[roprgm]

Summary

• align recording controls with confirmed browser audio capture
• treat empty and no-speech completion as successful while preserving existing input
• stop false recording states on capture end or microphone disconnection
• preserve valid transcript text when a later recognition error arrives
• document the predictable recording lifecycle and reorganize tests as a behavior specification

Root cause

The native adapter used a separate microphone stream for the waveform while the provider inferred recording state from the transcription object. A recognition service could therefore end while the waveform UI remained active. The provider also handled only rejected final-text promises, and the native adapter treated every recognition error as fatal even after valid text had been emitted.

Behavior

• checks speech-recognition support synchronously before requesting microphone access
• requests the microphone immediately and waits for audiostart before reporting recording
• exposes captureEnded so recording visuals stop during final-text processing
• handles microphone track termination and cleans up tracks and listeners
• returns to idle without an error when recognition ends with no text
• keeps permission, device, startup, and recognition failures retryable

Validation

• pnpm test — 41 tests passed
• pnpm check
• pnpm build

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
chat-voice-input	Ready	Preview, Comment	Jul 25, 2026 8:06pm