Android: voice mode is a control, and the only time the mic listens (#341)

[Khoality-dev]

Closes #341

The Android half of the Claude Design mockup "Kurisu - Voice Mode v1" (section 3), and the desktop's voice model (#253) with the same words: voice mode, the wake word, an interaction.

The model

• Voice mode is a control: the headset button in the chat's top bar, or the Chats screen's strip. This device remembers it (kurisu_voice_mode, the desktop's key). It is the only time the mic listens: CoreService records only in voice mode (ACTION_SET_VOICE_MODE, ACTION_RETRY_MIC), and with it off nothing said is used.
• It replaced "Always listen", which is gone from the ⋮ menu. AlwaysListenRemoval, a DataStore migration, deletes kurisu_asr_always_listen from the store rather than leaving it unread. Dictation is gone: speech without the wake word is ignored and never lands in the composer.
• The wake word starts an interaction, and only in voice mode.
• Each interaction is a new conversation. Its first message goes out with newConversation. CoreState.startNewInteraction clears the chat, the persona's cached conversation is forgotten as /clear does, and the transcript opens on a marker ("New conversation — You said “Kurisu” at 7:42 PM. The last one is in Chats.") that stays with the new conversation once the server creates it. The Chats screen opens the chat when an interaction starts, where it used to match raw transcripts itself.
• It ends 30 s after the last reply, finished streaming and speaking. Talking holds that window shut (setUserTalking, from the VAD's start of speech until what was said is transcribed), and it reopens full.

The UI (section 3)

• Top bar: the voice mode button is outlined when off and a filled blue circle with a ring when on, with an amber dot on a problem. The screen-reader state for the dot is "Needs attention". Under the model it says "· voice mode on", replacing "· mic …".
• Voice bar (VoiceBar), in place of the composer with the keyboard hidden. One state at a time from voiceBarPhase, in the desktop's order:
  • waiting (Say “Kurisu” to start);
  • listening, transcribing (placeholder lines), thinking, speaking (rings on the persona);
  • the 30-second window: the top line drains, with "Ends soon unless you say something". There is no countdown; the old "idle timeout in Ns" line and secondsUntil are gone.
• Problems, each with its fix:
  • no wake word set → Open Assistant;
  • microphone unavailable → Try again;
  • microphone access is off → Open settings, the app's Android settings page;
  • speech recognition didn't load → Retry.
• End voice mode is a 44 dp tonal button in the bottom-right in every state. A snackbar after ending says "Voice mode is off. The mic isn't listening."
• Colours: the voice colours are the design's, blue in dark mode too, where the theme's primary is grey.

Found on the way

• Turning voice mode on without RECORD_AUDIO would start a microphone foreground service without the permission, which Android 14+ refuses with a SecurityException. CoreService.requestVoiceMode checks first. Without access, the bar says "Microphone access is off" and the service is not touched. The chat retries when it comes back to the foreground (onResume), e.g. from Android settings.
• The voice bar's text had no colour of its own, so in dark mode outside a Scaffold it was near-black. It now sets LocalContentColor. Found in the emulator screenshots.

Tests

Written first (61e963c, d8fcb64, and the permission case), red before the implementation:

• VoiceInteractionManagerTest (16, rewritten): voice mode on and off; off, nothing is used; the wake word starts an interaction in a new conversation; no dictation; the window opens only after the reply, and talking holds it.
• VoiceBarPhaseTest (8), AlwaysListenRemovalTest (3).
• ChatViewModelVoiceModeTest (5): the control and its memory, the off notice, the new conversation's marker, and no mic access.
• ConversationsViewModelTest: an interaction opens the chat; a transcript alone does not.
• VoiceBarTest (instrumented, 14): the design's copy for every state, each problem's action, End voice mode in every state, the top-bar control and its dot, the marker. ChatInputTest loses its two voice cases, since the bar is no longer inside the composer.

Locally: testDevDebugUnitTest green; VoiceBarTest + ChatInputTest green on the thesys API 35 emulator. All twenty states were captured in light and dark and checked against the mockup.

Not covered

• The real microphone, VAD and wake word are still checked by hand, as CLAUDE.md says. No emulator test drives speech.
• docs/screens.md's screenshots predate this and still show the old strip; the text says so.

🤖 Generated with Claude Code

https://claude.ai/code/session_014L9pxjp1MTSJe5DZG9A9sv