Voice

Local-only push-to-talk dictation for OpenCode v2. Press f9, talk (pauses welcome), press f9 again — the transcript lands in a draft you can edit; takes pile up until you send them as one prompt.

No cloud, no temp audio files, no auto-send, no permission answering.

Install

Requirements: OpenCode v2, a local STT server (see docs/STT.md), macOS Microphone permission for your terminal.

mkdir -p ~/.config/opencode/plugins
git clone https://github.com/vaibhavgupta0786/opencode-voice ~/.config/opencode/plugins/voice
# restart the OpenCode TUI, open a session, press f9

Or via npm (after v0.3.0 is published):

opencode plugin add @vaibhavgupta0786/opencode-voice

f9 may be hijacked by macOS Fn keys — Ctrl+P → dictate (or /dictate) works the same.

Usage

Input	Action
f9	dictate a take (press to start, press to stop) — appended to the draft
f9 again	stop → transcribe → draft editor opens with all takes
f10	open the draft editor anytime (empty or not — hand-compose works too)
f12	wipe the draft (confirmation dialog)
/dictate, /voice-send, /voice-clear	same via slash command / palette

On MacBooks press F-keys with Fn. Plain F10 is the hardware mic-mute key and plain F11 is Show Desktop (macOS reserves it — which is why wipe lives on F12).

Windows: not yet supported — the bundled recorders target macOS (CoreAudio) and Linux (ALSA/PulseAudio). PRs welcome; the recorder source is opencode-dictate's recorder.c (miniaudio supports WASAPI).

Editor semantics (one dialog, always the whole draft):

• Opens after every take, with the cursor at the end and the view scrolled to the bottom — long drafts scroll while you type.
• Ctrl+Enter sends everything as one prompt (Cmd+Enter also works where the terminal delivers it). Esc closes and keeps your edits — they are folded back into the draft automatically, whichever way the dialog was closed.
• A failed send keeps the draft and lets you retry; switching sessions with the editor open refuses to deliver to the wrong chat.
• The draft is per session: dictate in chat A, switch to B, and each keeps its own takes. Switching back shows A's draft intact, take count and all.
• Switching chats mid-take discards that take: the recorder stops within half a second, nothing is transcribed or saved, and a toast says so. The drafts themselves are never touched.

A mic indicator lives in the prompt footer: 🎤 idle, 🔴 recording, 🟡 transcribing, 🎤·N = N takes in the draft. With a local STT server the mic carries a health dot — 🟢🎤 server up, 🔴🎤 server down (checked every 30 s, refreshed on every take; cloud STT endpoints show no dot). Takes cap at 60 s (raise via the maxMs option). Debug log: /tmp/opencode/voice-plugin.log.

Keeping the STT server alive

Dictation needs the local whisper server on 127.0.0.1:8080. If it ever dies (machine restart, crash), recovery is:

1. One-time opt-in: run /voice-setup (palette: "Install voice-check recovery skill"). This copies the bundled skill into your global skills directory — nothing is installed silently, ever.
2. Whenever voice breaks: type @voice-check (or just say "voice is dead") and the agent verifies the server, restarts it if needed, and reports back.

Manual recovery is one health check plus a restart — see below.

How it works

f9 -> bundled mic recorder (16 kHz PCM, stdout, in memory)
   -> silence/VAD state machine (toggle: pauses never end the take)
   -> POST WAV to local whisper at 127.0.0.1:8080
   -> draft (takes accumulate) -> editor -> session.prompt

See docs/STT.md (server + model choice) and docs/TROUBLESHOOTING.md.

Development

npm ci
npm test        # 12 unit tests, no mic needed
npm run typecheck

Pure logic in logic.ts, host/IO in tui.ts. See CONTRIBUTING.md.

Recorder binaries

bin/ ships prebuilt mic recorders built from opencode-dictate's recorder/recorder.c (miniaudio, MIT) — see NOTICE. SHA256 pinned in bin/SHA256SUMS and verified in CI. To rebuild from source, fetch recorder.c + miniaudio.h upstream and compile, e.g. on macOS: cc -O2 -o recorder recorder.c -framework CoreAudio -framework AudioToolbox -framework CoreFoundation.

Why this instead of X?

	This	renjfk/opencode-voice	opencode-dictate
Input	F9 press / F9 press (toggle)	hold ctrl+r	hands-free, open mic
Review before send	always (editable draft)	optional	auto-sends in conversation mode
Draft accumulation	per session	single	n/a (immediate send)
LLM cleanup pass	none needed	required (cloud API)	required for voice commands
TTS voice-out	roadmap	yes (Piper)	no
Answers permissions by voice	never	via LLM	yes (opt-out)
External audio deps	none (bundled recorders)	sox, whisper-cli	none (bundled)

Choose this if you want push-to-talk with edit-before-send, zero surprises (no auto-send, no permission answering), and nothing to install beyond a local whisper server.

Roadmap

• 🎤 mic button in the prompt footer
• Voice-out: spoken digest of long replies (not verbatim readout) + local TTS

License

MIT — see LICENSE.