code))) return;
+    setCopied(true);
+    if (copyResetRef.current) clearTimeout(copyResetRef.current);
+    copyResetRef.current = setTimeout(() => {
+      setCopied(false);
+      copyResetRef.current = null;
+    }, COPY_RESET_MS);
+  };
+
+  const handleDialogKeyDown = (event: KeyboardEvent<HTMLElement>) => {
+    if (variant !== "overlay") return;
+    if (event.key === "Escape") {
+      event.preventDefault();
+      onClose();
+      return;
+    }
+    if (event.key !== "Tab") return;
+    const focusable = getFocusableElements(event.currentTarget);
+    if (focusable.length === 0) {
+      event.preventDefault();
+      event.currentTarget.focus();
+      return;
+    }
+    const first = focusable[0];
+    const last = focusable[focusable.length - 1];
+    if (event.shiftKey && document.activeElement === first) {
+      event.preventDefault();
+      last.focus();
+    } else if (!event.shiftKey && document.activeElement === last) {
+      event.preventDefault();
+      first.focus();
+    }
+  };
+
+  const content = (
+    <section
+      ref={surfaceRef}
+      role={variant === "overlay" ? "dialog" : undefined}
+      aria-modal={variant === "overlay" ? true : undefined}
+      tabIndex={variant === "overlay" ? -1 : undefined}
+      onKeyDown={handleDialogKeyDown}
+      className={cn(
+        "relative flex min-h-0 flex-col bg-background",
+        variant === "panel"
+          ? "artifact-panel-shell mx-2 mt-[72px] mb-8 h-[calc(100%_-_104px)] overflow-visible rounded-[28px] border-t border-border/70 bg-card/95"
+          : "h-[min(92vh,900px)] w-[min(96vw,1200px)] overflow-hidden rounded-2xl border border-border shadow-xl",
+      )}
+      aria-label={`${artifact.title} artifact`}
+    >
+      <header
+        className={cn(
+          "relative flex shrink-0 items-center justify-between gap-3 px-2.5 py-2",
+          variant === "panel" && "rounded-t-[28px]",
+        )}
+      >
+        <div
+          className="flex items-center gap-1 rounded-full bg-muted/40 p-0.5"
+          role="tablist"
+          aria-label="Artifact view"
+        >
+          {(["preview", "source"] as const).map((mode) => {
+            const isPreview = mode === "preview";
+            const Icon = isPreview ? EyeIcon : CodeToggleIcon;
+            return (
+              <button
+                key={mode}
+                type="button"
+                role="tab"
+                disabled={isLoadingArtifact && !isPreview}
+                onClick={() => setViewMode(mode)}
+                className={cn(
+                  "flex size-8 items-center justify-center rounded-full text-muted-foreground transition-colors",
+                  effectiveViewMode === mode
+                    ? "bg-background text-foreground shadow-sm"
+                    : "hover:bg-background/70 hover:text-foreground",
+                  isLoadingArtifact &&
+                    !isPreview &&
+                    "cursor-not-allowed opacity-50",
+                )}
+                aria-label={
+                  isPreview ? "Preview artifact" : "View artifact source"
+                }
+                aria-selected={effectiveViewMode === mode}
+                aria-pressed={effectiveViewMode === mode}
+                title={
+                  isPreview
+                    ? "Preview"
+                    : isLoadingArtifact
+                      ? "Source available when generation finishes"
+                      : "Source"
+                }
+              >
+                <Icon className="size-4" />
+              </button>
+            );
+          })}
+        </div>
+        <div className="flex shrink-0 items-center gap-1">
+          <Button
+            type="button"
+            variant="ghost"
+            size="icon"
+            className="size-8"
+            disabled={isLoadingArtifact || !hasArtifactCode}
+            onClick={() => downloadTextFile(filename, artifact.code)}
+            aria-label="Download artifact HTML"
+          >
+            <DownloadIcon className="size-4" />
+          </Button>
+          <Button
+            type="button"
+            variant="ghost"
+            size="icon"
+            className="size-8"
+            disabled={isLoadingArtifact || !hasArtifactCode}
+            onClick={handleCopy}
+            aria-label="Copy artifact HTML"
+          >
+            {copied ? (
+              <CheckIcon className="size-4" />
+            ) : (
+              <CopyIcon className="size-4" />
+            )}
+          </Button>
+          {variant === "panel" && onOpenFullscreen ? (
+            <Button
+              type="button"
+              variant="ghost"
+              size="icon"
+              className="size-8"
+              onClick={onOpenFullscreen}
+              aria-label="Open artifact fullscreen"
+            >
+              <Maximize2Icon className="size-4" />
+            </Button>
+          ) : null}
+          <Button
+            type="button"
+            variant="ghost"
+            size="icon"
+            className="size-8"
+            onClick={onClose}
+            aria-label="Close artifact"
+          >
+            <XIcon className="size-4" />
+          </Button>
+        </div>
+        {isLoadingArtifact ? (
+          <ArtifactLoadingLine />
+        ) : (
+          <div className="absolute inset-x-0 bottom-0 h-px bg-border" />
+        )}
+      </header>
+
+      <div
+        className={cn(
+          "min-h-0 flex-1 overflow-hidden bg-background",
+          variant === "panel" && "rounded-b-[28px]",
+        )}
+      >
+        {isLoadingArtifact ? (
+          <ArtifactGeneratingPanel />
+        ) : effectiveViewMode === "preview" ? (
+          <ArtifactHtmlFrame
+            key={artifact.id}
+            code={artifact.code}
+            title={artifact.title}
+            fill={true}
+            className="h-full"
+          />
+        ) : (
+          <div className="h-full overflow-auto text-xs leading-relaxed [&_[data-streamdown=code-block]]:!my-0 [&_[data-streamdown=code-block]]:!gap-0 [&_[data-streamdown=code-block]]:!rounded-none [&_[data-streamdown=code-block]]:!border-0 [&_[data-streamdown=code-block]]:!bg-transparent [&_[data-streamdown=code-block]]:!p-0 [&_[data-streamdown=code-block-body]]:!border-0 [&_[data-streamdown=code-block-body]]:!bg-transparent [&_[data-streamdown=code-block-body]]:!p-0 [&_pre]:!m-0 [&_pre]:!bg-transparent [&_pre]:!p-0 [&_pre]:text-xs [&_pre]:leading-relaxed [&_code]:text-xs">
+            <Streamdown
+              mode="streaming"
+              plugins={{ code: artifactSourceCodePlugin }}
+              controls={{ code: false }}
+              shikiTheme={[unslothLightTheme, unslothDarkTheme]}
+            >
+              {sourceMarkdown}
+            </Streamdown>
+          </div>
+        )}
+      </div>
+    </section>
+  );
+
+  if (variant === "overlay") {
+    return (
+      <div
+        className="fixed inset-0 z-50 flex items-center justify-center bg-background/80 p-4 backdrop-blur-sm"
+        onMouseDown={(event) => {
+          if (event.target === event.currentTarget) onClose();
+        }}
+      >
+        {content}
+      </div>
+    );
+  }
+
+  return content;
+}

studio/frontend/src/features/chat/artifacts/html-frame.tsx

@@ -0,0 +1,100 @@
+// 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 { getAuthToken } from "@/features/auth";
+import { apiUrl } from "@/lib/api-base";
+import { cn } from "@/lib/utils";
+import { useCallback, useEffect, useMemo, useRef, useState } from "react";
+import { useChatRuntimeStore } from "../stores/chat-runtime-store";
+import { hashArtifactCode } from "./types";
+
+const HTML_FRAME_DEFAULT_HEIGHT = 400;
+const HTML_FRAME_MAX_HEIGHT = 900;
+
+export type ArtifactViewMode = "preview" | "source";
+export const ARTIFACT_VIEW_MODES: readonly ArtifactViewMode[] = [
+  "preview",
+  "source",
+];
+
+export function isArtifactViewMode(value: string): value is ArtifactViewMode {
+  return (ARTIFACT_VIEW_MODES as readonly string[]).includes(value);
+}
+
+export function buildArtifactSrcDoc(code: string): string {
+  const resizeScript = `<script>(()=>{const post=()=>parent.postMessage({chatArtifactHeight:document.documentElement.scrollHeight},"*");new ResizeObserver(post).observe(document.documentElement);window.addEventListener("load",post);post();})();</script>`;
+  return `${code}\n${resizeScript}`;
+}
+
+// Preview iframes intentionally omit allow-downloads: generated artifacts can
+// offer their own UI, but downloads must go through Studio's explicit
+// copy/download controls outside the no-same-origin sandbox.
+export function ArtifactHtmlFrame({
+  code,
+  title = "HTML artifact preview",
+  className,
+  fill = false,
+}: {
+  code: string;
+  title?: string;
+  className?: string;
+  fill?: boolean;
+}) {
+  const iframeRef = useRef<HTMLIFrameElement>(null);
+  const allowNetworkAccess = useChatRuntimeStore(
+    (state) => state.allowArtifactNetworkAccess,
+  );
+  const [height, setHeight] = useState(HTML_FRAME_DEFAULT_HEIGHT);
+  const artifactHtml = useMemo(() => buildArtifactSrcDoc(code), [code]);
+  const src = useMemo(() => {
+    const query = new URLSearchParams({ v: hashArtifactCode(code) });
+    if (allowNetworkAccess) {
+      const token = getAuthToken();
+      if (token) {
+        query.set("allow_network", "1");
+        query.set("token", token);
+      }
+    }
+    return apiUrl(`/api/inference/artifact-preview-frame?${query.toString()}`);
+  }, [allowNetworkAccess, code]);
+  const postArtifactHtml = useCallback(() => {
+    // The sandboxed frame intentionally has an opaque origin ("null").
+    // A wildcard target is required here;
+    // the payload is sent only to this iframe's contentWindow.
+    iframeRef.current?.contentWindow?.postMessage(
+      { type: "unsloth:artifact-html", html: artifactHtml },
+      "*",
+    );
+  }, [artifactHtml]);
+
+  useEffect(() => {
+    const handler = (event: MessageEvent) => {
+      if (event.source !== iframeRef.current?.contentWindow) return;
+      if (event.origin !== "null") return;
+      if (typeof event.data?.chatArtifactHeight !== "number") return;
+      setHeight(
+        Math.min(
+          Math.max(event.data.chatArtifactHeight, 160),
+          HTML_FRAME_MAX_HEIGHT,
+        ),
+      );
+    };
+    window.addEventListener("message", handler);
+    return () => window.removeEventListener("message", handler);
+  }, [postArtifactHtml]);
+
+  return (
+    <iframe
+      ref={iframeRef}
+      src={src}
+      sandbox="allow-scripts"
+      referrerPolicy="no-referrer"
+      onLoad={postArtifactHtml}
+      className={cn("block w-full border-0 bg-background", className)}
+      style={{ height: fill ? "100%" : height }}
+      title={title}
+    />
+  );
+}

studio/frontend/src/features/chat/artifacts/store.ts

@@ -0,0 +1,107 @@
+// 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 { create } from "zustand";
+import type { ChatArtifact, ChatArtifactSurface } from "./types";
+
+const autoOpenedArtifactIds = new Set<string>();
+
+export function hasAutoOpenedArtifact(artifactId: string): boolean {
+  return autoOpenedArtifactIds.has(artifactId);
+}
+
+export function rememberAutoOpenedArtifact(artifactId: string): void {
+  autoOpenedArtifactIds.add(artifactId);
+}
+
+export function clearAutoOpenedArtifacts(): void {
+  autoOpenedArtifactIds.clear();
+}
+
+type ChatArtifactsState = {
+  artifactsById: Record<string, ChatArtifact>;
+  selectedArtifactId: string | null;
+  surface: ChatArtifactSurface;
+  openArtifact: (
+    artifact: ChatArtifact,
+    options?: { surface?: ChatArtifactSurface },
+  ) => void;
+  updateArtifact: (artifact: ChatArtifact) => void;
+  closeArtifactSurface: () => void;
+  clearArtifactsForThread: (threadId: string | null | undefined) => void;
+  clearOrphanedArtifacts: () => void;
+  resetArtifacts: () => void;
+};
+
+export const useChatArtifactsStore = create<ChatArtifactsState>((set) => ({
+  artifactsById: {},
+  selectedArtifactId: null,
+  surface: "panel",
+  openArtifact: (artifact, options) =>
+    set((state) => ({
+      artifactsById: {
+        ...state.artifactsById,
+        [artifact.id]: artifact,
+      },
+      selectedArtifactId: artifact.id,
+      surface: options?.surface ?? state.surface,
+    })),
+  updateArtifact: (artifact) =>
+    set((state) =>
+      state.artifactsById[artifact.id]
+        ? {
+            artifactsById: {
+              ...state.artifactsById,
+              [artifact.id]: artifact,
+            },
+          }
+        : state,
+    ),
+  closeArtifactSurface: () =>
+    set({ selectedArtifactId: null, surface: "panel" }),
+  clearArtifactsForThread: (threadId) =>
+    set((state) => {
+      if (!threadId) return state;
+      const artifactsById = Object.fromEntries(
+        Object.entries(state.artifactsById).filter(
+          ([, artifact]) => artifact.threadId !== threadId,
+        ),
+      );
+      const selected = state.selectedArtifactId
+        ? artifactsById[state.selectedArtifactId]
+        : null;
+      return {
+        artifactsById,
+        selectedArtifactId: selected ? selected.id : null,
+      };
+    }),
+  clearOrphanedArtifacts: () =>
+    set((state) => {
+      const artifactsById = Object.fromEntries(
+        Object.entries(state.artifactsById).filter(
+          ([, artifact]) => artifact.threadId != null,
+        ),
+      );
+      const selected = state.selectedArtifactId
+        ? artifactsById[state.selectedArtifactId]
+        : null;
+      return {
+        artifactsById,
+        selectedArtifactId: selected ? selected.id : null,
+      };
+    }),
+  resetArtifacts: () =>
+    set({
+      artifactsById: {},
+      selectedArtifactId: null,
+      surface: "panel",
+    }),
+}));
+
+export function useSelectedChatArtifact(): ChatArtifact | null {
+  return useChatArtifactsStore((state) =>
+    state.selectedArtifactId
+      ? (state.artifactsById[state.selectedArtifactId] ?? null)
+      : null,
+  );
+}

studio/frontend/src/features/chat/artifacts/types.ts

@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+export type ChatArtifactSource = "tool" | "fence";
+export type ChatArtifactSurface = "panel" | "overlay";
+
+export interface ChatArtifact {
+  id: string;
+  title: string;
+  code: string;
+  source: ChatArtifactSource;
+  sourceMessageId?: string | null;
+  sourceToolCallId?: string | null;
+  threadId?: string | null;
+  isStreaming?: boolean;
+  createdAt: number;
+}
+
+export interface ChatArtifactInput {
+  title?: string | null;
+  code: string;
+  source: ChatArtifactSource;
+  sourceMessageId?: string | null;
+  sourceToolCallId?: string | null;
+  threadId?: string | null;
+  isStreaming?: boolean;
+}
+
+const DEFAULT_ARTIFACT_TITLE = "HTML artifact";
+
+export function normalizeArtifactTitle(title?: string | null): string {
+  const trimmed = title?.trim();
+  return trimmed && trimmed.length > 0 ? trimmed : DEFAULT_ARTIFACT_TITLE;
+}
+
+export function hashArtifactCode(code: string): string {
+  let hash = 5381;
+  for (let i = 0; i < code.length; i += 1) {
+    hash = ((hash << 5) + hash) ^ code.charCodeAt(i);
+  }
+  return (hash >>> 0).toString(36);
+}
+
+export function createArtifactId(input: ChatArtifactInput): string {
+  const threadSegment = input.threadId || "no-thread";
+  const messageSegment = input.sourceMessageId || "transient";
+  // Backend tool call IDs (call_0, call_1, …) reset per request, so
+  // the message ID is needed to scope them to a specific turn.
+  const parts = [input.source, threadSegment, messageSegment];
+
+  if (input.source === "tool" && input.sourceToolCallId) {
+    parts.push(input.sourceToolCallId);
+  } else {
+    parts.push(hashArtifactCode(input.code));
+  }
+
+  return parts.join(":");
+}
+
+export function createChatArtifact(input: ChatArtifactInput): ChatArtifact {
+  return {
+    id: createArtifactId(input),
+    title: normalizeArtifactTitle(input.title),
+    code: input.code,
+    source: input.source,
+    sourceMessageId: input.sourceMessageId ?? null,
+    sourceToolCallId: input.sourceToolCallId ?? null,
+    threadId: input.threadId ?? null,
+    isStreaming: input.isStreaming,
+    createdAt: Date.now(),
+  };
+}
+
+export function getArtifactFilename(
+  artifact: Pick<ChatArtifact, "title">,
+): string {
+  const slug = artifact.title
+    .trim()
+    .toLowerCase()
+    .replace(/[^a-z0-9]+/g, "-")
+    .replace(/^-+|-+$/g, "")
+    .slice(0, 48);
+  return `${slug || "artifact"}.html`;
+}

studio/frontend/src/features/chat/chat-mcp-servers-dialog.tsx

@@ -66,23 +66,40 @@ function headersToObject(rows: HeaderRow[]): Record<string, string> | undefined
   return Object.keys(out).length > 0 ? out : undefined;
 }
 
-function isValidUrl(url: string): boolean {
-  const trimmed = url.trim();
+// A non-HTTP address is a local stdio command. Case-insensitive to match the
+// backend's is_stdio(), so all layers split http-vs-command identically.
+function isHttpAddress(value: string): boolean {
+  const trimmed = value.trim().toLowerCase();
+  return trimmed.startsWith("http://") || trimmed.startsWith("https://");
+}
+
+function isValidAddress(value: string): boolean {
+  const trimmed = value.trim();
   if (!trimmed) return false;
-  try {
-    const parsed = new URL(trimmed);
-    return parsed.protocol === "http:" || parsed.protocol === "https:";
-  } catch {
-    return false;
+  if (isHttpAddress(trimmed)) {
+    try {
+      const parsed = new URL(trimmed);
+      return parsed.protocol === "http:" || parsed.protocol === "https:";
+    } catch {
+      return false;
+    }
   }
+  // Anything else is treated as a local command (stdio); the backend gates
+  // whether stdio servers are allowed on this host. Reject other URL schemes
+  // only when the command itself is a URL; "://" is fine inside an argument
+  // (e.g. a database connection string passed to the server).
+  return !trimmed.split(/\s+/)[0].includes("://");
 }
 
 function HeadersEditor({
   rows,
   onChange,
+  stdio,
 }: {
   rows: HeaderRow[];
   onChange: (rows: HeaderRow[]) => void;
+  // stdio servers reuse this editor for environment variables instead of headers.
+  stdio: boolean;
 }) {
   const update = (id: string, patch: Partial<HeaderRow>) =>
     onChange(rows.map((row) => (row.id === id ? { ...row, ...patch } : row)));
@@ -91,19 +108,41 @@ function HeadersEditor({
   const remove = (id: string) =>
     onChange(rows.filter((row) => row.id !== id));
 
+  const copy = stdio
+    ? {
+        label: "Environment variables",
+        add: "Add variable",
+        keyPlaceholder: "Variable name",
+        valuePlaceholder: "Variable value",
+        remove: "Remove variable",
+      }
+    : {
+        label: "Custom headers",
+        add: "Add header",
+        keyPlaceholder: "Header name",
+        valuePlaceholder: "Header value",
+        remove: "Remove header",
+      };
+
   return (
     <>
       <div className="flex items-center justify-between">
-        <Label className="text-sm">Custom headers</Label>
+        <Label className="text-sm">{copy.label}</Label>
         <Button type="button" variant="ghost" size="sm" onClick={add}>
           <HugeiconsIcon icon={PlusSignIcon} size={14} />
-          Add header
+          {copy.add}
         </Button>
       </div>
       {rows.length === 0 ? (
         <div className="text-xs text-muted-foreground">
-          Optional. Add an <code>Authorization</code> header here for servers
-          that require auth.
+          {stdio ? (
+            "Optional. Environment variables passed to the server process."
+          ) : (
+            <>
+              Optional. Add an <code>Authorization</code> header here for servers
+              that require auth.
+            </>
+          )}
         </div>
       ) : (
         <div className="flex flex-col gap-2">
@@ -111,12 +150,12 @@ function HeadersEditor({
             <div key={row.id} className="flex items-center gap-2">
               <Input
                 value={row.key}
-                placeholder="Header name"
+                placeholder={copy.keyPlaceholder}
                 onChange={(e) => update(row.id, { key: e.target.value })}
               />
               <Input
                 value={row.value}
-                placeholder="Header value"
+                placeholder={copy.valuePlaceholder}
                 onChange={(e) => update(row.id, { value: e.target.value })}
               />
               <Button
@@ -124,7 +163,7 @@ function HeadersEditor({
                 variant="ghost"
                 size="icon"
                 onClick={() => remove(row.id)}
-                aria-label="Remove header"
+                aria-label={copy.remove}
               >
                 <HugeiconsIcon icon={Delete02Icon} size={14} />
               </Button>
@@ -139,6 +178,8 @@ function HeadersEditor({
 export interface ChatMcpServersDialogProps {
   open: boolean;
   onOpenChange: (open: boolean) => void;
+  /** Open straight into the "add server" form instead of the list view. */
+  openToCreate?: boolean;
 }
 
 type View =
@@ -149,6 +190,7 @@ type View =
 export function ChatMcpServersDialog({
   open,
   onOpenChange,
+  openToCreate = false,
 }: ChatMcpServersDialogProps) {
   const [servers, setServers] = useState<McpServerConfig[]>([]);
   const [loading, setLoading] = useState(false);
@@ -175,7 +217,12 @@ export function ChatMcpServersDialog({
   useEffect(() => {
     if (!open) return;
     refresh();
-  }, [open, refresh]);
+    // Land on the create form when opened via "Add custom MCP".
+    if (openToCreate) {
+      setView({ kind: "create" });
+      setForm(EMPTY_FORM);
+    }
+  }, [open, openToCreate, refresh]);
 
   function startCreate() {
     setView({ kind: "create" });
@@ -199,8 +246,8 @@ export function ChatMcpServersDialog({
 
   async function testConnection() {
     const trimmedUrl = form.url.trim();
-    if (!isValidUrl(trimmedUrl)) {
-      toast.error("Enter a valid http:// or https:// URL first");
+    if (!isValidAddress(trimmedUrl)) {
+      toast.error("Enter an http(s):// URL or a local command first");
       return;
     }
     setTesting(true);
@@ -236,11 +283,11 @@ export function ChatMcpServersDialog({
       return;
     }
     if (!trimmedUrl) {
-      toast.error("URL is required");
+      toast.error("URL or command is required");
       return;
     }
-    if (!isValidUrl(trimmedUrl)) {
-      toast.error("URL must start with http:// or https://");
+    if (!isValidAddress(trimmedUrl)) {
+      toast.error("Enter an http(s):// URL or a local command");
       return;
     }
     setSaving(true);
@@ -331,6 +378,9 @@ export function ChatMcpServersDialog({
   }
 
   const showForm = view.kind !== "list";
+  // A local stdio command uses env vars, not headers or OAuth.
+  const addressIsCommand =
+    form.url.trim() !== "" && !isHttpAddress(form.url);
 
   return (
     <Dialog open={open} onOpenChange={onOpenChange}>
@@ -338,7 +388,7 @@ export function ChatMcpServersDialog({
         <DialogHeader>
           <DialogTitle>MCP Servers</DialogTitle>
           <DialogDescription>
-            Register remote MCP servers.
+            Register remote (HTTP) or local (stdio command) MCP servers.
           </DialogDescription>
         </DialogHeader>
 
@@ -356,40 +406,47 @@ export function ChatMcpServersDialog({
               />
             </div>
             <div className="grid gap-2">
-              <Label htmlFor="mcp-url">URL</Label>
+              <Label htmlFor="mcp-url">URL or command</Label>
               <Input
                 id="mcp-url"
                 value={form.url}
                 onChange={(e) =>
                   setForm((prev) => ({ ...prev, url: e.target.value }))
                 }
-                placeholder="https://example.com/mcp"
+                placeholder="https://example.com/mcp or npx -y @modelcontextprotocol/server-filesystem /tmp"
               />
+              <span className="text-xs text-muted-foreground">
+                An http(s) URL for a remote server, or a local command to run an
+                stdio server (desktop app only).
+              </span>
             </div>
 
-            <div className="flex items-start justify-between gap-3">
-              <div className="flex flex-col gap-0.5">
-                <Label className="text-sm" htmlFor="mcp-oauth">
-                  Use OAuth sign-in
-                </Label>
-                <span className="text-xs text-muted-foreground">
-                  For servers that require browser-based authentication
-                  (GitHub, Linear, etc.). A browser window will open on first
-                  connect.
-                </span>
+            {!addressIsCommand && (
+              <div className="flex items-start justify-between gap-3">
+                <div className="flex flex-col gap-0.5">
+                  <Label className="text-sm" htmlFor="mcp-oauth">
+                    Use OAuth sign-in
+                  </Label>
+                  <span className="text-xs text-muted-foreground">
+                    For servers that require browser-based authentication
+                    (GitHub, Linear, etc.). A browser window will open on first
+                    connect.
+                  </span>
+                </div>
+                <Switch
+                  id="mcp-oauth"
+                  checked={form.useOauth}
+                  onCheckedChange={(useOauth) =>
+                    setForm((prev) => ({ ...prev, useOauth }))
+                  }
+                />
               </div>
-              <Switch
-                id="mcp-oauth"
-                checked={form.useOauth}
-                onCheckedChange={(useOauth) =>
-                  setForm((prev) => ({ ...prev, useOauth }))
-                }
-              />
-            </div>
+            )}
 
             <HeadersEditor
               rows={form.headers}
               onChange={(headers) => setForm((prev) => ({ ...prev, headers }))}
+              stdio={addressIsCommand}
             />
 
             <div className="flex items-center justify-between gap-2 pt-2">
@@ -415,7 +472,7 @@ export function ChatMcpServersDialog({
             </div>
           </div>
         ) : (
-          <div className="flex flex-col gap-3">
+          <div className="flex min-w-0 flex-col gap-3">
             <div className="flex justify-end">
               <Button size="sm" onClick={startCreate}>
                 <HugeiconsIcon icon={PlusSignIcon} size={14} />

studio/frontend/src/features/chat/chat-providers-dialog.tsx

@@ -391,9 +391,22 @@ export function ChatProvidersSettings({
             const updatedAt = Number.isFinite(Date.parse(config.updated_at))
               ? Date.parse(config.updated_at)
               : Date.now();
+            const registryEntry =
+              registryRows.find((entry) => entry.provider_type === uiProviderType) ??
+              registryRows.find((entry) => entry.provider_type === config.provider_type);
+            const defaultModels = pruneProviderModelIds(
+              uiProviderType,
+              registryEntry?.default_models ?? [],
+            );
+            const savedModels = existing?.models ?? [];
+            const savedAvailableModels = existing?.availableModels ?? [];
             const existingModels = pruneProviderModelIds(
               uiProviderType,
-              existing?.models ?? [],
+              savedModels.length > 0 ? savedModels : defaultModels,
+            );
+            const existingAvailableModels = pruneProviderModelIds(
+              uiProviderType,
+              savedAvailableModels.length > 0 ? savedAvailableModels : defaultModels,
             );
             return {
               id: config.id,
@@ -401,7 +414,7 @@ export function ChatProvidersSettings({
               name: config.display_name,
               baseUrl: config.base_url ?? "",
               models: existingModels,
-              availableModels: existing?.availableModels ?? [],
+              availableModels: existingAvailableModels,
               enablePromptCaching: supportsProviderPromptCaching(uiProviderType)
                 ? (existing?.enablePromptCaching ?? true)
                 : undefined,

studio/frontend/src/features/chat/chat-settings-sheet.tsx

@@ -69,7 +69,7 @@ import {
 } from "@hugeicons/core-free-icons";
 import { Add01Icon, Delete02Icon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { ChevronDown } from "lucide-react";
+import { ChevronDown, ExternalLink } from "lucide-react";
 import { Tooltip as TooltipPrimitive } from "radix-ui";
 import { type CSSProperties, Fragment, type ReactNode } from "react";
 import {
@@ -100,7 +100,6 @@ import {
   toPresetParams,
 } from "./presets/preset-policy";
 import {
-  EXTERNAL_MAX_OUTPUT_TOKENS,
   type ProviderCapabilities,
   getExternalMaxOutputTokens,
   getExternalMinOutputTokens,
@@ -108,8 +107,6 @@ import {
   providerSupportsFastMode,
 } from "./provider-capabilities";
 import { useChatRuntimeStore } from "./stores/chat-runtime-store";
-import { ChatMcpServersDialog } from "./chat-mcp-servers-dialog";
-import { listMcpServers } from "./api/mcp-servers-api";
 import type { InferenceParams } from "./types/runtime";
 
 function ragSourceLabel(
@@ -391,11 +388,19 @@ function saveCollapsibleOpen(label: string, open: boolean) {
 
 function CollapsibleSection({
   label,
+  labelHref,
   children,
   defaultOpen = false,
   first = false,
 }: {
   label: string;
+  /**
+   * When set, the label text becomes an external link (e.g. to the feature's
+   * GitHub PR) instead of part of the collapse toggle. The chevron still
+   * toggles open/close, so we render the two as siblings rather than nesting
+   * an <a> inside the <button> (invalid HTML).
+   */
+  labelHref?: string;
   children?: ReactNode;
   defaultOpen?: boolean;
   first?: boolean;
@@ -405,31 +410,59 @@ function CollapsibleSection({
     return Object.hasOwn(saved, label) ? saved[label] : defaultOpen;
   });
 
+  const toggle = () => {
+    const next = !open;
+    setOpen(next);
+    saveCollapsibleOpen(label, next);
+  };
+
+  const headerClasses = cn(
+    "flex w-full items-center justify-between text-[12px] font-medium normal-case tracking-[0.04em] text-nav-fg-muted transition-colors focus-visible:outline-none focus-visible:ring-0",
+    first ? "pt-4 pb-5" : "py-5",
+  );
+
   return (
     <div
       className={cn(
         !first && "border-t border-black/[0.13] dark:border-white/[0.09]",
       )}
     >
-      <button
-        type="button"
-        onClick={() => {
-          const next = !open;
-          setOpen(next);
-          saveCollapsibleOpen(label, next);
-        }}
-        className={cn(
-          "flex w-full cursor-pointer items-center justify-between text-[12px] font-medium normal-case tracking-[0.04em] text-nav-fg-muted transition-colors hover:text-nav-fg focus-visible:outline-none focus-visible:ring-0",
-          first ? "pt-4 pb-5" : "py-5",
-        )}
-      >
-        <span className="leading-none">{label}</span>
-        <span className="flex shrink-0 items-center leading-none">
-          <ChevronDown
-            className={cn("size-3.5", open ? "rotate-0" : "-rotate-90")}
-          />
-        </span>
-      </button>
+      {labelHref ? (
+        <div className={headerClasses}>
+          <a
+            href={labelHref}
+            target="_blank"
+            rel="noopener noreferrer"
+            className="inline-flex cursor-pointer items-center gap-1 leading-none transition-colors hover:text-nav-fg"
+          >
+            <span>{label}</span>
+            <ExternalLink className="size-3" />
+          </a>
+          <button
+            type="button"
+            onClick={toggle}
+            aria-label={open ? `Collapse ${label}` : `Expand ${label}`}
+            className="flex shrink-0 cursor-pointer items-center leading-none transition-colors hover:text-nav-fg"
+          >
+            <ChevronDown
+              className={cn("size-3.5", open ? "rotate-0" : "-rotate-90")}
+            />
+          </button>
+        </div>
+      ) : (
+        <button
+          type="button"
+          onClick={toggle}
+          className={cn("cursor-pointer hover:text-nav-fg", headerClasses)}
+        >
+          <span className="leading-none">{label}</span>
+          <span className="flex shrink-0 items-center leading-none">
+            <ChevronDown
+              className={cn("size-3.5", open ? "rotate-0" : "-rotate-90")}
+            />
+          </span>
+        </button>
+      )}
       {open && <div className="pb-7">{children}</div>}
     </div>
   );
@@ -604,16 +637,6 @@ export function ChatSettingsPanel({
     (s) => s.modelRequiresTrustRemoteCode,
   );
   const currentCheckpoint = params.checkpoint;
-  const currentModelIsMultimodal = useChatRuntimeStore((s) => {
-    if (s.loadedIsMultimodal) return true;
-    const m = s.models.find((m) => m.id === currentCheckpoint);
-    return (
-      Boolean(m?.isVision) ||
-      Boolean(m?.isAudio) ||
-      Boolean(m?.hasAudioInput) ||
-      m?.audioType === "audio_vlm"
-    );
-  });
   const ggufContextLength = useChatRuntimeStore((s) => s.ggufContextLength);
   const ggufMaxContextLength = useChatRuntimeStore(
     (s) => s.ggufMaxContextLength,
@@ -861,45 +884,50 @@ export function ChatSettingsPanel({
   const previewTarget = usePreviewStore((s) => s.target);
   const previewStatus = usePreviewStore((s) => s.status);
 
+  const settingsScrollRef = useRef<HTMLDivElement>(null);
+
   const settingsContent = (
     <>
-      <div className="aui-thread-viewport relative h-full overflow-y-auto">
-        <div className="sticky top-0 z-10 flex h-[48px] items-start gap-2 bg-panel-surface pl-[18px] pr-[14px] pt-[11px]">
-          {rightSlotUsesSheet ? (
-            <span className="flex h-[34px] flex-1 items-center text-[15px] font-semibold tracking-[-0.01em] dark:tracking-[0.015em] text-nav-fg">
-              Configuration
+      <div
+        ref={settingsScrollRef}
+        className="relative h-full overflow-y-auto"
+      >
+      <div className="sticky top-0 z-10 flex h-[48px] items-start gap-2 bg-panel-surface pl-[18px] pr-[16px] pt-[11px]">
+        {rightSlotUsesSheet ? (
+          <span className="flex h-[34px] flex-1 items-center text-[15px] font-semibold tracking-[0em] dark:tracking-[0.015em] text-nav-fg">
+            Run settings
+          </span>
+        ) : (
+          <>
+            <span className="flex h-[34px] flex-1 items-center text-[15px] font-semibold tracking-[0em] dark:tracking-[0.015em] text-nav-fg">
+              Run settings
             </span>
-          ) : (
-            <>
-              <span className="flex h-[34px] flex-1 items-center text-[15px] font-semibold tracking-[-0.01em] dark:tracking-[0.015em] text-nav-fg">
-                Configuration
-              </span>
-              <Tooltip>
-                <TooltipPrimitive.Trigger asChild={true}>
-                  <button
-                    type="button"
-                    onClick={() => onOpenChange?.(false)}
-                    className="flex h-[34px] w-[34px] items-center justify-center rounded-[12px] 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-2 focus-visible:ring-ring"
-                    aria-label="Close configuration"
-                  >
-                    <HugeiconsIcon
-                      icon={LayoutAlignRightIcon}
-                      strokeWidth={1.75}
-                      className="size-icon"
-                    />
-                  </button>
-                </TooltipPrimitive.Trigger>
-                <TooltipContent
-                  side="bottom"
-                  sideOffset={6}
-                  className="tooltip-compact"
+            <Tooltip>
+              <TooltipPrimitive.Trigger asChild={true}>
+                <button
+                  type="button"
+                  onClick={() => onOpenChange?.(false)}
+                  className="flex h-[34px] w-[34px] cursor-pointer items-center justify-center rounded-[12px] 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-2 focus-visible:ring-ring"
+                  aria-label="Close run settings"
                 >
-                  Close configuration
-                </TooltipContent>
-              </Tooltip>
-            </>
-          )}
-        </div>
+                  <HugeiconsIcon
+                    icon={LayoutAlignRightIcon}
+                    strokeWidth={1.75}
+                    className="size-icon"
+                  />
+                </button>
+              </TooltipPrimitive.Trigger>
+              <TooltipContent
+                side="bottom"
+                sideOffset={6}
+                className="tooltip-compact"
+              >
+                Close run settings
+              </TooltipContent>
+            </Tooltip>
+          </>
+        )}
+      </div>
 
         <div className="px-[18px] pt-3">
           {hasModelContent && (
@@ -1876,7 +1904,7 @@ export function ChatSettingsPanel({
                 }
                 max={
                   isExternalModel
-                    ? EXTERNAL_MAX_OUTPUT_TOKENS
+                    ? getExternalMaxOutputTokens(externalProviderType)
                     : isGguf && ggufContextLength
                       ? ggufContextLength
                       : 32768
@@ -1904,12 +1932,6 @@ export function ChatSettingsPanel({
               </div>
             </CollapsibleSection>
           )}
-
-          {!isExternalModel ? (
-            <CollapsibleSection label="MCP Servers">
-              <McpServersSection />
-            </CollapsibleSection>
-          ) : null}
         </div>
       </div>
       <Dialog
@@ -2010,7 +2032,7 @@ export function ChatSettingsPanel({
         >
           <SheetHeader className="sr-only">
             <SheetTitle>
-              {slotShowsPreview ? "Document preview" : "Configuration"}
+              {slotShowsPreview ? "Document preview" : "Run settings"}
             </SheetTitle>
             <SheetDescription>
               {slotShowsPreview
@@ -2189,74 +2211,6 @@ function AutoHealToolCallsToggle() {
   );
 }
 
-function McpServersSection() {
-  const mcpEnabledForChat = useChatRuntimeStore((s) => s.mcpEnabledForChat);
-  const setMcpEnabledForChat = useChatRuntimeStore(
-    (s) => s.setMcpEnabledForChat,
-  );
-  const [enabledServerCount, setEnabledServerCount] = useState<number | null>(
-    null,
-  );
-  const [dialogOpen, setDialogOpen] = useState(false);
-  const [refreshTick, setRefreshTick] = useState(0);
-
-  useEffect(() => {
-    let cancelled = false;
-    listMcpServers()
-      .then((rows) => {
-        if (cancelled) return;
-        setEnabledServerCount(rows.filter((row) => row.is_enabled).length);
-      })
-      .catch(() => {
-        if (!cancelled) setEnabledServerCount(0);
-      });
-    return () => {
-      cancelled = true;
-    };
-  }, [refreshTick]);
-
-  return (
-    <div className="flex flex-col gap-3 pt-1">
-      <div className="flex items-center justify-between gap-3">
-        <div className="flex min-w-0 items-center gap-1.5">
-          <span className="min-w-0 text-[13px] font-medium leading-[1.25] tracking-nav text-nav-fg">
-            Use MCP Servers
-          </span>
-          <InfoHint>
-            When on, every server marked enabled in the manage dialog is
-            attached to this chat's tool list.
-          </InfoHint>
-        </div>
-        <Switch
-          className="panel-switch"
-          checked={mcpEnabledForChat}
-          onCheckedChange={setMcpEnabledForChat}
-          disabled={enabledServerCount === 0 && !mcpEnabledForChat}
-        />
-      </div>
-      <div className="flex items-center justify-between">
-        <span className="text-[11px] text-muted-foreground">
-          {enabledServerCount === null
-            ? "Loading…"
-            : enabledServerCount === 0
-              ? "No servers configured"
-              : `${enabledServerCount} server${enabledServerCount === 1 ? "" : "s"} enabled`}
-        </span>
-        <Button variant="ghost" size="sm" onClick={() => setDialogOpen(true)}>
-          Manage…
-        </Button>
-      </div>
-      <ChatMcpServersDialog
-        open={dialogOpen}
-        onOpenChange={(next) => {
-          setDialogOpen(next);
-          if (!next) setRefreshTick((tick) => tick + 1);
-        }}
-      />
-    </div>
-  );
-}
-
 function ChatTemplateFields() {
   const defaultTemplate = useChatRuntimeStore((s) => s.defaultChatTemplate);
   const override = useChatRuntimeStore((s) => s.chatTemplateOverride);

studio/frontend/src/features/chat/components/chat-search-dialog.tsx

@@ -8,7 +8,6 @@ import {
   CommandGroup,
   CommandList,
 } from "@/components/ui/command";
-import { useTrainingRuntimeStore } from "@/features/training";
 import { Cancel01Icon, Message01Icon, SearchIcon } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { useNavigate } from "@tanstack/react-router";
@@ -36,7 +35,6 @@ export function ChatSearchDialog() {
   useEffect(() => {
     const handler = (e: KeyboardEvent) => {
       if (!(e.metaKey || e.ctrlKey) || e.key.toLowerCase() !== "k") return;
-      if (useTrainingRuntimeStore.getState().isTrainingRunning) return;
       const el = document.activeElement as HTMLElement | null;
       const tag = el?.tagName;
       if (tag === "INPUT" || tag === "TEXTAREA" || el?.isContentEditable) return;
@@ -51,10 +49,10 @@ export function ChatSearchDialog() {
     <CommandDialog
       open={isOpen}
       onOpenChange={setOpen}
-      className="shadow-border corner-squircle w-[635px] max-w-[calc(100%-2rem)] gap-0 p-0 sm:max-w-[635px]"
+      className="corner-squircle top-[25%] w-[635px] max-w-[calc(100%-2rem)] gap-0 p-0 sm:max-w-[635px] border-0 ring-0 shadow-[0_10px_34px_rgba(0,0,0,0.12)] dark:border dark:border-border dark:ring-0 dark:shadow-none"
       overlayClassName="bg-transparent"
     >
-      <Command className="rounded-none p-0">
+      <Command className="rounded-4xl p-0">
         <div className="flex items-center gap-3 border-b border-border/40 px-4 py-3">
           <HugeiconsIcon
             icon={SearchIcon}
@@ -74,7 +72,7 @@ export function ChatSearchDialog() {
             <HugeiconsIcon icon={Cancel01Icon} strokeWidth={2} className="size-4" />
           </button>
         </div>
-        <CommandList className="max-h-[420px] p-1">
+        <CommandList className="cmd-native-scrollbar max-h-[420px] p-1">
           <CommandEmpty className="py-6 text-center text-xs text-muted-foreground">
             {loading
               ? "Loading…"
@@ -92,12 +90,18 @@ export function ChatSearchDialog() {
                     to: "/chat",
                     search:
                       item.type === "single"
-                        ? { thread: item.id }
-                        : { compare: item.id },
+                        ? {
+                            thread: item.id,
+                            ...(item.projectId ? { project: item.projectId } : {}),
+                          }
+                        : {
+                            compare: item.id,
+                            ...(item.projectId ? { project: item.projectId } : {}),
+                          },
                   });
                   close();
                 }}
-                className="relative flex cursor-default select-none items-center gap-3 rounded-lg px-3 py-2.5 text-sm outline-hidden data-selected:bg-muted data-selected:text-foreground"
+                className="relative flex cursor-pointer select-none items-center gap-3 rounded-lg px-3 py-2.5 text-sm outline-hidden data-selected:bg-muted data-selected:text-foreground"
               >
                 <HugeiconsIcon
                   icon={Message01Icon}

studio/frontend/src/features/chat/components/project-switcher.tsx

@@ -0,0 +1,120 @@
+// 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 {
+  DropdownMenu,
+  DropdownMenuContent,
+  DropdownMenuItem,
+  DropdownMenuSeparator,
+  DropdownMenuTrigger,
+} from "@/components/ui/dropdown-menu";
+import {
+  ArrowDown01Icon,
+  Folder01Icon,
+  Tick02Icon,
+} from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import type { ReactElement } from "react";
+import type { ProjectRecord } from "../types";
+
+export function ProjectSwitcher({
+  currentProject,
+  projects,
+  isLoading,
+  onSelectProject,
+  onViewAllProjects,
+}: {
+  currentProject: ProjectRecord | null;
+  projects: ProjectRecord[];
+  isLoading: boolean;
+  onSelectProject: (projectId: string) => void;
+  onViewAllProjects: () => void;
+}): ReactElement {
+  const showLoadingRow = isLoading && projects.length === 0;
+  const showEmptyRow = !isLoading && projects.length === 0;
+  const label = currentProject?.name ?? (isLoading ? "Project" : "Projects");
+
+  return (
+    <DropdownMenu>
+      <DropdownMenuTrigger asChild={true}>
+        <button
+          type="button"
+          aria-label={
+            currentProject
+              ? `Project: ${currentProject.name}. Switch project`
+              : isLoading
+                ? "Loading project"
+                : "Pick a project"
+          }
+          className="-mx-1 flex h-[34px] shrink-0 items-center gap-2 rounded-[10px] px-1.5 transition-colors hover:bg-[#ececec] focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring dark:hover:bg-[#2d2e32]"
+        >
+          <HugeiconsIcon
+            icon={Folder01Icon}
+            strokeWidth={1.75}
+            className="size-icon shrink-0 text-foreground/70"
+          />
+          <span className="flex min-w-0 flex-1 items-baseline">
+            <span className="min-w-0 flex max-w-[150px] flex-1 items-baseline truncate font-heading text-[16px] font-medium leading-tight text-black dark:text-white">
+              {label}
+            </span>
+          </span>
+          <span className="flex size-4 shrink-0 items-center justify-center">
+            <HugeiconsIcon
+              icon={ArrowDown01Icon}
+              strokeWidth={1.75}
+              className="relative top-0.5 size-3.5 text-muted-foreground"
+              aria-hidden={true}
+            />
+          </span>
+        </button>
+      </DropdownMenuTrigger>
+      <DropdownMenuContent
+        side="bottom"
+        align="start"
+        sideOffset={6}
+        className="app-user-menu menu-soft-surface ring-0 min-w-56 max-w-72 max-h-72 py-2 font-heading rounded-[14px] border-0"
+      >
+        {showLoadingRow ? (
+          <DropdownMenuItem disabled={true} className="text-muted-foreground">
+            Loading…
+          </DropdownMenuItem>
+        ) : null}
+        {showEmptyRow ? (
+          <DropdownMenuItem disabled={true} className="text-muted-foreground">
+            No projects yet
+          </DropdownMenuItem>
+        ) : null}
+        {projects.map((project) => {
+          const isActive = currentProject?.id === project.id;
+          return (
+            <DropdownMenuItem
+              key={project.id}
+              onSelect={() => onSelectProject(project.id)}
+              className="justify-between"
+            >
+              <span className="flex min-w-0 items-center gap-2">
+                <HugeiconsIcon
+                  icon={Folder01Icon}
+                  strokeWidth={1.75}
+                  className="size-icon shrink-0 text-foreground/70"
+                />
+                <span className="truncate">{project.name}</span>
+              </span>
+              {isActive ? (
+                <HugeiconsIcon
+                  icon={Tick02Icon}
+                  strokeWidth={2}
+                  className="size-icon shrink-0 text-foreground/80"
+                />
+              ) : null}
+            </DropdownMenuItem>
+          );
+        })}
+        <DropdownMenuSeparator />
+        <DropdownMenuItem onSelect={onViewAllProjects}>
+          View all projects
+        </DropdownMenuItem>
+      </DropdownMenuContent>
+    </DropdownMenu>
+  );
+}

studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts

@@ -1123,15 +1123,15 @@ export function useChatModelRuntime() {
     ],
   );
 
-  const ejectModel = useCallback(async () => {
+  const ejectModel = useCallback(async (): Promise<boolean> => {
     if (!params.checkpoint) {
-      return;
+      return false;
     }
     setModelsError(null);
     if (isExternalModelId(params.checkpoint)) {
       clearCheckpoint();
       await refresh();
-      return;
+      return true;
     }
     try {
       async function performUnload(): Promise<void> {
@@ -1140,17 +1140,21 @@ export function useChatModelRuntime() {
         await refresh();
       }
 
-      await toast.promise(performUnload(), {
+      const unloadPromise = performUnload();
+      toast.promise(unloadPromise, {
         loading: "Unloading model",
         success: { message: "Model unloaded", duration: 1200 },
         error: (err) =>
           err instanceof Error ? err.message : "Failed to unload model",
         description: "Releases VRAM and resets inference state.",
       });
+      await unloadPromise;
+      return true;
     } catch (error) {
       const message =
         error instanceof Error ? error.message : "Failed to unload model";
       setModelsError(message);
+      return false;
     }
   }, [clearCheckpoint, params.checkpoint, refresh, setModelsError]);
 

studio/frontend/src/features/chat/hooks/use-chat-projects.ts

@@ -0,0 +1,97 @@
+// 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 { useEffect, useState } from "react";
+import { CHAT_HISTORY_UPDATED_EVENT } from "../api/chat-api";
+import type { ProjectRecord } from "../types";
+import {
+  createStoredChatProject,
+  deleteStoredChatProject,
+  isExpectedBackgroundChatStorageError,
+  listStoredChatProjects,
+  moveStoredChatItemToProject,
+  updateStoredChatProject,
+} from "../utils/chat-history-storage";
+import type { SidebarItem } from "./use-chat-sidebar-items";
+
+let cachedProjects: ProjectRecord[] = [];
+
+export function useChatProjects(): {
+  projects: ProjectRecord[];
+  isLoading: boolean;
+  hasLoaded: boolean;
+} {
+  const [projects, setProjects] = useState<ProjectRecord[]>(cachedProjects);
+  const [isLoading, setIsLoading] = useState(cachedProjects.length === 0);
+  const [hasLoaded, setHasLoaded] = useState(cachedProjects.length > 0);
+
+  useEffect(() => {
+    let cancelled = false;
+
+    async function load() {
+      if (!cancelled) setIsLoading(true);
+      try {
+        const next = await listStoredChatProjects({ includeArchived: false });
+        cachedProjects = next;
+        if (!cancelled) setProjects(next);
+      } catch (error) {
+        if (isExpectedBackgroundChatStorageError(error)) {
+          return;
+        }
+        if (!cancelled) throw error;
+      } finally {
+        if (!cancelled) {
+          setHasLoaded(true);
+          setIsLoading(false);
+        }
+      }
+    }
+
+    const onHistoryUpdated = () => {
+      void load();
+    };
+
+    void load();
+    window.addEventListener(CHAT_HISTORY_UPDATED_EVENT, onHistoryUpdated);
+    return () => {
+      cancelled = true;
+      window.removeEventListener(CHAT_HISTORY_UPDATED_EVENT, onHistoryUpdated);
+    };
+  }, []);
+
+  return { projects, isLoading, hasLoaded };
+}
+
+export async function createChatProject(name: string): Promise<ProjectRecord> {
+  return createStoredChatProject(name);
+}
+
+export async function renameChatProject(
+  projectId: string,
+  name: string,
+): Promise<void> {
+  const trimmed = name.trim();
+  if (!trimmed) throw new Error("Project name is required.");
+  await updateStoredChatProject(projectId, { name: trimmed });
+}
+
+export async function updateChatProjectInstructions(
+  projectId: string,
+  instructions: string,
+): Promise<void> {
+  await updateStoredChatProject(projectId, { instructions: instructions.trim() });
+}
+
+export async function deleteChatProject(
+  projectId: string,
+  args: { deleteFiles?: boolean } = {},
+): Promise<void> {
+  await deleteStoredChatProject(projectId, args);
+}
+
+export async function moveChatItemToProject(
+  item: SidebarItem,
+  projectId: string | null,
+): Promise<void> {
+  await moveStoredChatItemToProject(item, projectId);
+}

studio/frontend/src/features/chat/hooks/use-chat-search-index.ts

@@ -2,7 +2,7 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { useEffect, useRef, useState } from "react";
-import { CHAT_HISTORY_UPDATED_EVENT } from "../api/chat-api";
+import { batchListChatMessages, CHAT_HISTORY_UPDATED_EVENT } from "../api/chat-api";
 import type { MessageRecord } from "../types";
 import {
   listStoredChatMessages,
@@ -15,6 +15,7 @@ export interface ChatSearchItem {
   title: string;
   preview: string;
   createdAt: number;
+  projectId?: string | null;
 }
 
 const THREAD_LIMIT = 200;
@@ -68,6 +69,7 @@ async function buildIndex(): Promise<ChatSearchItem[]> {
           id: t.pairId,
           title: t.title,
           createdAt: t.createdAt,
+          projectId: t.projectId ?? null,
         },
         threadIds: [t.id],
       });
@@ -78,6 +80,7 @@ async function buildIndex(): Promise<ChatSearchItem[]> {
           id: t.id,
           title: t.title,
           createdAt: t.createdAt,
+          projectId: t.projectId ?? null,
         },
         threadIds: [t.id],
       });
@@ -87,26 +90,34 @@ async function buildIndex(): Promise<ChatSearchItem[]> {
   const allThreadIds = Array.from(itemThreadIds.values()).flatMap(
     (e) => e.threadIds,
   );
-  const storedMessagesByThread = await Promise.all(
-    allThreadIds.map(async (threadId) => ({
-      threadId,
-      messages: await listStoredChatMessages(threadId),
-    })),
+  let messagesByThread = await batchListChatMessages(allThreadIds).catch(
+    () => new Map<string, MessageRecord[]>(),
   );
-  const messages = storedMessagesByThread.flatMap((entry) => entry.messages);
 
-  const byThreadId = new Map<string, MessageRecord[]>();
-  for (const m of messages) {
-    const arr = byThreadId.get(m.threadId);
-    if (arr) arr.push(m);
-    else byThreadId.set(m.threadId, [m]);
+  // Legacy-only chats can exist before server-side history import finishes.
+  // Fill just the missing ids from the legacy-aware path instead of issuing
+  // one request per thread up front.
+  const missingThreadIds = allThreadIds.filter(
+    (threadId) => !messagesByThread.has(threadId),
+  );
+  if (missingThreadIds.length > 0) {
+    const legacyEntries = await Promise.all(
+      missingThreadIds.map(async (threadId) => [
+        threadId,
+        await listStoredChatMessages(threadId).catch(() => []),
+      ] as const),
+    );
+    messagesByThread = new Map(messagesByThread);
+    for (const [threadId, messages] of legacyEntries) {
+      messagesByThread.set(threadId, messages);
+    }
   }
 
   const results: ChatSearchItem[] = [];
   for (const { item, threadIds } of itemThreadIds.values()) {
     const merged: MessageRecord[] = [];
     for (const tid of threadIds) {
-      const arr = byThreadId.get(tid);
+      const arr = messagesByThread.get(tid);
       if (arr) merged.push(...arr);
     }
     if (merged.length === 0) {
@@ -141,6 +152,7 @@ export function useChatSearchIndex(enabled: boolean): {
     if (!enabled) {
       // Clear stale results so the next open doesn't flash old items.
       setItems([]);
+      setLoading(false);
       return;
     }
     let cancelled = false;

studio/frontend/src/features/chat/hooks/use-chat-sidebar-items.ts

@@ -4,6 +4,7 @@
 import { useEffect, useState } from "react";
 import { CHAT_HISTORY_UPDATED_EVENT } from "../api/chat-api";
 import { useChatRuntimeStore } from "../stores/chat-runtime-store";
+import { useChatArtifactsStore } from "../artifacts/store";
 import type { ThreadRecord } from "../types";
 import {
   deleteStoredChatThreads,
@@ -23,6 +24,7 @@ export interface SidebarItem {
   id: string;
   title: string;
   createdAt: number;
+  projectId?: string | null;
 }
 
 export function groupThreads(threads: ThreadRecord[]): SidebarItem[] {
@@ -43,6 +45,7 @@ export function groupThreads(threads: ThreadRecord[]): SidebarItem[] {
         id: t.pairId,
         title: t.title,
         createdAt: t.createdAt,
+        projectId: t.projectId ?? null,
       });
     } else if (!t.pairId) {
       items.push({
@@ -50,6 +53,7 @@ export function groupThreads(threads: ThreadRecord[]): SidebarItem[] {
         id: t.id,
         title: t.title,
         createdAt: t.createdAt,
+        projectId: t.projectId ?? null,
       });
     }
   }
@@ -62,18 +66,32 @@ export function groupThreads(threads: ThreadRecord[]): SidebarItem[] {
 // discards stale responses.
 const SIDEBAR_REFRESH_DEBOUNCE_MS = 300;
 
-export function useChatSidebarItems() {
+export function useChatSidebarItems(options?: {
+  projectId?: string | null;
+  enabled?: boolean;
+  requireMessages?: boolean;
+}) {
   const [allThreads, setAllThreads] = useState<ThreadRecord[]>([]);
+  const enabled = options?.enabled ?? true;
+  const requireMessages = options?.requireMessages ?? true;
 
   useEffect(() => {
+    if (!enabled) {
+      return;
+    }
+
     let cancelled = false;
     let pendingTimer: ReturnType<typeof setTimeout> | null = null;
     let requestSeq = 0;
 
     async function doLoad(seq: number) {
       try {
-        const threads = await listStoredChatThreadsWithMessages({
+        const listThreads = requireMessages
+          ? listStoredChatThreadsWithMessages
+          : listStoredChatThreads;
+        const threads = await listThreads({
           includeArchived: false,
+          projectId: options?.projectId,
         });
         // Discard the response if a newer request was scheduled while we
         // were in flight, or if the effect was torn down.
@@ -106,7 +124,7 @@ export function useChatSidebarItems() {
       if (pendingTimer !== null) clearTimeout(pendingTimer);
       window.removeEventListener(CHAT_HISTORY_UPDATED_EVENT, load);
     };
-  }, []);
+  }, [enabled, options?.projectId, requireMessages]);
 
   const items = groupThreads(allThreads ?? []);
   const canCompare = useChatRuntimeStore((s) => Boolean(s.params.checkpoint));
@@ -157,6 +175,10 @@ export async function deleteChatItem(
   // generating against a thread that no longer exists.
   for (const id of threadIds) cancelIfRunning(id);
 
+  const artifactStore = useChatArtifactsStore.getState();
+  for (const id of threadIds) artifactStore.clearArtifactsForThread(id);
+  artifactStore.clearOrphanedArtifacts();
+
   // Optimistic tombstone: hide immediately; roll back on backend error.
   markChatThreadsDeleted(threadIds);
   notifyChatHistoryUpdated();

studio/frontend/src/features/chat/index.ts

@@ -21,7 +21,13 @@ export { useChatSearchStore } from "./stores/chat-search-store";
 export { useChatModelRuntime } from "./hooks/use-chat-model-runtime";
 export { ChatSearchDialog } from "./components/chat-search-dialog";
 export { setTrainingCompareHandoff } from "./lib/training-compare-handoff";
+export type { ProjectRecord } from "./types";
 export { clearAllChats, countAllChats } from "./utils/clear-all-chats";
+export { ArtifactCard } from "./artifacts/artifact-card";
+export {
+  useChatArtifactsStore,
+  useSelectedChatArtifact,
+} from "./artifacts/store";
 export { downloadChatExport } from "./utils/export-chat-history";
 export {
   deleteChatItem,
@@ -29,3 +35,11 @@ export {
   useChatSidebarItems,
   type SidebarItem,
 } from "./hooks/use-chat-sidebar-items";
+export {
+  createChatProject,
+  deleteChatProject,
+  moveChatItemToProject,
+  renameChatProject,
+  updateChatProjectInstructions,
+  useChatProjects,
+} from "./hooks/use-chat-projects";

studio/frontend/src/features/chat/mcp-composer-button.tsx

@@ -0,0 +1,355 @@
+// 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 {
+  Cancel01Icon,
+  McpServerIcon,
+  Tick02Icon,
+} from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { useCallback, useEffect, useState } from "react";
+import { toast } from "sonner";
+
+import {
+  DropdownMenu,
+  DropdownMenuContent,
+  DropdownMenuItem,
+  DropdownMenuLabel,
+  DropdownMenuSeparator,
+  DropdownMenuTrigger,
+} from "@/components/ui/dropdown-menu";
+import {
+  Tooltip,
+  TooltipContent,
+  TooltipTrigger,
+} from "@/components/ui/tooltip";
+import { cn } from "@/lib/utils";
+
+import {
+  type McpServerConfig,
+  createMcpServer,
+  listMcpServers,
+  updateMcpServer,
+} from "./api/mcp-servers-api";
+import { ChatMcpServersDialog } from "./chat-mcp-servers-dialog";
+import { useChatRuntimeStore } from "./stores/chat-runtime-store";
+
+type McpPreset = {
+  id: string;
+  displayName: string; // stored row name
+  url: string;
+  label?: string; // dropdown text, if different from displayName
+  hint?: string; // shown when the row is highlighted
+  disablesWebSearch?: boolean; // turn the built-in Search pill off when enabled
+};
+
+// Keyless remote MCP presets (rate-limited free tiers, no API key).
+// Hugging Face runs anonymously; add a token via "Add custom MCP".
+const MCP_PRESETS: readonly McpPreset[] = [
+  {
+    id: "context7",
+    displayName: "Context7",
+    url: "https://mcp.context7.com/mcp",
+    label: "Context7 (Realtime Docs)",
+  },
+  {
+    id: "exa",
+    displayName: "Exa",
+    url: "https://mcp.exa.ai/mcp",
+    label: "Exa (Semantic Search)",
+    hint: "Enabling Exa will disable default search",
+    disablesWebSearch: true,
+  },
+  {
+    id: "huggingface",
+    displayName: "Hugging Face",
+    url: "https://huggingface.co/mcp",
+  },
+] as const;
+
+// mcp_servers has no UNIQUE(url); dedupe by normalized URL so a preset
+// toggle reuses its row instead of creating duplicates.
+function normalizeMcpUrl(url: string): string {
+  return (url || "").trim().toLowerCase().replace(/\/+$/, "");
+}
+
+// Static, so it is not rebuilt on every render.
+const PRESET_URLS = new Set(MCP_PRESETS.map((p) => normalizeMcpUrl(p.url)));
+
+export function McpComposerButton() {
+  const modelLoaded = useChatRuntimeStore(
+    (s) => !!s.params.checkpoint && !s.modelLoading,
+  );
+  const supportsTools = useChatRuntimeStore((s) => s.supportsTools);
+  const mcpEnabledForChat = useChatRuntimeStore((s) => s.mcpEnabledForChat);
+  const setMcpEnabledForChat = useChatRuntimeStore(
+    (s) => s.setMcpEnabledForChat,
+  );
+  const setToolsEnabled = useChatRuntimeStore((s) => s.setToolsEnabled);
+
+  const [servers, setServers] = useState<McpServerConfig[]>([]);
+  const [dialogOpen, setDialogOpen] = useState(false);
+  const [menuOpen, setMenuOpen] = useState(false);
+  const [pendingUrl, setPendingUrl] = useState<string | null>(null);
+  const [hintKey, setHintKey] = useState<string | null>(null);
+
+  // mcp_enabled only applies on the local tool-capable send path; grey out otherwise.
+  const usable = modelLoaded && supportsTools;
+
+  // Keep the per-chat flag in step with whether any server is enabled. Reads the
+  // store directly so the callback stays stable (no refetch loop on mount).
+  const reconcileFlag = useCallback(
+    (rows: McpServerConfig[]) => {
+      const anyEnabled = rows.some((s) => s.is_enabled);
+      const current = useChatRuntimeStore.getState().mcpEnabledForChat;
+      if (anyEnabled && !current) setMcpEnabledForChat(true);
+      else if (!anyEnabled && current) setMcpEnabledForChat(false);
+    },
+    [setMcpEnabledForChat],
+  );
+
+  const refresh = useCallback(async () => {
+    try {
+      const rows = await listMcpServers();
+      setServers(rows);
+      reconcileFlag(rows);
+    } catch {
+      // Keep prior state if the list call fails.
+    }
+  }, [reconcileFlag]);
+
+  // Initial load reconciles the pill with already-enabled servers (also on open).
+  useEffect(() => {
+    void refresh();
+  }, [refresh]);
+
+  const enabledUrls = new Set(
+    servers.filter((s) => s.is_enabled).map((s) => normalizeMcpUrl(s.url)),
+  );
+  // Non-preset servers, shown below the presets so they stay toggleable.
+  const customServers = servers.filter(
+    (s) => !PRESET_URLS.has(normalizeMcpUrl(s.url)),
+  );
+  const enabledCount = servers.filter((s) => s.is_enabled).length;
+  const active = mcpEnabledForChat && enabledCount > 0;
+
+  async function toggleServer(args: {
+    url: string;
+    displayName: string;
+    checked: boolean;
+    existing?: McpServerConfig;
+    disablesWebSearch?: boolean;
+  }) {
+    const norm = normalizeMcpUrl(args.url);
+    if (pendingUrl === norm) return; // guard rapid double-clicks
+    setPendingUrl(norm);
+    try {
+      if (args.checked) {
+        // Reuse the already-loaded row, else create one.
+        if (args.existing) {
+          if (!args.existing.is_enabled) {
+            await updateMcpServer(args.existing.id, { isEnabled: true });
+          }
+        } else {
+          await createMcpServer({
+            displayName: args.displayName,
+            url: args.url,
+            isEnabled: true,
+          });
+        }
+        setMcpEnabledForChat(true);
+        // Exa is a search server; turn off the built-in Web Search to avoid overlap.
+        if (args.disablesWebSearch) setToolsEnabled(false);
+      } else if (args.existing) {
+        await updateMcpServer(args.existing.id, { isEnabled: false });
+      }
+      await refresh();
+    } catch (err) {
+      toast.error("Failed to update MCP server", {
+        description: err instanceof Error ? err.message : String(err),
+      });
+    } finally {
+      setPendingUrl(null);
+    }
+  }
+
+  // One dropdown row. Enabled rows get a green underlay and a tick that
+  // becomes an X on hover so a click removes them. A hint shows as a tooltip
+  // driven by row hover; the tooltip anchor is pointer-events-none so the whole
+  // row stays clickable (a Radix TooltipTrigger would swallow the select).
+  const renderRow = (opts: {
+    key: string;
+    label: string;
+    url: string;
+    displayName: string;
+    enabled: boolean;
+    existing?: McpServerConfig;
+    hint?: string;
+    disablesWebSearch?: boolean;
+  }) => (
+    <DropdownMenuItem
+      key={opts.key}
+      disabled={pendingUrl === normalizeMcpUrl(opts.url)}
+      onSelect={(e) => {
+        e.preventDefault();
+        void toggleServer({
+          url: opts.url,
+          displayName: opts.displayName,
+          checked: !opts.enabled,
+          existing: opts.existing,
+          disablesWebSearch: opts.disablesWebSearch,
+        });
+      }}
+      onPointerEnter={opts.hint ? () => setHintKey(opts.key) : undefined}
+      onPointerLeave={
+        opts.hint
+          ? () => setHintKey((k) => (k === opts.key ? null : k))
+          : undefined
+      }
+      className={cn(
+        "group/mcp relative flex items-center justify-between gap-2",
+        opts.enabled &&
+          "bg-emerald-500/10 data-[highlighted]:bg-emerald-500/20",
+      )}
+    >
+      <span className="truncate">{opts.label}</span>
+      {opts.enabled ? (
+        <span className="flex size-4 shrink-0 items-center justify-center text-emerald-600 dark:text-emerald-400">
+          <HugeiconsIcon
+            icon={Tick02Icon}
+            className="size-4 group-data-[highlighted]/mcp:hidden"
+            strokeWidth={2}
+          />
+          <HugeiconsIcon
+            icon={Cancel01Icon}
+            className="hidden size-4 text-foreground group-data-[highlighted]/mcp:block"
+            strokeWidth={2}
+          />
+        </span>
+      ) : null}
+      {opts.hint ? (
+        <Tooltip open={hintKey === opts.key}>
+          <TooltipTrigger asChild={true}>
+            <span
+              aria-hidden={true}
+              className="pointer-events-none absolute inset-y-0 right-0 w-0"
+            />
+          </TooltipTrigger>
+          <TooltipContent side="right">{opts.hint}</TooltipContent>
+        </Tooltip>
+      ) : null}
+    </DropdownMenuItem>
+  );
+
+  return (
+    <>
+      {usable ? (
+        <DropdownMenu
+          open={menuOpen}
+          onOpenChange={(open) => {
+            setMenuOpen(open);
+            if (open) void refresh();
+          }}
+        >
+          <DropdownMenuTrigger asChild={true}>
+            <button
+              type="button"
+              className="composer-pill-btn"
+              data-active={active ? "true" : "false"}
+              aria-label="MCP servers"
+            >
+              <HugeiconsIcon
+                icon={McpServerIcon}
+                className="size-3.5"
+                strokeWidth={2}
+              />
+              <span>MCP</span>
+            </button>
+          </DropdownMenuTrigger>
+          <DropdownMenuContent align="start" className="w-64">
+            <div className="flex items-center justify-between pr-1">
+              <DropdownMenuLabel>MCP Servers</DropdownMenuLabel>
+              <button
+                type="button"
+                aria-label="Close"
+                onClick={() => setMenuOpen(false)}
+                className="rounded-md p-1 text-muted-foreground hover:bg-accent hover:text-foreground"
+              >
+                <HugeiconsIcon
+                  icon={Cancel01Icon}
+                  className="size-3.5"
+                  strokeWidth={2}
+                />
+              </button>
+            </div>
+            {MCP_PRESETS.map((preset) => {
+              const norm = normalizeMcpUrl(preset.url);
+              return renderRow({
+                key: preset.id,
+                label: preset.label ?? preset.displayName,
+                url: preset.url,
+                displayName: preset.displayName,
+                enabled: enabledUrls.has(norm),
+                existing: servers.find((s) => normalizeMcpUrl(s.url) === norm),
+                hint: preset.hint,
+                disablesWebSearch: preset.disablesWebSearch,
+              });
+            })}
+            {customServers.length > 0 ? <DropdownMenuSeparator /> : null}
+            {customServers.map((server) =>
+              renderRow({
+                key: server.id,
+                label: server.display_name,
+                url: server.url,
+                displayName: server.display_name,
+                enabled: server.is_enabled,
+                existing: server,
+              }),
+            )}
+            <DropdownMenuSeparator />
+            <DropdownMenuItem
+              onSelect={() => {
+                setMenuOpen(false);
+                setDialogOpen(true);
+              }}
+            >
+              Add custom MCP
+            </DropdownMenuItem>
+          </DropdownMenuContent>
+        </DropdownMenu>
+      ) : (
+        <Tooltip>
+          <TooltipTrigger asChild={true}>
+            {/* Not disabled, so the tooltip still fires on hover. */}
+            <button
+              type="button"
+              className="composer-pill-btn cursor-not-allowed opacity-40"
+              data-active="false"
+              aria-disabled={true}
+              aria-label="MCP servers"
+            >
+              <HugeiconsIcon
+                icon={McpServerIcon}
+                className="size-3.5"
+                strokeWidth={2}
+              />
+              <span>MCP</span>
+            </button>
+          </TooltipTrigger>
+          <TooltipContent>
+            MCP works with local tool-capable models
+          </TooltipContent>
+        </Tooltip>
+      )}
+      <ChatMcpServersDialog
+        open={dialogOpen}
+        onOpenChange={(next) => {
+          setDialogOpen(next);
+          // Resync after managing servers.
+          if (!next) void refresh();
+        }}
+        openToCreate={true}
+      />
+    </>
+  );
+}

studio/frontend/src/features/chat/projects-page.tsx

@@ -0,0 +1,401 @@
+// 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 { Button } from "@/components/ui/button";
+import {
+  Dialog,
+  DialogContent,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+} from "@/components/ui/dialog";
+import {
+  DropdownMenu,
+  DropdownMenuContent,
+  DropdownMenuItem,
+  DropdownMenuTrigger,
+} from "@/components/ui/dropdown-menu";
+import { Input } from "@/components/ui/input";
+import { Skeleton } from "@/components/ui/skeleton";
+import {
+  Select,
+  SelectContent,
+  SelectItem,
+  SelectTrigger,
+  SelectValue,
+} from "@/components/ui/select";
+import { toast } from "@/lib/toast";
+import {
+  createChatProject,
+  deleteChatProject,
+  renameChatProject,
+  useChatProjects,
+  useChatRuntimeStore,
+  type ProjectRecord,
+} from "@/features/chat";
+import {
+  Delete02Icon,
+  Edit03Icon,
+  FolderAddIcon,
+  Search01Icon,
+} from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
+import { MoreHorizontalIcon } from "lucide-react";
+import { useNavigate } from "@tanstack/react-router";
+import { useMemo, useState } from "react";
+
+type SortMode = "activity" | "name";
+
+function formatUpdatedAgo(ts: number): string {
+  const diff = Date.now() - ts;
+  if (!Number.isFinite(diff) || diff < 0) return "just now";
+  const s = Math.floor(diff / 1000);
+  if (s < 60) return "just now";
+  const m = Math.floor(s / 60);
+  if (m < 60) return `${m} minute${m === 1 ? "" : "s"} ago`;
+  const h = Math.floor(m / 60);
+  if (h < 24) return `${h} hour${h === 1 ? "" : "s"} ago`;
+  const d = Math.floor(h / 24);
+  if (d < 30) return `${d} day${d === 1 ? "" : "s"} ago`;
+  const mo = Math.floor(d / 30);
+  if (mo < 12) return `${mo} month${mo === 1 ? "" : "s"} ago`;
+  const y = Math.floor(mo / 12);
+  return `${y} year${y === 1 ? "" : "s"} ago`;
+}
+
+export function ProjectsPage() {
+  const navigate = useNavigate();
+  const { projects, hasLoaded } = useChatProjects();
+
+  const [query, setQuery] = useState("");
+  const [sortMode, setSortMode] = useState<SortMode>("activity");
+
+  const [creating, setCreating] = useState(false);
+  const [nameDraft, setNameDraft] = useState("");
+  const [renaming, setRenaming] = useState<ProjectRecord | null>(null);
+  const [renameDraft, setRenameDraft] = useState("");
+  const [deleting, setDeleting] = useState<ProjectRecord | null>(null);
+
+  const visibleProjects = useMemo(() => {
+    const trimmed = query.trim().toLowerCase();
+    const filtered = trimmed
+      ? projects.filter((p) => p.name.toLowerCase().includes(trimmed))
+      : projects.slice();
+    filtered.sort((a, b) =>
+      sortMode === "name"
+        ? a.name.localeCompare(b.name)
+        : b.updatedAt - a.updatedAt,
+    );
+    return filtered;
+  }, [projects, query, sortMode]);
+
+  function openProject(projectId: string) {
+    const runtime = useChatRuntimeStore.getState();
+    runtime.setActiveThreadId(null);
+    runtime.setActiveProjectId(projectId);
+    navigate({ to: "/chat", search: { project: projectId } });
+  }
+
+  async function commitCreate() {
+    const name = nameDraft.trim();
+    if (!name) return;
+    try {
+      const project = await createChatProject(name);
+      setCreating(false);
+      setNameDraft("");
+      openProject(project.id);
+    } catch (err) {
+      toast.error("Failed to create project", {
+        description: err instanceof Error ? err.message : undefined,
+      });
+    }
+  }
+
+  async function commitRename() {
+    const target = renaming;
+    const name = renameDraft.trim();
+    if (!target || !name || name === target.name) {
+      setRenaming(null);
+      return;
+    }
+    setRenaming(null);
+    try {
+      await renameChatProject(target.id, name);
+    } catch (err) {
+      toast.error("Failed to rename project", {
+        description: err instanceof Error ? err.message : undefined,
+      });
+    }
+  }
+
+  async function commitDelete() {
+    const target = deleting;
+    if (!target) return;
+    setDeleting(null);
+    try {
+      await deleteChatProject(target.id);
+    } catch (err) {
+      toast.error("Failed to delete project", {
+        description: err instanceof Error ? err.message : undefined,
+      });
+    }
+  }
+
+  return (
+    <main className="mx-auto w-full max-w-7xl px-4 py-8 font-heading sm:px-6">
+      <div className="flex flex-wrap items-center justify-between gap-4">
+        <h1 className="text-2xl font-semibold tracking-tight text-foreground">
+          Projects
+        </h1>
+        <div className="flex items-center gap-3">
+          <div className="flex items-center gap-2">
+            <span className="text-sm text-muted-foreground">Sort by</span>
+            <Select
+              value={sortMode}
+              onValueChange={(v) => setSortMode(v as SortMode)}
+            >
+              <SelectTrigger className="h-9 w-[130px]">
+                <SelectValue />
+              </SelectTrigger>
+              <SelectContent>
+                <SelectItem value="activity">Activity</SelectItem>
+                <SelectItem value="name">Name</SelectItem>
+              </SelectContent>
+            </Select>
+          </div>
+          <Button
+            onClick={() => {
+              setNameDraft("");
+              setCreating(true);
+            }}
+          >
+            New project
+          </Button>
+        </div>
+      </div>
+
+      <div className="relative mt-6">
+        <span className="pointer-events-none absolute left-3 top-1/2 -translate-y-1/2 text-muted-foreground">
+          <HugeiconsIcon icon={Search01Icon} strokeWidth={1.75} className="size-icon" />
+        </span>
+        <Input
+          value={query}
+          onChange={(e) => setQuery(e.target.value)}
+          placeholder="Search projects..."
+          className="h-11 pl-10"
+          aria-label="Search projects"
+        />
+      </div>
+
+      {!hasLoaded ? (
+        <div className="mt-6 grid grid-cols-1 gap-4 sm:grid-cols-2 lg:grid-cols-3">
+          {Array.from({ length: 6 }).map((_, index) => (
+            <div
+              key={index}
+              className="min-h-[160px] rounded-[14px] border border-border/70 bg-card p-5"
+            >
+              <Skeleton className="h-5 w-2/3 rounded-[6px]" />
+              <Skeleton className="mt-3 h-4 w-full rounded-[6px]" />
+              <Skeleton className="mt-2 h-4 w-4/5 rounded-[6px]" />
+              <Skeleton className="mt-12 h-3 w-24 rounded-[6px]" />
+            </div>
+          ))}
+        </div>
+      ) : visibleProjects.length === 0 ? (
+        <div className="mt-16 flex flex-col items-center justify-center gap-2 text-center text-muted-foreground">
+          <p className="text-sm">
+            {projects.length === 0
+              ? "No projects yet."
+              : "No projects match your search."}
+          </p>
+          {projects.length === 0 && (
+            <Button
+              variant="outline"
+              className="mt-2"
+              onClick={() => {
+                setNameDraft("");
+                setCreating(true);
+              }}
+            >
+              <HugeiconsIcon icon={FolderAddIcon} strokeWidth={1.75} className="size-icon" />
+              Create your first project
+            </Button>
+          )}
+        </div>
+      ) : (
+        <div className="mt-6 grid grid-cols-1 gap-4 sm:grid-cols-2 lg:grid-cols-3">
+          {visibleProjects.map((project) => (
+            <div
+              key={project.id}
+              role="button"
+              tabIndex={0}
+              onClick={() => openProject(project.id)}
+              onKeyDown={(e) => {
+                if (e.key === "Enter" || e.key === " ") {
+                  e.preventDefault();
+                  openProject(project.id);
+                }
+              }}
+              className="group/project-card relative flex min-h-[160px] cursor-pointer flex-col rounded-[14px] border border-border/70 bg-card p-5 text-left transition-colors hover:border-border hover:bg-accent/40 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
+            >
+              <div className="flex items-start justify-between gap-2">
+                <h2 className="truncate pr-2 text-[16px] font-semibold text-foreground">
+                  {project.name}
+                </h2>
+                <DropdownMenu>
+                  <DropdownMenuTrigger asChild>
+                    <button
+                      type="button"
+                      onClick={(e) => e.stopPropagation()}
+                      aria-label="Project options"
+                      className="-mr-1 -mt-1 inline-flex size-7 shrink-0 items-center justify-center rounded-[8px] text-muted-foreground opacity-0 transition-opacity hover:bg-accent hover:text-foreground focus-visible:opacity-100 group-hover/project-card:opacity-100"
+                    >
+                      <MoreHorizontalIcon strokeWidth={1.75} className="size-icon" />
+                    </button>
+                  </DropdownMenuTrigger>
+                  <DropdownMenuContent
+                    side="bottom"
+                    align="end"
+                    sideOffset={4}
+                    onClick={(e) => e.stopPropagation()}
+                    onKeyDown={(e) => e.stopPropagation()}
+                    className="app-user-menu menu-soft-surface menu-flat-destructive ring-0 w-44 py-2 font-heading rounded-[14px] border-0"
+                  >
+                    <DropdownMenuItem
+                      onSelect={() => {
+                        setRenameDraft(project.name);
+                        setRenaming(project);
+                      }}
+                    >
+                      <HugeiconsIcon icon={Edit03Icon} strokeWidth={1.75} className="size-icon" />
+                      <span>Rename</span>
+                    </DropdownMenuItem>
+                    <DropdownMenuItem
+                      variant="destructive"
+                      onSelect={() => setDeleting(project)}
+                    >
+                      <HugeiconsIcon icon={Delete02Icon} strokeWidth={1.75} className="size-icon" />
+                      <span>Delete</span>
+                    </DropdownMenuItem>
+                  </DropdownMenuContent>
+                </DropdownMenu>
+              </div>
+              {project.instructions ? (
+                <p className="mt-2 line-clamp-3 text-sm text-muted-foreground">
+                  {project.instructions}
+                </p>
+              ) : null}
+              <span className="mt-auto pt-4 text-xs text-muted-foreground">
+                Updated {formatUpdatedAgo(project.updatedAt)}
+              </span>
+            </div>
+          ))}
+        </div>
+      )}
+
+      {/* Create project */}
+      <Dialog
+        open={creating}
+        onOpenChange={(open) => {
+          if (!open) setCreating(false);
+        }}
+      >
+        <DialogContent className="corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-md">
+          <DialogHeader>
+            <DialogTitle>New project</DialogTitle>
+          </DialogHeader>
+          <Input
+            value={nameDraft}
+            onChange={(e) => setNameDraft(e.target.value)}
+            onKeyDown={(e) => {
+              if (e.key === "Enter") {
+                e.preventDefault();
+                void commitCreate();
+              }
+            }}
+            autoFocus
+            maxLength={120}
+            placeholder="Project name"
+            aria-label="Project name"
+            className="focus-visible:border-input focus-visible:ring-0"
+          />
+          <DialogFooter className="flex-wrap gap-2 sm:justify-end">
+            <Button type="button" variant="ghost" onClick={() => setCreating(false)}>
+              Cancel
+            </Button>
+            <Button type="button" onClick={() => void commitCreate()} disabled={!nameDraft.trim()}>
+              Create
+            </Button>
+          </DialogFooter>
+        </DialogContent>
+      </Dialog>
+
+      {/* Rename project */}
+      <Dialog
+        open={renaming !== null}
+        onOpenChange={(open) => {
+          if (!open) setRenaming(null);
+        }}
+      >
+        <DialogContent className="corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-md">
+          <DialogHeader>
+            <DialogTitle>Rename project</DialogTitle>
+          </DialogHeader>
+          <Input
+            value={renameDraft}
+            onChange={(e) => setRenameDraft(e.target.value)}
+            onKeyDown={(e) => {
+              if (e.key === "Enter") {
+                e.preventDefault();
+                void commitRename();
+              }
+            }}
+            autoFocus
+            maxLength={120}
+            placeholder="Project name"
+            aria-label="Project name"
+            className="focus-visible:border-input focus-visible:ring-0"
+          />
+          <DialogFooter className="flex-wrap gap-2 sm:justify-end">
+            <Button type="button" variant="ghost" onClick={() => setRenaming(null)}>
+              Cancel
+            </Button>
+            <Button
+              type="button"
+              onClick={() => void commitRename()}
+              disabled={!renameDraft.trim() || renameDraft.trim() === renaming?.name}
+            >
+              Save
+            </Button>
+          </DialogFooter>
+        </DialogContent>
+      </Dialog>
+
+      {/* Delete project */}
+      <Dialog
+        open={deleting !== null}
+        onOpenChange={(open) => {
+          if (!open) setDeleting(null);
+        }}
+      >
+        <DialogContent className="menu-flat-destructive corner-squircle border border-border/60 bg-background/98 shadow-none sm:max-w-md">
+          <DialogHeader>
+            <DialogTitle>Delete project</DialogTitle>
+          </DialogHeader>
+          <p className="text-sm text-muted-foreground">
+            Are you sure you want to delete <em>{deleting?.name}</em>? Chats in this
+            project will be moved back to Recents.
+          </p>
+          <DialogFooter className="flex-wrap gap-2 sm:justify-end">
+            <Button type="button" variant="ghost" onClick={() => setDeleting(null)}>
+              Cancel
+            </Button>
+            <Button type="button" variant="destructive" onClick={() => void commitDelete()}>
+              Delete
+            </Button>
+          </DialogFooter>
+        </DialogContent>
+      </Dialog>
+    </main>
+  );
+}

studio/frontend/src/features/chat/runtime-provider.tsx

@@ -66,6 +66,7 @@ import { syncExportedRepositoryToBackend } from "./utils/delete-thread-message";
 import { getImageInputUnavailableReason } from "./utils/image-input-support";
 
 const pendingHistoryAppendByMessageId = new Map<string, Promise<void>>();
+const pendingRunStartReadyByMessageId = new Map<string, Promise<void>>();
 
 type TitleResponse = {
   choices?: Array<{
@@ -534,10 +535,12 @@ export async function ensureThreadRecord({
   threadId,
   modelType,
   pairId,
+  projectId,
 }: {
   threadId: string;
   modelType: ModelType;
   pairId?: string;
+  projectId?: string | null;
 }): Promise<void> {
   if (isChatThreadDeleted(threadId)) {
     return;
@@ -556,6 +559,7 @@ export async function ensureThreadRecord({
     modelType,
     modelId: currentModelId,
     pairId,
+    projectId: projectId ?? null,
     archived: false,
     createdAt: Date.now(),
   };
@@ -579,6 +583,8 @@ export async function ensureThreadRecord({
 function createStudioDbAdapter(
   modelType: ModelType,
   pairId?: string,
+  projectId?: string | null,
+  listThreads = true,
 ): unstable_RemoteThreadListAdapter {
   return {
     async fetch(remoteId: string) {
@@ -594,9 +600,16 @@ function createStudioDbAdapter(
     },
 
     async list() {
+      if (!listThreads) {
+        return { threads: [] };
+      }
       let threads: ThreadRecord[];
       try {
-        threads = await listStoredChatThreads({ modelType, pairId });
+        threads = await listStoredChatThreads({
+          modelType,
+          pairId,
+          ...(projectId !== undefined ? { projectId } : {}),
+        });
       } catch (error) {
         if (!isExpectedBackgroundChatStorageError(error)) {
           throw error;
@@ -615,7 +628,7 @@ function createStudioDbAdapter(
     },
 
     async initialize(threadId: string) {
-      await ensureThreadRecord({ threadId, modelType, pairId });
+      await ensureThreadRecord({ threadId, modelType, pairId, projectId });
       return { remoteId: threadId, externalId: undefined };
     },
 
@@ -696,7 +709,7 @@ function createStudioDbAdapter(
           const running = useChatRuntimeStore.getState().runningByThreadId;
           if (running[paired.id]) {
             setTimeout(() => {
-              void createStudioDbAdapter(modelType, pairId).generateTitle(
+              void createStudioDbAdapter(modelType, pairId, projectId).generateTitle(
                 remoteId,
                 messages,
               );
@@ -740,6 +753,22 @@ function trackHistoryAppend(
   return write;
 }
 
+function trackRunStartReady(
+  messageId: string,
+  ready: Promise<void>,
+): Promise<void> {
+  pendingRunStartReadyByMessageId.set(messageId, ready);
+  const cleanup = () => {
+    setTimeout(() => {
+      if (pendingRunStartReadyByMessageId.get(messageId) === ready) {
+        pendingRunStartReadyByMessageId.delete(messageId);
+      }
+    }, 30_000);
+  };
+  ready.then(cleanup, cleanup);
+  return ready;
+}
+
 async function waitForRunStartHistoryAppend(
   messages: Parameters<ChatModelAdapter["run"]>[0]["messages"],
 ): Promise<void> {
@@ -747,20 +776,22 @@ async function waitForRunStartHistoryAppend(
   if (!lastMessage || lastMessage.role !== "user") {
     return;
   }
-  const write = pendingHistoryAppendByMessageId.get(lastMessage.id);
-  if (!write) {
+  const ready =
+    pendingRunStartReadyByMessageId.get(lastMessage.id) ??
+    pendingHistoryAppendByMessageId.get(lastMessage.id);
+  if (!ready) {
     return;
   }
-  let didPersist = false;
+  let didBecomeReady = false;
   try {
-    await write;
-    didPersist = true;
+    await ready;
+    didBecomeReady = true;
   } finally {
     if (
-      didPersist &&
-      pendingHistoryAppendByMessageId.get(lastMessage.id) === write
+      didBecomeReady &&
+      pendingRunStartReadyByMessageId.get(lastMessage.id) === ready
     ) {
-      pendingHistoryAppendByMessageId.delete(lastMessage.id);
+      pendingRunStartReadyByMessageId.delete(lastMessage.id);
     }
   }
 }
@@ -783,7 +814,10 @@ function createPersistedRunAdapter(adapter: ChatModelAdapter): ChatModelAdapter
   };
 }
 
-function useStudioRuntimeAdapters(): StudioRuntimeAdapters {
+function useStudioRuntimeAdapters(
+  modelType: ModelType,
+  pairId?: string,
+): StudioRuntimeAdapters {
   const aui = useAui();
 
   const history = useMemo<ThreadHistoryAdapter>(
@@ -871,14 +905,22 @@ function useStudioRuntimeAdapters(): StudioRuntimeAdapters {
       },
 
       append({ parentId, message }: ExportedMessageRepositoryItem) {
+        const initializeThread = aui.threadListItem().initialize();
+        trackRunStartReady(message.id, initializeThread.then(() => undefined));
         const write = (async () => {
-          const { remoteId } = await aui.threadListItem().initialize();
+          const { remoteId } = await initializeThread;
           if (isChatThreadDeleted(remoteId)) {
             await deleteStoredChatThreads([remoteId]);
             return;
           }
           // Keep single-chat runtime state in sync once a new chat is first
           // persisted. Compare panes intentionally do not write global activeThreadId.
+          if (modelType === "base" && !pairId) {
+            const store = useChatRuntimeStore.getState();
+            if (store.activeThreadId !== remoteId) {
+              store.setActiveThreadId(remoteId);
+            }
+          }
           const thread = await getStoredChatThread(remoteId);
           if (thread) {
             await ensureStoredChatThread(remoteId, thread);
@@ -915,7 +957,7 @@ function useStudioRuntimeAdapters(): StudioRuntimeAdapters {
         return trackHistoryAppend(message.id, write);
       },
     }),
-    [aui],
+    [aui, modelType, pairId],
   );
 
   const dictation = useMemo(
@@ -947,8 +989,11 @@ function useStudioRuntimeAdapters(): StudioRuntimeAdapters {
 
 const chatAdapter = createOpenAIStreamAdapter();
 
-function useRuntimeHook(): ReturnType<typeof useLocalRuntime> {
-  const adapters = useStudioRuntimeAdapters();
+function useRuntimeHook(
+  modelType: ModelType,
+  pairId?: string,
+): ReturnType<typeof useLocalRuntime> {
+  const adapters = useStudioRuntimeAdapters(modelType, pairId);
   const persistedChatAdapter = useMemo(
     () => createPersistedRunAdapter(chatAdapter),
     [],
@@ -956,6 +1001,12 @@ function useRuntimeHook(): ReturnType<typeof useLocalRuntime> {
   return useLocalRuntime(persistedChatAdapter, { adapters });
 }
 
+function createRuntimeHook(modelType: ModelType, pairId?: string) {
+  return function useConfiguredRuntimeHook(): ReturnType<typeof useLocalRuntime> {
+    return useRuntimeHook(modelType, pairId);
+  };
+}
+
 function ThreadAutoSwitch({
   threadId,
   syncActiveThreadId = true,
@@ -1133,20 +1184,28 @@ export function ChatRuntimeProvider({
   children,
   modelType = "base",
   pairId,
+  projectId,
   initialThreadId,
   newThreadNonce,
   syncActiveThreadId = true,
+  listThreads = true,
 }: {
   children: ReactNode;
   modelType?: ModelType;
   pairId?: string;
+  projectId?: string | null;
   initialThreadId?: string;
   newThreadNonce?: string;
   syncActiveThreadId?: boolean;
+  listThreads?: boolean;
 }): ReactElement {
+  const runtimeHook = useMemo(
+    () => createRuntimeHook(modelType, pairId),
+    [modelType, pairId],
+  );
   const runtime = useRemoteThreadListRuntime({
-    runtimeHook: useRuntimeHook,
-    adapter: createStudioDbAdapter(modelType, pairId),
+    runtimeHook,
+    adapter: createStudioDbAdapter(modelType, pairId, projectId, listThreads),
   });
 
   const aui = useAui({});

studio/frontend/src/features/chat/shared-composer.tsx

@@ -43,8 +43,12 @@ import { useRagStore } from "@/features/rag/stores/rag-store";
 import { acquireIndexSlot, releaseIndexSlot } from "./utils/rag-index-queue";
 import { toast } from "@/lib/toast";
 import { loadModel, validateModel } from "./api/chat-api";
-import { parseExternalModelId, providerTypeSupportsVision } from "./external-providers";
+import {
+  parseExternalModelId,
+  providerTypeSupportsVision,
+} from "./external-providers";
 import { useExternalProvidersStore } from "./stores/external-providers-store";
+import { McpComposerButton } from "./mcp-composer-button";
 import {
   type ReasoningEffort,
   useChatRuntimeStore,
@@ -53,6 +57,7 @@ import {
   getExternalReasoningCapabilities,
   providerSupportsBuiltinCodeExecution,
   providerSupportsBuiltinImageGeneration,
+  providerSupportsBuiltinWebSearch,
   providerSupportsBuiltinWebFetch,
 } from "./provider-capabilities";
 import {
@@ -77,9 +82,9 @@ export type CompareMessagePart =
 export interface CompareHandle {
   append: (content: CompareMessagePart[]) => void;
   /** Append a user message without triggering generation. */
-  appendMessage: (content: CompareMessagePart[]) => void;
+  appendMessage: (content: CompareMessagePart[]) => Promise<string | null>;
   /** Trigger generation on the current thread (after appendMessage). */
-  startRun: () => void;
+  startRun: (parentId?: string | null) => void;
   cancel: () => void;
   isRunning: () => boolean;
   /** Returns a promise that resolves when the current or next run finishes. */
@@ -114,6 +119,7 @@ type PendingDoc = {
   errorMessage?: string;
 };
 const MAX_IMAGE_SIZE = 20 * 1024 * 1024;
+const COMPARE_APPEND_MESSAGE_TIMEOUT_MS = 10_000;
 
 function isNativeComposing(event: Event) {
   return "isComposing" in event && (event as InputEvent).isComposing === true;
@@ -133,7 +139,10 @@ function fileToBase64DataURL(file: File): Promise<string> {
   });
 }
 
-function formatReasoningEffortLabel(level: ReasoningEffort, modelId?: string): string {
+function formatReasoningEffortLabel(
+  level: ReasoningEffort,
+  modelId?: string,
+): string {
   if (level === "max") return "Max";
   if (level === "xhigh") {
     const normalized = modelId?.trim().toLowerCase() ?? "";
@@ -169,7 +178,12 @@ function useDictation(
   const start = useCallback(() => {
     const SpeechRecognitionAPI =
       typeof window !== "undefined" &&
-      (window.SpeechRecognition ?? (window as unknown as { webkitSpeechRecognition?: typeof SpeechRecognition }).webkitSpeechRecognition);
+      (window.SpeechRecognition ??
+        (
+          window as unknown as {
+            webkitSpeechRecognition?: typeof SpeechRecognition;
+          }
+        ).webkitSpeechRecognition);
     if (!SpeechRecognitionAPI) {
       return;
     }
@@ -215,7 +229,11 @@ function useDictation(
 
   const supported =
     typeof window !== "undefined" &&
-    !!(window.SpeechRecognition ?? (window as unknown as { webkitSpeechRecognition?: unknown }).webkitSpeechRecognition);
+    !!(
+      window.SpeechRecognition ??
+      (window as unknown as { webkitSpeechRecognition?: unknown })
+        .webkitSpeechRecognition
+    );
 
   return { isDictating, start, stop, supported };
 }
@@ -245,6 +263,7 @@ export function RegisterCompareHandle({
 }): ReactElement | null {
   const handlesRef = useContext(CompareHandlesContext);
   const aui = useAui();
+  const pendingAppendWaitersRef = useRef<Set<() => void>>(new Set());
 
   useEffect(() => {
     if (!handlesRef) {
@@ -254,13 +273,79 @@ export function RegisterCompareHandle({
     currentHandles[name] = {
       // fixes occasional reorder on reload.
       append: (content) =>
-        aui.thread().append({ role: "user", content, createdAt: new Date() } as never),
-      appendMessage: (content) =>
-        aui.thread().append({ role: "user", content, createdAt: new Date(), startRun: false } as never),
-      startRun: () => {
+        aui
+          .thread()
+          .append({ role: "user", content, createdAt: new Date() } as never),
+      appendMessage: (content) => {
+        const thread = aui.thread();
+        const beforeIds = new Set(
+          thread.getState().messages.map((message) => message.id),
+        );
+        thread.append({
+          role: "user",
+          content,
+          createdAt: new Date(),
+          startRun: false,
+        } as never);
+
+        const findAppendedUserMessageId = () => {
+          const messages = thread.getState().messages;
+          for (let index = messages.length - 1; index >= 0; index -= 1) {
+            const message = messages[index];
+            if (beforeIds.has(message.id) || message.role !== "user") {
+              continue;
+            }
+            return message.id;
+          }
+          return null;
+        };
+
+        const appendedId = findAppendedUserMessageId();
+        if (appendedId) {
+          return Promise.resolve(appendedId);
+        }
+
+        return new Promise<string | null>((resolve) => {
+          const startedAt = Date.now();
+          let settled = false;
+          let timer: number | null = null;
+          let cancel: (() => void) | null = null;
+          const cleanup = () => {
+            if (timer !== null) {
+              window.clearTimeout(timer);
+              timer = null;
+            }
+            if (cancel) {
+              pendingAppendWaitersRef.current.delete(cancel);
+            }
+          };
+          const finish = (messageId: string | null) => {
+            if (settled) return;
+            settled = true;
+            cleanup();
+            resolve(messageId);
+          };
+          cancel = () => finish(null);
+          const poll = () => {
+            timer = null;
+            const messageId = findAppendedUserMessageId();
+            if (
+              messageId ||
+              Date.now() - startedAt >= COMPARE_APPEND_MESSAGE_TIMEOUT_MS
+            ) {
+              finish(messageId);
+              return;
+            }
+            timer = window.setTimeout(poll, 16);
+          };
+          pendingAppendWaitersRef.current.add(cancel);
+          timer = window.setTimeout(poll, 0);
+        });
+      },
+      startRun: (parentId) => {
         const msgs = aui.thread().getState().messages;
-        const lastId = msgs.length > 0 ? msgs[msgs.length - 1].id : null;
-        aui.thread().startRun({ parentId: lastId });
+        const fallbackId = msgs.length > 0 ? msgs[msgs.length - 1].id : null;
+        aui.thread().startRun({ parentId: parentId ?? fallbackId });
       },
       cancel: () => aui.thread().cancelRun(),
       isRunning: () => aui.thread().getState().isRunning,
@@ -278,6 +363,10 @@ export function RegisterCompareHandle({
         }),
     };
     return () => {
+      for (const cancel of pendingAppendWaitersRef.current) {
+        cancel();
+      }
+      pendingAppendWaitersRef.current.clear();
       delete currentHandles[name];
     };
   }, [handlesRef, name, aui]);
@@ -300,7 +389,8 @@ function PendingImageThumb({
     setSrc(url);
     return () => URL.revokeObjectURL(url);
   }, [file]);
-  if (!src) return <div className="size-14 animate-pulse rounded-[14px] bg-muted" />;
+  if (!src)
+    return <div className="size-14 animate-pulse rounded-[14px] bg-muted" />;
   return (
     <div className="relative size-14 shrink-0 overflow-hidden rounded-[14px] border border-foreground/20 bg-muted">
       <img src={src} alt={file.name} className="h-full w-full object-cover" />
@@ -335,7 +425,10 @@ export function SharedComposer({
   const [running, setRunning] = useState(false);
   const [comparing, setComparing] = useState(false);
   const [pendingImages, setPendingImages] = useState<PendingImage[]>([]);
-  const [pendingAudio, setPendingAudio] = useState<{ name: string; base64: string } | null>(null);
+  const [pendingAudio, setPendingAudio] = useState<{
+    name: string;
+    base64: string;
+  } | null>(null);
   const [pendingDocs, setPendingDocs] = useState<PendingDoc[]>([]);
   const [dragging, setDragging] = useState(false);
   const [isComposing, setIsComposing] = useState(false);
@@ -368,10 +461,16 @@ export function SharedComposer({
   const setReasoningEnabled = useChatRuntimeStore((s) => s.setReasoningEnabled);
   const reasoningStyle = useChatRuntimeStore((s) => s.reasoningStyle);
   const reasoningEffort = useChatRuntimeStore((s) => s.reasoningEffort);
-  const supportsReasoningOff = useChatRuntimeStore((s) => s.supportsReasoningOff);
-  const reasoningEffortLevels = useChatRuntimeStore((s) => s.reasoningEffortLevels);
+  const supportsReasoningOff = useChatRuntimeStore(
+    (s) => s.supportsReasoningOff,
+  );
+  const reasoningEffortLevels = useChatRuntimeStore(
+    (s) => s.reasoningEffortLevels,
+  );
   const setReasoningEffort = useChatRuntimeStore((s) => s.setReasoningEffort);
-  const supportsPreserveThinking = useChatRuntimeStore((s) => s.supportsPreserveThinking);
+  const supportsPreserveThinking = useChatRuntimeStore(
+    (s) => s.supportsPreserveThinking,
+  );
   const preserveThinking = useChatRuntimeStore((s) => s.preserveThinking);
   const setPreserveThinking = useChatRuntimeStore((s) => s.setPreserveThinking);
   const supportsTools = useChatRuntimeStore((s) => s.supportsTools);
@@ -388,6 +487,8 @@ export function SharedComposer({
   );
   const ragToolEnabled = useChatRuntimeStore((s) => s.ragToolEnabled);
   const setRagToolEnabled = useChatRuntimeStore((s) => s.setRagToolEnabled);
+  const artifactsEnabled = useChatRuntimeStore((s) => s.artifactsEnabled);
+  const setArtifactsEnabled = useChatRuntimeStore((s) => s.setArtifactsEnabled);
   const webFetchToolsEnabled = useChatRuntimeStore(
     (s) => s.webFetchToolsEnabled,
   );
@@ -523,6 +624,7 @@ export function SharedComposer({
   // it needs the tool-calling loop. No external-builtin equivalent —
   // gate purely on supportsTools (mirrors web/code without a builtin).
   const ragDisabled = !modelLoaded || !supportsTools;
+  const artifactDisabled = !modelLoaded;
   // Fetch pill: Anthropic-only (web_fetch_20250910 / web_fetch_20260209).
   const webFetchDisabled = !modelLoaded || !supportsBuiltinWebFetch;
   const showWebFetchPill = supportsBuiltinWebFetch;
@@ -530,12 +632,17 @@ export function SharedComposer({
   // reference `toolsDisabled` (rare; both pills used it before).
   const toolsDisabled = codeDisabled;
   const setPendingAudioStore = useChatRuntimeStore((s) => s.setPendingAudio);
-  const clearPendingAudioStore = useChatRuntimeStore((s) => s.clearPendingAudio);
-
-  const { isDictating, start: startDictation, stop: stopDictation, supported: dictationSupported } = useDictation(
-    setText,
+  const clearPendingAudioStore = useChatRuntimeStore(
+    (s) => s.clearPendingAudio,
   );
 
+  const {
+    isDictating,
+    start: startDictation,
+    stop: stopDictation,
+    supported: dictationSupported,
+  } = useDictation(setText);
+
   useEffect(() => {
     const id = setInterval(() => {
       const handles = handlesRef.current;
@@ -552,8 +659,10 @@ export function SharedComposer({
     ta.style.height = "auto";
     const styles = window.getComputedStyle(ta);
     const lineHeight = parseFloat(styles.lineHeight) || 20;
-    const paddingY = parseFloat(styles.paddingTop) + parseFloat(styles.paddingBottom);
-    const borderY = parseFloat(styles.borderTopWidth) + parseFloat(styles.borderBottomWidth);
+    const paddingY =
+      parseFloat(styles.paddingTop) + parseFloat(styles.paddingBottom);
+    const borderY =
+      parseFloat(styles.borderTopWidth) + parseFloat(styles.borderBottomWidth);
     const maxHeight = lineHeight * 6 + paddingY + borderY;
     const next = Math.min(ta.scrollHeight, maxHeight);
     ta.style.height = `${next}px`;
@@ -886,12 +995,17 @@ export function SharedComposer({
     // LoraCompare and single-pane chats are unaffected.
     if (hasCompareHandles && !isGeneralizedCompare) {
       toast.error("Pick a model in each pane to compare", {
-        description: "Use the model dropdown above each pane, then send your prompt.",
+        description:
+          "Use the model dropdown above each pane, then send your prompt.",
       });
       return;
     }
 
-    if (pendingImages.length > 0 && !isGeneralizedCompare && imageUnavailableReason) {
+    if (
+      pendingImages.length > 0 &&
+      !isGeneralizedCompare &&
+      imageUnavailableReason
+    ) {
       // Single mode: the loaded model's runtime capability is known
       // here. Compare mode defers — each ensureModelLoaded below sets
       // loadedIsMultimodal for its side, and the chat-adapter's
@@ -931,20 +1045,96 @@ export function SharedComposer({
       const maxSeqLength = store.params.maxSeqLength;
       const trustRemoteCode = store.params.trustRemoteCode ?? false;
       const chatTemplateOverride = store.chatTemplateOverride;
-      const effectiveChatTemplateOverride =
-        chatTemplateOverride?.trim() ? chatTemplateOverride : null;
+      const effectiveChatTemplateOverride = chatTemplateOverride?.trim()
+        ? chatTemplateOverride
+        : null;
 
       function modelDisplayName(id: string): string {
+        const external = parseExternalModelId(id);
+        if (external) return external.modelId;
         const parts = id.split("/");
         return parts[parts.length - 1] || id;
       }
 
       // Helper: load a model and update store checkpoint
-      async function ensureModelLoaded(sel: CompareModelSelection): Promise<string> {
+      async function ensureModelLoaded(
+        sel: CompareModelSelection,
+      ): Promise<string> {
+        const external = parseExternalModelId(sel.id);
+        if (external) {
+          const externalStore = useExternalProvidersStore.getState();
+          if (!externalStore.connectionsEnabled) {
+            throw new Error(
+              "Connections are disabled. Turn on Enable connections in Settings -> Connections to use hosted models.",
+            );
+          }
+          const provider = externalStore.providers.find(
+            (p) => p.id === external.providerId,
+          );
+          if (!provider) {
+            throw new Error(
+              "Connection not found. Open Settings -> Connections and add it again.",
+            );
+          }
+
+          const reasoningCaps = getExternalReasoningCapabilities(
+            provider.providerType,
+            external.modelId,
+            {
+              isReasoningProvider: provider.isReasoningModel === true,
+              baseUrl: provider.baseUrl ?? null,
+            },
+          );
+          const supportsBuiltinWebSearch = providerSupportsBuiltinWebSearch(
+            provider.providerType,
+            external.modelId,
+            provider.baseUrl,
+          );
+          const supportsBuiltinCodeExecution =
+            providerSupportsBuiltinCodeExecution(
+              provider.providerType,
+              external.modelId,
+              provider.baseUrl,
+            );
+          const supportsBuiltinImageGeneration =
+            providerSupportsBuiltinImageGeneration(
+              provider.providerType,
+              external.modelId,
+              provider.baseUrl,
+            );
+          const supportsBuiltinWebFetch = providerSupportsBuiltinWebFetch(
+            provider.providerType,
+          );
+          const currentStore = useChatRuntimeStore.getState();
+          currentStore.setCheckpoint(sel.id, null);
+          useChatRuntimeStore.setState({
+            activeGgufVariant: null,
+            ggufContextLength: null,
+            ggufMaxContextLength: null,
+            ggufNativeContextLength: null,
+            activeNativePathToken: null,
+            supportsReasoning: reasoningCaps.supportsReasoning,
+            reasoningAlwaysOn: reasoningCaps.reasoningAlwaysOn,
+            reasoningStyle: reasoningCaps.reasoningStyle,
+            supportsReasoningOff: reasoningCaps.supportsReasoningOff,
+            reasoningEffortLevels: reasoningCaps.reasoningEffortLevels,
+            supportsPreserveThinking: false,
+            supportsTools: false,
+            supportsBuiltinWebSearch,
+            supportsBuiltinCodeExecution,
+            supportsBuiltinImageGeneration,
+            supportsBuiltinWebFetch,
+            loadedIsMultimodal:
+              providerTypeSupportsVision(provider.providerType) === true,
+          });
+          return "external";
+        }
+
         const currentStore = useChatRuntimeStore.getState();
         const isAlreadyActive =
           currentStore.params.checkpoint === sel.id &&
-          (currentStore.activeGgufVariant ?? null) === (sel.ggufVariant ?? null);
+          (currentStore.activeGgufVariant ?? null) ===
+            (sel.ggufVariant ?? null);
         if (!isAlreadyActive) {
           const validation = await validateModel({
             model_path: sel.id,
@@ -1022,9 +1212,19 @@ export function SharedComposer({
       const handle1 = handlesRef.current["model1"];
       const handle2 = handlesRef.current["model2"];
 
-      // Show user messages immediately on both sides
-      if (handle1) handle1.appendMessage(content);
-      if (handle2) handle2.appendMessage(content);
+      // Show user messages immediately on both sides and keep the ids
+      // so delayed model loads can start the run from the intended turn.
+      const [parentId1, parentId2] = await Promise.all([
+        handle1 ? handle1.appendMessage(content) : Promise.resolve(null),
+        handle2 ? handle2.appendMessage(content) : Promise.resolve(null),
+      ]);
+      if ((handle1 && !parentId1) || (handle2 && !parentId2)) {
+        toast.error("Compare failed", {
+          description:
+            "The prompt could not be added to both compare panes. Try sending it again.",
+        });
+        return;
+      }
 
       const name1 = model1?.id ? modelDisplayName(model1.id) : "";
       const name2 = model2?.id ? modelDisplayName(model2.id) : "";
@@ -1034,25 +1234,42 @@ export function SharedComposer({
       try {
         // Side 1: load → generate → wait
         if (handle1 && model1?.id) {
-          toast("Loading Model 1…", { id: toastId, description: name1, duration: Infinity });
+          toast("Loading Model 1…", {
+            id: toastId,
+            description: name1,
+            duration: Infinity,
+          });
           const status1 = await ensureModelLoaded(model1);
-          toast("Generating with Model 1…", { id: toastId, description: `${name1} (${status1})`, duration: Infinity });
+          toast("Generating with Model 1…", {
+            id: toastId,
+            description: `${name1} (${status1})`,
+            duration: Infinity,
+          });
           const done = handle1.waitForRunEnd();
-          handle1.startRun();
+          handle1.startRun(parentId1);
           await done;
         }
 
         // Side 2: load → generate → wait
         if (handle2 && model2?.id) {
-          const needsLoad = model2.id.toLowerCase() !== (model1?.id || "").toLowerCase()
-            || (model2.ggufVariant ?? "") !== (model1?.ggufVariant ?? "");
+          const needsLoad =
+            model2.id.toLowerCase() !== (model1?.id || "").toLowerCase() ||
+            (model2.ggufVariant ?? "") !== (model1?.ggufVariant ?? "");
           if (needsLoad) {
-            toast("Loading Model 2…", { id: toastId, description: name2, duration: Infinity });
+            toast("Loading Model 2…", {
+              id: toastId,
+              description: name2,
+              duration: Infinity,
+            });
           }
           const status2 = await ensureModelLoaded(model2);
-          toast("Generating with Model 2…", { id: toastId, description: `${name2} (${status2})`, duration: Infinity });
+          toast("Generating with Model 2…", {
+            id: toastId,
+            description: `${name2} (${status2})`,
+            duration: Infinity,
+          });
           const done = handle2.waitForRunEnd();
-          handle2.startRun();
+          handle2.startRun(parentId2);
           await done;
         }
 
@@ -1189,7 +1406,10 @@ export function SharedComposer({
               <span className="max-w-48 truncate">{pendingAudio.name}</span>
               <button
                 type="button"
-                onClick={() => { setPendingAudio(null); clearPendingAudioStore(); }}
+                onClick={() => {
+                  setPendingAudio(null);
+                  clearPendingAudioStore();
+                }}
                 className="flex size-4 items-center justify-center rounded-full hover:bg-destructive hover:text-destructive-foreground"
                 aria-label="Remove audio"
               >
@@ -1287,130 +1507,136 @@ export function SharedComposer({
           )}
           {showReasoningControl ? (
             effectiveReasoningStyle === "reasoning_effort" ? (
-            <DropdownMenu>
-              <DropdownMenuTrigger asChild={true}>
-                <button
-                  type="button"
-                  disabled={reasoningDisabled}
-                  className={cn(
-                    "flex items-center gap-1.5 rounded-full px-1.5 py-1.5 text-[13px] font-medium text-muted-foreground/70 transition-colors",
-                    reasoningDisabled
-                      ? "cursor-not-allowed opacity-40"
-                      : effectiveReasoningVisualEnabled
-                        ? "text-primary hover:bg-primary/10 dark:hover:bg-white/[0.08]"
-                        : "hover:bg-primary/10 dark:hover:bg-white/[0.08]",
-                  )}
-                  aria-label={thinkEffortAriaLabel({
-                    modelLoaded,
-                    reasoningDisabled,
-                    reasoningEffort,
-                  })}
-                >
-                  {effectiveReasoningVisualEnabled ? (
-                    <LightbulbIcon className="size-3.5" />
-                  ) : (
-                    <LightbulbOffIcon className="size-3.5" />
-                  )}
-                  <span>
-                    Think:{" "}
-                    {effectiveReasoningVisualEnabled
-                      ? formatReasoningEffortLabel(
-                          reasoningEffort,
-                          externalSelection?.modelId,
-                        )
-                      : formatReasoningDisabledLabel(
-                          effectiveSupportsReasoningOff,
-                          isExternalOpenAIReasoning,
-                          checkpoint,
-                        )}
-                  </span>
-                </button>
-              </DropdownMenuTrigger>
-              <DropdownMenuContent align="end">
-                {effectiveSupportsReasoningOff && (
-                  <DropdownMenuItem
-                    onSelect={() => {
-                      setReasoningEnabled(false);
-                      applyQwenThinkingParams(false);
-                    }}
-                  >
-                    {formatReasoningDisabledLabel(
-                      effectiveSupportsReasoningOff,
-                      isExternalOpenAIReasoning,
-                      checkpoint,
+              <DropdownMenu>
+                <DropdownMenuTrigger asChild={true}>
+                  <button
+                    type="button"
+                    disabled={reasoningDisabled}
+                    className={cn(
+                      "flex items-center gap-1.5 rounded-full px-1.5 py-1.5 text-[13px] font-medium text-muted-foreground/70 transition-colors",
+                      reasoningDisabled
+                        ? "cursor-not-allowed opacity-40"
+                        : effectiveReasoningVisualEnabled
+                          ? "text-primary hover:bg-primary/10 dark:hover:bg-white/[0.08]"
+                          : "hover:bg-primary/10 dark:hover:bg-white/[0.08]",
                     )}
-                    {!effectiveReasoningVisualEnabled ? " \u2713" : ""}
-                  </DropdownMenuItem>
-                )}
-                {effectiveReasoningEffortLevels
-                  .filter((level) => level !== "none")
-                  .map((level) => (
-                  <DropdownMenuItem
-                    key={level}
-                    onSelect={() => {
-                      setReasoningEffort(level);
-                      setReasoningEnabled(true);
-                      applyQwenThinkingParams(true);
-                      // Mutual exclusion: turning thinking on for a
-                      // Kimi model forces the web_search builtin off.
-                      if (isKimiExternal && toolsEnabled) {
-                        setToolsEnabled(false, { persist: false });
-                      }
-                    }}
+                    aria-label={thinkEffortAriaLabel({
+                      modelLoaded,
+                      reasoningDisabled,
+                      reasoningEffort,
+                    })}
                   >
-                    {formatReasoningEffortLabel(level, externalSelection?.modelId)}
-                    {effectiveReasoningVisualEnabled && reasoningEffort === level ? " \u2713" : ""}
-                  </DropdownMenuItem>
-                ))}
-              </DropdownMenuContent>
-            </DropdownMenu>
-          ) : (
-            <button
-              type="button"
-              disabled={reasoningDisabled || reasoningLockedOn}
-              aria-disabled={reasoningDisabled || reasoningLockedOn}
-              title={
-                reasoningLockedOn
-                  ? "This model requires reasoning to stay on."
-                  : undefined
-              }
-              onClick={() => {
-                if (reasoningLockedOn) return;
-                const next = !reasoningEnabled;
-                setReasoningEnabled(next);
-                applyQwenThinkingParams(next);
-                // Mutual exclusion: Kimi's $web_search builtin
-                // requires thinking off, so turning thinking on flips
-                // the Search pill off (and vice versa).
-                if (isKimiExternal && next && toolsEnabled) {
-                  setToolsEnabled(false, { persist: false });
+                    {effectiveReasoningVisualEnabled ? (
+                      <LightbulbIcon className="size-3.5" />
+                    ) : (
+                      <LightbulbOffIcon className="size-3.5" />
+                    )}
+                    <span>
+                      Think:{" "}
+                      {effectiveReasoningVisualEnabled
+                        ? formatReasoningEffortLabel(
+                            reasoningEffort,
+                            externalSelection?.modelId,
+                          )
+                        : formatReasoningDisabledLabel(
+                            effectiveSupportsReasoningOff,
+                            isExternalOpenAIReasoning,
+                            checkpoint,
+                          )}
+                    </span>
+                  </button>
+                </DropdownMenuTrigger>
+                <DropdownMenuContent align="end">
+                  {effectiveSupportsReasoningOff && (
+                    <DropdownMenuItem
+                      onSelect={() => {
+                        setReasoningEnabled(false);
+                        applyQwenThinkingParams(false);
+                      }}
+                    >
+                      {formatReasoningDisabledLabel(
+                        effectiveSupportsReasoningOff,
+                        isExternalOpenAIReasoning,
+                        checkpoint,
+                      )}
+                      {!effectiveReasoningVisualEnabled ? " \u2713" : ""}
+                    </DropdownMenuItem>
+                  )}
+                  {effectiveReasoningEffortLevels
+                    .filter((level) => level !== "none")
+                    .map((level) => (
+                      <DropdownMenuItem
+                        key={level}
+                        onSelect={() => {
+                          setReasoningEffort(level);
+                          setReasoningEnabled(true);
+                          applyQwenThinkingParams(true);
+                          // Mutual exclusion: turning thinking on for a
+                          // Kimi model forces the web_search builtin off.
+                          if (isKimiExternal && toolsEnabled) {
+                            setToolsEnabled(false, { persist: false });
+                          }
+                        }}
+                      >
+                        {formatReasoningEffortLabel(
+                          level,
+                          externalSelection?.modelId,
+                        )}
+                        {effectiveReasoningVisualEnabled &&
+                        reasoningEffort === level
+                          ? " \u2713"
+                          : ""}
+                      </DropdownMenuItem>
+                    ))}
+                </DropdownMenuContent>
+              </DropdownMenu>
+            ) : (
+              <button
+                type="button"
+                disabled={reasoningDisabled || reasoningLockedOn}
+                aria-disabled={reasoningDisabled || reasoningLockedOn}
+                title={
+                  reasoningLockedOn
+                    ? "This model requires reasoning to stay on."
+                    : undefined
                 }
-              }}
-              className={cn(
-                "flex items-center gap-1.5 rounded-full px-1.5 py-1.5 text-[13px] font-medium text-muted-foreground/70 transition-colors",
-                reasoningLockedOn
-                  ? "cursor-not-allowed text-primary"
-                  : reasoningDisabled
-                    ? "cursor-not-allowed opacity-40"
-                    : effectiveReasoningEnabled
-                      ? "text-primary hover:bg-primary/10 dark:hover:bg-white/[0.08]"
-                      : "hover:bg-primary/10 dark:hover:bg-white/[0.08]",
-              )}
-              aria-label={thinkToggleAriaLabel({
-                reasoningLockedOn,
-                modelLoaded,
-                reasoningDisabled,
-                effectiveReasoningEnabled,
-              })}
-            >
-              {reasoningLockedOn ||
-              (effectiveReasoningEnabled && !reasoningDisabled) ? (
-                <LightbulbIcon className="size-3.5" />
-              ) : (
-                <LightbulbOffIcon className="size-3.5" />
-              )}
-              <span>Think</span>
-            </button>
+                onClick={() => {
+                  if (reasoningLockedOn) return;
+                  const next = !reasoningEnabled;
+                  setReasoningEnabled(next);
+                  applyQwenThinkingParams(next);
+                  // Mutual exclusion: Kimi's $web_search builtin
+                  // requires thinking off, so turning thinking on flips
+                  // the Search pill off (and vice versa).
+                  if (isKimiExternal && next && toolsEnabled) {
+                    setToolsEnabled(false, { persist: false });
+                  }
+                }}
+                className={cn(
+                  "flex items-center gap-1.5 rounded-full px-1.5 py-1.5 text-[13px] font-medium text-muted-foreground/70 transition-colors",
+                  reasoningLockedOn
+                    ? "cursor-not-allowed text-primary"
+                    : reasoningDisabled
+                      ? "cursor-not-allowed opacity-40"
+                      : effectiveReasoningEnabled
+                        ? "cursor-pointer text-primary hover:bg-primary/10 dark:hover:bg-white/[0.08]"
+                        : "cursor-pointer hover:bg-primary/10 dark:hover:bg-white/[0.08]",
+                )}
+                aria-label={thinkToggleAriaLabel({
+                  reasoningLockedOn,
+                  modelLoaded,
+                  reasoningDisabled,
+                  effectiveReasoningEnabled,
+                })}
+              >
+                {reasoningLockedOn ||
+                (effectiveReasoningEnabled && !reasoningDisabled) ? (
+                  <LightbulbIcon className="size-3.5" />
+                ) : (
+                  <LightbulbOffIcon className="size-3.5" />
+                )}
+                <span>Think</span>
+              </button>
             )
           ) : null}
           {supportsPreserveThinking && (
@@ -1423,11 +1649,13 @@ export function SharedComposer({
                 !modelLoaded
                   ? "cursor-not-allowed opacity-40"
                   : preserveThinking
-                    ? "text-primary hover:bg-primary/10 dark:hover:bg-white/[0.08]"
-                    : "hover:bg-primary/10 dark:hover:bg-white/[0.08]",
+                    ? "cursor-pointer text-primary hover:bg-primary/10 dark:hover:bg-white/[0.08]"
+                    : "cursor-pointer hover:bg-primary/10 dark:hover:bg-white/[0.08]",
               )}
               aria-label={
-                preserveThinking ? "Disable preserve think" : "Enable preserve think"
+                preserveThinking
+                  ? "Disable preserve think"
+                  : "Enable preserve think"
               }
             >
               {preserveThinking && modelLoaded ? (
@@ -1456,7 +1684,9 @@ export function SharedComposer({
             }}
             className="composer-pill-btn"
             data-active={toolsEnabled && !searchDisabled ? "true" : "false"}
-            aria-label={toolsEnabled ? "Disable web search" : "Enable web search"}
+            aria-label={
+              toolsEnabled ? "Disable web search" : "Enable web search"
+            }
           >
             <GlobeIcon className="size-3.5" />
             <span>Search</span>
@@ -1467,7 +1697,11 @@ export function SharedComposer({
             onClick={() => setCodeToolsEnabled(!codeToolsEnabled)}
             className="composer-pill-btn"
             data-active={codeToolsEnabled && !codeDisabled ? "true" : "false"}
-            aria-label={codeToolsEnabled ? "Disable code execution" : "Enable code execution"}
+            aria-label={
+              codeToolsEnabled
+                ? "Disable code execution"
+                : "Enable code execution"
+            }
           >
             <CodeToggleIcon className="size-3.5" />
             <span>Code</span>
@@ -1478,9 +1712,13 @@ export function SharedComposer({
               disabled={imageDisabled}
               onClick={() => setImageToolsEnabled(!imageToolsEnabled)}
               className="composer-pill-btn"
-              data-active={imageToolsEnabled && !imageDisabled ? "true" : "false"}
+              data-active={
+                imageToolsEnabled && !imageDisabled ? "true" : "false"
+              }
               aria-label={
-                imageToolsEnabled ? "Disable image generation" : "Enable image generation"
+                imageToolsEnabled
+                  ? "Disable image generation"
+                  : "Enable image generation"
               }
             >
               <HugeiconsIcon
@@ -1520,6 +1758,22 @@ export function SharedComposer({
               <span>RAG</span>
             </button>
           )}
+          <button
+            type="button"
+            disabled={artifactDisabled}
+            onClick={() => setArtifactsEnabled(!artifactsEnabled)}
+            className="composer-pill-btn"
+            data-active={
+              artifactsEnabled && !artifactDisabled ? "true" : "false"
+            }
+            aria-label={
+              artifactsEnabled ? "Disable artifacts" : "Enable artifacts"
+            }
+          >
+            <FileTextIcon className="size-3.5" />
+            <span>Artifacts</span>
+          </button>
+          <McpComposerButton />
           {showWebFetchPill && (
             <button
               type="button"

studio/frontend/src/features/chat/stores/chat-runtime-store.ts

@@ -29,6 +29,11 @@ export const CHAT_TOOLS_ENABLED_KEY = "unsloth_chat_tools_enabled";
 export const CHAT_CODE_TOOLS_ENABLED_KEY = "unsloth_chat_code_tools_enabled";
 export const CHAT_IMAGE_TOOLS_ENABLED_KEY = "unsloth_chat_image_tools_enabled";
 export const CHAT_RAG_TOOL_ENABLED_KEY = "unsloth_chat_rag_tool_enabled";
+export const CHAT_ARTIFACTS_ENABLED_KEY = "unsloth_chat_artifacts_enabled";
+export const CHAT_COLLAPSE_HTML_ARTIFACTS_KEY =
+  "unsloth_chat_collapse_html_artifacts";
+export const CHAT_ALLOW_ARTIFACT_NETWORK_ACCESS_KEY =
+  "unsloth_chat_allow_artifact_network_access";
 export const CHAT_MCP_ENABLED_KEY = "unsloth_chat_mcp_enabled";
 export const CHAT_WEB_FETCH_TOOLS_ENABLED_KEY =
   "unsloth_chat_web_fetch_tools_enabled";
@@ -303,6 +308,9 @@ type ChatRuntimeStore = {
   ragToolEnabled: boolean;
   codeToolsEnabled: boolean;
   imageToolsEnabled: boolean;
+  artifactsEnabled: boolean;
+  collapseHtmlArtifacts: boolean;
+  allowArtifactNetworkAccess: boolean;
   mcpEnabledForChat: boolean;
   /**
    * Fetch pill state, independent of `toolsEnabled` (Search). Only
@@ -327,6 +335,7 @@ type ChatRuntimeStore = {
   chatTemplateOverride: string | null;
   loadedChatTemplateOverride: string | null;
   activeThreadId: string | null;
+  activeProjectId: string | null;
   settingsPanelOpen: boolean;
   pendingAudioBase64: string | null;
   pendingAudioName: string | null;
@@ -371,6 +380,7 @@ type ChatRuntimeStore = {
   setModelsError: (error: string | null) => void;
   setCheckpoint: (modelId: string, ggufVariant?: string | null) => void;
   setActiveThreadId: (threadId: string | null) => void;
+  setActiveProjectId: (projectId: string | null) => void;
   setSettingsPanelOpen: (open: boolean) => void;
   clearCheckpoint: () => void;
   setReasoningEnabled: (
@@ -384,6 +394,12 @@ type ChatRuntimeStore = {
   setToolsEnabled: (enabled: boolean, options?: { persist?: boolean }) => void;
   setCodeToolsEnabled: (enabled: boolean) => void;
   setImageToolsEnabled: (enabled: boolean) => void;
+  setArtifactsEnabled: (
+    enabled: boolean,
+    options?: { persist?: boolean },
+  ) => void;
+  setCollapseHtmlArtifacts: (enabled: boolean) => void;
+  setAllowArtifactNetworkAccess: (enabled: boolean) => void;
   setMcpEnabledForChat: (enabled: boolean) => void;
   setWebFetchToolsEnabled: (enabled: boolean) => void;
   setToolStatus: (status: string | null) => void;
@@ -424,6 +440,8 @@ type ScalarSettingKey =
   | "autoTitle"
   | "reasoningEffort"
   | "preserveThinking"
+  | "collapseHtmlArtifacts"
+  | "allowArtifactNetworkAccess"
   | "autoHealToolCalls"
   | "maxToolCallsPerMessage"
   | "toolCallTimeout"
@@ -465,6 +483,8 @@ const SCALAR_SETTING_KEYS = [
   "autoTitle",
   "reasoningEffort",
   "preserveThinking",
+  "collapseHtmlArtifacts",
+  "allowArtifactNetworkAccess",
   "autoHealToolCalls",
   "maxToolCallsPerMessage",
   "toolCallTimeout",
@@ -659,6 +679,12 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   ragToolEnabled: loadBool(CHAT_RAG_TOOL_ENABLED_KEY, false),
   codeToolsEnabled: loadBool(CHAT_CODE_TOOLS_ENABLED_KEY, false),
   imageToolsEnabled: loadBool(CHAT_IMAGE_TOOLS_ENABLED_KEY, false),
+  artifactsEnabled: loadBool(CHAT_ARTIFACTS_ENABLED_KEY, false),
+  collapseHtmlArtifacts: loadBool(CHAT_COLLAPSE_HTML_ARTIFACTS_KEY, false),
+  allowArtifactNetworkAccess: loadBool(
+    CHAT_ALLOW_ARTIFACT_NETWORK_ACCESS_KEY,
+    false,
+  ),
   mcpEnabledForChat: loadBool(CHAT_MCP_ENABLED_KEY, false),
   webFetchToolsEnabled: loadBool(CHAT_WEB_FETCH_TOOLS_ENABLED_KEY, false),
   toolStatus: null,
@@ -678,6 +704,7 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   chatTemplateOverride: null,
   loadedChatTemplateOverride: null,
   activeThreadId: null,
+  activeProjectId: null,
   settingsPanelOpen: false,
   pendingAudioBase64: null,
   pendingAudioName: null,
@@ -859,6 +886,7 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
     }),
   setActiveThreadId: (activeThreadId) =>
     set({ activeThreadId, contextUsage: null }),
+  setActiveProjectId: (activeProjectId) => set({ activeProjectId }),
   setSettingsPanelOpen: (settingsPanelOpen) => set({ settingsPanelOpen }),
   clearCheckpoint: () => {
     // Mirror setCheckpoint's persistence behavior: dropping the
@@ -893,6 +921,8 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       toolsEnabled: false,
       codeToolsEnabled: false,
       imageToolsEnabled: false,
+      artifactsEnabled: false,
+      mcpEnabledForChat: false,
       webFetchToolsEnabled: false,
       toolStatus: null,
       kvCacheDtype: null,
@@ -1015,6 +1045,36 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       saveBool(CHAT_IMAGE_TOOLS_ENABLED_KEY, imageToolsEnabled);
       return { imageToolsEnabled };
     }),
+  setArtifactsEnabled: (artifactsEnabled, options) =>
+    set(() => {
+      if (options?.persist !== false) {
+        saveBool(CHAT_ARTIFACTS_ENABLED_KEY, artifactsEnabled);
+      }
+      return { artifactsEnabled };
+    }),
+  setCollapseHtmlArtifacts: (collapseHtmlArtifacts) =>
+    set((state) => {
+      saveBool(CHAT_COLLAPSE_HTML_ARTIFACTS_KEY, collapseHtmlArtifacts);
+      setScalarSettingVersion(
+        "collapseHtmlArtifacts",
+        collapseHtmlArtifacts,
+        state.collapseHtmlArtifacts,
+      );
+      return { collapseHtmlArtifacts };
+    }),
+  setAllowArtifactNetworkAccess: (allowArtifactNetworkAccess) =>
+    set((state) => {
+      saveBool(
+        CHAT_ALLOW_ARTIFACT_NETWORK_ACCESS_KEY,
+        allowArtifactNetworkAccess,
+      );
+      setScalarSettingVersion(
+        "allowArtifactNetworkAccess",
+        allowArtifactNetworkAccess,
+        state.allowArtifactNetworkAccess,
+      );
+      return { allowArtifactNetworkAccess };
+    }),
   setMcpEnabledForChat: (mcpEnabledForChat) =>
     set(() => {
       saveBool(CHAT_MCP_ENABLED_KEY, mcpEnabledForChat);

studio/frontend/src/features/chat/thread-sidebar.tsx

@@ -42,7 +42,11 @@ export function ThreadSidebar({
   const { items } = useChatSidebarItems();
   const storeThreadId = useChatRuntimeStore((s) => s.activeThreadId);
   const activeId =
-    view.mode === "single" ? (view.threadId ?? storeThreadId) : view.pairId;
+    view.mode === "single"
+      ? (view.threadId ?? storeThreadId)
+      : view.mode === "compare"
+        ? view.pairId
+        : view.projectId;
 
   function viewForItem(item: SidebarItem): ChatView {
     return item.type === "single"

studio/frontend/src/features/chat/types.ts

@@ -4,8 +4,28 @@
 export type ModelType = "base" | "lora" | "model1" | "model2";
 
 export type ChatView =
-  | { mode: "single"; threadId?: string; newThreadNonce?: string }
-  | { mode: "compare"; pairId: string };
+  | {
+      mode: "project";
+      projectId: string;
+    }
+  | {
+      mode: "single";
+      threadId?: string;
+      newThreadNonce?: string;
+      projectId?: string | null;
+    }
+  | { mode: "compare"; pairId: string; projectId?: string | null };
+
+export interface ProjectRecord {
+  id: string;
+  name: string;
+  instructions?: string;
+  rootPath?: string | null;
+  sandboxPath?: string | null;
+  archived: boolean;
+  createdAt: number;
+  updatedAt: number;
+}
 
 export interface ThreadRecord {
   id: string;
@@ -13,6 +33,7 @@ export interface ThreadRecord {
   modelType: ModelType;
   modelId?: string;
   pairId?: string;
+  projectId?: string | null;
   archived: boolean;
   createdAt: number;
   /**

studio/frontend/src/features/chat/utils/chat-history-storage.ts

@@ -4,22 +4,32 @@
 import {
   buildBackendChatExport,
   clearBackendChats,
+  deleteChatProject,
   deleteChatThreads,
+  getChatProject,
   getChatMessage,
   getChatThread,
   batchListChatMessages,
+  listChatProjects,
   listChatImportLedger,
   listChatMessages,
   listChatThreads,
   notifyChatHistoryUpdated,
   recordChatImportLedger,
+  saveChatProject,
   saveChatMessage,
   saveChatThread,
   syncChatMessages,
+  updateChatProject,
   updateChatThread,
 } from "../api/chat-api";
 import { db, DEXIE_DB_NAME } from "../db";
-import type { MessageRecord, ModelType, ThreadRecord } from "../types";
+import type {
+  MessageRecord,
+  ModelType,
+  ProjectRecord,
+  ThreadRecord,
+} from "../types";
 import {
   isChatThreadDeleted,
   markChatThreadsDeleted,
@@ -28,6 +38,7 @@ import {
 type ThreadListArgs = {
   modelType?: ModelType;
   pairId?: string;
+  projectId?: string | null;
   includeArchived?: boolean;
 };
 
@@ -45,6 +56,7 @@ interface ExportedChat {
   exportedAt: string;
   version: 1;
   threadCount: number;
+  projects?: unknown[];
   threads: unknown[];
   messages: unknown[];
 }
@@ -82,6 +94,8 @@ function matchesThreadListArgs(
   return (
     !isChatThreadDeleted(thread.id) &&
     (!args.pairId || thread.pairId === args.pairId) &&
+    (args.projectId === undefined ||
+      (thread.projectId ?? null) === args.projectId) &&
     (!args.modelType || thread.modelType === args.modelType) &&
     (args.includeArchived !== false || !thread.archived)
   );
@@ -584,6 +598,72 @@ export async function listStoredChatThreadsWithMessages(
   return entries.filter((e) => e.hasContent).map((e) => e.thread);
 }
 
+export async function listStoredChatProjects(
+  args: { includeArchived?: boolean } = {},
+): Promise<ProjectRecord[]> {
+  return listChatProjects(args);
+}
+
+export async function getStoredChatProject(
+  projectId: string,
+): Promise<ProjectRecord | null> {
+  return getChatProject(projectId);
+}
+
+export async function createStoredChatProject(
+  name: string,
+): Promise<ProjectRecord> {
+  const trimmed = name.trim();
+  if (!trimmed) {
+    throw new Error("Project name is required.");
+  }
+  const now = Date.now();
+  return saveChatProject({
+    id: crypto.randomUUID(),
+    name: trimmed,
+    instructions: "",
+    archived: false,
+    createdAt: now,
+    updatedAt: now,
+  });
+}
+
+export async function updateStoredChatProject(
+  projectId: string,
+  patch: Partial<ProjectRecord>,
+): Promise<ProjectRecord> {
+  return updateChatProject(projectId, {
+    ...patch,
+    updatedAt: patch.updatedAt ?? Date.now(),
+  });
+}
+
+export async function deleteStoredChatProject(
+  projectId: string,
+  args: { deleteFiles?: boolean } = {},
+): Promise<void> {
+  await deleteChatProject(projectId, args);
+}
+
+export async function moveStoredChatItemToProject(
+  item: { type: "single" | "compare"; id: string },
+  projectId: string | null,
+): Promise<void> {
+  const threadIds =
+    item.type === "single"
+      ? [item.id]
+      : (await listStoredChatThreads({
+          pairId: item.id,
+          includeArchived: true,
+        })).map((thread) => thread.id);
+
+  await Promise.all(
+    Array.from(new Set(threadIds)).map((threadId) =>
+      updateStoredChatThread(threadId, { projectId }),
+    ),
+  );
+}
+
 export async function saveStoredChatMessage(
   message: MessageRecord,
 ): Promise<MessageRecord> {
@@ -767,6 +847,7 @@ export async function buildStoredChatExport(): Promise<ExportedChat> {
     exportedAt: new Date().toISOString(),
     version: 1,
     threadCount: threads.length,
+    projects: backend?.projects ?? [],
     threads,
     messages,
   };

studio/frontend/src/features/chat/utils/chat-settings-storage.ts

@@ -28,6 +28,9 @@ const CHAT_ACTIVE_PRESET_KEY = "unsloth_chat_active_preset";
 const CHAT_ACTIVE_PRESET_SOURCE_KEY = "unsloth_chat_active_preset_source";
 const REASONING_EFFORT_KEY = "unsloth_reasoning_effort";
 const PRESERVE_THINKING_KEY = "unsloth_preserve_thinking";
+const COLLAPSE_HTML_ARTIFACTS_KEY = "unsloth_chat_collapse_html_artifacts";
+const ALLOW_ARTIFACT_NETWORK_ACCESS_KEY =
+  "unsloth_chat_allow_artifact_network_access";
 const CHAT_PRESETS_KEY = "unsloth_chat_custom_presets";
 const LEGACY_CHAT_SYSTEM_PROMPTS_KEY = "unsloth_chat_system_prompts";
 const LEGACY_CHAT_SETTINGS_IMPORT_KEY =
@@ -216,6 +219,10 @@ function sanitizeChatSettings(value: unknown): PersistedChatSettings {
   const reasoningEffort = sanitizeReasoningEffort(value.reasoningEffort);
   const autoTitle = sanitizeBool(value.autoTitle);
   const preserveThinking = sanitizeBool(value.preserveThinking);
+  const collapseHtmlArtifacts = sanitizeBool(value.collapseHtmlArtifacts);
+  const allowArtifactNetworkAccess = sanitizeBool(
+    value.allowArtifactNetworkAccess,
+  );
   const autoHealToolCalls = sanitizeBool(value.autoHealToolCalls);
   const maxToolCallsPerMessage = sanitizeInt(value.maxToolCallsPerMessage, 1);
   const toolCallTimeout = sanitizeInt(value.toolCallTimeout, 1);
@@ -230,6 +237,12 @@ function sanitizeChatSettings(value: unknown): PersistedChatSettings {
   if (reasoningEffort) settings.reasoningEffort = reasoningEffort;
   if (preserveThinking !== undefined)
     settings.preserveThinking = preserveThinking;
+  if (collapseHtmlArtifacts !== undefined) {
+    settings.collapseHtmlArtifacts = collapseHtmlArtifacts;
+  }
+  if (allowArtifactNetworkAccess !== undefined) {
+    settings.allowArtifactNetworkAccess = allowArtifactNetworkAccess;
+  }
   if (autoHealToolCalls !== undefined) {
     settings.autoHealToolCalls = autoHealToolCalls;
   }
@@ -290,6 +303,8 @@ export function isEmptyChatSettings(settings: PersistedChatSettings): boolean {
     settings.autoTitle === undefined &&
     settings.reasoningEffort === undefined &&
     settings.preserveThinking === undefined &&
+    settings.collapseHtmlArtifacts === undefined &&
+    settings.allowArtifactNetworkAccess === undefined &&
     settings.autoHealToolCalls === undefined &&
     settings.maxToolCallsPerMessage === undefined &&
     settings.toolCallTimeout === undefined
@@ -318,6 +333,8 @@ export function loadLegacyChatSettings(): PersistedChatSettings {
   );
   const autoTitle = loadBool(AUTO_TITLE_KEY);
   const preserveThinking = loadBool(PRESERVE_THINKING_KEY);
+  const collapseHtmlArtifacts = loadBool(COLLAPSE_HTML_ARTIFACTS_KEY);
+  const allowArtifactNetworkAccess = loadBool(ALLOW_ARTIFACT_NETWORK_ACCESS_KEY);
   const autoHealToolCalls = loadBool(AUTO_HEAL_TOOL_CALLS_KEY);
   const maxToolCallsPerMessage = loadInt(MAX_TOOL_CALLS_KEY, 1);
   const toolCallTimeout = loadInt(TOOL_CALL_TIMEOUT_KEY, 1);
@@ -336,6 +353,12 @@ export function loadLegacyChatSettings(): PersistedChatSettings {
   if (reasoningEffort) settings.reasoningEffort = reasoningEffort;
   if (preserveThinking !== undefined)
     settings.preserveThinking = preserveThinking;
+  if (collapseHtmlArtifacts !== undefined) {
+    settings.collapseHtmlArtifacts = collapseHtmlArtifacts;
+  }
+  if (allowArtifactNetworkAccess !== undefined) {
+    settings.allowArtifactNetworkAccess = allowArtifactNetworkAccess;
+  }
   if (autoHealToolCalls !== undefined) {
     settings.autoHealToolCalls = autoHealToolCalls;
   }

studio/frontend/src/features/native-intents/index.ts

@@ -0,0 +1,11 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+export { NativeModelChip } from "./components/native-model-chip";
+export { NativeModelDropOverlay } from "./components/native-model-drop-overlay";
+export { useNativeIntentStore } from "./store";
+export type { NativeIntent } from "./types";
+export { useChooseNativeModel } from "./use-native-dialogs";
+export { useNativeModelDrop } from "./use-native-drop";
+export type { NativeModelDropState } from "./use-native-drop";
+export { useNativePathLeasesSupported } from "./use-native-readiness";

studio/frontend/src/features/profile/utils/avatar-initials.ts

@@ -7,7 +7,7 @@ export function initialsFromName(name: string): string {
   return trimmed[0]!.toUpperCase();
 }
 
-/** Default blue background for avatar fallback (readable white text). */
+/** Default Unsloth-green background for avatar fallback (readable white text). */
 export function avatarBgStyle(): { backgroundColor: string } {
-  return { backgroundColor: "hsl(217 58% 48%)" };
+  return { backgroundColor: "#14b789" };
 }

studio/frontend/src/features/settings/tabs/chat-tab.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 { Switch } from "@/components/ui/switch";
 import { toast } from "@/lib/toast";
 import {
   Dialog,
@@ -15,6 +16,7 @@ import {
   clearAllChats,
   countAllChats,
   downloadChatExport,
+  useChatRuntimeStore,
 } from "@/features/chat";
 import { useT } from "@/i18n";
 import { Delete02Icon, Download02Icon } from "@hugeicons/core-free-icons";
@@ -29,10 +31,26 @@ export function ChatTab() {
   const [count, setCount] = useState<number | null>(null);
   const [exporting, setExporting] = useState(false);
   const [clearing, setClearing] = useState(false);
+  const collapseHtmlArtifacts = useChatRuntimeStore(
+    (state) => state.collapseHtmlArtifacts,
+  );
+  const setCollapseHtmlArtifacts = useChatRuntimeStore(
+    (state) => state.setCollapseHtmlArtifacts,
+  );
+  const allowArtifactNetworkAccess = useChatRuntimeStore(
+    (state) => state.allowArtifactNetworkAccess,
+  );
+  const setAllowArtifactNetworkAccess = useChatRuntimeStore(
+    (state) => state.setAllowArtifactNetworkAccess,
+  );
+  const hydratePersistedSettings = useChatRuntimeStore(
+    (state) => state.hydratePersistedSettings,
+  );
 
   useEffect(() => {
     void countAllChats().then(setCount);
-  }, []);
+    void hydratePersistedSettings();
+  }, [hydratePersistedSettings]);
 
   const handleExport = async () => {
     setExporting(true);
@@ -111,6 +129,31 @@ export function ChatTab() {
         </p>
       </header>
 
+      <SettingsSection title={t("settings.chat.artifacts.title")}>
+        <SettingsRow
+          label={t("settings.chat.artifacts.collapseHtmlBlocks")}
+          description={t(
+            "settings.chat.artifacts.collapseHtmlBlocksDescription",
+          )}
+        >
+          <Switch
+            checked={collapseHtmlArtifacts}
+            onCheckedChange={setCollapseHtmlArtifacts}
+          />
+        </SettingsRow>
+        <SettingsRow
+          label={t("settings.chat.artifacts.allowNetworkAccess")}
+          description={t(
+            "settings.chat.artifacts.allowNetworkAccessDescription",
+          )}
+        >
+          <Switch
+            checked={allowArtifactNetworkAccess}
+            onCheckedChange={setAllowArtifactNetworkAccess}
+          />
+        </SettingsRow>
+      </SettingsSection>
+
       <SettingsSection title={t("settings.chat.data")}>
         <SettingsRow
           label={t("settings.chat.exportHistory")}

studio/frontend/src/features/studio/studio-page.tsx

@@ -53,6 +53,10 @@ export function StudioPage(): ReactElement {
   const selectedHistoryRunId = useTrainingRuntimeStore((s) => s.selectedHistoryRunId);
   const setSelectedHistoryRunId = useTrainingRuntimeStore((s) => s.setSelectedHistoryRunId);
 
+  const setCurrentRunViewActive = useTrainingRuntimeStore(
+    (s) => s.setCurrentRunViewActive,
+  );
+
   useEffect(() => {
     return () => setSelectedHistoryRunId(null);
   }, [setSelectedHistoryRunId]);
@@ -67,6 +71,13 @@ export function StudioPage(): ReactElement {
         ? "configure"
         : requestedTab;
 
+  // Mirror "Current Run" tab state into the store so the sidebar can highlight
+  // the run this view refers to. Cleared on unmount (leaving the studio page).
+  useEffect(() => {
+    setCurrentRunViewActive(activeTab === "current-run");
+    return () => setCurrentRunViewActive(false);
+  }, [activeTab, setCurrentRunViewActive]);
+
   const { setPinned } = useSidebar();
   const pinSidebar = useCallback(() => setPinned(true), [setPinned]);
 

studio/frontend/src/features/studio/training-start-overlay.tsx

@@ -37,6 +37,12 @@ import { useT } from "@/i18n";
 
 const HF_REPO_REGEX = /^[A-Za-z0-9._-]+\/[A-Za-z0-9._-]+$/;
 
+// Tracks which jobs have already played the terminal intro animation. The
+// overlay unmounts when you navigate away from the training page, so without
+// this its typing/fade-in would replay on every return even though the run
+// itself is still going. Module-level so it survives remounts.
+const animatedJobs = new Set<string>();
+
 function formatBytes(n: number): string {
   if (n <= 0) return "0 B";
   if (n < 1024) return `${n} B`;
@@ -254,6 +260,7 @@ export function TrainingStartOverlay({
   const { stopTrainingRun, dismissTrainingRun } = useTrainingActions();
   const isStarting = useTrainingRuntimeStore((s) => s.isStarting);
   const phase = useTrainingRuntimeStore((s) => s.phase);
+  const jobId = useTrainingRuntimeStore((s) => s.jobId);
   const startModelName = useTrainingRuntimeStore((s) => s.startModelName);
   const startDatasetName = useTrainingRuntimeStore((s) => s.startDatasetName);
   const startFromResume = useTrainingRuntimeStore((s) => s.startFromResume);
@@ -298,6 +305,16 @@ export function TrainingStartOverlay({
     }
   }, [isStarting]);
 
+  // Play the intro animation only the first time we mount for a given job.
+  // On later remounts (e.g. leaving the training page and coming back) the
+  // terminal renders its final state instantly so the logs don't restart.
+  const alreadyAnimated = jobId != null && animatedJobs.has(jobId);
+  useEffect(() => {
+    if (jobId != null) {
+      animatedJobs.add(jobId);
+    }
+  }, [jobId]);
+
   return (
     <div className="pointer-events-none absolute inset-0 z-30 flex items-center justify-center rounded-2xl bg-background/45 backdrop-blur-[1px]">
       <div className="pointer-events-auto relative flex w-[860px] max-w-[calc(100%-2rem)] flex-col items-center gap-4">
@@ -311,7 +328,7 @@ export function TrainingStartOverlay({
             <Button
               variant="ghost"
               size="icon"
-              className="absolute right-3 top-3 z-10 size-7 cursor-pointer rounded-full text-muted-foreground/60 hover:bg-destructive/10 hover:text-destructive"
+              className="absolute right-3 top-3 z-10 size-7 cursor-pointer rounded-full text-muted-foreground/90 hover:bg-destructive/10 hover:text-destructive"
               onClick={() => setCancelDialogOpen(true)}
               disabled={cancelRequested}
             >
@@ -349,6 +366,7 @@ export function TrainingStartOverlay({
           <Terminal
             className="w-full min-h-[390px] rounded-2xl px-7 py-6 text-left"
             startOnView={false}
+            instant={alreadyAnimated}
           >
           <TypingAnimation
             duration={36}

studio/frontend/src/features/training/stores/training-runtime-store.ts

@@ -46,6 +46,7 @@ const initialState: TrainingRuntimeState = {
   resetGeneration: 0,
   stopRequested: false,
   selectedHistoryRunId: null,
+  currentRunViewActive: false,
 };
 
 function sortSeries(points: TrainingSeriesPoint[]): TrainingSeriesPoint[] {
@@ -182,6 +183,9 @@ export const useTrainingRuntimeStore = create<TrainingRuntimeStore>()((set) => (
   setSelectedHistoryRunId: (selectedHistoryRunId) =>
     set({ selectedHistoryRunId }),
 
+  setCurrentRunViewActive: (currentRunViewActive) =>
+    set({ currentRunViewActive }),
+
   applyStatus: (payload) =>
     set((state) => {
       const metricHistory = applyMetricHistoryFromStatus(payload);

studio/frontend/src/features/training/types/runtime.ts

@@ -105,6 +105,9 @@ export interface TrainingRuntimeState {
   resetGeneration: number;
   stopRequested: boolean;
   selectedHistoryRunId: string | null;
+  // True while the studio "Current Run" tab is the active view, so the sidebar
+  // can highlight which run row the current run refers to (the active job).
+  currentRunViewActive: boolean;
 }
 
 export interface TrainingRuntimeActions {
@@ -127,6 +130,7 @@ export interface TrainingRuntimeActions {
   setStartQueued: (jobId: string, message: string) => void;
   setRuntimeError: (message: string) => void;
   setSelectedHistoryRunId: (id: string | null) => void;
+  setCurrentRunViewActive: (value: boolean) => void;
 }
 
 export type TrainingRuntimeStore = TrainingRuntimeState & TrainingRuntimeActions;

studio/frontend/src/i18n/locales/en.ts

@@ -166,6 +166,15 @@ export const en = {
     chat: {
       title: "Chat",
       description: "Manage your chat history stored on this device.",
+      artifacts: {
+        title: "Artifacts",
+        collapseHtmlBlocks: "Collapse HTML blocks",
+        collapseHtmlBlocksDescription:
+          "Artifacts mode collapses full HTML fallback automatically. Turn this on to also collapse full fenced HTML documents when Artifacts is off.",
+        allowNetworkAccess: "Allow artifact network access",
+        allowNetworkAccessDescription:
+          "Let artifact previews load scripts, styles, fonts, media, fetch, and WebSocket resources from HTTP(S) CDNs. Keep off for fully offline previews.",
+      },
       data: "Data",
       exportHistory: "Export chat history",
       exportHistoryDescription:

studio/frontend/src/index.css

@@ -14,6 +14,51 @@
 
 @custom-variant dark (&:is(.dark *));
 
+@layer components {
+	/* Sidebar scroll area (Gemini-style). The BOTTOM edge always fades so the
+	   last rows dissolve into the profile footer (no divider line). Once
+	   scrolled, the TOP edge also fades so rows dissolve as they pass under the
+	   pinned New Chat / Search header. */
+	/* Top edge: when scrolled, rows dissolve as they pass under the pinned
+	   New Chat / Search header (mask — fades content to transparent). */
+	.sidebar-scroll-fade.is-scrolled {
+		-webkit-mask-image: linear-gradient(to bottom, transparent 0, #000 14px);
+		mask-image: linear-gradient(to bottom, transparent 0, #000 14px);
+	}
+	/* Bottom edge: Gemini-style sticky gradient OVERLAY painted in the sidebar
+	   background colour, so the last rows wash into the profile footer. Uses a
+	   zero-height sticky anchor with an absolutely-positioned gradient so it
+	   never takes layout space. */
+	.sidebar-bottom-fade {
+		position: sticky;
+		bottom: 0;
+		height: 0;
+		z-index: 1;
+		pointer-events: none;
+	}
+	.sidebar-bottom-fade::after {
+		content: "";
+		position: absolute;
+		inset-inline: 0;
+		bottom: 0;
+		height: 40px;
+		background: linear-gradient(to top, var(--sidebar), transparent);
+	}
+}
+
+@layer base {
+	h1,
+	h2,
+	h3,
+	[data-slot="dialog-title"],
+	[data-slot="alert-dialog-title"] {
+		font-family: "Hellix", "Space Grotesk Variable", var(--font-sans) !important;
+		font-weight: 500 !important;
+	}
+}
+
+
+
 @font-face {
 	font-family: "Hellix";
 	src: url("/fonts/Hellix-Regular.woff") format("woff");
@@ -82,13 +127,15 @@
 	--chart-4: oklch(0.6926 0.1112 346.5775);
 	--chart-5: oklch(0.7497 0.1003 85.0057);
 	--radius: 1.1rem;
-	--sidebar: #f9faf9;
+	/* Match the page background so the sidebar reads as one surface with the
+	   content; the faint --sidebar-border on its right edge is the separator. */
+	--sidebar: oklch(1 0 0);
 	--sidebar-foreground: oklch(0.1281 0.0179 169.2764);
 	--sidebar-primary: #17b88b;
 	--sidebar-primary-foreground: oklch(1 0 0);
 	--sidebar-accent: oklch(0.96 0.0279 166.55);
 	--sidebar-accent-foreground: oklch(0.2868 0.0649 159.9823);
-	--sidebar-border: oklch(0.945 0.0101 164.8536);
+	--sidebar-border: oklch(0.92 0 0);
 	--sidebar-ring: #17b88b;
 	--destructive-foreground: oklch(1 0 0);
 	--font-sans: "Inter Variable", ui-sans-serif, sans-serif, system-ui;
@@ -121,7 +168,7 @@
 	/*    0px 0px 0px 0px hsl(0 0% 0% / 0),*/
 	/*    0px 8px 10px 0px hsl(0 0% 0% / 0);*/
 	/*--shadow-2xl: 0px 0px 0px 0px hsl(0 0% 0% / 0);*/
-	--tracking-normal: -0.01em;
+	--tracking-normal: 0em;
 
 	/* Hex (not OKLCH) so the rendered surface matches the design mockup pixel-for-pixel. */
 	--nav-fg: #383835;
@@ -202,7 +249,8 @@
 	--chart-3: oklch(0.7554 0.1285 197.339);
 	--chart-4: oklch(0.7503 0.1199 346.7805);
 	--chart-5: oklch(0.799 0.1196 84.6633);
-	--sidebar: #18181a;
+	/* Match the page background (dark); --sidebar-border is the right-edge separator. */
+	--sidebar: #1f2023;
 	--sidebar-foreground: #ececee;
 	--sidebar-primary: #17b88b;
 	--sidebar-primary-foreground: oklch(1 0 0);
@@ -306,8 +354,8 @@
 	--font-mono: JetBrains Mono, monospace;
 	--font-serif: Source Serif 4, serif;
 	--radius: 1.1rem;
-	--tracking-tighter: calc(var(--tracking-normal) - 0.05em);
-	--tracking-tight: calc(var(--tracking-normal) - 0.025em);
+	--tracking-tighter: 0em;
+	--tracking-tight: 0em;
 	--tracking-wide: calc(var(--tracking-normal) + 0.025em);
 	--tracking-wider: calc(var(--tracking-normal) + 0.05em);
 	--tracking-widest: calc(var(--tracking-normal) + 0.1em);
@@ -439,7 +487,7 @@
 	h5,
 	h6 {
 		font-family: var(--font-sans);
-		letter-spacing: -0.02em;
+		letter-spacing: 0;
 	}
 }
 
@@ -451,15 +499,14 @@
 	   on global antialiased font-smoothing to neutralize the bloom. */
 	.font-heading {
 		font-family: var(--font-heading);
-		letter-spacing: -0.01em;
+		letter-spacing: 0;
 	}
 
-	/* Dark mode loosens tracking to offset optical bloom on dark surfaces. */
 	.tracking-nav {
-		letter-spacing: 0.015em;
+		letter-spacing: 0;
 	}
 	.dark .tracking-nav {
-		letter-spacing: 0.03em;
+		letter-spacing: 0;
 	}
 
 	.nav-icon-btn {
@@ -486,6 +533,10 @@
 	.sidebar-nav-btn:hover,
 	.sidebar-nav-btn[data-active="true"],
 	.sidebar-nav-btn[data-state="open"],
+	.group\/project-item:hover .sidebar-nav-btn,
+	.group\/project-item:has(.sidebar-row-action[data-state="open"]) .sidebar-nav-btn,
+	.group\/project-chat-item:hover .sidebar-nav-btn,
+	.group\/project-chat-item:has(.sidebar-row-action[data-state="open"]) .sidebar-nav-btn,
 	.group\/recent-item:hover .sidebar-nav-btn,
 	.group\/recent-item:has(.sidebar-row-action[data-state="open"]) .sidebar-nav-btn,
 	.group\/run-item:hover .sidebar-nav-btn,
@@ -496,6 +547,10 @@
 	.dark .sidebar-nav-btn:hover,
 	.dark .sidebar-nav-btn[data-active="true"],
 	.dark .sidebar-nav-btn[data-state="open"],
+	.dark .group\/project-item:hover .sidebar-nav-btn,
+	.dark .group\/project-item:has(.sidebar-row-action[data-state="open"]) .sidebar-nav-btn,
+	.dark .group\/project-chat-item:hover .sidebar-nav-btn,
+	.dark .group\/project-chat-item:has(.sidebar-row-action[data-state="open"]) .sidebar-nav-btn,
 	.dark .group\/recent-item:hover .sidebar-nav-btn,
 	.dark .group\/recent-item:has(.sidebar-row-action[data-state="open"]) .sidebar-nav-btn,
 	.dark .group\/run-item:hover .sidebar-nav-btn,
@@ -504,7 +559,7 @@
 	}
 
 	.sidebar-row-action {
-		@apply absolute top-0 bottom-0 right-0 inline-flex items-center justify-end pl-2 pr-1.5 opacity-0 pointer-events-none outline-none;
+		@apply absolute top-0 bottom-0 right-0 inline-flex cursor-pointer items-center justify-end pl-2 pr-1.5 opacity-0 pointer-events-none outline-none;
 	}
 	.sidebar-row-action[data-state="open"] {
 		@apply opacity-100 pointer-events-auto;
@@ -530,12 +585,22 @@
 	}
 
 	.sidebar-sticky-label {
-		@apply sticky top-0 z-20 rounded-none bg-sidebar pt-0 pb-1.5 pl-[18px] pr-4 text-[13px]! font-medium normal-case tracking-[0.04em] text-nav-fg-muted focus-visible:ring-0! focus-visible:outline-none shadow-[0_-8px_0_0_var(--sidebar)] transition-shadow duration-150;
+		@apply rounded-none bg-sidebar pt-0 pb-[8px] pl-[18px] pr-4 text-[14.5px]! leading-[17px] font-medium normal-case focus-visible:ring-0! focus-visible:outline-none transition-shadow duration-150;
+		/* Muted section-header gray, matching Gemini's "Notebooks"/"Recents".
+		   Lightened from #5f6368 so the label reads as a header, clearly
+		   lighter than the near-black nav items. */
+		color: #80868b;
+		letter-spacing: 0;
 	}
-	.sidebar-sticky-label.is-scrolled {
-		@apply shadow-[0_-8px_0_0_var(--sidebar),0_0.5px_0_0_var(--sidebar-border)];
+	.sidebar-sticky-label-following {
+		@apply pt-[21px];
+	}
+	.dark .sidebar-sticky-label {
+		/* Muted gray, clearly dimmer than the near-white nav items (#ececee) so
+		   "Train"/"Recents" read as section headers — like Gemini's dark mode. */
+		color: #9aa0a6;
+		letter-spacing: 0;
 	}
-
 	/* Neutral panel input surface — sidesteps the green cast on
 	   `--input` / `--border` (both have a small chroma at hue ~165 in
 	   light mode). Same value drives the preset input pill, the system
@@ -698,7 +763,8 @@
 		border-color: rgb(255 255 255 / 0.12) !important;
 	}
 
-	.app-user-menu [data-slot="dropdown-menu-item"] {
+	.app-user-menu [data-slot="dropdown-menu-item"],
+	.app-user-menu [data-slot="dropdown-menu-sub-trigger"] {
 		height: 32px;
 		padding: 0 0.625rem !important;
 		gap: 8.5px !important;
@@ -706,28 +772,38 @@
 		font-weight: 500;
 		font-size: 14.5px;
 		line-height: 19px;
-		letter-spacing: 0.015em;
+		letter-spacing: 0;
 		color: var(--nav-fg);
 	}
-	.dark .app-user-menu [data-slot="dropdown-menu-item"] {
-		letter-spacing: 0.03em;
+	.dark .app-user-menu [data-slot="dropdown-menu-item"],
+	.dark .app-user-menu [data-slot="dropdown-menu-sub-trigger"] {
+		letter-spacing: 0;
 	}
-	.app-user-menu [data-slot="dropdown-menu-item"] svg {
+	.app-user-menu [data-slot="dropdown-menu-item"] svg,
+	.app-user-menu [data-slot="dropdown-menu-sub-trigger"] svg {
 		width: 19px !important;
 		height: 19px !important;
 		flex-shrink: 0;
 	}
-	.app-user-menu [data-slot="dropdown-menu-item"]:focus {
+	.app-user-menu [data-slot="dropdown-menu-item"]:focus,
+	.app-user-menu [data-slot="dropdown-menu-sub-trigger"]:focus,
+	.app-user-menu [data-slot="dropdown-menu-sub-trigger"][data-state="open"] {
 		background-color: var(--nav-surface-hover);
 		color: #000;
 	}
-	.dark .app-user-menu [data-slot="dropdown-menu-item"]:focus {
+	.dark .app-user-menu [data-slot="dropdown-menu-item"]:focus,
+	.dark .app-user-menu [data-slot="dropdown-menu-sub-trigger"]:focus,
+	.dark .app-user-menu [data-slot="dropdown-menu-sub-trigger"][data-state="open"] {
 		color: #fff;
 	}
-	.app-user-menu [data-slot="dropdown-menu-item"]:focus * {
+	.app-user-menu [data-slot="dropdown-menu-item"]:focus *,
+	.app-user-menu [data-slot="dropdown-menu-sub-trigger"]:focus *,
+	.app-user-menu [data-slot="dropdown-menu-sub-trigger"][data-state="open"] * {
 		color: #000 !important;
 	}
-	.dark .app-user-menu [data-slot="dropdown-menu-item"]:focus * {
+	.dark .app-user-menu [data-slot="dropdown-menu-item"]:focus *,
+	.dark .app-user-menu [data-slot="dropdown-menu-sub-trigger"]:focus *,
+	.dark .app-user-menu [data-slot="dropdown-menu-sub-trigger"][data-state="open"] * {
 		color: #fff !important;
 	}
 
@@ -800,7 +876,7 @@
 	}
 
 	.composer-pill-btn {
-		@apply flex items-center gap-1.5 rounded-full px-1.5 py-1.5 text-[13px] font-medium text-muted-foreground/70 transition-colors hover:bg-primary/10 dark:hover:bg-white/[0.08] disabled:cursor-not-allowed disabled:opacity-40;
+		@apply flex cursor-pointer items-center gap-1.5 rounded-full px-1.5 py-1.5 text-[13px] font-medium text-muted-foreground/70 transition-colors hover:bg-primary/10 dark:hover:bg-white/[0.08] disabled:cursor-not-allowed disabled:opacity-40;
 	}
 	.composer-pill-btn[data-active="true"] {
 		color: var(--primary);
@@ -815,10 +891,161 @@
 	}
 
 	.composer-footer-note {
-		@apply mt-1.5 text-center text-[11px] tracking-[0.04em] text-muted-foreground;
+		@apply mt-1.5 text-center text-[11px] tracking-[0em] text-muted-foreground;
 		font-family: var(--font-sans);
 	}
 
+	.artifact-panel-shell {
+		box-shadow: 0 2px 8px -2px rgba(27, 27, 31, 0.16);
+	}
+
+	.dark .artifact-panel-shell {
+		box-shadow: none;
+	}
+
+	.chat-artifact-split[data-artifact-layout-animating="true"] > [data-panel] {
+		transition:
+			flex-basis 260ms var(--ease-out-cubic),
+			flex-grow 260ms var(--ease-out-cubic),
+			flex-shrink 260ms var(--ease-out-cubic);
+		will-change: flex-basis, flex-grow;
+	}
+
+	.chat-artifact-pop-surface {
+		pointer-events: none;
+		opacity: 0;
+		transform: scale(0.965) translateY(8px);
+		transform-origin: center center;
+		transition:
+			opacity 180ms var(--ease-out-cubic),
+			transform 220ms var(--ease-out-cubic);
+		will-change: opacity, transform;
+	}
+
+	.chat-artifact-pop-surface[data-artifact-surface-visible="true"] {
+		pointer-events: auto;
+		opacity: 1;
+		transform: scale(1) translateY(0);
+	}
+
+	@media (prefers-reduced-motion: reduce) {
+		.chat-artifact-pop-surface {
+			transition: none;
+			transform: none;
+		}
+	}
+
+	.artifact-card-shimmer {
+		background: linear-gradient(
+			105deg,
+			transparent 0%,
+			color-mix(in oklch, var(--muted-foreground) 4%, transparent) 40%,
+			color-mix(in oklch, var(--muted-foreground) 8%, transparent) 50%,
+			color-mix(in oklch, var(--muted-foreground) 4%, transparent) 60%,
+			transparent 100%
+		);
+		transform: translateX(-120%);
+		animation: artifact-card-shimmer 1.55s var(--ease-out-cubic) infinite;
+	}
+
+	.artifact-loading-line {
+		width: 48%;
+		background: color-mix(in oklch, var(--primary) 88%, transparent);
+		transform: translate3d(-125%, 0, 0) scaleX(0.78);
+		transform-origin: center center;
+		animation: artifact-loading-line 1.7s linear infinite;
+	}
+
+	.artifact-panel-shell::before,
+	.artifact-panel-shell::after {
+		content: "";
+		position: absolute;
+		z-index: 20;
+		pointer-events: none;
+		transition:
+			border-color 150ms var(--ease-out-cubic),
+			background-color 150ms var(--ease-out-cubic),
+			box-shadow 150ms var(--ease-out-cubic),
+			transform 150ms var(--ease-out-cubic);
+	}
+
+	.artifact-panel-shell::before {
+		inset: 0;
+		border: 2px solid transparent;
+		border-radius: inherit;
+		-webkit-mask: linear-gradient(90deg, #000 0 24px, transparent 24px);
+		mask: linear-gradient(90deg, #000 0 24px, transparent 24px);
+	}
+
+	.artifact-panel-shell::after {
+		top: 50%;
+		left: -2px;
+		height: 28px;
+		width: 4px;
+		border-radius: 999px;
+		background: transparent;
+		transform: translateY(-50%);
+	}
+
+	[data-slot="resizable-handle"]:hover
+		+ [data-slot="resizable-panel"]
+		.artifact-panel-shell::before,
+	[data-slot="resizable-handle"]:active
+		+ [data-slot="resizable-panel"]
+		.artifact-panel-shell::before,
+	[data-slot="resizable-handle"][data-resize-handle-state="hover"]
+		+ [data-slot="resizable-panel"]
+		.artifact-panel-shell::before,
+	[data-slot="resizable-handle"][data-resize-handle-state="drag"]
+		+ [data-slot="resizable-panel"]
+		.artifact-panel-shell::before {
+		border-color: color-mix(in oklch, var(--primary) 58%, var(--border));
+	}
+
+	[data-slot="resizable-handle"]:hover
+		+ [data-slot="resizable-panel"]
+		.artifact-panel-shell::after,
+	[data-slot="resizable-handle"]:active
+		+ [data-slot="resizable-panel"]
+		.artifact-panel-shell::after,
+	[data-slot="resizable-handle"][data-resize-handle-state="hover"]
+		+ [data-slot="resizable-panel"]
+		.artifact-panel-shell::after,
+	[data-slot="resizable-handle"][data-resize-handle-state="drag"]
+		+ [data-slot="resizable-panel"]
+		.artifact-panel-shell::after {
+		background: color-mix(in oklch, var(--primary) 58%, var(--border));
+		box-shadow: none;
+		transform: translateY(-50%) scaleY(1.06);
+	}
+
+	@keyframes artifact-card-shimmer {
+		to {
+			transform: translateX(120%);
+		}
+	}
+
+	@keyframes artifact-loading-line {
+		0% {
+			transform: translate3d(-125%, 0, 0) scaleX(0.78);
+		}
+		35% {
+			transform: translate3d(-18%, 0, 0) scaleX(0.96);
+		}
+		62% {
+			transform: translate3d(72%, 0, 0) scaleX(1.12);
+		}
+		82% {
+			transform: translate3d(132%, 0, 0) scaleX(1.26);
+		}
+		94% {
+			transform: translate3d(188%, 0, 0) scaleX(1.36);
+		}
+		100% {
+			transform: translate3d(220%, 0, 0) scaleX(1.42);
+		}
+	}
+
 	/* Fine-tuning Studio: equal default height, expandable when needed (md+) */
 	.min-h-studio-config-column {
 		@apply md:min-h-[470px];
@@ -1135,24 +1362,16 @@
 	background: #23252a;
 }
 
-[data-sidebar="content"] {
-	scrollbar-color: oklch(0.5 0 0 / 0.22) var(--sidebar);
-}
 
-.dark [data-sidebar="content"] {
-	scrollbar-color: oklch(0.72 0 0 / 0.25) var(--sidebar);
+/* Search list: re-enable the native scrollbar (CommandList defaults to no-scrollbar). */
+.cmd-native-scrollbar {
+	scrollbar-width: thin;
+	-ms-overflow-style: auto;
 }
-
-[data-sidebar="content"]::-webkit-scrollbar-track {
-	background: var(--sidebar);
-}
-
-[data-sidebar="content"]::-webkit-scrollbar-thumb {
-	background: oklch(0.5 0 0 / 0.22);
-}
-
-.dark [data-sidebar="content"]::-webkit-scrollbar-thumb {
-	background: oklch(0.72 0 0 / 0.25);
+.cmd-native-scrollbar::-webkit-scrollbar {
+	display: block;
+	width: 8px;
+	height: 8px;
 }
 
 .preview-scrollbar {

studio/install_llama_prebuilt.py

@@ -169,8 +169,17 @@ DEFAULT_MAX_MACOS_RELEASE_FALLBACKS = env_int(
     16,
     minimum = 1,
 )
+# Deterministic macOS pin. At b9428 ggml-org's macOS runner moved to macOS 26
+# (Tahoe), so b9428+ prebuilts only load on macOS 26+. b9415 is the last build
+# stamped below 26 (arm64 minos 14, x64 minos 13.3); loads on macOS 13.3/14/15/26.
+_PINNED_MACOS_FALLBACK_TAG = "b9415"
+_PINNED_MACOS_LATEST_FLOOR = (26, 0)
 FORCE_COMPILE_DEFAULT_REF = os.environ.get("UNSLOTH_LLAMA_FORCE_COMPILE_REF", "master")
 
+# sm_103 (B300 / GB300 Blackwell Ultra) is not built natively but runs on the
+# bundled base compute_100 PTX, which the driver JIT-compiles forward to sm_103.
+# It is listed in every bundle that ships the sm_100 build (the "newer" and
+# "portable" classes) so those hosts get a prebuilt instead of a source compile.
 DIRECT_LINUX_BUNDLE_PROFILES: dict[str, dict[str, Any]] = {
     "cuda12-older": {
         "runtime_line": "cuda12",
@@ -183,7 +192,7 @@ DIRECT_LINUX_BUNDLE_PROFILES: dict[str, dict[str, Any]] = {
     "cuda12-newer": {
         "runtime_line": "cuda12",
         "coverage_class": "newer",
-        "supported_sms": ["86", "89", "90", "100", "120"],
+        "supported_sms": ["86", "89", "90", "100", "103", "120"],
         "min_sm": 86,
         "max_sm": 120,
         "rank": 20,
@@ -191,7 +200,7 @@ DIRECT_LINUX_BUNDLE_PROFILES: dict[str, dict[str, Any]] = {
     "cuda12-portable": {
         "runtime_line": "cuda12",
         "coverage_class": "portable",
-        "supported_sms": ["70", "75", "80", "86", "89", "90", "100", "120"],
+        "supported_sms": ["70", "75", "80", "86", "89", "90", "100", "103", "120"],
         "min_sm": 70,
         "max_sm": 120,
         "rank": 30,
@@ -207,7 +216,7 @@ DIRECT_LINUX_BUNDLE_PROFILES: dict[str, dict[str, Any]] = {
     "cuda13-newer": {
         "runtime_line": "cuda13",
         "coverage_class": "newer",
-        "supported_sms": ["86", "89", "90", "100", "120"],
+        "supported_sms": ["86", "89", "90", "100", "103", "120"],
         "min_sm": 86,
         "max_sm": 120,
         "rank": 50,
@@ -215,13 +224,81 @@ DIRECT_LINUX_BUNDLE_PROFILES: dict[str, dict[str, Any]] = {
     "cuda13-portable": {
         "runtime_line": "cuda13",
         "coverage_class": "portable",
-        "supported_sms": ["75", "80", "86", "89", "90", "100", "120"],
+        "supported_sms": ["75", "80", "86", "89", "90", "100", "103", "120"],
         "min_sm": 75,
         "max_sm": 120,
         "rank": 60,
     },
 }
 
+# Lowest CUDA major we ship prebuilts for, and the highest major we probe for
+# installed runtime libraries. Detection and runtime-line derivation are
+# generated per major so a new toolkit (cuda14, ...) needs no code change while
+# llama.cpp keeps the cudart64_<major>.dll / libcudart.so.<major> naming.
+_MIN_CUDA_MAJOR = 12
+_MAX_PROBE_CUDA_MAJOR = 19
+
+# Last ggml-org release whose Windows win-cuda-13 build is still sub-13.3
+# (cuda-13.1, b9360, 2026-05-27). Upstream bumped win-cuda-13 to 13.3 at b9365
+# and now ships only cuda-12.4 + cuda-13.3. cuda-12.4 predates Blackwell (ggml
+# compiles sm_120 only at toolkit >= 12.8), so a Blackwell host on a 13.0/13.1/13.2
+# driver is gated off 13.3 and would drop to a CPU-only 12.4 build. b9360 is
+# immutable, so we pin its cuda-13.1 build (plus paired cudart) as a GPU
+# fallback for exactly those hosts. See unslothai/unsloth#5887.
+_PINNED_BLACKWELL_FALLBACK_TAG = "b9360"
+_PINNED_BLACKWELL_FALLBACK_RUNTIME = "13.1"
+# Floor at 13.0: b9360 ships native sm_120a SASS (no PTX/JIT) and a bundled
+# cuda-13.1 cudart, both of which run on a CUDA 13.0 r580+ driver via CUDA
+# minor-version compatibility, so the mainstream 13.0 Blackwell branch is covered.
+_PINNED_BLACKWELL_DRIVER_FLOOR = (13, 0)
+_BLACKWELL_MIN_SM = 120
+# ggml compiles Blackwell sm_120 only at toolkit >= 12.8, so an in-release
+# windows-cuda build at or above this already covers Blackwell and makes the
+# older pinned 13.1 fallback unnecessary (cuda-12.4 is below it).
+_BLACKWELL_MIN_TOOLKIT = (12, 8)
+_PINNED_BLACKWELL_LLAMA_SHA256 = (
+    "31ddb8b42d7ab4a47cab8c48c397519f580ca502df7e73f3ab396eacc16c8e8d"
+)
+_PINNED_BLACKWELL_CUDART_SHA256 = (
+    "f96935e7e385e3b2d0189239077c10fe8fd7e95690fea4afec455b1b6c7e3f18"
+)
+
+
+def _cuda_runtime_lines_for_major(major: int) -> list[str]:
+    """Runtime lines a driver of this CUDA major can use, newest major first
+    down to the minimum we ship. A driver runs its own major and any older one
+    (backward compatibility)."""
+    return [f"cuda{m}" for m in range(major, _MIN_CUDA_MAJOR - 1, -1)]
+
+
+def _resolve_linux_bundle_profile(bundle_profile: str) -> "dict[str, Any] | None":
+    """Profile (runtime line + sm coverage) for a linux-x64-cuda<major>-<class>
+    bundle. Known majors use their published coverage; an unknown future major
+    reuses the newest known major's coverage for the same class as a forward
+    default, with the post-build GPU smoke test as the backstop."""
+    known = DIRECT_LINUX_BUNDLE_PROFILES.get(bundle_profile)
+    if known is not None:
+        return known
+    m = re.fullmatch(
+        r"cuda(?P<major>\d+)-(?P<klass>older|newer|portable)", bundle_profile
+    )
+    if not m:
+        return None
+    base_key = max(
+        (
+            k
+            for k, v in DIRECT_LINUX_BUNDLE_PROFILES.items()
+            if v["coverage_class"] == m.group("klass")
+        ),
+        key = lambda k: int(re.match(r"cuda(\d+)-", k).group(1)),
+        default = None,
+    )
+    if base_key is None:
+        return None
+    profile = dict(DIRECT_LINUX_BUNDLE_PROFILES[base_key])
+    profile["runtime_line"] = f"cuda{m.group('major')}"
+    return profile
+
 
 @dataclass
 class HostInfo:
@@ -769,6 +846,26 @@ def windows_cuda_asset_aliases(
     return aliases
 
 
+def _published_windows_cuda_runtime(
+    upstream_assets: dict[str, str], major: int, driver: tuple[int, int] | None
+) -> str | None:
+    """Highest cuda-<major>.<minor> published upstream that `driver` can run by
+    default CUDA compatibility, i.e. (major, minor) <= driver. None if nothing
+    qualifies. Gating on the driver (not just the major) keeps a 13.3 build off
+    a driver that only advertises 13.1, where it would otherwise rely on the
+    unguaranteed minor-version-compatibility path."""
+    if driver is None:
+        return None
+    best: int | None = None
+    for name in upstream_assets:
+        m = re.search(r"-bin-win-cuda-(\d+)\.(\d+)-x64\.zip$", name)
+        if m and int(m.group(1)) == major:
+            minor = int(m.group(2))
+            if (major, minor) <= driver and (best is None or minor > best):
+                best = minor
+    return f"{major}.{best}" if best is not None else None
+
+
 def format_byte_count(num_bytes: float) -> str:
     units = ["B", "KiB", "MiB", "GiB", "TiB"]
     value = float(num_bytes)
@@ -1225,7 +1322,7 @@ def parse_direct_linux_release_bundle(
     inferred_labels: list[str] = []
 
     linux_asset_re = re.compile(
-        r"^app-(?P<label>.+)-(?P<target>linux-x64(?:-cpu)?|linux-x64-(?:cuda12|cuda13)-(?:older|newer|portable))\.tar\.gz$"
+        r"^app-(?P<label>.+)-(?P<target>linux-x64(?:-cpu)?|linux-x64-cuda\d+-(?:older|newer|portable))\.tar\.gz$"
     )
     for asset_name in sorted(assets):
         match = linux_asset_re.fullmatch(asset_name)
@@ -1250,7 +1347,7 @@ def parse_direct_linux_release_bundle(
             continue
 
         bundle_profile = target.removeprefix("linux-x64-")
-        profile = DIRECT_LINUX_BUNDLE_PROFILES.get(bundle_profile)
+        profile = _resolve_linux_bundle_profile(bundle_profile)
         if profile is None:
             continue
         artifacts.append(
@@ -1312,7 +1409,16 @@ def direct_linux_release_plan(
 
     attempts: list[AssetChoice] = []
     if host.has_usable_nvidia:
-        selection = linux_cuda_choice_from_release(host, bundle)
+        # Prefer the cudart major Studio loads at runtime (torch's bundled
+        # libcudart), not the newest detected on disk. Without this a stray
+        # cuda13 runtime outranks the torch cuda12 the binary links against.
+        torch_preference = detect_torch_cuda_runtime_preference(host)
+        selection = linux_cuda_choice_from_release(
+            host,
+            bundle,
+            preferred_runtime_line = torch_preference.runtime_line,
+            selection_preamble = torch_preference.selection_log,
+        )
         if selection is not None:
             attempts.extend(selection.attempts)
     if host.has_rocm and not host.has_usable_nvidia:
@@ -1400,6 +1506,11 @@ def direct_upstream_release_plan(
                     torch_preference.selection_log,
                 )
             )
+            # Blackwell on a 13.1/13.2 driver: prefer the pinned cuda-13.1 GPU
+            # build over the CPU-only cuda-12.4 the in-release gating leaves.
+            pinned = _pinned_windows_cuda_fallback(host, attempts)
+            if pinned is not None:
+                attempts.insert(0, pinned)
         elif host.has_rocm:
             lemonade_choice = resolve_lemonade_rocm_choice(
                 host, "windows", "windows-hip", llama_tag = requested_tag
@@ -1526,6 +1637,24 @@ def direct_upstream_release_plan(
     )
 
 
+def pinned_macos_release_tag(host: HostInfo, repo: str) -> str | None:
+    """Pin b9415 (the last upstream macOS build that loads below macOS 26) for a
+    known pre-26 host on ggml-org upstream; return None to keep latest selection.
+    The unslothai/llama.cpp fork ships its own prebuilts (arm64 minos 14, x64
+    minos 13.3) and needs no pin, so this is a no-op there and for macOS 26+,
+    unknown version, non-macOS."""
+    if repo != UPSTREAM_REPO:
+        return None
+    if not host.is_macos:
+        return None
+    version = host.macos_version
+    if version is None:
+        return None
+    if version >= _PINNED_MACOS_LATEST_FLOOR:
+        return None
+    return _PINNED_MACOS_FALLBACK_TAG
+
+
 def resolve_simple_install_release_plans(
     llama_tag: str,
     host: HostInfo,
@@ -1536,18 +1665,27 @@ def resolve_simple_install_release_plans(
 ) -> tuple[str, list[InstallReleasePlan]]:
     repo = published_repo or DEFAULT_PUBLISHED_REPO
     requested_tag = normalized_requested_llama_tag(llama_tag)
+    # The unslothai/llama.cpp fork ships only linux-x64 bundles. An arm64 Linux
+    # host with a GPU (GH200/GB200/DGX Spark) routes here; it must not install an
+    # x64 binary, so fall back to a source build that targets the GPU rather than
+    # selecting the wrong arch (or silently dropping to a CPU arm64 build).
+    if host.is_linux and not host.is_x86_64 and repo == DEFAULT_PUBLISHED_REPO:
+        raise PrebuiltFallback(
+            f"{repo} ships only linux-x64 prebuilts; "
+            f"{host.machine or 'non-x64'} Linux falls back to source build"
+        )
     allow_older_release_fallback = (
         requested_tag == "latest" and not published_release_tag
     )
+    # macOS: pin the last upstream build that loads on a pre-26 host instead of
+    # fetching the latest (macOS 26 only) build and walking back release by
+    # release. No-op on macOS 26+, unknown version, non-macOS, and the fork.
+    if allow_older_release_fallback:
+        pinned_macos = pinned_macos_release_tag(host, repo)
+        if pinned_macos is not None:
+            requested_tag = pinned_macos
+            allow_older_release_fallback = False
     release_limit = max(1, max_release_fallbacks)
-    # macOS may need to walk past a run of too-new prebuilts. Only when the host
-    # version is known; otherwise keep the default (cannot tell up front).
-    if (
-        host.is_macos
-        and allow_older_release_fallback
-        and host.macos_version is not None
-    ):
-        release_limit = max(release_limit, DEFAULT_MAX_MACOS_RELEASE_FALLBACKS)
     plans: list[InstallReleasePlan] = []
     last_error: PrebuiltFallback | None = None
 
@@ -1767,8 +1905,8 @@ def linux_runtime_dirs_for_required_libraries(
 
 def detected_linux_runtime_lines() -> tuple[list[str], dict[str, list[str]]]:
     line_requirements = {
-        "cuda13": ["libcudart.so.13", "libcublas.so.13"],
-        "cuda12": ["libcudart.so.12", "libcublas.so.12"],
+        f"cuda{m}": [f"libcudart.so.{m}", f"libcublas.so.{m}"]
+        for m in range(_MAX_PROBE_CUDA_MAJOR, _MIN_CUDA_MAJOR - 1, -1)
     }
     detected: list[str] = []
     runtime_dirs: dict[str, list[str]] = {}
@@ -2951,6 +3089,46 @@ def detect_host() -> HostInfo:
     )
 
 
+def _normalize_forwarded_gfx(value: str | None) -> str | None:
+    """Extract a single gfx token from a forwarded --rocm-gfx / env value.
+    setup.sh/setup.ps1 already picked the active GPU, so take the token as-is
+    without re-applying visible-device selection. Ignore anything malformed."""
+    if not value:
+        return None
+    m = re.search(r"gfx[1-9][0-9a-z]{2,3}", value.lower())
+    return m.group(0) if m else None
+
+
+def _apply_host_overrides(
+    host: HostInfo,
+    *,
+    override_has_rocm: bool = False,
+    override_rocm_gfx: str | None = None,
+    force_cpu: bool = False,
+) -> HostInfo:
+    """Fold setup.sh/setup.ps1's forwarded detection into the host profile.
+    A forwarded gfx (--rocm-gfx or UNSLOTH_ROCM_GFX_ARCH) is authoritative and
+    implies ROCm: the installer's own hipinfo/amd-smi probe can miss the arch on
+    amd-smi-only hosts or when setup inferred it from the GPU name, leaving
+    rocm_gfx_target None and no lemonade prebuilt selected. force_cpu is the
+    opposite explicit signal (arm64 Linux GPU host whose source build failed):
+    drop GPU attributes so the CPU prebuilt for this OS/arch is selected."""
+    if force_cpu:
+        return dataclasses_replace(
+            host,
+            has_usable_nvidia = False,
+            has_physical_nvidia = False,
+            has_rocm = False,
+            rocm_gfx_target = None,
+        )
+    gfx = _normalize_forwarded_gfx(override_rocm_gfx)
+    if gfx:
+        return dataclasses_replace(host, has_rocm = True, rocm_gfx_target = gfx)
+    if override_has_rocm and not host.has_rocm:
+        return dataclasses_replace(host, has_rocm = True)
+    return host
+
+
 def pick_windows_cuda_runtime(host: HostInfo) -> str | None:
     if not host.driver_cuda_version:
         return None
@@ -2966,17 +3144,21 @@ def compatible_linux_runtime_lines(host: HostInfo) -> list[str]:
     if not host.driver_cuda_version:
         return []
     major, _minor = host.driver_cuda_version
-    if major >= 13:
-        return ["cuda13", "cuda12"]
-    if major >= 12:
-        return ["cuda12"]
-    return []
+    if major < _MIN_CUDA_MAJOR:
+        return []
+    return _cuda_runtime_lines_for_major(major)
 
 
 def windows_runtime_line_info() -> dict[str, tuple[str, ...]]:
+    # Generated per CUDA major (newest first) so a new toolkit is detected
+    # without a code change while the cudart64_<major>.dll naming holds.
     return {
-        "cuda13": ("cudart64_13*.dll", "cublas64_13*.dll", "cublasLt64_13*.dll"),
-        "cuda12": ("cudart64_12*.dll", "cublas64_12*.dll", "cublasLt64_12*.dll"),
+        f"cuda{m}": (
+            f"cudart64_{m}*.dll",
+            f"cublas64_{m}*.dll",
+            f"cublasLt64_{m}*.dll",
+        )
+        for m in range(_MAX_PROBE_CUDA_MAJOR, _MIN_CUDA_MAJOR - 1, -1)
     }
 
 
@@ -2993,12 +3175,13 @@ def detected_windows_runtime_lines() -> tuple[list[str], dict[str, list[str]]]:
 
 
 def compatible_windows_runtime_lines(host: HostInfo) -> list[str]:
-    driver_runtime = pick_windows_cuda_runtime(host)
-    if driver_runtime == "13.1":
-        return ["cuda13", "cuda12"]
-    if driver_runtime == "12.4":
-        return ["cuda12"]
-    return []
+    if not host.driver_cuda_version:
+        return []
+    major, minor = host.driver_cuda_version
+    # cuda12 prebuilts need a 12.4+ driver; cuda13+ any minor of the major.
+    if major < _MIN_CUDA_MAJOR or (major == _MIN_CUDA_MAJOR and minor < 4):
+        return []
+    return _cuda_runtime_lines_for_major(major)
 
 
 def runtime_line_from_cuda_version(cuda_version: str | None) -> str | None:
@@ -3075,7 +3258,6 @@ def windows_cuda_attempts(
     selection_preamble: Iterable[str] = (),
 ) -> list[AssetChoice]:
     selection_log = list(selection_preamble)
-    runtime_by_line = {"cuda12": "12.4", "cuda13": "13.1"}
     driver_runtime = pick_windows_cuda_runtime(host)
     detected_runtime_lines, runtime_dirs = detected_windows_runtime_lines()
     compatible_runtime_lines = compatible_windows_runtime_lines(host)
@@ -3118,12 +3300,7 @@ def windows_cuda_attempts(
             selection_log.append(
                 "windows_cuda_selection: detected CUDA runtime DLLs were incompatible with the reported driver"
             )
-        fallback_runtime_lines = (
-            ["cuda13", "cuda12"]
-            if driver_runtime == "13.1"
-            else (["cuda12"] if driver_runtime == "12.4" else [])
-        )
-        normal_runtime_lines = fallback_runtime_lines
+        normal_runtime_lines = compatible_runtime_lines
 
     runtime_order: list[str] = []
     if preferred_runtime_line and preferred_runtime_line in normal_runtime_lines:
@@ -3147,6 +3324,13 @@ def windows_cuda_attempts(
         for runtime_line in normal_runtime_lines
         if runtime_line not in runtime_order
     )
+    # Keep every driver-compatible line reachable as a fallback, so a line gated
+    # out by the driver version still drops to an older major (cuda13 -> cuda12).
+    runtime_order.extend(
+        runtime_line
+        for runtime_line in compatible_runtime_lines
+        if runtime_line not in runtime_order
+    )
     selection_log.append(
         "windows_cuda_selection: normal_runtime_order="
         + (",".join(normal_runtime_lines) if normal_runtime_lines else "none")
@@ -3158,7 +3342,18 @@ def windows_cuda_attempts(
 
     attempts: list[AssetChoice] = []
     for runtime_line in runtime_order:
-        runtime = runtime_by_line[runtime_line]
+        major = int(runtime_line.removeprefix("cuda"))
+        # Track whatever minor llama.cpp actually ships for this major
+        # (cuda13 -> 13.1, 13.3, ...). Skip the line when the release has no
+        # matching asset instead of guessing a now-missing name.
+        runtime = _published_windows_cuda_runtime(
+            upstream_assets, major, host.driver_cuda_version
+        )
+        if runtime is None:
+            selection_log.append(
+                f"windows_cuda_selection: no driver-supported asset for {runtime_line}"
+            )
+            continue
         selected_name = None
         asset_url = None
         for candidate_name in windows_cuda_upstream_asset_names(llama_tag, runtime):
@@ -3213,6 +3408,110 @@ def windows_cuda_attempts(
     return attempts
 
 
+def _windows_cuda_attempt_covers_blackwell(attempt: AssetChoice) -> bool:
+    """True if an in-release windows-cuda attempt is built with a toolkit that
+    covers Blackwell sm_120 (>= 12.8), read from its asset name's CUDA minor."""
+    if attempt.install_kind != "windows-cuda":
+        return False
+    m = re.search(r"-bin-win-cuda-(\d+)\.(\d+)-x64\.zip$", attempt.name)
+    return (
+        m is not None and (int(m.group(1)), int(m.group(2))) >= _BLACKWELL_MIN_TOOLKIT
+    )
+
+
+def _pinned_windows_cuda_fallback(
+    host: HostInfo, existing_cuda_attempts: list[AssetChoice]
+) -> AssetChoice | None:
+    """Pinned GPU fallback for a Blackwell host the in-release build gates off.
+    Upstream stopped publishing a sub-13.3 Windows cuda13 build after b9360, and
+    cuda-12.4 cannot offload sm_120, so a 13.1/13.2 driver would land on CPU.
+    b9360's cuda-13.1 build is immutable and runs on those drivers. Returns None
+    (dormant) whenever the in-release selection already offers a Blackwell-capable
+    build (toolkit >= 12.8, e.g. a runnable cuda13/cuda14), so it self-disables
+    once upstream ships a driver-runnable build again.
+
+    The b9360 binary reuses the current release's source tree and convert scripts
+    and is recorded via binary_release_tag, the same binary/source split used for
+    the lemonade prebuilt."""
+    if not (host.is_windows and host.is_x86_64 and host.has_usable_nvidia):
+        return None
+    driver = host.driver_cuda_version
+    if driver is None or driver < _PINNED_BLACKWELL_DRIVER_FLOOR:
+        return None
+    caps = normalize_compute_caps(host.compute_caps)
+    if not caps or int(caps[-1]) < _BLACKWELL_MIN_SM:
+        return None
+    if any(
+        _windows_cuda_attempt_covers_blackwell(attempt)
+        for attempt in existing_cuda_attempts
+    ):
+        return None
+    tag = _PINNED_BLACKWELL_FALLBACK_TAG
+    runtime = _PINNED_BLACKWELL_FALLBACK_RUNTIME
+    base = (
+        f"https://github.com/{UPSTREAM_REPO}/releases/download/"
+        f"{urllib.parse.quote(tag, safe = '')}"
+    )
+    name = f"llama-{tag}-bin-win-cuda-{runtime}-x64.zip"
+    cudart_name = f"cudart-llama-bin-win-cuda-{runtime}-x64.zip"
+    return AssetChoice(
+        repo = UPSTREAM_REPO,
+        tag = tag,
+        name = name,
+        url = f"{base}/{name}",
+        source_label = "upstream",
+        install_kind = "windows-cuda",
+        runtime_line = "cuda13",
+        runtime_name = cudart_name,
+        runtime_url = f"{base}/{cudart_name}",
+        expected_sha256 = _PINNED_BLACKWELL_LLAMA_SHA256,
+        runtime_sha256 = _PINNED_BLACKWELL_CUDART_SHA256,
+        selection_log = [
+            f"windows_cuda_selection: pinned {tag} cuda-{runtime} Blackwell GPU "
+            f"fallback (in-release cuda13 gated off by driver "
+            f"{driver[0]}.{driver[1]})"
+        ],
+    )
+
+
+def _augment_checksums_with_pin(
+    checksums: ApprovedReleaseChecksums, pin: AssetChoice
+) -> ApprovedReleaseChecksums:
+    """Add the pin's own verified hashes to a copy of the approved checksums so
+    apply_approved_hashes keeps it on the published path (b9360 is not in the
+    release manifest)."""
+    artifacts = dict(checksums.artifacts)
+    if pin.expected_sha256:
+        artifacts[pin.name] = ApprovedArtifactHash(
+            asset_name = pin.name,
+            sha256 = pin.expected_sha256,
+            repo = pin.repo,
+            kind = "prebuilt",
+        )
+    if pin.runtime_name and pin.runtime_sha256:
+        artifacts[pin.runtime_name] = ApprovedArtifactHash(
+            asset_name = pin.runtime_name,
+            sha256 = pin.runtime_sha256,
+            repo = pin.repo,
+            kind = "prebuilt",
+        )
+    return dataclasses_replace(checksums, artifacts = artifacts)
+
+
+def _with_pinned_windows_cuda_fallback(
+    host: HostInfo,
+    attempts: list[AssetChoice],
+    checksums: ApprovedReleaseChecksums,
+) -> tuple[list[AssetChoice], ApprovedReleaseChecksums]:
+    """Insert the Blackwell pin ahead of the Windows CUDA attempts and keep it
+    through apply_approved_hashes, or return the inputs unchanged when dormant.
+    Gives the published install path the same GPU fallback as the simple path."""
+    pin = _pinned_windows_cuda_fallback(host, attempts)
+    if pin is None:
+        return attempts, checksums
+    return [pin, *attempts], _augment_checksums_with_pin(checksums, pin)
+
+
 def published_windows_cuda_attempts(
     host: HostInfo,
     release: PublishedReleaseBundle,
@@ -3220,13 +3519,26 @@ def published_windows_cuda_attempts(
     selection_preamble: Iterable[str] = (),
 ) -> list[AssetChoice]:
     selection_log = list(release.selection_log) + list(selection_preamble)
-    runtime_by_line = {"cuda12": "12.4", "cuda13": "13.1"}
+    # Seed the runtime-line ordering from the real published windows-cuda minors
+    # (their names encode the minor), so a future CUDA major published here is
+    # ordered too instead of a hardcoded cuda12/cuda13 pair. Keys mirror the
+    # upstream naming so windows_cuda_attempts can match them; fall back to the
+    # long-standing default when the release lists no windows-cuda asset.
+    published_minors: list[str] = []
+    for artifact in release.artifacts:
+        if artifact.install_kind != "windows-cuda":
+            continue
+        m = re.search(r"-bin-win-cuda-(\d+\.\d+)-x64\.zip$", artifact.asset_name)
+        if m:
+            published_minors.append(m.group(1))
+    if not published_minors:
+        published_minors = ["12.4", "13.1"]
     runtime_order = windows_cuda_attempts(
         host,
         release.upstream_tag,
         {
-            f"llama-{release.upstream_tag}-bin-win-cuda-{runtime}-x64.zip": "published"
-            for runtime in runtime_by_line.values()
+            f"llama-{release.upstream_tag}-bin-win-cuda-{minor}-x64.zip": "published"
+            for minor in published_minors
         },
         preferred_runtime_line,
         selection_log,
@@ -3255,11 +3567,20 @@ def published_windows_cuda_attempts(
             asset_url = release.assets.get(artifact.asset_name)
             if not asset_url:
                 continue
-            # See windows_cuda_attempts: pair the cudart bundle.
+            am = re.search(r"-bin-win-cuda-(\d+)\.(\d+)-x64\.zip$", artifact.asset_name)
+            # Gate the real published minor against the driver, so a published
+            # windows-cuda artifact can never bypass the driver-version gate.
+            if (
+                am is not None
+                and host.driver_cuda_version is not None
+                and (int(am.group(1)), int(am.group(2))) > host.driver_cuda_version
+            ):
+                continue
+            # See windows_cuda_attempts: pair the cudart bundle for the real minor.
             runtime_archive_name: str | None = None
             runtime_archive_url: str | None = None
-            if artifact.asset_name.startswith("llama-"):
-                runtime = runtime_by_line[runtime_line]
+            if am is not None and artifact.asset_name.startswith("llama-"):
+                runtime = f"{am.group(1)}.{am.group(2)}"
                 cudart_name = f"cudart-llama-bin-win-cuda-{runtime}-x64.zip"
                 cudart_url = release.assets.get(cudart_name)
                 if cudart_url and cudart_url != asset_url:
@@ -3812,18 +4133,23 @@ def resolve_release_asset_choice(
             torch_preference.selection_log,
         )
         if published_attempts:
+            pin_attempts, pin_checksums = _with_pinned_windows_cuda_fallback(
+                host, published_attempts, checksums
+            )
             try:
-                return apply_approved_hashes(published_attempts, checksums)
+                return apply_approved_hashes(pin_attempts, pin_checksums)
             except PrebuiltFallback as exc:
                 log(
                     "published Windows CUDA assets ignored for install planning: "
                     f"{release.repo}@{release.release_tag} ({exc})"
                 )
         upstream_assets = github_release_assets(UPSTREAM_REPO, llama_tag)
-        return apply_approved_hashes(
+        upstream_attempts, upstream_checksums = _with_pinned_windows_cuda_fallback(
+            host,
             resolve_windows_cuda_choices(host, llama_tag, upstream_assets),
             checksums,
         )
+        return apply_approved_hashes(upstream_attempts, upstream_checksums)
 
     published_choice: AssetChoice | None = None
     if host.is_windows and host.is_x86_64:
@@ -5065,9 +5391,10 @@ def preflight_macos_installed_binaries(
     install_dir: Path,
     host: HostInfo,
 ) -> None:
-    """Reject a macos prebuilt whose minimum-OS is newer than the host so the
-    release walk-back advances to the newest compatible release. No-op when the
-    host macOS version is unknown (runtime validation remains the backstop)."""
+    """Reject a macos prebuilt whose minimum-OS is newer than the host. The
+    upstream selector pins a loadable release up front, so here this is the
+    post-download backstop; the published/fork path also uses it to advance the
+    walk-back. No-op when the host macOS version is unknown (runtime validates)."""
     if not host.is_macos or host.macos_version is None:
         return
     issues = macos_binary_minos_issues(binaries, install_dir, host)
@@ -6177,10 +6504,16 @@ def install_prebuilt(
     *,
     simple_policy: bool = False,
     override_has_rocm: bool = False,
+    override_rocm_gfx: str | None = None,
+    force_cpu: bool = False,
 ) -> None:
     host = detect_host()
-    if override_has_rocm and not host.has_rocm:
-        host = dataclasses_replace(host, has_rocm = True)
+    host = _apply_host_overrides(
+        host,
+        override_has_rocm = override_has_rocm,
+        override_rocm_gfx = override_rocm_gfx,
+        force_cpu = force_cpu,
+    )
     choice: AssetChoice | None = None
     try:
         with install_lock(install_lock_path(install_dir)):
@@ -6325,6 +6658,26 @@ def parse_args() -> argparse.Namespace:
             "so the HIP llama.cpp prebuilt is selected even when hipinfo is not on PATH."
         ),
     )
+    parser.add_argument(
+        "--rocm-gfx",
+        default = os.environ.get("UNSLOTH_ROCM_GFX_ARCH"),
+        help = (
+            "Forward the AMD gfx target (e.g. gfx1151) that setup.ps1/setup.sh "
+            "resolved, so the lemonade HIP prebuilt is selected even when the "
+            "installer's own hipinfo/amd-smi probe cannot report it. Implies "
+            "--has-rocm. Defaults to the UNSLOTH_ROCM_GFX_ARCH environment variable."
+        ),
+    )
+    parser.add_argument(
+        "--cpu-fallback",
+        action = "store_true",
+        default = False,
+        help = (
+            "Select the CPU prebuilt for this OS/arch even when a GPU is present. "
+            "setup.sh uses this as a last resort for arm64 Linux GPU hosts whose "
+            "source build failed (no arm64 CUDA prebuilt exists anywhere)."
+        ),
+    )
     resolve_group = parser.add_mutually_exclusive_group()
     resolve_group.add_argument(
         "--resolve-llama-tag",
@@ -6446,6 +6799,8 @@ def main() -> int:
         published_release_tag = args.published_release_tag or "",
         simple_policy = args.simple_policy,
         override_has_rocm = args.has_rocm,
+        override_rocm_gfx = args.rocm_gfx,
+        force_cpu = args.cpu_fallback,
     )
     return EXIT_SUCCESS
 

studio/setup.ps1

@@ -50,6 +50,12 @@ if ($script:UnslothVerbose) {
     $env:UNSLOTH_VERBOSE = '1'
 }
 $script:LlamaCppDegraded = $false
+# CUDA toolkit state, published by Resolve-CudaToolkit. Only the Phase 4 source
+# build consumes these; the prebuilt path leaves them at these defaults.
+$script:CudaToolkitReady = $false
+$script:NvccPath = $null
+$script:CudaToolkitRoot = $null
+$script:CudaArch = $null
 
 # Detect if running from pip install (no frontend/ dir in studio)
 $FrontendDir = Join-Path $ScriptDir "frontend"
@@ -1050,7 +1056,12 @@ if ($vsResult) {
 # ============================================
 # 1e. CUDA Toolkit (nvcc for llama.cpp build + env vars)
 # ============================================
-if ($HasNvidiaSmi) {
+# Defined here but invoked lazily right before a Phase 4 source build; the
+# prebuilt llama.cpp path needs no local toolkit. With -RequireOrExit a source
+# build is committed, so hard-fail if no driver-compatible toolkit can be found
+# or installed. Without it, detection is best-effort and only sets the flag.
+function Resolve-CudaToolkit {
+    param([switch]$RequireOrExit)
 # IMPORTANT: The CUDA Toolkit version must be <= the max CUDA version the
 # NVIDIA driver supports.  nvidia-smi reports this as "CUDA Version: X.Y".
 # If we install a toolkit newer than the driver supports, llama-server will
@@ -1146,6 +1157,11 @@ if ($DriverMaxCuda) {
 
 # -- If incompatible toolkit is blocking, tell user to uninstall it --
 if (-not $NvccPath -and $IncompatibleToolkit) {
+    if (-not $RequireOrExit) {
+        substep "CUDA Toolkit $IncompatibleToolkit exceeds driver max $DriverMaxCuda -- skipping; prebuilt llama.cpp needs no local toolkit" "Yellow"
+        $script:CudaToolkitReady = $false
+        return
+    }
     Write-Host "" -ForegroundColor Red
     Write-Host "========================================================================" -ForegroundColor Red
     Write-Host "[ERROR] CUDA Toolkit $IncompatibleToolkit is installed but INCOMPATIBLE" -ForegroundColor Red
@@ -1163,8 +1179,8 @@ if (-not $NvccPath -and $IncompatibleToolkit) {
     exit 1
 }
 
-# -- No toolkit at all: install via winget --
-if (-not $NvccPath) {
+# -- No toolkit at all: install via winget (only when a source build needs it) --
+if (-not $NvccPath -and $RequireOrExit) {
     Write-Host "CUDA toolkit (nvcc) not found -- installing via winget..." -ForegroundColor Yellow
     $HasWinget = $null -ne (Get-Command winget -ErrorAction SilentlyContinue)
     if ($HasWinget) {
@@ -1223,6 +1239,11 @@ if (-not $NvccPath) {
 }
 
 if (-not $NvccPath) {
+    if (-not $RequireOrExit) {
+        substep "no driver-compatible CUDA Toolkit found -- skipping; prebuilt llama.cpp needs no local toolkit" "Yellow"
+        $script:CudaToolkitReady = $false
+        return
+    }
     Write-Host "[ERROR] CUDA Toolkit (nvcc) is required but could not be found or installed." -ForegroundColor Red
     if ($DriverMaxCuda) {
         Write-Host "        Install CUDA Toolkit $DriverMaxCuda from https://developer.nvidia.com/cuda-toolkit-archive" -ForegroundColor Yellow
@@ -1313,8 +1334,11 @@ substep "CudaToolkitDir = $CudaToolkitRoot\"
 if (-not $CudaArch) {
     substep "could not detect compute capability -- cmake will use defaults" "Yellow"
 }
-} else {
-    step "cuda" "skipped (no NVIDIA GPU detected)" "Yellow"
+# Publish the resolved toolkit to script scope for the Phase 4 build.
+$script:NvccPath = $NvccPath
+$script:CudaToolkitRoot = $CudaToolkitRoot
+$script:CudaArch = $CudaArch
+$script:CudaToolkitReady = $true
 }
 
 if ($HasROCm) {
@@ -2459,6 +2483,12 @@ if ($env:UNSLOTH_LLAMA_FORCE_COMPILE -eq "1") {
         )
         if ($HasROCm) {
             $prebuiltArgs += "--has-rocm"
+            # Forward the resolved gfx arch so the lemonade HIP prebuilt is picked
+            # even when the installer's own probe cannot report it (amd-smi-only
+            # hosts, name-inferred arch).
+            if ($script:ROCmGfxArch) {
+                $prebuiltArgs += @("--rocm-gfx", $script:ROCmGfxArch)
+            }
         }
         if ($env:UNSLOTH_LLAMA_RELEASE_TAG) {
             $prebuiltArgs += @("--published-release-tag", $env:UNSLOTH_LLAMA_RELEASE_TAG)
@@ -2581,7 +2611,8 @@ if ($NeedLlamaSourceBuild) {
 # We build:
 #   - llama-server:   for GGUF model inference (with HTTPS if OpenSSL available)
 #   - llama-quantize: for GGUF export quantization
-# Prerequisites (git, cmake, VS Build Tools, CUDA Toolkit) already installed in Phase 1.
+# Prerequisites git, cmake, VS Build Tools were installed in Phase 1; the CUDA
+# Toolkit is resolved lazily just below via Resolve-CudaToolkit (source build only).
 $OriginalLlamaCppDir = $LlamaCppDir
 $BuildDir = Join-Path $LlamaCppDir "build"
 $LlamaServerBin = Join-Path $BuildDir "bin\Release\llama-server.exe"
@@ -2627,6 +2658,10 @@ if (-not $NeedLlamaSourceBuild) {
     substep "Install CMake from https://cmake.org/download/ and re-run setup." "Yellow"
     $script:LlamaCppDegraded = $true
 } else {
+    # A source build is committed here. The CUDA toolkit is only needed now, so
+    # resolve (and winget-install if needed) it lazily, failing fast if no
+    # driver-compatible toolkit exists. The prebuilt path never reaches this.
+    if ($HasNvidiaSmi) { Resolve-CudaToolkit -RequireOrExit }
     Write-Host ""
     if ($HasNvidiaSmi) {
         substep "building llama.cpp with CUDA support..."

studio/setup.sh

@@ -857,6 +857,16 @@ else
     if [ -n "${UNSLOTH_LLAMA_RELEASE_TAG:-}" ]; then
         _PREBUILT_CMD+=(--published-release-tag "$UNSLOTH_LLAMA_RELEASE_TAG")
     fi
+    # Forward the gfx arch resolved above so the lemonade HIP prebuilt is picked
+    # even when the installer's own probe cannot report it (amd-smi-only hosts,
+    # name-inferred arch). Implies --has-rocm on the installer side.
+    if [ -n "${_setup_gfx:-}" ]; then
+        _PREBUILT_CMD+=(--rocm-gfx "$_setup_gfx")
+    elif [ "$_setup_amd_detected" = true ]; then
+        # AMD was detected but gfx resolution failed; tell the installer ROCm is
+        # present so it can still attempt a prebuilt. Mirrors setup.ps1 behaviour.
+        _PREBUILT_CMD+=(--has-rocm)
+    fi
     _PREBUILT_LOG="$(mktemp)"
     set +e
     if _is_verbose; then
@@ -1363,6 +1373,32 @@ else
 }
 fi  # end _SKIP_GGUF_BUILD check
 
+# ── arm64 Linux GPU: CPU prebuilt as a last resort ──
+# arm64 Linux with a GPU has no CUDA prebuilt anywhere (the unslothai fork is
+# x64 only; ggml-org ships no Linux CUDA build), so it source-builds for the
+# GPU above. If that produced no binary, install ggml-org's arm64 CPU prebuilt
+# instead of leaving the host without llama.cpp.
+if [ "$_LLAMA_CPP_DEGRADED" = true ] \
+        && [ "$_HOST_SYSTEM" = "Linux" ] \
+        && { [ "$_HOST_MACHINE" = "aarch64" ] || [ "$_HOST_MACHINE" = "arm64" ]; }; then
+    substep "GPU source build unavailable; trying ggml-org arm64 CPU prebuilt..."
+    _ARM64_CPU_CMD=(
+        python "$SCRIPT_DIR/install_llama_prebuilt.py"
+        --install-dir "$LLAMA_CPP_DIR"
+        --llama-tag "$_REQUESTED_LLAMA_TAG"
+        --published-repo "ggml-org/llama.cpp"
+        --simple-policy
+        --cpu-fallback
+    )
+    # Trust the installer's exit code: it validates the server before exiting 0,
+    # the same signal the primary prebuilt path above relies on.
+    if run_quiet_no_exit "arm64 CPU prebuilt" "${_ARM64_CPU_CMD[@]}"; then
+        step "llama.cpp" "arm64 CPU prebuilt installed (GPU build unavailable)" "$C_WARN"
+        _LLAMA_CPP_DEGRADED=false
+        print_installed_llama_prebuilt_release "$LLAMA_CPP_DIR"
+    fi
+fi
+
 # ── Footer ──
 if [ "$_LLAMA_ONLY" = "1" ]; then
     echo ""

tests/sh/test_mac_intel_compat.sh

@@ -563,6 +563,80 @@ else
     FAIL=$((FAIL + 1))
 fi
 
+echo ""
+echo "=== Apple Silicon x86_64 (Rosetta) venv rebuild ==="
+
+# Extract the real guard block from install.sh so we exercise the shipped logic
+# (comment header down to its column-0 closing fi).
+_GUARD_FILE=$(mktemp)
+awk '/Guard against two independent Apple Silicon venv problems/{f=1} f{print} f&&/^fi$/{exit}' \
+    "$INSTALL_SH" > "$_GUARD_FILE"
+
+if [ ! -s "$_GUARD_FILE" ]; then
+    echo "  FAIL: could not extract Apple Silicon venv guard from install.sh"
+    FAIL=$((FAIL + 1))
+else
+    # Runner: stub uv (via run_install_cmd) + a fake venv python, source the
+    # guard, then print "<final_arch> <final_ver> | <recreate_selectors>".
+    # The stub maps a uv arm64 selector to the interpreter uv would produce:
+    # cpython-3.12-* -> arm64 3.12.7, cpython-3.13-* -> arm64 $REBUILD_313_VERSION.
+    _RUNNER=$(mktemp)
+    cat > "$_RUNNER" << 'RUNNER_EOF'
+GUARD="$1"; VENV_DIR="$2"
+make_python() {  # dir machine version
+    mkdir -p "$1/bin"
+    printf '#!/usr/bin/env bash\necho "%s %s"\n' "$2" "$3" > "$1/bin/python"
+    chmod +x "$1/bin/python"
+}
+RECREATE_LOG=$(mktemp); : > "$RECREATE_LOG"
+run_install_cmd() {
+    shift  # drop the human label
+    if [ "$1" = "uv" ] && [ "$2" = "venv" ]; then
+        dir="$3"; sel=""; shift 3
+        while [ $# -gt 0 ]; do [ "$1" = "--python" ] && { sel="$2"; shift; }; shift; done
+        echo "$sel" >> "$RECREATE_LOG"
+        case "$sel" in
+            *3.12-macos-aarch64*) make_python "$dir" arm64 "3.12.7" ;;
+            *3.13-macos-aarch64*) make_python "$dir" arm64 "${REBUILD_313_VERSION:-3.13.3}" ;;
+            *)                    make_python "$dir" arm64 "$sel" ;;
+        esac
+    fi
+}
+[ "$INIT_ARCH" != none ] && make_python "$VENV_DIR" "$INIT_ARCH" "$INIT_VER"
+PYTHON_VERSION="3.13"
+. "$GUARD" >&2  # guard's user-facing echoes go to stderr; keep stdout clean
+final="none"; [ -x "$VENV_DIR/bin/python" ] && final="$("$VENV_DIR/bin/python" -c x)"
+printf '%s | %s\n' "$final" "$(paste -sd, "$RECREATE_LOG" 2>/dev/null)"
+rm -f "$RECREATE_LOG"
+RUNNER_EOF
+
+    _run_guard() {  # _USER_PYTHON OS _ARCH INIT_ARCH INIT_VER REBUILD_313_VERSION
+        _vd=$(mktemp -d)
+        env _USER_PYTHON="$1" OS="$2" _ARCH="$3" INIT_ARCH="$4" INIT_VER="$5" \
+            REBUILD_313_VERSION="$6" bash "$_RUNNER" "$_GUARD_FILE" "$_vd/venv"
+        rm -rf "$_vd"
+    }
+
+    assert_eq "clean arm64 venv left untouched" \
+        "arm64 3.13.3 | " "$(_run_guard '' macos arm64 arm64 3.13.3 '')"
+    assert_eq "x86_64 venv rebuilt as arm64" \
+        "arm64 3.13.3 | cpython-3.13-macos-aarch64-none" \
+        "$(_run_guard '' macos arm64 x86_64 3.13.3 '')"
+    assert_eq "x86_64 venv that lands on 3.13.8 is rebuilt then downgraded to 3.12" \
+        "arm64 3.12.7 | cpython-3.13-macos-aarch64-none,cpython-3.12-macos-aarch64-none" \
+        "$(_run_guard '' macos arm64 x86_64 3.13.3 3.13.8)"
+    assert_eq "arm64 3.13.8 venv downgraded to 3.12" \
+        "arm64 3.12.7 | cpython-3.12-macos-aarch64-none" \
+        "$(_run_guard '' macos arm64 arm64 3.13.8 '')"
+    assert_eq "--python override skips the guard entirely" \
+        "x86_64 3.13.3 | " "$(_run_guard 3.11 macos arm64 x86_64 3.13.3 '')"
+    assert_eq "x86_64 host (Intel/Rosetta shell) is a no-op here" \
+        "x86_64 3.13.3 | " "$(_run_guard '' macos x86_64 x86_64 3.13.3 '')"
+
+    rm -f "$_RUNNER"
+fi
+rm -f "$_GUARD_FILE"
+
 echo ""
 echo "Results: $PASS passed, $FAIL failed"
 [ "$FAIL" -eq 0 ] || exit 1

tests/studio/install/test_install_llama_prebuilt_logic.py

@@ -438,6 +438,80 @@ def test_simple_linux_direct_release_uses_published_source_checksums_for_branch(
     assert exact_source is True
 
 
+def test_simple_linux_direct_release_honors_torch_cudart_preference(
+    monkeypatch: pytest.MonkeyPatch,
+):
+    # Regression: a Blackwell host (sm_120, driver 13.0) with BOTH cudart majors
+    # visible -- a stray cuda13 wheel plus torch's cuda12 -- must install the
+    # cuda12 build that matches the runtime torch, not the newest-major cuda13
+    # build (which loads no GPU and silently falls back to CPU).
+    release = {
+        "tag_name": "b9334",
+        "assets": [
+            {
+                "name": f"app-b9334-linux-x64-{profile}.tar.gz",
+                "browser_download_url": f"https://example.test/app-b9334-linux-x64-{profile}.tar.gz",
+            }
+            for profile in (
+                "cuda12-newer",
+                "cuda12-portable",
+                "cuda13-newer",
+                "cuda13-portable",
+            )
+        ],
+    }
+    # cuda13 detected first (newest-major order); both compatible with driver 13.0.
+    monkeypatch.setattr(
+        INSTALL_LLAMA_PREBUILT,
+        "detected_linux_runtime_lines",
+        lambda: (
+            ["cuda13", "cuda12"],
+            {
+                "cuda13": ["/usr/local/lib/python3.13/site-packages/nvidia/cu13/lib"],
+                "cuda12": [
+                    "/venv/lib/python3.13/site-packages/nvidia/cuda_runtime/lib"
+                ],
+            },
+        ),
+    )
+    host = HostInfo(
+        system = "Linux",
+        machine = "x86_64",
+        is_windows = False,
+        is_linux = True,
+        is_macos = False,
+        is_x86_64 = True,
+        is_arm64 = False,
+        nvidia_smi = "nvidia-smi",
+        driver_cuda_version = (13, 0),
+        compute_caps = ["120"],
+        visible_cuda_devices = None,
+        has_physical_nvidia = True,
+        has_usable_nvidia = True,
+    )
+
+    def first_asset_for_torch(line):
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "detect_torch_cuda_runtime_preference",
+            lambda h: INSTALL_LLAMA_PREBUILT.CudaRuntimePreference(
+                runtime_line = line, selection_log = []
+            ),
+        )
+        plan = INSTALL_LLAMA_PREBUILT.direct_linux_release_plan(
+            release, host, "unslothai/llama.cpp", "latest"
+        )
+        return plan.attempts[0]
+
+    # torch reports cuda12 (the cu128 runtime) -> install the cuda12 build.
+    primary = first_asset_for_torch("cuda12")
+    assert primary.name == "app-b9334-linux-x64-cuda12-newer.tar.gz"
+    assert primary.runtime_line == "cuda12"
+
+    # torch unavailable -> unchanged newest-major fallback (documents the residual).
+    assert first_asset_for_torch(None).name == "app-b9334-linux-x64-cuda13-newer.tar.gz"
+
+
 @pytest.mark.parametrize(
     "mutate, expected_match",
     [

tests/studio/install/test_macos_version_compat.py

@@ -238,32 +238,46 @@ def _fake_macos_releases(tags):
     ]
 
 
-class TestMacosReleaseWalkback:
-    """A known-version macOS host must generate enough older-release plans to
-    walk back past a run of too-new prebuilts; unknown-version and non-macOS
-    hosts keep the conservative 2-release default."""
+class TestMacosReleasePin:
+    """A known pre-26 macOS host deterministically pins the last upstream release
+    whose prebuilt loads on it (b9415) instead of walking back release by release;
+    macOS 26+ and unknown-version hosts keep normal latest selection with the
+    conservative 2-release default."""
 
-    TAGS = [f"b{n}" for n in range(9437, 9400, -1)]  # 37 newest-first releases
+    TAGS = [f"b{n}" for n in range(9442, 9400, -1)]  # newest-first, includes b9415
 
     def _patch_releases(self, monkeypatch):
-        monkeypatch.setattr(
-            ILP,
-            "iter_release_payloads_by_time",
-            lambda repo, published_release_tag, requested_tag: _fake_macos_releases(
-                self.TAGS
-            ),
-        )
+        def fake_iter(repo, published_release_tag, requested_tag):
+            # The real iterator yields only the requested tag when one is pinned.
+            if requested_tag and requested_tag != "latest":
+                return _fake_macos_releases([requested_tag])
+            return _fake_macos_releases(self.TAGS)
 
-    def test_known_macos_host_walks_back_deeper(self, monkeypatch):
+        monkeypatch.setattr(ILP, "iter_release_payloads_by_time", fake_iter)
+
+    def test_pre26_host_pins_b9415(self, monkeypatch):
         self._patch_releases(monkeypatch)
-        _tag, plans = ILP.resolve_simple_install_release_plans(
+        tag, plans = ILP.resolve_simple_install_release_plans(
             "latest",
             make_macos_host((14, 0)),
             "ggml-org/llama.cpp",
             "",
         )
-        assert len(plans) == ILP.DEFAULT_MAX_MACOS_RELEASE_FALLBACKS
-        assert len(plans) > ILP.DEFAULT_MAX_PREBUILT_RELEASE_FALLBACKS
+        assert tag == ILP._PINNED_MACOS_FALLBACK_TAG == "b9415"
+        assert len(plans) == 1
+        assert plans[0].release_tag == "b9415"
+
+    def test_tahoe_host_takes_latest(self, monkeypatch):
+        self._patch_releases(monkeypatch)
+        tag, plans = ILP.resolve_simple_install_release_plans(
+            "latest",
+            make_macos_host((26, 0)),
+            "ggml-org/llama.cpp",
+            "",
+        )
+        assert tag == "latest"
+        assert plans[0].release_tag == self.TAGS[0]  # newest release
+        assert len(plans) == ILP.DEFAULT_MAX_PREBUILT_RELEASE_FALLBACKS
 
     def test_unknown_macos_host_uses_default(self, monkeypatch):
         self._patch_releases(monkeypatch)

tests/studio/install/test_rocm_support.py

@@ -35,6 +35,8 @@ AssetChoice = prebuilt_mod.AssetChoice
 PrebuiltFallback = prebuilt_mod.PrebuiltFallback
 resolve_upstream_asset_choice = prebuilt_mod.resolve_upstream_asset_choice
 runtime_patterns_for_choice = prebuilt_mod.runtime_patterns_for_choice
+_apply_host_overrides = prebuilt_mod._apply_host_overrides
+_normalize_forwarded_gfx = prebuilt_mod._normalize_forwarded_gfx
 
 # install_python_stack.py
 _STACK_PATH = PACKAGE_ROOT / "studio" / "install_python_stack.py"
@@ -2598,5 +2600,100 @@ class TestHipSdkInstalledButDeviceInaccessible:
         assert "GPU not ROCm-accessible" in source
 
 
+# =============================================================================
+# TEST: --rocm-gfx forwarding -- setup.sh/setup.ps1 hand their resolved gfx arch
+# to install_llama_prebuilt.py so the lemonade HIP prebuilt is selected even when
+# the installer's own hipinfo/amd-smi probe cannot report it.
+# =============================================================================
+
+_SETUP_SH_PATH = PACKAGE_ROOT / "studio" / "setup.sh"
+
+
+class TestNormalizeForwardedGfx:
+    """A forwarded gfx string is reduced to a single clean gfx token."""
+
+    def test_plain_token(self):
+        assert _normalize_forwarded_gfx("gfx1151") == "gfx1151"
+
+    def test_uppercase_normalized(self):
+        assert _normalize_forwarded_gfx("GFX1151") == "gfx1151"
+
+    def test_extracts_from_noise(self):
+        assert _normalize_forwarded_gfx("gcnArchName: gfx942") == "gfx942"
+
+    def test_malformed_is_ignored(self):
+        assert _normalize_forwarded_gfx("not-a-gpu") is None
+
+    def test_empty_and_none(self):
+        assert _normalize_forwarded_gfx("") is None
+        assert _normalize_forwarded_gfx(None) is None
+
+
+class TestApplyHostOverrides:
+    """Forwarded ROCm detection is folded into the host profile correctly."""
+
+    def test_forwarded_gfx_fills_empty_probe(self):
+        # amd-smi-only / name-inferred host: installer probe found no gfx.
+        host = rocm_host(rocm_gfx_target = None)
+        out = _apply_host_overrides(host, override_rocm_gfx = "gfx1151")
+        assert out.has_rocm is True
+        assert out.rocm_gfx_target == "gfx1151"
+
+    def test_forwarded_gfx_implies_rocm(self):
+        # A CPU-looking host with a forwarded gfx is an AMD host.
+        out = _apply_host_overrides(cpu_host(), override_rocm_gfx = "gfx1200")
+        assert out.has_rocm is True
+        assert out.rocm_gfx_target == "gfx1200"
+
+    def test_forwarded_gfx_is_authoritative(self):
+        # setup already applied visible-device selection; its value wins.
+        host = rocm_host(rocm_gfx_target = "gfx1100")
+        out = _apply_host_overrides(host, override_rocm_gfx = "gfx1151")
+        assert out.rocm_gfx_target == "gfx1151"
+
+    def test_has_rocm_only_keeps_probe_gfx(self):
+        out = _apply_host_overrides(cpu_host(), override_has_rocm = True)
+        assert out.has_rocm is True
+        assert out.rocm_gfx_target is None
+
+    def test_malformed_forwarded_gfx_falls_back_to_has_rocm(self):
+        out = _apply_host_overrides(
+            cpu_host(), override_has_rocm = True, override_rocm_gfx = "junk"
+        )
+        assert out.has_rocm is True
+        assert out.rocm_gfx_target is None
+
+    def test_no_overrides_leaves_host_unchanged(self):
+        host = nvidia_host()
+        assert _apply_host_overrides(host) is host
+
+
+class TestRocmGfxForwarding:
+    """setup.sh / setup.ps1 forward their resolved gfx; the installer accepts it."""
+
+    def test_installer_exposes_rocm_gfx_arg(self):
+        source = _PREBUILT_PATH.read_text(encoding = "utf-8")
+        assert '"--rocm-gfx"' in source
+        # Defaults to the env override so a standalone run still works.
+        assert 'os.environ.get("UNSLOTH_ROCM_GFX_ARCH")' in source
+
+    def test_setup_sh_forwards_rocm_gfx(self):
+        source = _SETUP_SH_PATH.read_text(encoding = "utf-8")
+        assert "--rocm-gfx" in source
+        assert '"$_setup_gfx"' in source
+
+    def test_setup_sh_forwards_has_rocm(self):
+        # When AMD is detected but gfx resolution fails, setup.sh must still
+        # forward --has-rocm so the installer knows ROCm is present.
+        source = _SETUP_SH_PATH.read_text(encoding = "utf-8")
+        assert "--has-rocm" in source
+        assert "_setup_amd_detected" in source
+
+    def test_setup_ps1_forwards_rocm_gfx(self):
+        source = _SETUP_PS1_PATH.read_text(encoding = "utf-8")
+        assert "--rocm-gfx" in source
+        assert "$script:ROCmGfxArch" in source
+
+
 if __name__ == "__main__":
     pytest.main([__file__, "-v"])

tests/studio/install/test_selection_logic.py

@@ -52,6 +52,9 @@ compatible_windows_runtime_lines = (
 runtime_line_from_cuda_version = INSTALL_LLAMA_PREBUILT.runtime_line_from_cuda_version
 apply_approved_hashes = INSTALL_LLAMA_PREBUILT.apply_approved_hashes
 linux_cuda_choice_from_release = INSTALL_LLAMA_PREBUILT.linux_cuda_choice_from_release
+parse_direct_linux_release_bundle = (
+    INSTALL_LLAMA_PREBUILT.parse_direct_linux_release_bundle
+)
 windows_cuda_attempts = INSTALL_LLAMA_PREBUILT.windows_cuda_attempts
 resolve_upstream_asset_choice = INSTALL_LLAMA_PREBUILT.resolve_upstream_asset_choice
 resolve_requested_install_tag = INSTALL_LLAMA_PREBUILT.resolve_requested_install_tag
@@ -74,6 +77,18 @@ windows_cuda_upstream_asset_names = (
     INSTALL_LLAMA_PREBUILT.windows_cuda_upstream_asset_names
 )
 env_int = INSTALL_LLAMA_PREBUILT.env_int
+direct_upstream_release_plan = INSTALL_LLAMA_PREBUILT.direct_upstream_release_plan
+_pinned_windows_cuda_fallback = INSTALL_LLAMA_PREBUILT._pinned_windows_cuda_fallback
+CudaRuntimePreference = INSTALL_LLAMA_PREBUILT.CudaRuntimePreference
+published_windows_cuda_attempts = INSTALL_LLAMA_PREBUILT.published_windows_cuda_attempts
+_windows_cuda_attempt_covers_blackwell = (
+    INSTALL_LLAMA_PREBUILT._windows_cuda_attempt_covers_blackwell
+)
+resolve_release_asset_choice = INSTALL_LLAMA_PREBUILT.resolve_release_asset_choice
+pinned_macos_release_tag = INSTALL_LLAMA_PREBUILT.pinned_macos_release_tag
+resolve_simple_install_release_plans = (
+    INSTALL_LLAMA_PREBUILT.resolve_simple_install_release_plans
+)
 
 
 # ---------------------------------------------------------------------------
@@ -397,6 +412,44 @@ class TestCompatibleLinuxRuntimeLines:
         host = make_host(driver_cuda_version = (13, 0))
         assert compatible_linux_runtime_lines(host) == ["cuda13", "cuda12"]
 
+    def test_future_major_derives_lines(self):
+        # A future major (14.x) offers cuda14 first, then older majors.
+        host = make_host(driver_cuda_version = (14, 0))
+        assert compatible_linux_runtime_lines(host) == ["cuda14", "cuda13", "cuda12"]
+
+
+class TestParseDirectLinuxReleaseBundle:
+    def _release(self, *targets):
+        names = [f"app-bTEST-linux-x64-{t}.tar.gz" for t in targets]
+        return {
+            "tag_name": "bTEST",
+            "assets": [
+                {"name": n, "browser_download_url": "https://x/" + n} for n in names
+            ],
+        }
+
+    def _cuda_artifact(self, bundle):
+        return [a for a in bundle.artifacts if a.install_kind == "linux-cuda"][0]
+
+    def test_parses_known_cuda13_bundle(self):
+        bundle = parse_direct_linux_release_bundle(
+            "unslothai/llama.cpp", self._release("cuda13-newer")
+        )
+        assert bundle is not None
+        assert self._cuda_artifact(bundle).runtime_line == "cuda13"
+
+    def test_parses_future_cuda_major_with_forward_profile(self):
+        # A future major name parses and inherits the newest known major's
+        # coverage for the same class as a forward default.
+        bundle = parse_direct_linux_release_bundle(
+            "unslothai/llama.cpp", self._release("cuda14-newer")
+        )
+        assert bundle is not None
+        art = self._cuda_artifact(bundle)
+        assert art.runtime_line == "cuda14"
+        assert art.coverage_class == "newer"
+        assert art.max_sm == 120  # inherited from cuda13-newer
+
 
 # ===========================================================================
 # G. pick_windows_cuda_runtime + compatible_windows_runtime_lines
@@ -442,6 +495,10 @@ class TestCompatibleWindowsRuntimeLines:
         host = make_host(driver_cuda_version = (13, 0))
         assert compatible_windows_runtime_lines(host) == ["cuda13", "cuda12"]
 
+    def test_future_major_derives_lines(self):
+        host = make_host(driver_cuda_version = (14, 0))
+        assert compatible_windows_runtime_lines(host) == ["cuda14", "cuda13", "cuda12"]
+
 
 # ===========================================================================
 # H. runtime_line_from_cuda_version
@@ -1195,6 +1252,79 @@ class TestLinuxCudaChoiceFromRelease:
         assert result is None
 
 
+def make_profile_artifact(asset_name, profile_name, **overrides):
+    profile = INSTALL_LLAMA_PREBUILT.DIRECT_LINUX_BUNDLE_PROFILES[profile_name]
+    defaults = dict(
+        runtime_line = profile["runtime_line"],
+        coverage_class = profile["coverage_class"],
+        supported_sms = [str(value) for value in profile["supported_sms"]],
+        min_sm = int(profile["min_sm"]),
+        max_sm = int(profile["max_sm"]),
+        bundle_profile = profile_name,
+        rank = int(profile["rank"]),
+    )
+    defaults.update(overrides)
+    return make_artifact(asset_name, **defaults)
+
+
+class TestBlackwellUltraSm103Coverage:
+    """sm_103 (B300 / GB300) runs on the bundled base compute_100 PTX via JIT."""
+
+    def test_profiles_list_sm103_wherever_sm100_is_shipped(self):
+        for (
+            name,
+            profile,
+        ) in INSTALL_LLAMA_PREBUILT.DIRECT_LINUX_BUNDLE_PROFILES.items():
+            sms = {str(value) for value in profile["supported_sms"]}
+            if "100" in sms:
+                assert "103" in sms, name
+            else:
+                assert "103" not in sms, name
+
+    def test_b300_selects_cuda13_newer_prebuilt(self, monkeypatch):
+        mock_linux_runtime(monkeypatch, ["cuda13"])
+        host = make_host(compute_caps = ["103"], driver_cuda_version = (13, 0))
+        art = make_profile_artifact("cuda13-newer.tar.gz", "cuda13-newer")
+        release = make_release([art])
+        result = linux_cuda_choice_from_release(host, release)
+        assert result is not None
+        assert result.primary.name == "cuda13-newer.tar.gz"
+
+    def test_b300_selects_cuda12_newer_prebuilt(self, monkeypatch):
+        mock_linux_runtime(monkeypatch, ["cuda12"])
+        host = make_host(compute_caps = ["103"], driver_cuda_version = (12, 8))
+        art = make_profile_artifact("cuda12-newer.tar.gz", "cuda12-newer")
+        release = make_release([art])
+        result = linux_cuda_choice_from_release(host, release)
+        assert result is not None
+        assert result.primary.name == "cuda12-newer.tar.gz"
+
+    def test_b300_reported_as_decimal_normalizes_and_matches(self, monkeypatch):
+        mock_linux_runtime(monkeypatch, ["cuda13"])
+        host = make_host(compute_caps = ["10.3"], driver_cuda_version = (13, 0))
+        art = make_profile_artifact("cuda13-portable.tar.gz", "cuda13-portable")
+        release = make_release([art])
+        result = linux_cuda_choice_from_release(host, release)
+        assert result is not None
+
+    def test_b300_falls_back_to_portable_when_only_portable_present(self, monkeypatch):
+        mock_linux_runtime(monkeypatch, ["cuda13"])
+        host = make_host(compute_caps = ["103"], driver_cuda_version = (13, 0))
+        art = make_profile_artifact("cuda13-portable.tar.gz", "cuda13-portable")
+        release = make_release([art])
+        result = linux_cuda_choice_from_release(host, release)
+        assert result is not None
+        assert result.primary.name == "cuda13-portable.tar.gz"
+
+    def test_older_bundle_still_rejects_b300(self, monkeypatch):
+        mock_linux_runtime(monkeypatch, ["cuda13"])
+        host = make_host(compute_caps = ["103"], driver_cuda_version = (13, 0))
+        art = make_profile_artifact("cuda13-older.tar.gz", "cuda13-older")
+        release = make_release([art])
+        result = linux_cuda_choice_from_release(host, release)
+        assert result is None
+
+
 # ===========================================================================
 # L. resolve_install_attempts
 # ===========================================================================
@@ -1777,12 +1907,23 @@ class TestWindowsCudaAttempts:
         assert result[0].runtime_line == "cuda13"
         assert result[1].runtime_line == "cuda12"
 
-    def test_driver_13_0_cuda13_dlls_selects_cuda13_asset(self, monkeypatch):
+    def test_driver_below_published_minor_is_gated_to_cuda12(self, monkeypatch):
+        # A 13.0 driver cannot run a 13.1 build (forward minor), so it is gated
+        # out of cuda13 and falls back to the cuda12 build it can run, even when
+        # only the cuda13 runtime libs are detected.
         mock_windows_runtime(monkeypatch, ["cuda13"])
         host = make_host(system = "Windows", machine = "AMD64", driver_cuda_version = (13, 0))
         assets = self._upstream("13.1", "12.4")
         result = windows_cuda_attempts(host, self.TAG, assets, None)
-        assert len(result) == 1
+        assert result[0].runtime_line == "cuda12"
+        assert result[0].name == f"llama-{self.TAG}-bin-win-cuda-12.4-x64.zip"
+
+    def test_driver_at_published_minor_selects_cuda13(self, monkeypatch):
+        # A 13.1 driver matches the published 13.1 build exactly.
+        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
+        host = make_host(system = "Windows", machine = "AMD64", driver_cuda_version = (13, 1))
+        assets = self._upstream("13.1", "12.4")
+        result = windows_cuda_attempts(host, self.TAG, assets, None)
         assert result[0].runtime_line == "cuda13"
         assert result[0].name == f"llama-{self.TAG}-bin-win-cuda-13.1-x64.zip"
 
@@ -1885,6 +2026,468 @@ class TestWindowsCudaAttempts:
             assert attempt.runtime_url is None
             assert attempt.runtime_name is None
 
+    def test_tracks_upstream_cuda13_minor_bump(self, monkeypatch):
+        # ggml-org bumped the published Windows cuda13 build 13.1 -> 13.3; the
+        # selector must follow it instead of the old hardcoded 13.1 (#5861).
+        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
+        host = make_host(system = "Windows", machine = "AMD64", driver_cuda_version = (13, 3))
+        assets = self._upstream("13.3", "12.4")
+        result = windows_cuda_attempts(host, self.TAG, assets, None)
+        assert result[0].runtime_line == "cuda13"
+        assert result[0].name == f"llama-{self.TAG}-bin-win-cuda-13.3-x64.zip"
+
+    def test_cuda13_minor_bump_pairs_matching_cudart(self, monkeypatch):
+        # The paired cudart bundle must track the same bumped minor.
+        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
+        host = make_host(system = "Windows", machine = "AMD64", driver_cuda_version = (13, 3))
+        assets = {
+            f"llama-{self.TAG}-bin-win-cuda-13.3-x64.zip": "https://example.com/llama-13.3",
+            "cudart-llama-bin-win-cuda-13.3-x64.zip": "https://example.com/cudart-13.3",
+            f"llama-{self.TAG}-bin-win-cuda-12.4-x64.zip": "https://example.com/llama-12.4",
+            "cudart-llama-bin-win-cuda-12.4-x64.zip": "https://example.com/cudart-12.4",
+        }
+        result = windows_cuda_attempts(host, self.TAG, assets, None)
+        assert result[0].name == f"llama-{self.TAG}-bin-win-cuda-13.3-x64.zip"
+        assert result[0].runtime_name == "cudart-llama-bin-win-cuda-13.3-x64.zip"
+
+    def test_driver_below_published_minor_does_not_get_newer_build(self, monkeypatch):
+        # ggml-org ships only cuda-13.3; a 13.1 driver cannot run it (forward
+        # minor), so it is gated to the cuda-12.4 build instead of an
+        # unguaranteed 13.3. A 13.3 driver still gets 13.3 (see other tests).
+        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
+        host = make_host(system = "Windows", machine = "AMD64", driver_cuda_version = (13, 1))
+        assets = self._upstream("13.3", "12.4")
+        result = windows_cuda_attempts(host, self.TAG, assets, None)
+        assert result[0].runtime_line == "cuda12"
+        assert result[0].name == f"llama-{self.TAG}-bin-win-cuda-12.4-x64.zip"
+
+    def test_tracks_future_cuda13_minor(self, monkeypatch):
+        # A later within-major bump (13.4) is tracked the same as 13.3.
+        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
+        host = make_host(system = "Windows", machine = "AMD64", driver_cuda_version = (13, 4))
+        assets = self._upstream("13.4", "12.4")
+        result = windows_cuda_attempts(host, self.TAG, assets, None)
+        assert result[0].name == f"llama-{self.TAG}-bin-win-cuda-13.4-x64.zip"
+
+    def test_new_cuda_major_selected_when_published(self, monkeypatch):
+        # A new CUDA major (14.x) driver picks the published cuda14 build.
+        mock_windows_runtime(monkeypatch, ["cuda14", "cuda13", "cuda12"])
+        host = make_host(system = "Windows", machine = "AMD64", driver_cuda_version = (14, 0))
+        assets = self._upstream("14.0", "13.3", "12.4")
+        result = windows_cuda_attempts(host, self.TAG, assets, None)
+        assert result[0].runtime_line == "cuda14"
+        assert result[0].name == f"llama-{self.TAG}-bin-win-cuda-14.0-x64.zip"
+
+    def test_new_cuda_major_degrades_to_published_cuda13(self, monkeypatch):
+        # A 14.x driver with no cuda14 build runs the newest published cuda13
+        # build via backward compatibility.
+        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
+        host = make_host(system = "Windows", machine = "AMD64", driver_cuda_version = (14, 0))
+        assets = self._upstream("13.3", "12.4")
+        result = windows_cuda_attempts(host, self.TAG, assets, None)
+        assert result[0].name == f"llama-{self.TAG}-bin-win-cuda-13.3-x64.zip"
+
+
+# ===========================================================================
+# N.1b. _pinned_windows_cuda_fallback -- pinned b9360 cuda-13.1 Blackwell fallback
+# ===========================================================================
+
+
+class TestPinnedBlackwellCudaFallback:
+    """A Blackwell host on a 13.0/13.1/13.2 driver, gated off the in-release 13.3
+    build, gets the pinned immutable b9360 cuda-13.1 GPU build instead of the
+    CPU-only cuda-12.4 drop. The pin is dormant for everyone else."""
+
+    TAG = "b8508"
+
+    def _win_host(self, driver, caps):
+        return make_host(
+            system = "Windows",
+            machine = "AMD64",
+            driver_cuda_version = driver,
+            compute_caps = caps,
+        )
+
+    def test_pin_offered_for_driver_13_1_blackwell(self):
+        pin = _pinned_windows_cuda_fallback(self._win_host((13, 1), ["120"]), [])
+        assert pin is not None
+        assert pin.tag == "b9360"
+        assert pin.runtime_line == "cuda13"
+        assert pin.name == "llama-b9360-bin-win-cuda-13.1-x64.zip"
+        assert pin.runtime_name == "cudart-llama-bin-win-cuda-13.1-x64.zip"
+        assert pin.url.endswith("/b9360/llama-b9360-bin-win-cuda-13.1-x64.zip")
+        assert pin.runtime_url.endswith("/b9360/cudart-llama-bin-win-cuda-13.1-x64.zip")
+        assert pin.install_kind == "windows-cuda"
+        assert pin.expected_sha256 and len(pin.expected_sha256) == 64
+        assert pin.runtime_sha256 and len(pin.runtime_sha256) == 64
+
+    def test_pin_offered_for_driver_13_2(self):
+        assert (
+            _pinned_windows_cuda_fallback(self._win_host((13, 2), ["120"]), [])
+            is not None
+        )
+
+    def test_pin_offered_for_sm121_variant(self):
+        # sm_121 is Blackwell-family and also needs toolkit >= 12.8.
+        assert (
+            _pinned_windows_cuda_fallback(self._win_host((13, 1), ["121"]), [])
+            is not None
+        )
+
+    def test_pin_uses_max_of_multi_gpu_caps(self):
+        assert (
+            _pinned_windows_cuda_fallback(self._win_host((13, 1), ["86", "120"]), [])
+            is not None
+        )
+
+    @pytest.mark.parametrize("sm", ["89", "90", "100"])
+    def test_pin_not_offered_to_non_blackwell(self, sm):
+        # Ada/Hopper run the cuda-12.4 build fine; the pin must not fire.
+        assert _pinned_windows_cuda_fallback(self._win_host((13, 1), [sm]), []) is None
+
+    def test_pin_offered_for_driver_13_0(self):
+        # b9360 is native sm_120a SASS (no JIT) and ships a cuda-13.1 cudart,
+        # both of which run on a 13.0 r580+ driver via CUDA minor-version
+        # compatibility. 13.0 is the mainstream Blackwell branch, so it must fire.
+        assert (
+            _pinned_windows_cuda_fallback(self._win_host((13, 0), ["120"]), [])
+            is not None
+        )
+
+    def test_pin_not_offered_below_floor(self):
+        # 12.x predates Blackwell entirely; the pin stays dormant below 13.0.
+        assert (
+            _pinned_windows_cuda_fallback(self._win_host((12, 9), ["120"]), []) is None
+        )
+
+    def test_pin_not_offered_without_driver(self):
+        assert _pinned_windows_cuda_fallback(self._win_host(None, ["120"]), []) is None
+
+    def test_pin_not_offered_on_linux(self):
+        host = make_host(
+            system = "Linux",
+            machine = "x86_64",
+            driver_cuda_version = (13, 1),
+            compute_caps = ["120"],
+        )
+        assert _pinned_windows_cuda_fallback(host, []) is None
+
+    def test_pin_dormant_when_cuda13_attempt_present(self, monkeypatch):
+        # A runnable in-release cuda13 build makes the pin unnecessary.
+        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
+        host = self._win_host((13, 1), ["120"])
+        assets = {
+            f"llama-{self.TAG}-bin-win-cuda-13.1-x64.zip": "https://example.com/13.1",
+            f"llama-{self.TAG}-bin-win-cuda-12.4-x64.zip": "https://example.com/12.4",
+        }
+        existing = windows_cuda_attempts(host, self.TAG, assets, None)
+        assert any(a.runtime_line == "cuda13" for a in existing)
+        assert _pinned_windows_cuda_fallback(host, existing) is None
+
+    def _win_cuda_attempt(self, minor):
+        major = minor.split(".")[0]
+        return AssetChoice(
+            repo = UPSTREAM_REPO,
+            tag = self.TAG,
+            name = f"llama-{self.TAG}-bin-win-cuda-{minor}-x64.zip",
+            url = "https://example.com/x",
+            source_label = "upstream",
+            install_kind = "windows-cuda",
+            runtime_line = f"cuda{major}",
+        )
+
+    def test_pin_dormant_when_runnable_cuda14_present(self, monkeypatch):
+        # A future Blackwell host with an in-release cuda14 build (no cuda13)
+        # must not get the older b9360 13.1 pin ahead of the runnable cuda14.
+        mock_windows_runtime(monkeypatch, ["cuda14", "cuda12"])
+        host = self._win_host((14, 0), ["120"])
+        assets = {
+            f"llama-{self.TAG}-bin-win-cuda-14.0-x64.zip": "https://example.com/14.0",
+            f"llama-{self.TAG}-bin-win-cuda-12.4-x64.zip": "https://example.com/12.4",
+        }
+        existing = windows_cuda_attempts(host, self.TAG, assets, None)
+        assert any(a.runtime_line == "cuda14" for a in existing)
+        assert _pinned_windows_cuda_fallback(host, existing) is None
+
+    def test_pin_dormant_when_runnable_cuda12_8_present(self):
+        # A cuda-12.8 build also covers Blackwell, so the pin defers to it.
+        host = self._win_host((13, 1), ["120"])
+        existing = [self._win_cuda_attempt("12.8")]
+        assert _pinned_windows_cuda_fallback(host, existing) is None
+
+    def test_pin_fires_when_only_cuda12_4_present(self):
+        # cuda-12.4 does not cover Blackwell, so the pin still fires.
+        host = self._win_host((13, 1), ["120"])
+        existing = [self._win_cuda_attempt("12.4")]
+        assert _pinned_windows_cuda_fallback(host, existing) is not None
+
+    @pytest.mark.parametrize(
+        "minor, covers",
+        [
+            ("12.4", False),
+            ("12.8", True),
+            ("13.1", True),
+            ("13.3", True),
+            ("14.0", True),
+        ],
+    )
+    def test_attempt_covers_blackwell(self, minor, covers):
+        assert (
+            _windows_cuda_attempt_covers_blackwell(self._win_cuda_attempt(minor))
+            is covers
+        )
+
+    def test_attempt_covers_blackwell_ignores_non_cuda_kind(self):
+        cpu = AssetChoice(
+            repo = UPSTREAM_REPO,
+            tag = self.TAG,
+            name = f"llama-{self.TAG}-bin-win-cpu-x64.zip",
+            url = "https://example.com/x",
+            source_label = "upstream",
+            install_kind = "windows-cpu",
+        )
+        assert _windows_cuda_attempt_covers_blackwell(cpu) is False
+
+
+# ===========================================================================
+# N.1c. direct_upstream_release_plan -- pinned Blackwell fallback ordering
+# ===========================================================================
+
+
+class TestDirectUpstreamBlackwellPin:
+    """End to end: the pin lands ahead of cuda-12.4 on the simple/upstream path
+    a Blackwell Windows host actually uses, and stays absent once a runnable
+    in-release cuda13 build exists."""
+
+    TAG = "b9365"
+
+    def _release(self):
+        names = [
+            f"llama-{self.TAG}-bin-win-cuda-13.3-x64.zip",
+            "cudart-llama-bin-win-cuda-13.3-x64.zip",
+            f"llama-{self.TAG}-bin-win-cuda-12.4-x64.zip",
+            "cudart-llama-bin-win-cuda-12.4-x64.zip",
+            f"llama-{self.TAG}-bin-win-cpu-x64.zip",
+        ]
+        return {
+            "tag_name": self.TAG,
+            "assets": [
+                {"name": n, "browser_download_url": f"https://example.com/{n}"}
+                for n in names
+            ],
+        }
+
+    def _no_torch(self, monkeypatch):
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "detect_torch_cuda_runtime_preference",
+            lambda host: CudaRuntimePreference(runtime_line = None, selection_log = []),
+        )
+
+    def test_blackwell_13_1_prepends_pin(self, monkeypatch):
+        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
+        self._no_torch(monkeypatch)
+        host = make_host(
+            system = "Windows",
+            machine = "AMD64",
+            driver_cuda_version = (13, 1),
+            compute_caps = ["120"],
+        )
+        plan = direct_upstream_release_plan(
+            self._release(), host, UPSTREAM_REPO, "latest"
+        )
+        order = [(a.tag, a.runtime_line or a.install_kind) for a in plan.attempts]
+        assert order == [
+            ("b9360", "cuda13"),
+            (self.TAG, "cuda12"),
+            (self.TAG, "windows-cpu"),
+        ]
+        assert plan.attempts[0].name == "llama-b9360-bin-win-cuda-13.1-x64.zip"
+        # Direct/upstream path stays unverified-by-manifest (no approved hashes).
+        assert plan.approved_checksums.artifacts == {}
+
+    def test_blackwell_13_3_no_pin(self, monkeypatch):
+        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
+        self._no_torch(monkeypatch)
+        host = make_host(
+            system = "Windows",
+            machine = "AMD64",
+            driver_cuda_version = (13, 3),
+            compute_caps = ["120"],
+        )
+        plan = direct_upstream_release_plan(
+            self._release(), host, UPSTREAM_REPO, "latest"
+        )
+        assert "b9360" not in [a.tag for a in plan.attempts]
+        assert plan.attempts[0].tag == self.TAG
+        assert plan.attempts[0].runtime_line == "cuda13"
+        assert plan.attempts[0].name == f"llama-{self.TAG}-bin-win-cuda-13.3-x64.zip"
+
+
+# ===========================================================================
+# N.1d. published_windows_cuda_attempts -- version-dynamic ordering seed
+# ===========================================================================
+
+
+class TestPublishedWindowsCudaAttemptsDynamicMajor:
+    """The published-path ordering seed is derived from the release's real
+    published minors, so a future CUDA major published here is selectable
+    instead of being hidden by a hardcoded cuda12/cuda13 seed."""
+
+    TAG = "b8508"
+
+    def _win_cuda_artifact(self, minor, runtime_line):
+        return make_artifact(
+            f"llama-{self.TAG}-bin-win-cuda-{minor}-x64.zip",
+            install_kind = "windows-cuda",
+            runtime_line = runtime_line,
+            max_sm = 120,
+        )
+
+    def _release(self, minors_lines):
+        artifacts = [self._win_cuda_artifact(m, line) for m, line in minors_lines]
+        return make_release(artifacts, upstream_tag = self.TAG)
+
+    def test_future_cuda14_published_is_selected(self, monkeypatch):
+        # With the dynamic seed a 14.x driver reaches a published cuda14 build;
+        # the old hardcoded cuda12/cuda13 seed would never order it (the cuda14
+        # line would be skipped for want of a 14.x asset in the seed).
+        mock_windows_runtime(monkeypatch, ["cuda14", "cuda13", "cuda12"])
+        release = self._release(
+            [("14.0", "cuda14"), ("13.3", "cuda13"), ("12.4", "cuda12")]
+        )
+        host = make_host(
+            system = "Windows",
+            machine = "AMD64",
+            driver_cuda_version = (14, 0),
+            compute_caps = ["120"],
+        )
+        result = published_windows_cuda_attempts(host, release, None)
+        assert result[0].runtime_line == "cuda14"
+        assert result[0].name == f"llama-{self.TAG}-bin-win-cuda-14.0-x64.zip"
+
+    def test_cuda13_minor_selected_for_13_3_driver(self, monkeypatch):
+        # Existing behavior unchanged: a 13.3 driver gets the real 13.3 build.
+        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
+        release = self._release([("13.3", "cuda13"), ("12.4", "cuda12")])
+        host = make_host(
+            system = "Windows",
+            machine = "AMD64",
+            driver_cuda_version = (13, 3),
+            compute_caps = ["120"],
+        )
+        result = published_windows_cuda_attempts(host, release, None)
+        assert result[0].runtime_line == "cuda13"
+        assert result[0].name == f"llama-{self.TAG}-bin-win-cuda-13.3-x64.zip"
+
+    def test_below_minor_driver_gated_to_cuda12(self, monkeypatch):
+        # A 13.1 driver is gated off a published 13.3 and falls to cuda12.
+        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
+        release = self._release([("13.3", "cuda13"), ("12.4", "cuda12")])
+        host = make_host(
+            system = "Windows",
+            machine = "AMD64",
+            driver_cuda_version = (13, 1),
+            compute_caps = ["120"],
+        )
+        result = published_windows_cuda_attempts(host, release, None)
+        assert result[0].runtime_line == "cuda12"
+
+
+# ===========================================================================
+# N.1e. resolve_release_asset_choice -- pin on the published install path
+# ===========================================================================
+
+
+class TestResolveReleaseAssetChoicePin:
+    """The published (non --simple-policy) install path reaches the same b9360
+    Blackwell pin as the simple path, with its verified hash threaded."""
+
+    TAG = "b8508"
+
+    def _release(self, minors_lines):
+        artifacts = [
+            make_artifact(
+                f"llama-{self.TAG}-bin-win-cuda-{minor}-x64.zip",
+                install_kind = "windows-cuda",
+                runtime_line = line,
+                max_sm = 120,
+            )
+            for minor, line in minors_lines
+        ]
+        assets = {}
+        for minor, _line in minors_lines:
+            assets[f"llama-{self.TAG}-bin-win-cuda-{minor}-x64.zip"] = (
+                f"https://example.com/llama-{minor}"
+            )
+            assets[f"cudart-llama-bin-win-cuda-{minor}-x64.zip"] = (
+                f"https://example.com/cudart-{minor}"
+            )
+        return make_release(artifacts, upstream_tag = self.TAG, assets = assets)
+
+    def _checksums(self, minors):
+        names = []
+        for minor in minors:
+            names.append(f"llama-{self.TAG}-bin-win-cuda-{minor}-x64.zip")
+            names.append(f"cudart-llama-bin-win-cuda-{minor}-x64.zip")
+        return make_checksums(names)
+
+    def _no_torch(self, monkeypatch):
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "detect_torch_cuda_runtime_preference",
+            lambda host: CudaRuntimePreference(runtime_line = None, selection_log = []),
+        )
+
+    def test_pin_applied_on_published_path_for_13_1(self, monkeypatch):
+        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
+        self._no_torch(monkeypatch)
+        release = self._release([("13.3", "cuda13"), ("12.4", "cuda12")])
+        checksums = self._checksums(["12.4"])  # 13.3 gated off for a 13.1 driver
+        host = make_host(
+            system = "Windows",
+            machine = "AMD64",
+            driver_cuda_version = (13, 1),
+            compute_caps = ["120"],
+        )
+        result = resolve_release_asset_choice(host, self.TAG, release, checksums)
+        assert result[0].tag == "b9360"
+        assert result[0].name == "llama-b9360-bin-win-cuda-13.1-x64.zip"
+        # apply_approved_hashes threaded the pin's verified hash from the
+        # augmented checksums (the pin survives the approved-hash gate).
+        assert result[0].expected_sha256 and len(result[0].expected_sha256) == 64
+        assert result[0].runtime_sha256 and len(result[0].runtime_sha256) == 64
+        assert any(a.runtime_line == "cuda12" for a in result)
+
+    def test_pin_dormant_on_published_path_for_13_3(self, monkeypatch):
+        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
+        self._no_torch(monkeypatch)
+        release = self._release([("13.3", "cuda13"), ("12.4", "cuda12")])
+        checksums = self._checksums(["13.3", "12.4"])
+        host = make_host(
+            system = "Windows",
+            machine = "AMD64",
+            driver_cuda_version = (13, 3),
+            compute_caps = ["120"],
+        )
+        result = resolve_release_asset_choice(host, self.TAG, release, checksums)
+        assert "b9360" not in [a.tag for a in result]
+        assert result[0].name == f"llama-{self.TAG}-bin-win-cuda-13.3-x64.zip"
+
+    def test_pin_not_applied_for_non_blackwell(self, monkeypatch):
+        mock_windows_runtime(monkeypatch, ["cuda13", "cuda12"])
+        self._no_torch(monkeypatch)
+        release = self._release([("13.3", "cuda13"), ("12.4", "cuda12")])
+        checksums = self._checksums(["12.4"])
+        host = make_host(
+            system = "Windows",
+            machine = "AMD64",
+            driver_cuda_version = (13, 1),
+            compute_caps = ["89"],
+        )
+        result = resolve_release_asset_choice(host, self.TAG, release, checksums)
+        assert "b9360" not in [a.tag for a in result]
+
 
 # ===========================================================================
 # N.1. apply_approved_hashes -- runtime archive checksum threading
@@ -2109,3 +2712,288 @@ class TestResolveUpstreamAssetChoice:
         result = resolve_upstream_asset_choice(host, self.TAG)
         assert result.install_kind == "windows-cuda"
         assert result.name == cuda_name
+
+
+# ===========================================================================
+# N.2. Deterministic macOS prebuilt pin (b9415)
+# ===========================================================================
+
+
+def _macos_host(machine = "arm64", version = (15, 5)):
+    return make_host(
+        system = "Darwin",
+        machine = machine,
+        nvidia_smi = None,
+        driver_cuda_version = None,
+        compute_caps = [],
+        has_physical_nvidia = False,
+        has_usable_nvidia = False,
+        macos_version = version,
+    )
+
+
+class TestPinnedMacosReleaseTag:
+    """pinned_macos_release_tag: pin b9415 only for ggml-org upstream macOS hosts
+    below macOS 26; latest (None) for 26+, unknown version, the fork, non-macOS."""
+
+    def test_arm64_sequoia_pins_b9415(self):
+        host = _macos_host("arm64", (15, 5))
+        assert pinned_macos_release_tag(host, UPSTREAM_REPO) == "b9415"
+
+    def test_arm64_sonoma_pins_b9415(self):
+        host = _macos_host("arm64", (14, 7))
+        assert pinned_macos_release_tag(host, UPSTREAM_REPO) == "b9415"
+
+    def test_x64_ventura_13_3_pins_b9415(self):
+        # b9415's Intel slice is minos 13.3, so 13.3 Intel hosts still load it.
+        host = _macos_host("x86_64", (13, 3))
+        assert pinned_macos_release_tag(host, UPSTREAM_REPO) == "b9415"
+
+    def test_tahoe_26_0_takes_latest(self):
+        host = _macos_host("arm64", (26, 0))
+        assert pinned_macos_release_tag(host, UPSTREAM_REPO) is None
+
+    def test_tahoe_26_1_takes_latest(self):
+        host = _macos_host("arm64", (26, 1))
+        assert pinned_macos_release_tag(host, UPSTREAM_REPO) is None
+
+    def test_unknown_version_takes_latest(self):
+        host = _macos_host("arm64", None)
+        assert pinned_macos_release_tag(host, UPSTREAM_REPO) is None
+
+    def test_fork_repo_is_dormant(self):
+        # The unslothai/llama.cpp fork publishes its own minos-13.3 prebuilts.
+        host = _macos_host("arm64", (15, 5))
+        fork = INSTALL_LLAMA_PREBUILT.DEFAULT_PUBLISHED_REPO
+        assert pinned_macos_release_tag(host, fork) is None
+
+    def test_non_macos_host_is_dormant(self):
+        host = make_host(system = "Linux", machine = "x86_64")
+        assert pinned_macos_release_tag(host, UPSTREAM_REPO) is None
+
+
+class TestResolveSimpleMacosPin:
+    """End to end on the simple/upstream path macOS actually uses: a pre-26 host
+    deterministically resolves b9415 (no walk-back); a macOS 26 host takes the
+    latest release. Mirrors how setup.sh routes Darwin to ggml-org/llama.cpp."""
+
+    TAGS = ["b9442", "b9430", "b9428", "b9415"]  # newest-first feed
+
+    def _feed(self, monkeypatch):
+        calls = []
+
+        def _release(tag):
+            name = f"llama-{tag}-bin-macos-arm64.tar.gz"
+            return {
+                "tag_name": tag,
+                "assets": [
+                    {
+                        "name": name,
+                        "browser_download_url": f"https://example.com/{name}",
+                    }
+                ],
+            }
+
+        def fake_iter(repo, published_release_tag = "", requested_tag = ""):
+            calls.append((repo, published_release_tag, requested_tag))
+            # Emulate the real iterator: a specific tag yields only that release.
+            if requested_tag and requested_tag != "latest":
+                yield _release(requested_tag)
+                return
+            for tag in self.TAGS:
+                yield _release(tag)
+
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT, "iter_release_payloads_by_time", fake_iter
+        )
+        return calls
+
+    def test_pre26_host_pins_b9415_without_walkback(self, monkeypatch):
+        calls = self._feed(monkeypatch)
+        host = _macos_host("arm64", (15, 5))
+        requested_tag, plans = resolve_simple_install_release_plans(
+            "latest", host, "ggml-org/llama.cpp", ""
+        )
+        assert requested_tag == "b9415"
+        assert len(plans) == 1
+        assert plans[0].release_tag == "b9415"
+        assert plans[0].llama_tag == "b9415"
+        assert plans[0].attempts[0].install_kind == "macos-arm64"
+        assert plans[0].attempts[0].name == "llama-b9415-bin-macos-arm64.tar.gz"
+        # The pin overrode the requested tag before any release was fetched.
+        assert calls[0][2] == "b9415"
+        # Simple/upstream path stays unverified-by-manifest, exactly as before.
+        assert plans[0].approved_checksums.artifacts == {}
+
+    def test_tahoe_host_takes_latest_release(self, monkeypatch):
+        calls = self._feed(monkeypatch)
+        host = _macos_host("arm64", (26, 0))
+        requested_tag, plans = resolve_simple_install_release_plans(
+            "latest", host, "ggml-org/llama.cpp", ""
+        )
+        assert requested_tag == "latest"
+        assert plans[0].release_tag == "b9442"
+        # No pin: the iterator was asked for latest, not a specific tag.
+        assert calls[0][2] == "latest"
+
+
+# ===========================================================================
+# Linux arm64 + GPU must not install the x64-only fork bundle
+# ===========================================================================
+
+
+class TestLinuxArm64ForkFallsBackToSource:
+    """The unslothai/llama.cpp fork ships only linux-x64 bundles. An arm64
+    Linux host with a GPU (GH200/GB200/DGX Spark) routes to the fork and must
+    fall back to a source build instead of selecting an x64 binary."""
+
+    def test_arm64_nvidia_fork_raises_before_fetching_releases(self, monkeypatch):
+        # Guard fires before any release is fetched: poison the iterator to prove
+        # it is never called.
+        def _boom(*_a, **_k):
+            raise AssertionError("iterator must not run for arm64 fork hosts")
+
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT, "iter_release_payloads_by_time", _boom
+        )
+        host = make_host(system = "Linux", machine = "aarch64")
+        with pytest.raises(PrebuiltFallback, match = "linux-x64 prebuilts"):
+            resolve_simple_install_release_plans(
+                "latest", host, "unslothai/llama.cpp", ""
+            )
+
+    def test_x86_64_fork_is_not_blocked_by_the_arch_guard(self, monkeypatch):
+        # x64 host must pass the guard and reach the iterator (here empty, so it
+        # raises the generic message, not the arch one).
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "iter_release_payloads_by_time",
+            lambda *_a, **_k: iter(()),
+        )
+        host = make_host(system = "Linux", machine = "x86_64")
+        with pytest.raises(PrebuiltFallback) as exc:
+            resolve_simple_install_release_plans(
+                "latest", host, "unslothai/llama.cpp", ""
+            )
+        assert "linux-x64 prebuilts" not in str(exc.value)
+
+    def test_arm64_cpu_on_ggml_org_is_not_blocked(self, monkeypatch):
+        # CPU-only arm64 routes to ggml-org (not the fork), so the guard must not
+        # fire; it reaches the iterator (empty here -> generic message).
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "iter_release_payloads_by_time",
+            lambda *_a, **_k: iter(()),
+        )
+        host = make_host(
+            system = "Linux",
+            machine = "aarch64",
+            nvidia_smi = None,
+            driver_cuda_version = None,
+            compute_caps = [],
+            has_physical_nvidia = False,
+            has_usable_nvidia = False,
+        )
+        with pytest.raises(PrebuiltFallback) as exc:
+            resolve_simple_install_release_plans(
+                "latest", host, "ggml-org/llama.cpp", ""
+            )
+        assert "linux-x64 prebuilts" not in str(exc.value)
+
+
+# ===========================================================================
+# arm64 Linux GPU: CPU prebuilt fallback after a failed source build (--cpu-fallback)
+# ===========================================================================
+
+
+class TestCpuFallback:
+    """--cpu-fallback drops GPU attributes so the CPU prebuilt for the host's
+    OS/arch is selected, letting an arm64 GPU host install ggml-org's arm64 CPU
+    build as a last resort when its source build produced no binary."""
+
+    _SETUP_SH = PACKAGE_ROOT / "studio" / "setup.sh"
+
+    def _arm64_nvidia(self):
+        return make_host(
+            system = "Linux",
+            machine = "aarch64",
+            driver_cuda_version = (13, 0),
+            compute_caps = ["90"],
+            has_physical_nvidia = True,
+            has_usable_nvidia = True,
+        )
+
+    def test_force_cpu_drops_gpu_attrs_before_planning(self, monkeypatch, tmp_path):
+        captured = {}
+
+        def _capture(llama_tag, host, *a, **k):
+            captured["host"] = host
+            raise PrebuiltFallback("stop after capture")
+
+        monkeypatch.setattr(INSTALL_LLAMA_PREBUILT, "detect_host", self._arm64_nvidia)
+        monkeypatch.setattr(
+            INSTALL_LLAMA_PREBUILT,
+            "resolve_simple_install_release_plans",
+            _capture,
+        )
+        # install_prebuilt exits EXIT_FALLBACK on PrebuiltFallback; we only care
+        # about the host it handed to the resolver before that.
+        with pytest.raises(SystemExit):
+            INSTALL_LLAMA_PREBUILT.install_prebuilt(
+                install_dir = tmp_path / "llama",
+                llama_tag = "latest",
+                published_repo = "ggml-org/llama.cpp",
+                published_release_tag = "",
+                simple_policy = True,
+                force_cpu = True,
+            )
+        host = captured["host"]
+        assert host.has_usable_nvidia is False
+        assert host.has_physical_nvidia is False
+        assert host.has_rocm is False
+        # Arch is preserved so the arm64 CPU bundle (not x64) is chosen.
+        assert host.is_arm64 is True
+
+    def test_cpu_forced_arm64_selects_ubuntu_arm64(self):
+        tag = "b9444"
+        release = {
+            "tag_name": tag,
+            "assets": [
+                {
+                    "name": f"llama-{tag}-bin-ubuntu-arm64.tar.gz",
+                    "browser_download_url": f"https://x/llama-{tag}-bin-ubuntu-arm64.tar.gz",
+                },
+                {
+                    "name": f"llama-{tag}-bin-ubuntu-x64.tar.gz",
+                    "browser_download_url": f"https://x/llama-{tag}-bin-ubuntu-x64.tar.gz",
+                },
+            ],
+        }
+        # A GPU arm64 host cannot pick the CPU arm64 bundle on its own.
+        with pytest.raises(PrebuiltFallback):
+            direct_upstream_release_plan(
+                release, self._arm64_nvidia(), "ggml-org/llama.cpp", "latest"
+            )
+        # force_cpu drops the GPU attributes, so the CPU arm64 bundle is selected.
+        cpu_host = make_host(
+            system = "Linux",
+            machine = "aarch64",
+            nvidia_smi = None,
+            driver_cuda_version = None,
+            compute_caps = [],
+            has_physical_nvidia = False,
+            has_usable_nvidia = False,
+        )
+        plan = direct_upstream_release_plan(
+            release, cpu_host, "ggml-org/llama.cpp", "latest"
+        )
+        assert plan.attempts[0].install_kind == "linux-arm64"
+        assert plan.attempts[0].name == f"llama-{tag}-bin-ubuntu-arm64.tar.gz"
+
+    def test_setup_sh_has_arm64_cpu_prebuilt_fallback(self):
+        source = self._SETUP_SH.read_text(encoding = "utf-8")
+        assert "--cpu-fallback" in source
+        # Fallback targets ggml-org (the only repo with an arm64 Linux build) and
+        # is gated on a degraded source build for arm64.
+        assert "ggml-org/llama.cpp" in source
+        assert "_LLAMA_CPP_DEGRADED" in source

tests/studio/test_resolve_cuda_toolkit.ps1

@@ -0,0 +1,121 @@
+#!/usr/bin/env pwsh
+# Unit test for Resolve-CudaToolkit in studio/setup.ps1. No GPU required: the
+# detection helpers (nvidia-smi, nvcc, Find-Nvcc, ...) are stubbed so the real
+# function logic runs against a spoofed Blackwell sm_120 / driver 13.2 host.
+#
+# The function is extracted via AST and run in a child pwsh per scenario, because
+# the -RequireOrExit path calls `exit` (which would otherwise kill this harness).
+#
+# Run: pwsh -NoProfile -File tests/studio/test_resolve_cuda_toolkit.ps1
+
+$ErrorActionPreference = "Stop"
+$setupPath = [System.IO.Path]::Combine($PSScriptRoot, "..", "..", "studio", "setup.ps1")
+$setupPath = (Resolve-Path $setupPath).Path
+
+# --- Extract the function source (not the whole installer) ---
+$tokens = $null; $errors = $null
+$ast = [System.Management.Automation.Language.Parser]::ParseFile($setupPath, [ref]$tokens, [ref]$errors)
+if ($errors) { $errors | ForEach-Object { $_.ToString() }; throw "setup.ps1 has parse errors" }
+$fn = $ast.FindAll({ param($n)
+    $n -is [System.Management.Automation.Language.FunctionDefinitionAst] -and $n.Name -eq "Resolve-CudaToolkit"
+}, $true)
+if ($fn.Count -ne 1) { throw "expected exactly one Resolve-CudaToolkit, found $($fn.Count)" }
+$fnText = $fn[0].Extent.Text
+
+# --- Spoof executables: nvidia-smi reports driver max CUDA 13.2; nvcc 13.3 ---
+$work = Join-Path ([System.IO.Path]::GetTempPath()) ("rct_" + [guid]::NewGuid().ToString("N"))
+New-Item -ItemType Directory -Force -Path $work | Out-Null
+$smiFake  = Join-Path $work "nvidia-smi.ps1"
+$nvccFake = Join-Path $work "nvcc.ps1"
+Set-Content -LiteralPath $smiFake  -Value "'CUDA Version: 13.2'"
+Set-Content -LiteralPath $nvccFake -Value "'Cuda compilation tools, release 13.3, V13.3.0'"
+
+$failures = 0
+function Check($name, $cond) {
+    if ($cond) { Write-Host "  PASS  $name" }
+    else { Write-Host "  FAIL  $name" -ForegroundColor Red; $script:failures++ }
+}
+
+# Build + run one scenario in a child pwsh; returns @{ Exit; Out }.
+function Run-Case {
+    param([string]$FindMode, [bool]$Require)
+    $requireLit = if ($Require) { '$true' } else { '$false' }
+    $child = @"
+`$ErrorActionPreference = 'Continue'
+[Environment]::SetEnvironmentVariable('CUDA_PATH', `$null, 'Process')
+`$FindNvccMode = '$FindMode'
+`$NvccFake = '$nvccFake'
+function substep { param(`$m, `$c) Write-Host "  `$m" }
+function step    { param(`$l, `$v, `$c) Write-Host "[`$l] `$v" }
+function Add-ToUserPath { param(`$Directory, `$Position) `$true }
+function Refresh-Environment { }
+function Get-CudaComputeCapability { '120' }
+function Test-NvccArchSupport { param(`$NvccExe, `$Arch) `$true }
+function Get-NvccMaxArch { param(`$NvccExe) '120' }
+`$script:WingetCalled = `$false
+function winget { `$script:WingetCalled = `$true; 'no matching versions' }
+function Find-Nvcc {
+    param([string]`$MaxVersion = '')
+    switch (`$FindNvccMode) {
+        'compatible'   { return `$NvccFake }
+        'incompatible' { if (`$MaxVersion) { return `$null } else { return `$NvccFake } }
+        default        { return `$null }
+    }
+}
+`$NvidiaSmiExe  = '$smiFake'
+`$VsInstallPath = `$null
+`$HasNvidiaSmi  = `$true
+`$script:CudaToolkitReady = `$false
+`$script:NvccPath = `$null; `$script:CudaToolkitRoot = `$null; `$script:CudaArch = `$null
+
+$fnText
+
+if ($requireLit) { Resolve-CudaToolkit -RequireOrExit } else { Resolve-CudaToolkit }
+Write-Host ("RESULT ready={0} nvcc={1} winget={2}" -f `$script:CudaToolkitReady, `$script:NvccPath, `$script:WingetCalled)
+"@
+    $childFile = Join-Path $work ("case_" + [guid]::NewGuid().ToString("N") + ".ps1")
+    Set-Content -LiteralPath $childFile -Value $child
+    $out = & pwsh -NoProfile -File $childFile 2>&1 | Out-String
+    return @{ Exit = $LASTEXITCODE; Out = $out }
+}
+
+try {
+    Write-Host "Scenario 1: prebuilt path, too-new toolkit (no -RequireOrExit) -> defers, no exit"
+    $r = Run-Case -FindMode "incompatible" -Require $false
+    Check "exits 0 (not blocked)"        ($r.Exit -eq 0)
+    Check "CudaToolkitReady = false"     ($r.Out -match "ready=False")
+    Check "winget NOT called"            ($r.Out -match "winget=False")
+    Check "no INCOMPATIBLE error text"   (-not ($r.Out -match "INCOMPATIBLE"))
+
+    Write-Host "Scenario 2: forced source build, too-new toolkit (-RequireOrExit) -> hard exit"
+    $r = Run-Case -FindMode "incompatible" -Require $true
+    Check "exits non-zero"               ($r.Exit -ne 0)
+    Check "preserved INCOMPATIBLE error" ($r.Out -match "is installed but INCOMPATIBLE")
+
+    Write-Host "Scenario 3: compatible toolkit (-RequireOrExit) -> resolves, env set"
+    $r = Run-Case -FindMode "compatible" -Require $true
+    Check "exits 0"                      ($r.Exit -eq 0)
+    Check "CudaToolkitReady = true"      ($r.Out -match "ready=True")
+    Check "NvccPath published"           ($r.Out -match "nvcc=.*nvcc")
+
+    Write-Host "Scenario 4: no toolkit, prebuilt path (no -RequireOrExit) -> defers, no winget"
+    $r = Run-Case -FindMode "none" -Require $false
+    Check "exits 0"                      ($r.Exit -eq 0)
+    Check "CudaToolkitReady = false"     ($r.Out -match "ready=False")
+    Check "winget NOT called"            ($r.Out -match "winget=False")
+
+    Write-Host "Scenario 5: no toolkit, forced (-RequireOrExit) -> winget attempted then exit"
+    # The function exits before the RESULT line here, so assert on the winget-block
+    # marker in output rather than the flag.
+    $r = Run-Case -FindMode "none" -Require $true
+    Check "winget attempted"             ($r.Out -match "installing via winget")
+    Check "exits non-zero"               ($r.Exit -ne 0)
+    Check "preserved nvcc-required error" ($r.Out -match "CUDA Toolkit \(nvcc\) is required")
+}
+finally {
+    Remove-Item -Recurse -Force -LiteralPath $work -ErrorAction SilentlyContinue
+}
+
+Write-Host ""
+if ($failures -gt 0) { Write-Host "$failures check(s) FAILED" -ForegroundColor Red; exit 1 }
+Write-Host "All checks passed" -ForegroundColor Green

unsloth/models/_utils.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2026.5.8"
+__version__ = "2026.5.10"
 
 __all__ = [
     "SUPPORTS_BFLOAT16",
@@ -1216,8 +1216,6 @@ if is_openai_available():
 
 # =============================================
 # Get Flash Attention v2 if Ampere (RTX 30xx, A100)
-import bitsandbytes as bnb
-
 from transformers import AutoTokenizer
 from transformers.utils.import_utils import _is_package_available