Voice intake

Dictate a bug or a small change request on your phone. Mushi transcribes it, files a report, and — only for change requests, and only after you confirm the verbatim transcript — dispatches the fix and replies with the draft PR link in the same chat.

phone voice ──▶ transcript ──▶ sanitise + PII scrub ──▶ refuse privileged verbs
            ──▶ intent (create_report | open_draft_pr | unknown)
            ──▶ report (source = voice)
            ──▶ [open_draft_pr only] verbatim transcript + confirm token ──▶ you confirm
            ──▶ fix dispatched ──▶ draft PR ──▶ reply in Slack / Telegram / push

Four inboxes feed the same pipeline. Pick the one that matches your phone.

Enable it

Voice intake is off by default because speech is personal data.

iOS Shortcut

Two Shortcuts: one to speak, one to confirm.

The raw request, for reference:

curl -X POST "https://<ref>.supabase.co/functions/v1/api/v1/intake/voice" \
  -H "X-Mushi-Api-Key: mushi_xxx" \
  -H "Idempotency-Key: 4d2a…" \
  -H "Content-Type: application/json" \
  -d '{"source":"ios_shortcut","transcript":"the login button does nothing on Safari"}'

{
  "ok": true,
  "session": {
    "id": "…",
    "status": "created",
    "transcript": "the login button does nothing on Safari",
    "action": "create_report",
    "summary": "Login button unresponsive on Safari",
    "report_id": "…",
    "message": "Filed report: Login button unresponsive on Safari\n\nFrom: \"the login button does nothing on Safari\""
  }
}

A change request comes back as status: "awaiting_confirm" with a confirm_token; POST /v1/intake/voice/:id/confirm with { "token": "…" } dispatches, …/cancel discards. Tokens are single-use and expire after 10 minutes.

Telegram (the Android inbox)

Plain text messages work the same way. Voice notes are capped at 120 seconds and 20 MB.

Slack

Two ways in, both need the Mushi Slack app installed with files:read, commands and chat:write:

• Slash command: /mushi voice fix the footer alignment on checkout — the reply is ephemeral and the confirmation uses the same buttons as fix dispatch.
• Audio clip: record a Slack audio clip (or share an audio file) in a channel the app is in. Mushi downloads the clip with the bot token, transcribes it, and replies in a thread. Slack’s own transcript is ignored.

Installed console (PWA)

Install the console to your Home Screen (Chrome on Android: Install app; Safari on iOS: Add to Home Screen). The Voice page has a tap-to-talk button (max 120 s) and, on Android, the console appears in the share sheet for audio files — Google Recorder’s .m4a works. The recording is uploaded through a short-lived signed URL into the private voice-intake bucket and deleted after transcription. Confirmation is a button on the page; the PR link arrives as a web push notification when you granted permission.

What voice intake refuses

Voice text is untrusted and lossy — “delete” and “deploy” sound alike, 削除 and 消去 do too. A transcript that mentions any of these is refused outright, before any model sees it, and the verbatim transcript is echoed back so you can see what was heard:

Everything else that gets through is still constrained:

• Agents open draft PRs on a new branch only. Merging stays a human action in the console.
• The intent is extracted into a strict schema (create_report, open_draft_pr, unknown); the transcript is passed as delimited untrusted data and is never free-texted into an agent prompt.
• Invisible Unicode (zero-width characters, tag characters, bidi controls) and HTML comments are stripped before classification.
• A change request needs an explicit confirmation of the verbatim transcript: a single-use HMAC token bound to the session, transcript hash, action and expiry.
• Reporter (end-user) tokens cannot use the voice routes at all.

Privacy and retention

• Audio is deleted from the bucket as soon as the transcript is stored unless you set a retention window (0–365 days). The daily retention sweep enforces the window.
• Transcripts pass through the PII scrubber (emails, phone numbers, IPs, secret-shaped tokens) before they are written to the report.
• The report keeps the sanitised transcript, the detected language and the sha256 of the audio bytes as an audit trail; it never keeps the audio itself unless retention is on.
• Speech-to-text is a vendor hop to OpenAI on the project’s key. Mention it in your privacy policy if your users can reach a voice inbox.
• Per-project caps: 30 voice requests per minute, 120 audio minutes per day, plus the normal ingest burst limit.

API reference

403 VOICE_INTAKE_DISABLED means the project has not enabled the feature. See the error catalog for the rest.