Changelog¶

There are no tagged releases. This list is built from the git history of main, newest first. Dates are commit dates.

2026-10-09¶

README notes that the listener uses the en-US recognizer and where the relevance judge looks for MLX (~/.local/mlx-server/bin/python). (#4)

README gets a plain one-paragraph summary, a "What I built" table, a Known limits section, and fixes for claims that disagreed with the code. (#3)

Credits name Kyutai for Pocket TTS. README adds other voice tools worth knowing (whisper.cpp, RealtimeSTT, mlx-audio).

New: dictation/bin/speak-on-stop, a Claude Code Stop hook that reads the last reply aloud.
New: dictation/filter/relevance.py (tune out / tune in, pet-talk regexes, local LLM judge), filter.py and enroll.py (voiceprint).
dispatch runs the listener through the voiceprint and relevance filters, and forces arch -arm64 for the voiceprint filter.
inject drops noise markers and fragments under 3 words, kills TTS before typing, scopes Escape and Return to the bound window, and sends a bare newline when that window is not focused.
Added the 2.0 research logs: BARGE_IN_ATTEMPTS.md, EXPERIMENT_PLAN.md, RESEARCH_2026_05_21.md.

README section on Ohm, a separate browser-based version (not in this repo). Removed a public URL from it the same day.

Listener detects silent recognizer wedges (audio arriving, no events for 20 s) and the preventive recycle drops from 600 s to 300 s.
inject presses Return after typing so the line submits.
GitHub Sponsors enabled.

First public release: Listen.swift, MicProbe.swift, dictation, dispatch, inject, narrative-claude.sh, CLAUDE.md, MIT license.
dispatch never pastes into a dead Terminal window (shutdown flag instead of signals).
narrative-claude.sh made executable.
README rewritten, plus the three-repo stack table.