- /dictate + ctrl+shift+d toggle one shared session: press to start, press again
to stop, flush and paste into the prompt (a terminal gives no key-release event
for hold-to-talk without the Kitty keyboard protocol)
- client.recordMic(): spawns ffmpeg AVFoundation capture and appends s16le frames
as they arrive (no pacer — the mic is already 1x paced); stop() = commit + the
same quiescence drain transcribe() uses, since is_last_result never flips true
- enumerates mics by slicing the "AVFoundation audio devices:" block off stderr
(ffmpeg prints the table there, then fails opening the dummy input)
- --device takes an index or name, --lang / ASR_DICTATE_LANGUAGE, --submit sends
instead of pasting; asr_health now reports the mic inputs it enumerated
- capture death surfaces via the emptyGraceMs path with the TCC hint (System
Settings → Privacy & Security → Microphone) when ffmpeg exits non-zero at 0 bytes
- env.sh sources the gitignored .env so a shell gets ASR_NIM_URL / ASR_API_KEY
- asr_transcribe / asr_health tools + the /asr command, with live partials,
per-word timings, transcript export and TUI rendering
- zero-dependency realtime client (https + Node 22 global WebSocket):
mint transcription session -> subprotocol handshake -> event loop -> quiescence
- decodes anything ffmpeg supports (and native 16k mono/stereo pcm16/pcm_f32le wav)
to the 16 kHz mono PCM16 the NIM mandates
- credentials via ASR_API_KEY / ASR_NIM_URL or ~/.pi/agent/asr.json; never in source
- README documents the six places the published API reference diverges from the
deployed protocol (commit vs done, cumulative deltas, double .completed,
HTTP 403 at upgrade, ?token= not rejected, model reset on session.update)
Iulian
iulianp