[Khoality-dev]

Closes #253

The flow as diagrams: https://claude.ai/artifact/TJbMdQFqz9NdJPuYcC3u3K

The model (the owner's, agreed in session)

Three words, used the same way in the UI, the code and the docs: voice mode, the wake word, an interaction. Nothing is a "call" or an "interactive mode" any more.

• Voice mode is the only time the mic listens.
  • It is turned on and off from a headset button in the chat header, shown where the host can capture the microphone (a new microphone capability: always in Electron, a secure context in a browser).
  • It replaces the "Always listen" setting, and this device remembers it.
  • While it is on, the voice bar (VoiceModeBar) replaces the message box.
• The wake word starts an interaction, and only in voice mode. Speech without it is ignored while voice mode waits. With voice mode off nothing is heard: no dictation, and the wake word means nothing.
• An interaction sends everything said, with no wake word needed again.
  • Each interaction is a new conversation.
  • It ends 30 s after the assistant's last reply has finished (streaming and speaking) with nothing said since. Voice mode then waits for the wake word again.
  • The assistant's app_end_interaction tool ends it the same way.
• Push-to-talk (Ctrl+Space) is removed.

This changes #337's behaviour, merged earlier today: outside voice mode the wake word no longer starts anything.

Tests

New tests, each run first against the unfixed code and failed:

• state/voiceMode.test.ts (4): voice mode on waits without an interaction; off ends it and stops the mic; an interaction ending leaves voice mode on; the setting is remembered and restored.
• ui/chat/voiceMode.test.tsx (9), replacing wakeWord.test.tsx:
  • the wake word starts an interaction in a new conversation;
  • later messages go to that conversation, and the next interaction is another new conversation;
  • speech without the wake word is ignored;
  • the timeout fires 30 s after the reply finished, not while it streams or speaks, and starts over when you speak;
  • nothing is used with voice mode off.
• Desktop Playwright voiceMode.spec.ts (3), using Chromium's fake mic:
  • the header starts voice mode and the voice bar waits for the wake word;
  • both buttons end it, and it survives a reload;
  • Settings no longer has "Always listen".

After the fix:

• The new voice-mode Playwright specs, with settings.spec.ts: 13/13.
• settings.spec.ts's voice-picker locator is now getByRole('combobox', { name: 'Voice' }), because getByLabel('Voice') also matched "Start voice mode".
• Workspace npm test: all suites pass. npm run typecheck: clean.

Not covered

• The e2e build cannot load the VAD model (ort-wasm-simd-threaded.mjs over file://). The specs therefore check the controls and state, not real speech.
• Android is not changed here. It still enters its voice mode by the wake word and has no button. That is a follow-up issue.

🤖 Generated with Claude Code

https://claude.ai/code/session_014L9pxjp1MTSJe5DZG9A9sv