Private voice dogfood runbook (JOE-1113)¶

Desktop Local only. Feature flag features.voice default off.

Preconditions¶

1. Open Cowork Desktop on macOS (primary). Windows/Linux are best-effort (see packaging residual).
2. Set in open-cowork.config.json (or operator overlay):

{
  "features": {
    "voice": true
  }
}

1. Install tools (not pre-bundled in every CI package):
2. ffmpeg on PATH (or OPEN_COWORK_FFMPEG_PATH)
3. aurum CLI on PATH (or OPEN_COWORK_AURUM_BIN), model tiny-q5_1 cached
4. Optional: drop binaries under packaged resources/voice/ (see apps/desktop/resources/voice/README.md)
5. Grant microphone permission when prompted (host-owned, not renderer getUserMedia).
6. Workspace must be Desktop Local — Cloud Web / paired authorities stay not_supported.

Scenarios¶

A. Dictation PTT (JOE-1105 / JOE-1102)¶

1. Open Chat or Home composer.
2. Confirm mic control visible (data-testid="voice-ptt-button").
3. Click mic → status Listening… (live region).
4. Speak a short phrase; click again → Transcribing… then text injects into composer.
5. Cancel mid-listen (if exposed) must not send a prompt.
6. Partials may update the dictation segment while holding; final replaces them.

B. Hotkey (JOE-1110)¶

1. Focus Open Cowork (app-focused only — not OS-wide paste).
2. Default: Cmd+Shift+Space / Ctrl+Shift+Space.
3. Settings → Privacy → Private voice shows accelerator when flag on.
4. Avoid collision with command palette (Cmd/Ctrl+Shift+P).

C. Read-aloud (JOE-1103)¶

1. Complete an assistant message.
2. Read aloud on the bubble (when TTS ready).
3. Stop cancels immediately; starting PTT stops read-aloud.

D. Conversation mode (JOE-1107)¶

1. Toggle conversation (radio icon) on.
2. Mic starts a listen → final → agent prompt → stream → local TTS of reply.
3. Status cycles Listening / Transcribing / Thinking / Speaking.

E. Continuous VAD + barge-in (JOE-1104)¶

1. Conversation mode on → toggle continuous (activity icon).
2. Privacy red dot while mic armed.
3. After speak, mic re-arms; speak over TTS to barge-in (cancels speech + re-listens).
4. Continuous stays off by default; never silent always-on.

F. Assets offline (JOE-1109)¶

1. Settings → Privacy → Voice assets.
2. Refresh status / Ensure local model.
3. Missing model + no download env → clear fail-closed message (no silent network).

G. Security smoke (JOE-1111)¶

1. Enable voice, dictate once.
2. Confirm logs show textChars / chars only — no transcript body or PCM in main logs.
3. Cloud Web build: voice controls absent / not_supported.

Evidence to record¶

Paste evidence as a comment on JOE-1096.

Claim freeze (allowed vs forbidden)¶

Allowed (when flag on + tools present)¶

• “Desktop Local private voice (opt-in): on-device STT via Aurum, local OS TTS, PTT and optional conversation.”
• “Audio stays on-machine on the default path (local_only).”

Forbidden¶

• Cloud / multi-tenant / enterprise voice GA
• Browser / Cloud Web microphone for Studio
• Always-on continuous listen without explicit user enable
• “Ships fully offline in every package” without aurum + model present
• Aurum-as-TTS, cloud STT/TTS as default
• OS-wide Accessibility dictation into other apps (ZephyrFlow territory)