The configure action registered the provider but never called
set_active(), so the sidecar kept using the old/default provider.
Also updated the local provider default from localhost:8080 to
localhost:11434/v1 (Ollama). Added debug logging for configure.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Diarization: Audio.crop patch now pads short segments with zeros to
match the expected duration. pyannote batches embeddings with vstack
which requires uniform tensor sizes — the last segment of a file can
be shorter than the 10s window.
CI: Reordered sidecar workflow to check for python/ changes FIRST,
before bumping version or configuring git. All subsequent steps are
gated on has_changes. This prevents unnecessary version bumps and
build runs when only app code changes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
base_url was being set to 'http://localhost:11434/v1' by the frontend,
then LocalProvider appended another '/v1', resulting in '/v1/v1'.
Now the provider uses base_url directly (frontend already appends /v1).
Also fixed health check to hit Ollama root instead of /health.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The previous patch only replaced Audio.__call__ (segmentation), but
pyannote also calls Audio.crop during speaker embedding extraction.
crop loads a time segment of audio — patched to load full file via
soundfile then slice the tensor to the requested time range.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
soundfile was only a transitive dep of torchaudio but collect_all()
in PyInstaller can't bundle it if it's not installed. Adding it as
an explicit dependency ensures it's in the venv and bundled correctly.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
soundfile needs collect_all() to include libsndfile native library —
hiddenimports alone wasn't enough, causing 'No module named soundfile'
in the frozen sidecar. This is needed for the pyannote Audio patch
that bypasses torchaudio/torchcodec.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
torchaudio 2.10 unconditionally delegates load() to torchcodec, ignoring
the backend parameter. Since torchcodec is excluded from PyInstaller,
this broke our pyannote Audio monkey-patch.
Fix: replace torchaudio.load() with soundfile.read() + torch.from_numpy().
soundfile handles WAV natively (audio is pre-converted to WAV), has no
torchcodec dependency, and is already a transitive dependency.
Also added soundfile to PyInstaller hiddenimports.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
CSP: Add blob: to connect-src/img-src/media-src for wavesurfer.js audio
playback. Add http://tauri.localhost to default-src for devtools.
pyannote: sys.modules block didn't work — pyannote still uses AudioDecoder
unconditionally. New approach: monkey-patch Audio.__call__ in diarize.py
to use torchaudio.load() directly, bypassing the broken torchcodec path.
Patch runs once before pipeline loading.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
CSP: Add connect-src for ipc.localhost and asset.localhost so Tauri IPC
commands and local file loading (waveform, audio playback) work.
pyannote: Block torchcodec in sys.modules at startup so pyannote.audio
falls back to torchaudio for audio decoding. pyannote has a bug where
it uses AudioDecoder unconditionally even when torchcodec import fails.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
torchcodec is partially bundled but non-functional (missing FFmpeg DLLs),
causing pyannote.audio to try AudioDecoder which fails with NameError.
Excluding it forces pyannote to fall back to torchaudio for audio loading.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Sidecar now has its own version (1.0.0) and release lifecycle:
- Sidecar tags: sidecar-v1.0.0, sidecar-v1.0.1, etc.
- App tags: v0.2.x (unchanged)
- Sidecar workflow triggers only on python/** changes or manual dispatch
- App release no longer bumps python/pyproject.toml
Sidecar version tracked via sidecar-version.txt in app data dir:
- resolve_sidecar_path() reads version from file instead of CARGO_PKG_VERSION
- download_sidecar() fetches latest sidecar-v* release from Gitea API
- check_sidecar_update() compares local vs remote sidecar versions
- Version file written after successful download
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Default torch on PyPI is CPU-only on Windows. Must use PyTorch's own
package index (cu126) to get CUDA-enabled wheels. This also pins the
CUDA version on Linux for deterministic builds.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- transcribe: catch model load failures on CUDA and retry with CPU
- hardware detect: test CUDA runtime actually works (torch.zeros on cuda)
before recommending GPU, since CPU-only builds detect CUDA via driver
but lack cublas/cuDNN libraries
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Exclude ctranslate2.converters from PyInstaller bundle — these modules
import torch at module level causing circular import crashes, and are
only needed for model conversion (never used at runtime)
- Defer all heavy ML imports to first handler call instead of startup,
so the sidecar can send its ready message without loading torch/whisper
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- New release.yml: bumps patch version, commits with skip-ci marker, tags, creates Gitea release
- Build workflows now trigger on v* tags only (not branch push)
- Simplified upload steps: use tag directly, retry loop for release lookup
- Fix macOS: install jq if missing
- Sync python/pyproject.toml version to 0.2.0
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Tauri's externalBin only bundled the single sidecar executable, but
PyInstaller's onedir output requires companion DLLs and _internal/.
The binary was also renamed with a target triple suffix that
resolve_sidecar_path() didn't look for, causing it to fall back to
dev mode which used a compile-time CI path (CARGO_MANIFEST_DIR).
- Switch from externalBin to bundle.resources to include all sidecar files
- Pass Tauri resource_dir to sidecar manager for platform-aware path resolution
- Remove rename_binary() since externalBin target triple naming is no longer needed
- Remove broken production-to-dev fallback that could never work on user machines
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- uv pip --python works better with the venv directory path than the
python binary path (avoids "No virtual environment found" on Windows)
- Add .exe suffix to Windows python path for non-uv fallback
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- build_sidecar.py: pip_install() now includes 'install' in the command,
callers pass only package names (was doubling up as 'uv pip install install torch')
- CI: set shell: bash on uv steps so Windows doesn't try to use cmd.exe
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- CI: install uv via astral-sh/setup-uv, use uv to install Python
and run the build script (replaces setup-python which fails on
self-hosted macOS runners)
- build_sidecar.py: auto-detects uv and uses it for venv creation
and package installation (much faster), falls back to standard
venv + pip when uv is not available
- Remove .github/workflows duplicate
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Create /Users/runner directory on macOS before setup-python (permission fix)
- Use `python -m pip` everywhere instead of calling pip directly (Windows fix)
- Refactor build_sidecar.py to use pip_install() helper via python -m pip
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Move the blocking pipeline() call to a daemon thread and emit estimated
progress messages every 2 seconds from the main thread. The progress
estimate uses audio duration to calibrate the expected total time.
Also pass audio_duration_sec from PipelineService to DiarizeService.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
