Recover the mic when the audio device changes mid-session

[CansuKhon]

The bug

On macOS with Bluetooth headphones (AirPods Max here), the voice line transcribes
fine for the first several turns, then every recording after that fails:

||PaMacCore (AUHAL)|| Error on line 1322: err='-10851', msg=Audio Unit: Invalid Property Value
[ears] record/transcribe failed: PortAudioError('Error opening InputStream: Internal PortAudio error', -9986)
[ptt] (tap or empty — ignored)

Once it starts, it never recovers — every subsequent press fails the same way until
the process is restarted.

Why

PortAudio caches the device list when it initializes. A Bluetooth headset flips
between listen mode and mic mode every time the mic opens, which invalidates the
cached device. record_held opens a fresh InputStream per press, so every open
after the switch hits a device that no longer exists as PortAudio remembers it.

The failure mode is quiet in the worst way: the voice line stays up, the greeting
plays, the key registers, [ptt] recording still prints — and the agent answers
"my ears hit an error" on every single press. Everything looks healthy.

The fix

_input_stream() re-initializes PortAudio (which refreshes the device list) and
retries once on PortAudioError. Both capture paths use it — record_held for
push-to-talk and Ears.listen_once for the open mic.

The _terminate() call is guarded: it raises PortAudioError if PortAudio is
already down, which would otherwise turn the recovery path into a second crash.
(Found that one by testing the recovery, not by reading it.)

Testing

On macOS 15 (darwin 25.6.0), Python 3.12, sounddevice 0.5.6:

• Normal open: unchanged.
• Simulated device invalidation (sd._terminate() to reproduce the -9986 state),
  then _input_stream() — recovers, logs [ears] audio device changed — reopening the mic.
• Full record_held() through a deliberately broken device — recovers and returns
  transcribed text.
• Live push-to-talk after the patch on the machine that produced the original logs.

Found while setting up fullstack-agent. Happy to adjust the approach if you'd
rather handle it at a different layer.

🤖 Generated with Claude Code