[tiensonqin]

Summary

• Keep the recording UI simple: remember the transcription toggle and selected language, show language selection before recording, and display live partial/final text while recording.
• Feed saved audio and on-device transcription from the same microphone stream. Replace iOS post-recording analysis with AVAudioEngine + SpeechAnalyzer, and Android file replay with AudioRecord + AAC encoding + ML Kit streaming recognition.
• Move Android capture into a microphone foreground service with a wake lock so the session can continue when the activity backgrounds or the screen locks. Preserve the existing iOS background-audio entitlement.
• Prepare/download the selected language model before capture; preserve recorded audio if transcription fails and keep the existing asset/transcript import flow.

Validation

• Full iOS simulator app build, signing and native-runtime export verification passed.
• Existing Swift Testing suite: 111 tests passed (custom app-host runner; no swift test).
• Full Android assembleDebug including the native core, unit tests (84), and lintDebug passed.
• OCaml core: 1,294 tests passed. Android JNI boundary and native-packaging checks passed.
• Capture localization generation/check passed for all 25 locales.

Real microphone transcription, model installation on supported hardware, backgrounding and actual device lock-screen behavior still need device validation. Live iOS transcription requires iOS 26 and a supported device; Android transcription depends on ML Kit's on-device availability. No UI/lock-screen E2E claims are made by the build and unit-test results.

Link to Devin session: https://app.devin.ai/sessions/ed65e199ce3a442a87f5f63e02234d69
Open in Devin Desktop: https://app.devin.ai/desktop/session/ed65e199ce3a442a87f5f63e02234d69?variant=devin
Requested by: @tiensonqin

---

🟢 Devin fixed all 2 findings · Follow Devin's work

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[devin-ai-integration]

Devin Review found 2 findings. Devin automatically fixed all 2 findings.

Fixed by Devin (2)

• Activity recreation deletes active recordings — An Activity destruction during capture used to cancel the service and delete its file. Destruction now detaches only the permission request. Recording operations and pending completion live in an Activity-independent session scope, and the restored runtime resumes saving only into the original graph/target. The save/import handoff cannot be interrupted by runtime disposal. Android full build, 84 existing unit tests and lint passed.
• Orphaned recording blocks future recordings — Reworked request ownership so the recreated runtime always observes the pending session, even when another graph (or no graph) is restored. Cancel now reaches the common completion/cleanup path rather than leaving recordingRequest orphaned. Graph identity is checked only when attaching a non-null asset; a mismatched graph reports an error with the retained audio path instead of importing into the wrong graph. Android full build, unit tests and lint passed.

View findings in Devin Review · Follow Devin's work