NOTE blocks stay non-speech in frontend and backend VTT parsing; audio_info
keeps declared sizes for float WAV and restores stream position; ffmpeg shim
uses absolute targets, byte-verified copies, PATH priority, no path logging;
transcription guidance is localized through failure codes in all 21 locales.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
- model_manager: queued callers wait in short slices and re-check abandonment
- backend.ts/backend-gate: scrub child output (tokens, home paths) at source and at agent hand-off
- failure/error_journal: recognise OSError.winerror == 8 and [WinError 8] without English text
- subprocess_backend: generating heartbeats re-arm recv only, not model-load grace
- tests: event-based sync for voxcpm2 heartbeat test
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Install staged audio with a restorable backup before committing its row,
restore on any failure, serialize lock/unlock/consent file swaps, name the
FileNotFoundError skip (CodeQL), and drop the real-timer race in the offline
delete test.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Require tokenizer, feature-extractor settings and all weight shards before a
cached Whisper snapshot is used, including the first main-ref hit; speak the
last line-start SAY: of unclosed reasoning; keep quoted closing tags without
disabling reasoning removal when the source also contains the tag.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The pre-launch import probe killed at 30 s and treated a timeout as a broken
runtime (false setup_required / 'environment incomplete' on slow disks and
scanner-contended hosts). It now gets 180 s and a pure timeout is inconclusive.
The default startup budget doubles on small hosts and a backend that is still
printing extends the wait (capped at 3 budgets).
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
The liveness probe was a sync route that imported torch and queried the CUDA
driver on every probe from the shared worker pool, so a busy but healthy
backend missed its 1.5 s deadline and was reported as unresponsive.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
snapshot_download(repo, local_files_only=True) only resolves the main ref, which
catalogue installs never write, so cloning reported an installed speech-to-text
model as missing. The model fallback now also tries every cached snapshot of the
configured PyTorch Whisper, large-v3-turbo and large-v3, skipping truncated
ones. Adds end-to-end resolution tests (short, long, supplied, no ASR).
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Classify private per-chunk ASR exceptions into fixed public replies, raise a
typed MediaToolUnavailableError from the validated decoder, and expose the
resolved ffmpeg under the bare name so dependencies that exec 'ffmpeg' find
imageio-ffmpeg's renamed binary on every platform.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
torchaudio 2.9 removed info() and routes load/save through TorchCodec. Add
audio_io.audio_info (header-declared WAV values, soundfile fallback), move the
dub cache checks and gpu_sandbox onto the audio_io helpers, and gate the whole
class with an AST test.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Review findings on #2435: the 180 s recv deadline killed valid CPU/slow-GPU
renders (no frames during generate) and the 30-step clamp silently reduced
the 32/64-step quality presets.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Same class as #2483: clear/delete/prune history and unlock unlinked audio
before the transaction committed; consent re-record deleted the previous
recording first; lock copied over the live locked take in place. Stage files
and swap/unlink after commit.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Review findings on #2467: a late-failing abandoned load left _load_abandoned set;
a queued caller's timeout could brand another caller's live load as stuck; the
warm-path test had no assertions.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
