vox models command, CI/release workflows, neutral whisper-fixes reference

.github/workflows/ci.yml

@@ -0,0 +1,29 @@
+name: CI
+
+on:
+  pull_request:
+  push:
+    branches: [master]
+
+jobs:
+  test:
+    # mlx-whisper only installs on Apple Silicon
+    runs-on: macos-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: astral-sh/setup-uv@v5
+        with:
+          enable-cache: true
+
+      - name: Install dependencies
+        run: uv sync --frozen
+
+      - name: Lint
+        run: uv run ruff check
+
+      - name: Format check
+        run: uv run ruff format --check
+
+      - name: Test
+        run: uv run pytest -q

.github/workflows/release.yml

@@ -0,0 +1,74 @@
+name: Release
+
+on:
+  push:
+    branches: [master]
+    paths:
+      - pyproject.toml
+
+concurrency:
+  group: release
+  cancel-in-progress: false
+
+jobs:
+  test:
+    runs-on: macos-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: astral-sh/setup-uv@v5
+        with:
+          enable-cache: true
+
+      - name: Install dependencies
+        run: uv sync --frozen
+
+      - name: Lint
+        run: uv run ruff check
+
+      - name: Test
+        run: uv run pytest -q
+
+  publish:
+    needs: test
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+      id-token: write
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: astral-sh/setup-uv@v5
+
+      - name: Detect version change
+        id: version
+        run: |
+          LOCAL=$(python3 -c "import tomllib; print(tomllib.load(open('pyproject.toml','rb'))['project']['version'])")
+          PUBLISHED=$(curl -fsSL https://pypi.org/pypi/vox-transcribe/json \
+            | python3 -c "import json,sys; print(json.load(sys.stdin)['info']['version'])" 2>/dev/null || echo "0.0.0")
+          echo "local=$LOCAL" >> "$GITHUB_OUTPUT"
+          if [ "$LOCAL" != "$PUBLISHED" ]; then
+            echo "changed=true" >> "$GITHUB_OUTPUT"
+            echo "Version changed: $PUBLISHED -> $LOCAL"
+          else
+            echo "changed=false" >> "$GITHUB_OUTPUT"
+            echo "Version unchanged ($LOCAL), skipping publish"
+          fi
+
+      - name: Build distributions
+        if: steps.version.outputs.changed == 'true'
+        run: uv build
+
+      - name: Publish to PyPI
+        if: steps.version.outputs.changed == 'true'
+        uses: pypa/gh-action-pypi-publish@release/v1
+
+      - name: Tag and GitHub Release
+        if: steps.version.outputs.changed == 'true'
+        env:
+          GH_TOKEN: ${{ github.token }}
+          VERSION: ${{ steps.version.outputs.local }}
+        run: |
+          git tag "v$VERSION"
+          git push origin "v$VERSION"
+          gh release create "v$VERSION" --title "v$VERSION" --generate-notes

README.md

@@ -67,6 +67,10 @@ vox doctor
 # Download a model ahead of time
 vox init -m small
 
+# List available models per backend
+vox models
+vox models -b openai --format json
+
 # Get API schema (for AI agents)
 vox schema transcribe
 ```
@@ -80,6 +84,7 @@ vox schema transcribe
 | `vox channel <url>` | Batch transcribe a YouTube channel |
 | `vox init [-m MODEL]` | Download Whisper model |
 | `vox doctor` | Check dependency health |
+| `vox models [-b BACKEND]` | List transcription models per backend |
 | `vox schema [COMMAND]` | JSON schema for agent introspection |
 
 ## Models
@@ -139,7 +144,7 @@ Input (URL or file)
 
 ```bash
 uv sync
-uv run pytest              # Run tests (107 tests)
+uv run pytest              # Run tests (159 tests)
 uv run ruff check --fix    # Lint
 uv run ruff format         # Format
 uv run ty check            # Type check

skills/vox/SKILL.md

@@ -44,17 +44,18 @@ Apply in this order, highest volume first:
    isolated `Merci.`) emitted over silent stretches. Found in 73 of 227 files.
 2. **Collapse decoder repetition loops** — detect by run length (≥5 identical
    consecutive tokens, pinning at 221–223), never by vocabulary.
-3. **Fix domain vocabulary** — brand and instrument names are where Whisper fails,
-   not general anglicisms.
+3. **Fix proper nouns** — brand, product and person names are where Whisper fails,
+   not general anglicisms. Ask the user for a glossary rather than guessing.
 4. **Dedupe segment boundaries** — 1.29% of boundaries repeat a word, and this
    leaks into the concatenated `text` field.
 5. **Re-punctuate and recapitalize only where measured as degraded** — most files
    are fine; a minority have zero periods and no uppercase at all.
-6. **Normalize hyphenation and number formats** for consistency.
+6. **Restore diacritics and normalize hyphenation and number formats**.
 
-Do NOT strip filler words (Whisper already removes them: 296 `euh` per 1.8M words)
-and do NOT apply broad homophone rules (>90% false positives). Insert
-`[Speaker Name]:` markers only when speakers are clearly identifiable.
+Fix form, never content. Do NOT strip filler words (Whisper already removes them:
+296 `euh` per 1.8M words) and do NOT apply broad homophone rules (>90% false
+positives). Mark unintelligible passages `[inaudible]` rather than inventing, and
+insert `[Speaker Name]:` markers only when speakers are clearly identifiable.
 
 When editing `.srt`, preserve timestamps, sequence numbers and block count.
 
@@ -76,6 +77,11 @@ vox channel "https://www.youtube.com/@ChannelName" --years 2025 --summarizer non
 ```bash
 vox schema transcribe
 vox schema init
+vox schema models
+
+# Which models can I pass to --model?
+vox models --format json
+vox models -b openai --format json
 ```
 
 ## Commands
@@ -87,6 +93,7 @@ vox schema init
 | `vox channel <url>` | Batch transcribe a YouTube channel |
 | `vox init [-m MODEL] [-l LANG]` | Download Whisper model + check deps |
 | `vox doctor` | Check dependencies health |
+| `vox models [-b BACKEND]` | List models available per backend |
 | `vox schema [COMMAND]` | JSON schema for agent introspection |
 | `vox --version` | Print the installed version |