← Docs hub

On-device TTS lifecycle recovery

Date: 2026-07-31
Branch: feature_a2a_task_orchestration_bridge
Device: WRBA1M10KRDWA0925E00614

1. Symptom

After moving the on-device agent to the A2A task orchestration bridge branch:

• Cloud responses were produced, but TTS and the launcher speaking-state transition did not complete.
• Pressing the PUI conversation cancel button terminated the MaumAi process.
• The runtime reported lateinit property audioManager has not been initialized.

These were not separate TTS, launcher, and cancel defects. They shared one initialization boundary failure.

2. Root cause

DeviceCommunicator.initializeDeviceControlProxy() retained context.applicationContext so the communicator and Cloud event outbox would not retain a service instance.

The DeviceAgent connection callback still attempted to recover the service with:

(context as? ForegroundService)?.run { ... }

Because the retained context was an application context, the cast always failed. DeviceAgent was connected, but the callback that starts ForegroundService.initializeApp() did not run. STT, TTS, and AudioPlayManager therefore never reached a valid ready state.

3. Corrected lifecycle contract

The fixed boundary is:

1. DeviceCommunicator keeps application context for safe process lifetime.
2. DeviceAgent connection and disconnection are delivered through explicit callbacks, not by casting context back to ForegroundService.
3. ForegroundService starts model initialization after the connected callback.
4. AudioPlayManager becomes ready only after TTS initialization succeeds.
5. Every TTS and effect-sound entrypoint checks audio readiness before playback.
6. PUI cancellation is a normal coroutine cancellation path, not a playback failure. It releases the voice session, restores IDLE, and preserves the MaumAi process.

DeviceRuntimeLifecyclePolicy isolates the Android-independent readiness and callback decisions so they can be unit-tested without a service instance.

4. Deployment safety

The final system APK was platform-signed before replacement.

No Android debug certificate was used.

5. Verification

Source verification

./gradlew :app:testDebugUnitTest --rerun-tasks passed. The new policy tests cover:

• DeviceAgent callbacks without relying on an Android context cast.
• An already-ready audio manager.
• Lazy audio-manager creation after TTS becomes ready.
• Safe rejection while TTS is unavailable.

Real-device verification

After system APK replacement and reboot:

• DeviceAgent connected to MaumAi.
• LLM initialization started from the connection callback.
• STT initialized.
• TTS initialized at 22,050 Hz.
• Speech recognition initialized.
• AudioTrack and AudioPlayManager initialized.
• The text E2E input 안녕 produced a Cloud DEF response and audible TTS.
• The launcher entered SpeakingActivity.
• PUI cancel during a long TTS emitted STOP_PLAYING.
• The runtime logged TTS playback cancelled and TTS worker cancelled.
• MaumAi PID remained stable.
• The launcher returned to HomeActivity.
• The voice state returned to IDLE.
• No FATAL EXCEPTION or uninitialized audioManager failure occurred.

6. Scope

This recovery changes the on-device DeviceAgent-to-ForegroundService initialization and TTS cancellation boundary. It does not change DeviceAgent TaskManager source, Cloud route policy, family classification, or Planner decision rules.