Mobile Codex: allow repeated microphone dictation after a transcription completes

Summary

When using Codex from the mobile app and prompting via microphone dictation, the microphone can be used for the first prompt/transcription, but it cannot be started again afterward in the same mobile Codex flow. The user has to fall back to typing or leave/reopen the flow.

Observed behavior

1. Open Codex in the mobile app.
2. Use the microphone button to dictate a prompt.
3. Let the audio be transcribed and submit/use the prompt.
4. Try to use the microphone again for a follow-up prompt.

Result: the microphone cannot be used again. The UI appears to remain in a state where another recording session cannot be started.

Expected behavior

After a dictation is stopped, transcribed, and inserted/submitted, the composer should fully reset its voice-input state so the microphone button can be used again immediately for follow-up prompts.

A user should be able to complete a normal mobile Codex conversation hands-free or mostly hands-free:

• dictate an initial task
• send it
• dictate follow-up instructions
• send those
• repeat without restarting the app or switching to typed input

Why this matters

Mobile Codex usage is especially valuable when the user is away from a keyboard. If the microphone only works once per flow, voice prompting becomes unreliable for iterative coding sessions, approvals, corrections, and follow-up instructions.

Suggested enhancement

Reset or recreate the mobile voice input / recorder / transcription state after each completed transcription, including any stale recording, transcribing, permission, MediaRecorder, or native audio-session state. The microphone affordance should return to a ready state after completion, failure, cancellation, and send.

It would also help to add a lightweight regression test or QA path for repeated mobile dictation:

1. record + transcribe
2. send prompt
3. record + transcribe again
4. send follow-up

Notes

I searched existing openai/codex issues for related mobile microphone / voice-input reports before opening this. The closest issues I found were about different voice UX or desktop transcription failures, not repeated microphone use in the mobile Codex flow.