Voice intake

Route: /voice

Scenario: You are away from your desk and something is broken. You say what is wrong, read back the transcript the console shows you, tap Confirm, and a draft pull request lands while you are still holding your phone.

Voice intake turns a spoken note into a report, and optionally into a dispatched draft-PR fix. Everything passes through one gate: you always see the verbatim transcript before anything is dispatched.

Turning it on

Voice intake is off until you enable it. Go to Settings → Voice intake and turn on the toggle. In the same tab you set:

Setting	What it does
Audio retention	Days to keep the clip. `0` deletes it the moment the transcript exists.
Languages	Hints passed to the transcriber. Leave them all off to let it detect.
Telegram bot token	Optional. Enables the Telegram inbox described below.
GitHub user token	Optional. Needed only for GitHub-hosted cloud agent tasks.

Transcription uses your OpenAI key from Settings → API keys. Without one, audio intake fails with a message telling you so; typed intake still works.

The four ways in

Path	Where it works
Tap to talk	The `/voice` page in any browser, best installed to the Home Screen.
Upload a clip	Anywhere. Accepts ogg, m4a, mp3, webm and wav up to 25 MB.
Type it instead	Anywhere. Skips transcription entirely.
Share a voice memo	Android Chrome only, once the console is installed. Share the memo to Mushi.

Install the console to your Home Screen to get the share target, offline shell and push notifications. On iOS, notifications only work after installing — that is an Apple restriction, and the page says so where it matters.

The confirmation gate

Nothing dispatches on your voice alone. After transcription you see the exact words that were heard, what Mushi intends to do, and a one-line summary. Then you choose Confirm or Cancel.

The confirm token is bound to the session, the transcript hash, the intended action and the expiry. It is single-use and valid for ten minutes. Change any of those and the token stops verifying, so a replayed or forwarded link cannot dispatch anything.

Some phrases are refused outright and never reach a model, in English and Japanese: merge, deploy, delete, drop, force-push, production, rollback, revert, and anything about secrets or tokens. Voice is a good way to describe a bug. It is not a way to touch production.

The transcript is untrusted input. It is wrapped in explicit delimiters before any model sees it, and personally identifying strings are scrubbed before the text is stored. Read what is on screen before you confirm — that is the whole point of the gate.

Getting the answer back

When the draft pull request opens, or a confirmation is needed, or a fix fails, Mushi tells you on whichever surface you used:

Push to this phone — tap Notify this device on the voice page. Uses Web Push, so it reaches you with the console closed.
Slack — the transcript card carries Confirm and Cancel buttons.
Telegram — same two buttons, as an inline keyboard.

Telegram is the practical Android inbox: a voice memo sent to a bot arrives as a file the server can fetch.

Create a bot with @BotFather and paste its token into Settings → Voice intake.
Press Generate bind code and send /start <code> to your bot within ten minutes.
Press Connect webhook.

The binding is per chat and per project, and you can remove any chat from the same panel.

Recent voice requests

The list below the recorder shows every session with its status, and links through to the report or the pull request once they exist. It updates live.

Status	Meaning
`created`	A report was filed; no fix was requested.
`awaiting_confirm`	Waiting for you. Expires ten minutes after the transcript.
`confirmed`	You approved it; the agent has been dispatched.
`cancelled`	You declined, or the window closed. Nothing was dispatched.
`failed`	Transcription, a cap, or a refusal stopped it. The reason is shown.

Limits

Voice intake is capped per project so a stuck client cannot spend your transcription budget: 30 intakes per minute, a daily ceiling on transcribed audio minutes, 25 MB per clip, and a maximum clip length. Transcribed minutes and their cost are recorded per session so you can see what voice actually costs you.

Self-hosting

Set VAPID_PUBLIC_KEY, VAPID_PRIVATE_KEY and VAPID_SUBJECT as edge secrets before push will work; the page reports push as unavailable until you do. See Edge functions for the deploy list.