Skip to content

v0.4.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 20 Apr 01:40
· 140 commits to main since this release
v0.4.0
352eca4

Added — Voice pipeline

  • Bidirectional voice pipeline — full duplex audio (capture → STT → agent_loop → TTS → playback) with WebRTC echo cancellation
  • MCP shim — bridges mod3 tools through cogos kernel as MCP tool surface
  • Bus-mediated dashboard chat — dashboard chat goes through cogos kernel buses instead of in-process loop, so external observers see the same conversation events

Added — Bargein provider registry

  • Pluggable BargeinProvider interface (bargein/providers/base.py) — was a hardcoded SuperWhisper file watcher; now extensible
  • SuperWhisperProvider (bargein/providers/superwhisper.py) — first provider, opt-in via MOD3_BARGEIN_PROVIDERS=superwhisper. Absorbs the SuperWhisper SQLite + filesystem detection logic that was previously drifting in a sibling repo
  • BargeinRegistry (bargein/__init__.py) — registry + shared handle_bargein_start() helper, used by both legacy file watcher and provider dispatch
  • BargeinRegistry.wait_for_event() — synchronous wait primitive used by await_voice_input() to block on in-process registry events
  • New "superwhisper" value in BargeinSource literal

Added — From earlier work, never released

  • Queue-aware speak() returns with enriched metadata (PR #4)
  • SpeechQueue for serial playback (thread-safe)
  • User-state detection (held status when user is recording)
  • /v1/stop HTTP endpoint for playback control
  • vad_check MCP tool

Changed

  • Default MOD3_BARGEIN_PROVIDERS= (empty) preserves current behavior — no providers auto-start
  • await_voice_input() now waits on both BargeinRegistry events AND legacy /tmp/mod3-barge-in.json for backward compat

Fixed

  • Speaking lock ownership — (pid, job_id)-aware with idempotent re-acquire. Two overlapping mod3 processes can no longer falsely interrupt each other.
  • Bus subscriber endpoint — KernelBusSubscriber honors COGOS_ENDPOINT at call time (previously hardcoded localhost:6931)
  • Session-scoped reply routing — kernel replies with session_id get routed to the matching browser channel; older payloads fall back to broadcast
  • Signal path unification — mcp_shim.py reads from /tmp/mod3-barge-in.json (was orphan ~/.mod3_bargein_signal.json that nobody wrote to)
  • Held job zombie drain bug
  • Pyright type errors in Gate abstract class
  • Various ruff lint issues

Reviewed by

  • claude-opus-4-7 (interactive)
  • gpt-5.4 (peer review, 3 passes)

Notes on versioning

This release jumps from v0.1.0 to v0.4.0. An earlier v0.2.0 tag exists from before the cogos-dev org transition (no GitHub release was created). v0.3.0 was bumped in pyproject.toml and added to the CHANGELOG but never tagged. v0.4.0 captures everything since the last released version (v0.1.0).

Assets 2
Loading