Feature: opt-in local voice mode (/say TTS + /listen STT)

[NateFoxtrot]

What

Opt-in voice mode for the interactive TUI: /say <text> speaks the agent's reply through a local TTS server, /listen records up to 8s of speech, transcribes it, and drops the transcript into the editor as your next prompt.

Working implementation: https://github.com/NateFoxtrot/prime-agent/tree/feat/voice-mode (single commit, tsgo --noEmit + biome clean, based on current main).

Why

Hands-free field/homelab use: hear agent output while reading a monitor elsewhere, and reply by voice while both hands are on the task. Everything is local — no cloud TTS/STT, no new npm deps.

How it works

• /say <text> → POST {text, voice} to a local TTS HTTP endpoint (default http://127.0.0.1:9460/tts; any server returning a wav body works — I use a small Qwen3-TTS wrapper). Falls back to espeak-ng -w if the server is down. Playback via mpv.
• /listen → records via sox rec (16kHz mono wav), transcribes with faster-whisper (small, int8, CPU), inserts the transcript into the editor.

Config (env):

Var	Default	Purpose
PRIME_TTS_URL	http://127.0.0.1:9460/tts	TTS endpoint
PRIME_TTS_VOICE	serena	voice name passed through
PRIME_AUDIO_PLAYER	mpv	wav player
PRIME_VOICE	unset	0 disables both commands

All external tools fail silently — no TTS server, no sox, no whisper → the commands just no-op, never break the session.

Files: packages/coding-agent/src/core/voice.ts (new, ~95 lines, zero deps) + /say /listen handlers in interactive-mode.ts.

Happy to open a PR if there's interest — open to swapping faster-whisper for any local STT the project prefers (whisper.cpp, vosk) or gating the feature behind a settings flag instead of env.