Decouple from upstream fork: rename to computer, Forgejo Packages releases, fix stale tests

troed commented

2026-07-31 22:00:55 +02:00

Owner

Summary

Decouple computer from the upstream huggingface/speech-to-speech fork and switch releases to our own Forgejo Package registry.

Rename importable module speech_to_speech -> computer across src/, tests/, scripts/, configs and docs (zero tracked references remain)
Forgejo Actions: move workflows to .forgejo/workflows/; publish.yml now builds with uv build, checks with twine check --strict, and uploads to the PyPI-compatible registry at https://git.sync.wtf/api/packages/troed/pypi using the COMPUTER_PACKAGE secret
pyproject.toml: package name computer, computer/computer-client console scripts, openwakeword promoted to a core dependency
AGENTS.md: rewritten for the devel/main branch workflow and Forgejo Packages release flow
Cleanup: removed GitHub-only artifacts (star-history workflow, dependabot), unused Dockerfiles, demo/, archive/
Test fixes: 30 previously failing tests fixed (stale _setup_faster mock signatures missing quant, missing interrupt/barge-in attrs in the speculative-turns VAD test factory)

Test Plan

uv run pytest — 741 passed, 16 warnings (2 pre-existing errors from standalone scripts/test_live*.py requiring runtime --host/--port)
uv build + uvx twine check --strict pass
ruff/mypy clean on changed files (pre-existing issues unchanged)

## Summary Decouple `computer` from the upstream `huggingface/speech-to-speech` fork and switch releases to our own Forgejo Package registry. - **Rename importable module** `speech_to_speech` -> `computer` across `src/`, `tests/`, `scripts/`, configs and docs (zero tracked references remain) - **Forgejo Actions**: move workflows to `.forgejo/workflows/`; `publish.yml` now builds with `uv build`, checks with `twine check --strict`, and uploads to the PyPI-compatible registry at `https://git.sync.wtf/api/packages/troed/pypi` using the `COMPUTER_PACKAGE` secret - **pyproject.toml**: package name `computer`, `computer`/`computer-client` console scripts, `openwakeword` promoted to a core dependency - **AGENTS.md**: rewritten for the devel/main branch workflow and Forgejo Packages release flow - **Cleanup**: removed GitHub-only artifacts (star-history workflow, dependabot), unused Dockerfiles, `demo/`, `archive/` - **Test fixes**: 30 previously failing tests fixed (stale `_setup_faster` mock signatures missing `quant`, missing interrupt/barge-in attrs in the speculative-turns VAD test factory) ## Test Plan - [x] `uv run pytest` — 741 passed, 16 warnings (2 pre-existing errors from standalone `scripts/test_live*.py` requiring runtime `--host/--port`) - [x] `uv build` + `uvx twine check --strict` pass - [x] ruff/mypy clean on changed files (pre-existing issues unchanged)

troed added 4 commits

2026-07-31 22:00:55 +02:00

Fix NativeLLMSTTHandler.setup() rejecting unexpected kwargs from faster-whisper args dict 401e6dd9c2

Fix native-llm: add user message to chat when transcript is empty but audio present b956287747

When --stt native-llm is used, transcript is empty but audio_bytes are present.
The _on_transcription_completed condition was 'if transcript' which skipped
adding a user message, causing:
  1. Audio injected into wrong message or not at all → stale responses
  2. Consecutive assistant messages at end of history → llama-server 400 error

Fix: use placeholder text '[audio]' when audio is present but transcript empty.
The _serialize() method then finds the user message and injects the input_audio
content part correctly.

Fix native-llm audio: wrap raw PCM in WAV container before base64 afe8652646

llama-server requires a valid WAV file for input_audio, not bare int16 PCM
samples. Added _pcm_to_wav() helper that writes a RIFF/WAV header around the
raw PCM bytes (16kHz mono int16).

Decouple from upstream fork, rename to computer, publish via Forgejo Packages

CI / ruff (pull_request) Has been cancelled

Details

CI / mypy (pull_request) Has been cancelled

Details

CI / pytest (pull_request) Has been cancelled

Details

CI / package (pull_request) Has been cancelled

Details

CI / Install smoke (linux) (pull_request) Has been cancelled

Details

CI / Install smoke (macos-arm64) (pull_request) Has been cancelled

Details

e32e134d1b

- Rename importable module speech_to_speech -> computer (all src, tests, scripts, docs)
- Move workflows to .forgejo/ (Forgejo Actions), publish PyPI-compatible package to
  git.sync.wtf package registry instead of PyPI; drop GitHub-only artifacts
- Rewrite AGENTS.md for devel/main branch workflow + Forgejo Packages releases
- Remove unused files: Dockerfiles, demo/, archive/, star-history assets
- Make openwakeword a core dependency (wake word is core functionality)
- Fix 30 stale tests (qwen3 _setup_faster quant param, speculative-turns interrupt attrs)

troed added 1 commit

2026-07-31 22:01:45 +02:00

chore: remove .DS_Store from repo, add to .gitignore

CI / ruff (pull_request) Has been cancelled

Details

CI / mypy (pull_request) Has been cancelled

Details

CI / pytest (pull_request) Has been cancelled

Details

CI / package (pull_request) Has been cancelled

Details

CI / Install smoke (linux) (pull_request) Has been cancelled