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

# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.

# Three end-to-end smoke jobs that boot a freshly-installed Studio and
# exercise the surfaces real users hit through the OpenAI / Anthropic
# SDKs and curl, on the FREE windows-latest runner. Each job picks the
# smallest model that exercises the behaviour under test, primes
# HF_HOME via actions/cache, and shares the install.ps1 --local
# --no-torch bootstrap.
#
#   1. OpenAI, Anthropic API tests
#        gemma-3-270m-it UD-Q4_K_XL (~254 MiB).
#   2. Tool calling Tests
#        Qwen3.5-2B UD-Q4_K_XL (~890 MiB).
#   3. JSON, images
#        Qwen3-VL-2B-Instruct UD-IQ2_XXS + mmproj-F16 (~1.4 GiB total).
#        Within the 14 GB windows-latest SSD budget.

name: Windows Studio GGUF CI

on:
  pull_request:
    paths:
      - 'studio/**'
      - 'unsloth/**'
      - 'unsloth_cli/**'
      - 'install.ps1'
      - 'pyproject.toml'
      - '.github/workflows/studio-windows-inference-smoke.yml'
  push:
    branches: [main, pip]
  workflow_dispatch:

concurrency:
  group: ${{ github.workflow }}-${{ github.ref }}
  cancel-in-progress: true

permissions:
  contents: read

jobs:
  # ─────────────────────────────────────────────────────────────────────
  # Job 1: OpenAI, Anthropic API tests
  # ─────────────────────────────────────────────────────────────────────
  openai-anthropic:
    name: OpenAI, Anthropic API tests
    runs-on: windows-latest
    timeout-minutes: 30
    defaults:
      run:
        shell: bash
    env:
      GGUF_REPO: unsloth/gemma-3-270m-it-GGUF
      GGUF_VARIANT: UD-Q4_K_XL
      GGUF_FILE: gemma-3-270m-it-UD-Q4_K_XL.gguf
      STUDIO_PORT: '18888'
      HF_HOME: ${{ github.workspace }}/hf-cache
      # Force UTF-8 for stdio (Windows defaults to cp1252; hf
      # download / Studio CLI print "✓" checkmarks and crash
      # otherwise).
      PYTHONIOENCODING: utf-8
      PYTHONUTF8: '1'
    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          persist-credentials: false

      # Fast GPU-free gate: parse install.ps1 + setup.ps1 and run the PowerShell
      # unit tests (CUDA-toolkit + torch-flavor helpers) before the heavy GGUF smoke.
      - name: PowerShell installer unit tests
        shell: pwsh
        run: |
          foreach ($f in @('install.ps1', 'studio/setup.ps1')) {
            $errs = $null
            [void][System.Management.Automation.Language.Parser]::ParseFile(
              (Resolve-Path $f).Path, [ref]$null, [ref]$errs)
            if ($errs) { $errs | ForEach-Object { $_.ToString() }; exit 1 }
            Write-Host "$f parsed with no errors"
          }
          pwsh -NoProfile -File tests/studio/test_resolve_cuda_toolkit.ps1
          pwsh -NoProfile -File tests/studio/test_torch_flavor.ps1

      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
        with:
          node-version: '22'

      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
        with:
          python-version: '3.12'

      # Split restore + save (rather than the one-step actions/cache) so a
      # transient restore-side failure does not kill the whole job. v5 has a
      # known flake where it logs "Cache hit for: <key>" and then exits
      # non-zero without actually extracting the archive (see
      # actions/cache#1621 and github community discussion #163260).
      # continue-on-error on restore masks that failure so the Prime step
      # below can re-download from HF and the job keeps running. Save then
      # populates the cache key on a real miss only; cache keys are
      # immutable, so a corrupted cached entry persists until the -v1
      # suffix below is bumped.
      - name: Restore HF_HOME cache for ${{ env.GGUF_REPO }}
        id: cache-hf
        uses: actions/cache/restore@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
        continue-on-error: true
        with:
          path: hf-cache
          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2

      - name: Prime HF_HOME with the GGUF
        id: prime-hf
        # Run on a real cache miss AND on the silent-restore-failure mode
        # described above (outcome != success).
        if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
        env:
          HF_TOKEN: ${{ secrets.HF_TOKEN }}
        run: |
          python -m pip install --upgrade huggingface_hub
          mkdir -p hf-cache
          bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE"
          bash .github/scripts/hf-download-with-retry.sh ggml-org/models tinyllamas/stories260K.gguf

      - name: Save HF_HOME cache for ${{ env.GGUF_REPO }}
        # Only write a fresh cache entry when we actually rebuilt the
        # directory (Prime ran and succeeded). Skipping when Prime is
        # skipped avoids "already exists" save warnings on the happy path.
        if: always() && steps.prime-hf.outcome == 'success'
        uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
        with:
          path: hf-cache
          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-v2

      - name: Pre-install Windows tweaks (npm 11 + Defender exclusions)
        shell: pwsh
        # See studio-windows-update-smoke.yml for the full rationale.
        # tl;dr: setup.ps1 needs npm >=11 to skip a 35 s winget Node
        # reinstall, and Defender's real-time scan dominates the
        # frontend / uv-pip-extract steps.
        run: |
          $ProgressPreference = 'SilentlyContinue'
          Write-Host "npm version before upgrade: $(npm -v)"
          npm install -g 'npm@^11' 2>&1 | Out-Host
          Write-Host "npm version after upgrade: $(npm -v)"
          # NOTE: do NOT pre-create these directories. See
          # studio-windows-update-smoke.yml for the full rationale --
          # creating an empty studio/frontend/dist trips setup.ps1's
          # mtime-based staleness check into "frontend up to date, skip
          # rebuild" and Studio boots with an empty dist directory.
          # Add-MpPreference accepts paths that do not yet exist.
          foreach ($p in @(
            "$env:USERPROFILE\.unsloth",
            "$env:USERPROFILE\AppData\Local\uv",
            "$env:GITHUB_WORKSPACE\studio\frontend\node_modules",
            "$env:GITHUB_WORKSPACE\studio\frontend\dist"
          )) {
            try {
              Add-MpPreference -ExclusionPath $p -ErrorAction Stop
              Write-Host "Defender exclusion added: $p"
            } catch {
              Write-Host "Defender exclusion skipped ($($_.Exception.Message)): $p"
            }
          }

      - name: Install Studio (--local, --no-torch)
        shell: pwsh
        env:
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          HF_TOKEN: ${{ secrets.HF_TOKEN }}
        run: |
          New-Item -ItemType Directory -Force -Path logs | Out-Null
          # *>&1 captures Write-Host (Information stream) output;
          # plain 2>&1 does not. setup.ps1 emits "prebuilt installed
          # and validated" via Write-Host, and we grep for that.
          $ProgressPreference = 'SilentlyContinue'
          & ./install.ps1 --local --no-torch *>&1 | Tee-Object -FilePath logs/install.log

      - name: Assert install.ps1 used the Windows llama.cpp prebuilt
        run: |
          # Filesystem check; setup.ps1's stream output isn't captured.
          LLAMA_DIR=~/.unsloth/llama.cpp
          INFO="$LLAMA_DIR/UNSLOTH_PREBUILT_INFO.json"
          BIN="$LLAMA_DIR/build/bin/Release/llama-server.exe"
          if grep -q "falling back to source build" logs/install.log; then
            echo "::error::install.ps1 fell back to source-build llama.cpp on Windows."
            grep -E "llama-prebuilt|llama.cpp" logs/install.log | tail -60
            exit 1
          fi
          if [ ! -f "$INFO" ]; then
            echo "::error::no UNSLOTH_PREBUILT_INFO.json at $INFO."
            ls -la "$LLAMA_DIR" || true
            exit 1
          fi
          if [ ! -f "$BIN" ]; then
            echo "::error::no llama-server.exe at $BIN."
            ls -la "$LLAMA_DIR/build/bin" || true
            exit 1
          fi
          echo "install.ps1 installed the Windows prebuilt llama.cpp:"
          cat "$INFO"

      - name: Add Studio shim to GITHUB_PATH
        run: |
          SHIM_DIR=~/.unsloth/studio/bin
          if [ ! -f "$SHIM_DIR/unsloth.exe" ]; then
            echo "::error::unsloth.exe shim not found at $SHIM_DIR"
            ls -la ~/.unsloth/studio/ || true
            exit 1
          fi
          cygpath -w "$SHIM_DIR" >> "$GITHUB_PATH"

      - name: Install OpenAI + Anthropic Python SDKs
        run: python -m pip install 'openai>=1.50' 'anthropic>=0.40'

      - name: Reset auth + boot Studio (API-only)
        run: |
          unsloth studio reset-password
          mkdir -p logs
          UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
            > logs/studio.log 2>&1 &
          echo "STUDIO_PID=$!" >> "$GITHUB_ENV"

      - name: Wait for /api/health
        run: |
          for i in $(seq 1 180); do
            if curl -fs "http://127.0.0.1:${STUDIO_PORT}/api/health" > /tmp/health.json; then
              jq -e '.status == "healthy"' /tmp/health.json
              exit 0
            fi
            sleep 1
          done
          echo "Studio did not become healthy in 180s"
          tail -200 logs/studio.log
          exit 1

      - name: Password rotation (old must fail, new must work)
        run: |
          OLD=$(cat ~/.unsloth/studio/auth/.bootstrap_password)
          NEW="CIRotated-$(python -c 'import secrets; print(secrets.token_urlsafe(12))')"
          echo "::add-mask::$OLD"
          echo "::add-mask::$NEW"
          OLD_TOKEN=$(curl -fs -X POST "http://127.0.0.1:${STUDIO_PORT}/api/auth/login" \
            -H 'content-type: application/json' \
            -d "{\"username\":\"unsloth\",\"password\":\"$OLD\"}" | jq -r .access_token)
          [ -n "$OLD_TOKEN" ] && [ "$OLD_TOKEN" != "null" ] || { echo "bootstrap login failed"; exit 1; }
          curl -fs -X POST "http://127.0.0.1:${STUDIO_PORT}/api/auth/change-password" \
            -H "Authorization: Bearer $OLD_TOKEN" -H 'content-type: application/json' \
            -d "{\"current_password\":\"$OLD\",\"new_password\":\"$NEW\"}" > /dev/null
          OLD_STATUS=$(curl -s -o /dev/null -w '%{http_code}' \
            -X POST "http://127.0.0.1:${STUDIO_PORT}/api/auth/login" \
            -H 'content-type: application/json' \
            -d "{\"username\":\"unsloth\",\"password\":\"$OLD\"}")
          if [ "$OLD_STATUS" != "401" ]; then
            echo "::error::Login with old password returned $OLD_STATUS, expected 401"
            exit 1
          fi
          NEW_TOKEN=$(curl -fs -X POST "http://127.0.0.1:${STUDIO_PORT}/api/auth/login" \
            -H 'content-type: application/json' \
            -d "{\"username\":\"unsloth\",\"password\":\"$NEW\"}" | jq -r .access_token)
          [ -n "$NEW_TOKEN" ] && [ "$NEW_TOKEN" != "null" ] || { echo "new login failed"; exit 1; }
          echo "TOKEN=$NEW_TOKEN" >> "$GITHUB_ENV"
          echo "password rotation OK (old=401, new=200)"

      - name: Load the GGUF (HF repo + variant, served from HF_HOME cache)
        run: |
          # Retry the load step a few times so a transient TCP RST during
          # llama-server warm-up (Windows runner image churn,
          # windows-latest -> windows-2025-vs2026 rollout) doesn't fail
          # the whole job. The Studio backend's _wait_for_health now
          # catches httpx.ReadError too; this retry layer covers the
          # cases the backend can't recover from on its own.
          LOAD_OK=0
          for attempt in 1 2 3; do
            HTTP=$(curl -s -o /tmp/load.json -w '%{http_code}' \
              -X POST "http://127.0.0.1:${STUDIO_PORT}/api/inference/load" \
              -H "Authorization: Bearer $TOKEN" -H 'content-type: application/json' \
              --max-time 600 \
              -d "{\"model_path\":\"$GGUF_REPO\",\"gguf_variant\":\"$GGUF_VARIANT\",\"is_lora\":false,\"max_seq_length\":2048}")
            if [ "$HTTP" = "200" ]; then LOAD_OK=1; break; fi
            echo "::warning::/api/inference/load attempt $attempt returned $HTTP; response:"
            cat /tmp/load.json || true
            sleep 10
          done
          [ "$LOAD_OK" = "1" ] || { echo "::error::/api/inference/load failed 3 attempts"; exit 22; }
          jq '{status, display_name, is_gguf, context_length}' /tmp/load.json

      - name: Multi-turn determinism via OpenAI + Anthropic SDKs
        env:
          BASE_URL: http://127.0.0.1:18888
        run: |
          python - <<'PY'
          import json
          import os
          from openai import OpenAI
          from anthropic import Anthropic

          BASE = os.environ["BASE_URL"]
          KEY  = os.environ["TOKEN"]
          SEED = 3407

          PROMPTS = [
              "What is 1+1?",
              "What did I ask before?",
              "What is the capital of France?",
              "Repeat the city name",
          ]

          def run_openai():
              client = OpenAI(base_url = f"{BASE}/v1", api_key = KEY)
              history, replies = [], []
              for prompt in PROMPTS:
                  history.append({"role": "user", "content": prompt})
                  resp = client.chat.completions.create(
                      model       = "default",
                      messages    = history,
                      temperature = 0.0,
                      max_tokens  = 80,
                      seed        = SEED,
                      extra_body  = {"enable_thinking": False},
                  )
                  text = resp.choices[0].message.content or ""
                  replies.append(text)
                  history.append({"role": "assistant", "content": text})
              return replies

          def run_anthropic():
              client = Anthropic(
                  base_url        = BASE,
                  api_key         = "unused",
                  default_headers = {"Authorization": f"Bearer {KEY}"},
              )
              history, replies = [], []
              for prompt in PROMPTS:
                  history.append({"role": "user", "content": prompt})
                  msg = client.messages.create(
                      model       = "default",
                      max_tokens  = 80,
                      messages    = history,
                      temperature = 0.0,
                      extra_body  = {"seed": SEED, "enable_thinking": False},
                  )
                  text = "".join(b.text for b in msg.content if getattr(b, "type", None) == "text")
                  replies.append(text)
                  history.append({"role": "assistant", "content": text})
              return replies

          for label, runner in (("openai", run_openai), ("anthropic", run_anthropic)):
              first  = runner()
              second = runner()
              for i, (a, b) in enumerate(zip(first, second), start = 1):
                  print(f"[{label} turn {i}] {a!r}")
                  assert a, f"{label}: empty turn {i} response"
                  # Compare on stripped content: llama-server can vary
                  # trailing whitespace (specifically a final '\n') between
                  # otherwise-identical greedy runs depending on the
                  # batch-flush boundary at which the stream is closed. The
                  # generated tokens are identical; only the trailing
                  # whitespace differs. Keep the raw repr in the failure
                  # message so a real divergence is still legible.
                  assert a.strip() == b.strip(), (
                      f"{label} non-deterministic at turn {i} with temperature=0.0:\n"
                      f"  run1: {a!r}\n  run2: {b!r}"
                  )
              joined = " ".join(first).lower()
              assert "1" in first[0], f"{label}: turn-1 answer should contain '1', got {first[0]!r}"
              assert "paris" in joined, f"{label}: expected 'paris' somewhere in the four-turn transcript: {first}"
              print(f"[{label}] OK -- 4 turns, run1 == run2, history grounded")
          PY

      - name: Stop Studio
        if: always()
        # Run as cmd so we are not running through the Git Bash shell;
        # Git Bash on windows-latest has been observed to exit 143
        # (SIGTERM) from any inline kill/sleep block, masking a green
        # test run. The runner reclaims the Studio child process at
        # job end either way, so just emit a marker and exit 0.
        shell: cmd
        run: echo Stop Studio (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end)

      - 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
        # upload-artifact can pick it up. Crucial for diagnosing a
        # subprocess crash where Studio's traceback only shows the
        # symptom (httpx ReadError) but not the cause.
        run: |
          mkdir -p logs/llama-server
          cp -v ~/.unsloth/studio/logs/llama-server/*.log logs/llama-server/ 2>/dev/null || \
            echo "no llama-server logs to collect"

      - 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
          path: |
            logs/studio.log
            logs/install.log
            logs/llama-server/*.log
          retention-days: 7

  # ─────────────────────────────────────────────────────────────────────
  # Job 2: Tool calling Tests
  # ─────────────────────────────────────────────────────────────────────
  tool-calling:
    name: Tool calling Tests
    runs-on: windows-latest
    timeout-minutes: 30
    defaults:
      run:
        shell: bash
    env:
      # Tool calling is the highest-volume GGUF in this workflow
      # (Qwen3.5-2B at Q4_K_XL = ~1.28 GiB). The previous HF_HOME
      # cache stored xet chunks + blobs + snapshots = ~4.7 GiB --
      # 3.7x file-size inflation, dominating the post-step upload
      # (211 s on first run; subsequent runs hit the cache, but the
      # one-time cost recurs every time the cache key bumps). Use
      # main's `--local-dir gguf-cache` pattern: cache the flat .gguf
      # only, pass an absolute path to Studio's /api/inference/load.
      # The OpenAI/Anth and JSON+images jobs still cover the
      # gguf_variant resolution path.
      GGUF_REPO: unsloth/Qwen3.5-2B-GGUF
      GGUF_FILE: Qwen3.5-2B-UD-Q4_K_XL.gguf
      STUDIO_PORT: '18898'
      # Force UTF-8 for stdio (Windows defaults to cp1252; hf
      # download / Studio CLI print "✓" checkmarks and crash
      # otherwise).
      PYTHONIOENCODING: utf-8
      PYTHONUTF8: '1'
    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          persist-credentials: false

      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
        with:
          node-version: '22'

      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
        with:
          python-version: '3.12'

      # Split restore + save so a transient restore-side failure does not
      # kill the whole job. See the matching block in the tool-calling job
      # above for the full rationale (actions/cache#1621).
      - name: Restore GGUF model cache
        id: cache-gguf
        uses: actions/cache/restore@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
        continue-on-error: true
        with:
          path: gguf-cache
          key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1

      - name: Download GGUF if cache miss
        id: download-gguf
        if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
        env:
          HF_TOKEN: ${{ secrets.HF_TOKEN }}
        run: |
          python -m pip install --upgrade huggingface_hub
          mkdir -p gguf-cache
          bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE" gguf-cache

      - name: Save GGUF model cache
        if: always() && steps.download-gguf.outcome == 'success'
        uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
        with:
          path: gguf-cache
          key: ${{ runner.os }}-gguf-${{ env.GGUF_REPO }}-${{ env.GGUF_FILE }}-v1

      - name: Pre-install Windows tweaks (npm 11 + Defender exclusions)
        shell: pwsh
        # See studio-windows-update-smoke.yml for the full rationale.
        # tl;dr: setup.ps1 needs npm >=11 to skip a 35 s winget Node
        # reinstall, and Defender's real-time scan dominates the
        # frontend / uv-pip-extract steps.
        run: |
          $ProgressPreference = 'SilentlyContinue'
          Write-Host "npm version before upgrade: $(npm -v)"
          npm install -g 'npm@^11' 2>&1 | Out-Host
          Write-Host "npm version after upgrade: $(npm -v)"
          # NOTE: do NOT pre-create these directories. See
          # studio-windows-update-smoke.yml for the full rationale --
          # creating an empty studio/frontend/dist trips setup.ps1's
          # mtime-based staleness check into "frontend up to date, skip
          # rebuild" and Studio boots with an empty dist directory.
          # Add-MpPreference accepts paths that do not yet exist.
          foreach ($p in @(
            "$env:USERPROFILE\.unsloth",
            "$env:USERPROFILE\AppData\Local\uv",
            "$env:GITHUB_WORKSPACE\studio\frontend\node_modules",
            "$env:GITHUB_WORKSPACE\studio\frontend\dist"
          )) {
            try {
              Add-MpPreference -ExclusionPath $p -ErrorAction Stop
              Write-Host "Defender exclusion added: $p"
            } catch {
              Write-Host "Defender exclusion skipped ($($_.Exception.Message)): $p"
            }
          }

      - name: Install Studio (--local, --no-torch)
        shell: pwsh
        env:
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          HF_TOKEN: ${{ secrets.HF_TOKEN }}
        run: |
          New-Item -ItemType Directory -Force -Path logs | Out-Null
          # *>&1 captures Write-Host (Information stream) output;
          # plain 2>&1 does not. setup.ps1 emits "prebuilt installed
          # and validated" via Write-Host, and we grep for that.
          $ProgressPreference = 'SilentlyContinue'
          & ./install.ps1 --local --no-torch *>&1 | Tee-Object -FilePath logs/install.log

      - name: Assert install.ps1 used the Windows llama.cpp prebuilt
        run: |
          # Filesystem check; setup.ps1's stream output isn't captured.
          LLAMA_DIR=~/.unsloth/llama.cpp
          INFO="$LLAMA_DIR/UNSLOTH_PREBUILT_INFO.json"
          BIN="$LLAMA_DIR/build/bin/Release/llama-server.exe"
          if grep -q "falling back to source build" logs/install.log; then
            echo "::error::install.ps1 fell back to source-build llama.cpp on Windows."
            grep -E "llama-prebuilt|llama.cpp" logs/install.log | tail -60
            exit 1
          fi
          if [ ! -f "$INFO" ]; then
            echo "::error::no UNSLOTH_PREBUILT_INFO.json at $INFO."
            ls -la "$LLAMA_DIR" || true
            exit 1
          fi
          if [ ! -f "$BIN" ]; then
            echo "::error::no llama-server.exe at $BIN."
            ls -la "$LLAMA_DIR/build/bin" || true
            exit 1
          fi
          echo "install.ps1 installed the Windows prebuilt llama.cpp:"
          cat "$INFO"

      - name: Add Studio shim to GITHUB_PATH
        run: |
          SHIM_DIR=~/.unsloth/studio/bin
          if [ ! -f "$SHIM_DIR/unsloth.exe" ]; then
            echo "::error::unsloth.exe shim not found at $SHIM_DIR"
            ls -la ~/.unsloth/studio/ || true
            exit 1
          fi
          cygpath -w "$SHIM_DIR" >> "$GITHUB_PATH"

      - name: Reset auth + boot Studio (API-only, default tool policy)
        run: |
          unsloth studio reset-password
          mkdir -p logs
          UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
            > logs/studio.log 2>&1 &
          echo "STUDIO_PID=$!" >> "$GITHUB_ENV"

      - name: Wait for /api/health, log in, change password, load model
        run: |
          for i in $(seq 1 180); do
            if curl -fs "http://127.0.0.1:${STUDIO_PORT}/api/health" > /tmp/health.json; then
              jq -e '.status == "healthy"' /tmp/health.json && break
            fi
            sleep 1
          done
          jq -e '.status == "healthy"' /tmp/health.json
          OLD=$(cat ~/.unsloth/studio/auth/.bootstrap_password)
          NEW="CITool-$(python -c 'import secrets; print(secrets.token_urlsafe(12))')"
          echo "::add-mask::$OLD"
          echo "::add-mask::$NEW"
          OLD_TOKEN=$(curl -fs -X POST "http://127.0.0.1:${STUDIO_PORT}/api/auth/login" \
            -H 'content-type: application/json' \
            -d "{\"username\":\"unsloth\",\"password\":\"$OLD\"}" | jq -r .access_token)
          curl -fs -X POST "http://127.0.0.1:${STUDIO_PORT}/api/auth/change-password" \
            -H "Authorization: Bearer $OLD_TOKEN" -H 'content-type: application/json' \
            -d "{\"current_password\":\"$OLD\",\"new_password\":\"$NEW\"}" > /dev/null
          TOKEN=$(curl -fs -X POST "http://127.0.0.1:${STUDIO_PORT}/api/auth/login" \
            -H 'content-type: application/json' \
            -d "{\"username\":\"unsloth\",\"password\":\"$NEW\"}" | jq -r .access_token)
          echo "API_KEY=$TOKEN" >> "$GITHUB_ENV"
          # GITHUB_WORKSPACE on windows-latest is a Windows path with
          # backslashes ("D:\a\unsloth\unsloth"). Bash handles it as a
          # raw string, but we cannot embed `\a` etc. in JSON without
          # JSON-string-escaping every backslash. Replace `\` with `/`
          # via bash parameter expansion -- pathlib.Path on Windows
          # accepts forward slashes natively, so Studio's loader sees
          # a normal path.
          GGUF_PATH="${GITHUB_WORKSPACE//\\//}/gguf-cache/${GGUF_FILE}"
          ls -lh "$GGUF_PATH"
          # Retry: same rationale as the OpenAI/Anthropic job.
          LOAD_OK=0
          for attempt in 1 2 3; do
            HTTP=$(curl -s -o /tmp/load.json -w '%{http_code}' \
              -X POST "http://127.0.0.1:${STUDIO_PORT}/api/inference/load" \
              -H "Authorization: Bearer $TOKEN" -H 'content-type: application/json' \
              --max-time 600 \
              -d "{\"model_path\":\"$GGUF_PATH\",\"is_lora\":false,\"max_seq_length\":2048}")
            if [ "$HTTP" = "200" ]; then LOAD_OK=1; break; fi
            echo "::warning::/api/inference/load attempt $attempt returned $HTTP; response:"
            cat /tmp/load.json || true
            sleep 10
          done
          [ "$LOAD_OK" = "1" ] || { echo "::error::/api/inference/load failed 3 attempts"; exit 22; }
          jq '{status, display_name}' /tmp/load.json

      - name: Tool calling, server-side tools, thinking on/off
        env:
          BASE_URL: http://127.0.0.1:18898
        run: |
          python - <<'PY'
          import json
          import os
          import urllib.request

          BASE = os.environ["BASE_URL"]
          KEY  = os.environ["API_KEY"]
          SEED = 3407
          # Same temperature shim as the Mac job. Small Qwen3.5-2B
          # quants can degenerate at temperature=0; a small non-zero
          # temperature with a fixed seed keeps the test deterministic
          # while escaping the trap.
          TEMP = 0.2

          def post(path, body, *, timeout = 240):
              data = json.dumps(body).encode()
              req = urllib.request.Request(
                  f"{BASE}{path}",
                  data    = data,
                  method  = "POST",
                  headers = {
                      "Authorization": f"Bearer {KEY}",
                      "Content-Type": "application/json",
                  },
              )
              with urllib.request.urlopen(req, timeout = timeout) as resp:
                  return resp.status, json.loads(resp.read().decode())

          def post_sse(path, body, *, timeout = 600):
              body = {**body, "stream": True}
              data = json.dumps(body).encode()
              req = urllib.request.Request(
                  f"{BASE}{path}",
                  data    = data,
                  method  = "POST",
                  headers = {
                      "Authorization": f"Bearer {KEY}",
                      "Content-Type": "application/json",
                  },
              )
              parts = []
              with urllib.request.urlopen(req, timeout = timeout) as resp:
                  for raw in resp:
                      line = raw.decode().strip()
                      if not line.startswith("data: "):
                          continue
                      payload = line[6:]
                      if payload == "[DONE]":
                          break
                      try:
                          chunk = json.loads(payload)
                      except json.JSONDecodeError:
                          continue
                      for choice in chunk.get("choices", []):
                          delta = choice.get("delta", {}) or {}
                          if delta.get("content"):
                              parts.append(delta["content"])
              return "".join(parts)

          # ── 1. Standard OpenAI function calling ──────────────────────
          weather_tool = {
              "type": "function",
              "function": {
                  "name": "get_weather",
                  "description": "Get current weather for a city.",
                  "parameters": {
                      "type": "object",
                      "properties": {"city": {"type": "string"}},
                      "required": ["city"],
                  },
              },
          }

          status, data = post("/v1/chat/completions", {
              "messages":    [{"role": "user", "content": "What is the weather in Paris?"}],
              "tools":       [weather_tool],
              "tool_choice": "required",
              "stream":      False,
              "temperature": TEMP,
              "seed":        SEED,
              "max_tokens":  600,
          })
          assert status == 200, f"tool call status {status}: {data}"
          choice = data["choices"][0]
          tool_calls = (choice.get("message") or {}).get("tool_calls") or []
          if tool_calls:
              tc = tool_calls[0]
              assert tc["function"]["name"] == "get_weather", (
                  f"unexpected tool name: {tc['function']['name']!r}"
              )
              args = json.loads(tc["function"]["arguments"])
              assert args.get("city"), f"missing city arg: {args}"
              print(f"[tools] PASS function calling -> {tc['function']['name']}({args}) finish={choice.get('finish_reason')!r}")
          else:
              print(
                  f"[tools] WARN function calling: no tool_calls (finish_reason="
                  f"{choice.get('finish_reason')!r}); HTTP path OK, model output drift."
              )

          # ── 2. Server-side python tool ───────────────────────────────
          content = post_sse("/v1/chat/completions", {
              "messages":      [{"role": "user", "content": "What is 123 * 456? Use the python tool to compute it and tell me the number."}],
              "enable_tools":  True,
              "enabled_tools": ["python"],
              "session_id":    "ci-tool-calling-py",
              "temperature":   TEMP,
              "seed":          SEED,
              "max_tokens":    600,
          })
          if "56088" in content or "56,088" in content:
              print(f"[tools] PASS python tool ({len(content)} chars, found 56088)")
          else:
              assert content, "python tool: SSE stream empty"
              print(
                  f"[tools] WARN python tool: SSE OK ({len(content)} chars) but "
                  f"model didn't return 56088 -- model output drift"
              )

          # ── 3. Server-side bash (terminal) tool ──────────────────────
          # On Windows the terminal tool resolves to the system shell
          # (cmd.exe wrapper) and `echo hello-bash-tool` works the same
          # way it does on POSIX. The model still has to choose to
          # invoke the tool; assert non-empty SSE if it doesn't.
          content = post_sse("/v1/chat/completions", {
              "messages":      [{"role": "user", "content": "Use the terminal tool to run `echo hello-bash-tool` and tell me the exact output."}],
              "enable_tools":  True,
              "enabled_tools": ["terminal"],
              "session_id":    "ci-tool-calling-bash",
              "temperature":   TEMP,
              "seed":          SEED,
              "max_tokens":    600,
          })
          if "hello-bash-tool" in content:
              print(f"[tools] PASS terminal tool ({len(content)} chars)")
          else:
              assert content, "terminal tool: SSE stream empty"
              print(
                  f"[tools] WARN terminal tool: SSE OK ({len(content)} chars) but "
                  f"model didn't echo 'hello-bash-tool' -- model output drift"
              )

          # ── 4. Server-side web_search tool ───────────────────────────
          # DuckDuckGo can be flaky from CI runners; only assert that
          # the SSE stream opens and yields any data.
          try:
              content = post_sse("/v1/chat/completions", {
                  "messages":      [{"role": "user", "content": "Search the web for 'unsloth ai github' and summarise."}],
                  "enable_tools":  True,
                  "enabled_tools": ["web_search"],
                  "session_id":    "ci-tool-calling-web",
                  "temperature":   TEMP,
                  "seed":          SEED,
                  "max_tokens":    400,
              })
              print(f"[tools] PASS web_search stream ({len(content)} chars)")
          except Exception as exc:
              print(f"[tools] WARN web_search probe failed (non-blocking): {exc}")

          # ── 5. Thinking on / off ─────────────────────────────────────
          def thinking_call(enable):
              status, data = post("/v1/chat/completions", {
                  "messages":        [{"role": "user", "content": "Briefly: is 17 prime?"}],
                  "stream":          False,
                  "enable_thinking": enable,
                  "temperature":     TEMP,
                  "seed":            SEED,
                  "max_tokens":      300,
              })
              assert status == 200
              msg = data["choices"][0]["message"]
              raw = (msg.get("content") or "") + (msg.get("reasoning_content") or "")
              return raw

          on_text  = thinking_call(True)
          off_text = thinking_call(False)
          had_think_on = ("<think>" in on_text) or len(on_text) > 80
          if not had_think_on:
              print(
                  f"[tools] WARN enable_thinking=True produced no thinking signal: "
                  f"{on_text[:200]!r}"
              )
          assert "<think>" not in off_text, (
              f"enable_thinking=False but <think> still present: {off_text!r}"
          )
          print(f"[tools] PASS thinking on/off (on={len(on_text)} chars, off={len(off_text)} chars)")
          PY

      - name: Stop Studio
        if: always()
        # Run as cmd so we are not running through the Git Bash shell;
        # Git Bash on windows-latest has been observed to exit 143
        # (SIGTERM) from any inline kill/sleep block, masking a green
        # test run. The runner reclaims the Studio child process at
        # job end either way, so just emit a marker and exit 0.
        shell: cmd
        run: echo Stop Studio (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end)

      - 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
        # upload-artifact can pick it up. Crucial for diagnosing a
        # subprocess crash where Studio's traceback only shows the
        # symptom (httpx ReadError) but not the cause.
        run: |
          mkdir -p logs/llama-server
          cp -v ~/.unsloth/studio/logs/llama-server/*.log logs/llama-server/ 2>/dev/null || \
            echo "no llama-server logs to collect"

      - 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
          path: |
            logs/studio.log
            logs/install.log
            logs/llama-server/*.log
          retention-days: 7

  # ─────────────────────────────────────────────────────────────────────
  # Job 3: JSON, images
  # ─────────────────────────────────────────────────────────────────────
  json-images:
    name: JSON, images
    runs-on: windows-latest
    timeout-minutes: 35
    defaults:
      run:
        shell: bash
    env:
      GGUF_REPO: unsloth/Qwen3-VL-2B-Instruct-GGUF
      GGUF_VARIANT: UD-IQ2_XXS
      GGUF_FILE: Qwen3-VL-2B-Instruct-UD-IQ2_XXS.gguf
      MMPROJ_FILE: mmproj-F16.gguf
      STUDIO_PORT: '18899'
      HF_HOME: ${{ github.workspace }}/hf-cache
      # Force UTF-8 for stdio (Windows defaults to cp1252; hf
      # download / Studio CLI print "✓" checkmarks and crash
      # otherwise).
      PYTHONIOENCODING: utf-8
      PYTHONUTF8: '1'
    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd  # v6.0.2
        with:
          persist-credentials: false

      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e  # v6.4.0
        with:
          node-version: '22'

      - uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405  # v6.2.0
        with:
          python-version: '3.12'

      # Split restore + save so a transient restore-side failure does not
      # kill the whole job. See the matching block in the tool-calling job
      # for the full rationale (actions/cache#1621). This is the block that
      # actually broke in run 25713577488: "Cache hit for: <key>" was
      # logged, the step exited non-zero in ~0.3 s without extracting the
      # 3.4 GiB archive, and steps 6-15 were skipped.
      - name: Restore HF_HOME cache for ${{ env.GGUF_REPO }} (model + mmproj)
        id: cache-hf
        uses: actions/cache/restore@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
        continue-on-error: true
        with:
          path: hf-cache
          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-${{ env.MMPROJ_FILE }}-v2

      - name: Prime HF_HOME with the GGUF + mmproj
        id: prime-hf
        if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
        env:
          HF_TOKEN: ${{ secrets.HF_TOKEN }}
        run: |
          python -m pip install --upgrade huggingface_hub
          mkdir -p hf-cache
          bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$GGUF_FILE"
          bash .github/scripts/hf-download-with-retry.sh "$GGUF_REPO" "$MMPROJ_FILE"
          bash .github/scripts/hf-download-with-retry.sh ggml-org/models tinyllamas/stories260K.gguf

      - name: Save HF_HOME cache for ${{ env.GGUF_REPO }} (model + mmproj)
        if: always() && steps.prime-hf.outcome == 'success'
        uses: actions/cache/save@27d5ce7f107fe9357f9df03efb73ab90386fccae  # v5.0.5
        with:
          path: hf-cache
          key: ${{ runner.os }}-hf-${{ env.GGUF_REPO }}-${{ env.GGUF_VARIANT }}-${{ env.MMPROJ_FILE }}-v2

      - name: Pre-install Windows tweaks (npm 11 + Defender exclusions)
        shell: pwsh
        # See studio-windows-update-smoke.yml for the full rationale.
        # tl;dr: setup.ps1 needs npm >=11 to skip a 35 s winget Node
        # reinstall, and Defender's real-time scan dominates the
        # frontend / uv-pip-extract steps.
        run: |
          $ProgressPreference = 'SilentlyContinue'
          Write-Host "npm version before upgrade: $(npm -v)"
          npm install -g 'npm@^11' 2>&1 | Out-Host
          Write-Host "npm version after upgrade: $(npm -v)"
          # NOTE: do NOT pre-create these directories. See
          # studio-windows-update-smoke.yml for the full rationale --
          # creating an empty studio/frontend/dist trips setup.ps1's
          # mtime-based staleness check into "frontend up to date, skip
          # rebuild" and Studio boots with an empty dist directory.
          # Add-MpPreference accepts paths that do not yet exist.
          foreach ($p in @(
            "$env:USERPROFILE\.unsloth",
            "$env:USERPROFILE\AppData\Local\uv",
            "$env:GITHUB_WORKSPACE\studio\frontend\node_modules",
            "$env:GITHUB_WORKSPACE\studio\frontend\dist"
          )) {
            try {
              Add-MpPreference -ExclusionPath $p -ErrorAction Stop
              Write-Host "Defender exclusion added: $p"
            } catch {
              Write-Host "Defender exclusion skipped ($($_.Exception.Message)): $p"
            }
          }

      - name: Install Studio (--local, --no-torch)
        shell: pwsh
        env:
          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          HF_TOKEN: ${{ secrets.HF_TOKEN }}
        run: |
          New-Item -ItemType Directory -Force -Path logs | Out-Null
          # *>&1 captures Write-Host (Information stream) output;
          # plain 2>&1 does not. setup.ps1 emits "prebuilt installed
          # and validated" via Write-Host, and we grep for that.
          $ProgressPreference = 'SilentlyContinue'
          & ./install.ps1 --local --no-torch *>&1 | Tee-Object -FilePath logs/install.log

      - name: Assert install.ps1 used the Windows llama.cpp prebuilt
        run: |
          # Filesystem check; setup.ps1's stream output isn't captured.
          LLAMA_DIR=~/.unsloth/llama.cpp
          INFO="$LLAMA_DIR/UNSLOTH_PREBUILT_INFO.json"
          BIN="$LLAMA_DIR/build/bin/Release/llama-server.exe"
          if grep -q "falling back to source build" logs/install.log; then
            echo "::error::install.ps1 fell back to source-build llama.cpp on Windows."
            grep -E "llama-prebuilt|llama.cpp" logs/install.log | tail -60
            exit 1
          fi
          if [ ! -f "$INFO" ]; then
            echo "::error::no UNSLOTH_PREBUILT_INFO.json at $INFO."
            ls -la "$LLAMA_DIR" || true
            exit 1
          fi
          if [ ! -f "$BIN" ]; then
            echo "::error::no llama-server.exe at $BIN."
            ls -la "$LLAMA_DIR/build/bin" || true
            exit 1
          fi
          echo "install.ps1 installed the Windows prebuilt llama.cpp:"
          cat "$INFO"

      - name: Add Studio shim to GITHUB_PATH
        run: |
          SHIM_DIR=~/.unsloth/studio/bin
          if [ ! -f "$SHIM_DIR/unsloth.exe" ]; then
            echo "::error::unsloth.exe shim not found at $SHIM_DIR"
            ls -la ~/.unsloth/studio/ || true
            exit 1
          fi
          cygpath -w "$SHIM_DIR" >> "$GITHUB_PATH"

      - name: Install OpenAI + Anthropic Python SDKs
        run: python -m pip install 'openai>=1.50' 'anthropic>=0.40'

      - name: Reset auth + boot Studio (API-only)
        run: |
          unsloth studio reset-password
          mkdir -p logs
          UNSLOTH_API_ONLY=1 unsloth studio -H 127.0.0.1 -p "$STUDIO_PORT" \
            > logs/studio.log 2>&1 &
          echo "STUDIO_PID=$!" >> "$GITHUB_ENV"

      - name: Wait for /api/health, log in, change password, load model
        run: |
          for i in $(seq 1 180); do
            if curl -fs "http://127.0.0.1:${STUDIO_PORT}/api/health" > /tmp/health.json; then
              jq -e '.status == "healthy"' /tmp/health.json && break
            fi
            sleep 1
          done
          jq -e '.status == "healthy"' /tmp/health.json
          OLD=$(cat ~/.unsloth/studio/auth/.bootstrap_password)
          NEW="CIJson-$(python -c 'import secrets; print(secrets.token_urlsafe(12))')"
          echo "::add-mask::$OLD"
          echo "::add-mask::$NEW"
          OLD_TOKEN=$(curl -fs -X POST "http://127.0.0.1:${STUDIO_PORT}/api/auth/login" \
            -H 'content-type: application/json' \
            -d "{\"username\":\"unsloth\",\"password\":\"$OLD\"}" | jq -r .access_token)
          curl -fs -X POST "http://127.0.0.1:${STUDIO_PORT}/api/auth/change-password" \
            -H "Authorization: Bearer $OLD_TOKEN" -H 'content-type: application/json' \
            -d "{\"current_password\":\"$OLD\",\"new_password\":\"$NEW\"}" > /dev/null
          TOKEN=$(curl -fs -X POST "http://127.0.0.1:${STUDIO_PORT}/api/auth/login" \
            -H 'content-type: application/json' \
            -d "{\"username\":\"unsloth\",\"password\":\"$NEW\"}" | jq -r .access_token)
          echo "API_KEY=$TOKEN" >> "$GITHUB_ENV"
          # Retry: same rationale as the OpenAI/Anthropic and Tool calling jobs.
          LOAD_OK=0
          for attempt in 1 2 3; do
            HTTP=$(curl -s -o /tmp/load.json -w '%{http_code}' \
              -X POST "http://127.0.0.1:${STUDIO_PORT}/api/inference/load" \
              -H "Authorization: Bearer $TOKEN" -H 'content-type: application/json' \
              --max-time 900 \
              -d "{\"model_path\":\"$GGUF_REPO\",\"gguf_variant\":\"$GGUF_VARIANT\",\"is_lora\":false,\"max_seq_length\":2048}")
            if [ "$HTTP" = "200" ]; then LOAD_OK=1; break; fi
            echo "::warning::/api/inference/load attempt $attempt returned $HTTP; response:"
            cat /tmp/load.json || true
            sleep 10
          done
          [ "$LOAD_OK" = "1" ] || { echo "::error::/api/inference/load failed 3 attempts"; exit 22; }
          jq '{status, display_name, is_vision}' /tmp/load.json

      - name: JSON schema decoding + image input
        env:
          BASE_URL: http://127.0.0.1:18899
        run: |
          python - <<'PY'
          import base64
          import json
          import os
          import urllib.request
          from openai import OpenAI
          from anthropic import Anthropic

          BASE = os.environ["BASE_URL"]
          KEY  = os.environ["API_KEY"]
          SEED = 3407
          TEMP = 0.2

          def post(path, body, *, timeout = 240):
              req = urllib.request.Request(
                  f"{BASE}{path}",
                  data    = json.dumps(body).encode(),
                  method  = "POST",
                  headers = {
                      "Authorization": f"Bearer {KEY}",
                      "Content-Type":  "application/json",
                  },
              )
              with urllib.request.urlopen(req, timeout = timeout) as resp:
                  return resp.status, json.loads(resp.read().decode())

          # ── 1. response_format = json_object (JSON mode) ─────────────
          status, data = post("/v1/chat/completions", {
              "model":         "default",
              "messages":      [
                  {"role": "system", "content": 'Reply with a single JSON object of the form {"city": "...", "country": "..."}. Output ONLY the JSON, nothing else.'},
                  {"role": "user",   "content": "What is the capital of France?"},
              ],
              "temperature":     TEMP,
              "max_tokens":      600,
              "seed":            SEED,
              "stream":          False,
              "enable_thinking": False,
              "response_format": {"type": "json_object"},
          }, timeout = 600)
          assert status == 200, f"json status {status}: {data}"
          assert (
              isinstance(data.get("choices"), list)
              and data["choices"]
              and "message" in data["choices"][0]
          ), f"json response envelope malformed: {data}"
          content = (data["choices"][0]["message"].get("content") or "").strip()
          print(f"[json] raw json_object content: {content!r}")
          if content.startswith("```"):
              content = content.split("```", 2)[1]
              if content.startswith("json"):
                  content = content[4:]
              content = content.strip("`\n ")
          if content:
              try:
                  parsed = json.loads(content)
                  if "paris" in str(parsed.get("city", "")).lower():
                      print(f"[json] PASS json_object -> {parsed}")
                  else:
                      print(f"[json] WARN json_object decoded but city!=Paris: {parsed}")
              except json.JSONDecodeError as exc:
                  print(f"[json] WARN json_object content not parseable ({exc}); content={content!r}")
          else:
              print("[json] WARN json_object produced empty content")

          status2, data2 = post("/v1/chat/completions", {
              "model":         "default",
              "messages":      [{"role": "user", "content": "What is the capital of France? Answer with one word."}],
              "temperature":     TEMP,
              "max_tokens":      400,
              "seed":            SEED,
              "stream":          False,
              "enable_thinking": False,
          }, timeout = 600)
          assert status2 == 200, f"plain status {status2}: {data2}"
          plain = (data2["choices"][0]["message"].get("content") or "").lower()
          print(f"[json] plain capital-of-france reply: {plain!r}")
          if "paris" in plain:
              print("[json] PASS plain inference path (paris mentioned)")
          else:
              print(
                  f"[json] WARN plain inference returned no 'paris' -- "
                  f"model output drift. HTTP path validated separately above."
              )

          # ── 2. OpenAI image_url (data URI base64) ───────────────────
          PNG_64X64_RED_B64 = (
              "iVBORw0KGgoAAAANSUhEUgAAAEAAAABACAIAAAAlC+aJAAAAYklEQVR4nO3PMQ0AIADAMEAI/k"
              "UhBhEcDcmqYJtn7/GzpQNeNaA1oDWgNaA1oDWgNaA1oDWgNaA1oDWgNaA1oDWgNaA1oDWgNaA"
              "1oDWgNaA1oDWgNaA1oDWgNaA1oDWgNaA1oDWgNaBdCJ0BmMJ25zMAAAAASUVORK5CYII="
          )
          data_uri = f"data:image/png;base64,{PNG_64X64_RED_B64}"

          # On Windows + the Qwen3-VL mmproj, llama.cpp's vision
          # path runs on CPU (no Metal involvement). The wrapper is
          # kept for resilience but the vision path is expected to
          # work on Windows; an exception here is a real regression.
          client = OpenAI(base_url = f"{BASE}/v1", api_key = KEY)
          try:
              openai_resp = client.chat.completions.create(
                  model       = "default",
                  temperature = TEMP,
                  max_tokens  = 80,
                  seed        = SEED,
                  messages    = [{
                      "role": "user",
                      "content": [
                          {"type": "image_url", "image_url": {"url": data_uri}},
                          {"type": "text",      "text": "What colour dominates this image? Reply in one word."},
                      ],
                  }],
              )
              openai_text = (openai_resp.choices[0].message.content or "").lower()
              print(f"[image/openai] reply: {openai_text!r}")
              if openai_text:
                  print("[image/openai] PASS image_url accepted, non-empty response")
              else:
                  print("[image/openai] WARN image_url accepted but empty content")
          except Exception as exc:
              print(
                  f"[image/openai] WARN image_url SDK call raised: {type(exc).__name__}: "
                  f"{exc}. Studio successfully forwarded the request; failure here is "
                  f"upstream llama.cpp vision behaviour."
              )

          # ── 3. Anthropic source/base64 image ────────────────────────
          anthropic = Anthropic(
              base_url        = BASE,
              api_key         = "unused",
              default_headers = {"Authorization": f"Bearer {KEY}"},
          )
          try:
              a_msg = anthropic.messages.create(
                  model       = "default",
                  max_tokens  = 80,
                  temperature = TEMP,
                  extra_body  = {"seed": SEED},
                  messages    = [{
                      "role": "user",
                      "content": [
                          {
                              "type":   "image",
                              "source": {
                                  "type":       "base64",
                                  "media_type": "image/png",
                                  "data":       PNG_64X64_RED_B64,
                              },
                          },
                          {"type": "text", "text": "Describe this image briefly."},
                      ],
                  }],
              )
              a_text = "".join(b.text for b in a_msg.content if getattr(b, "type", None) == "text")
              print(f"[image/anthropic] reply: {a_text!r}")
              if a_text:
                  print("[image/anthropic] PASS source/base64 accepted, non-empty response")
              else:
                  print("[image/anthropic] WARN source/base64 accepted but empty content")
          except Exception as exc:
              print(
                  f"[image/anthropic] WARN anthropic image SDK call raised: "
                  f"{type(exc).__name__}: {exc}. Likely upstream llama.cpp vision "
                  f"behaviour, NOT a Studio regression."
              )
          PY

      - name: Stop Studio
        if: always()
        # Run as cmd so we are not running through the Git Bash shell;
        # Git Bash on windows-latest has been observed to exit 143
        # (SIGTERM) from any inline kill/sleep block, masking a green
        # test run. The runner reclaims the Studio child process at
        # job end either way, so just emit a marker and exit 0.
        shell: cmd
        run: echo Stop Studio (no-op; runner reclaims STUDIO_PID=%STUDIO_PID% at job end)

      - 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
        # upload-artifact can pick it up. Crucial for diagnosing a
        # subprocess crash where Studio's traceback only shows the
        # symptom (httpx ReadError) but not the cause.
        run: |
          mkdir -p logs/llama-server
          cp -v ~/.unsloth/studio/logs/llama-server/*.log logs/llama-server/ 2>/dev/null || \
            echo "no llama-server logs to collect"

      - 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
          path: |
            logs/studio.log
            logs/install.log
            logs/llama-server/*.log
          retention-days: 7