Voice exit leaves hidden overlay owning the thread; primary UI remains thinking

Desktop package: OpenAI.Codex 26.903.9818.0, running as ChatGPT.exe on Windows. Built-in Check for Updates reports up to date.

User impact: after ending voice, the main UI can remain thinking indefinitely and submitted text/replies appear missing. Force-quit/relaunch recovers it. Repeated during the same day.

Observed sequence on September 10, 2026 (UTC):

• 20:05:10: avatarOverlay became thread owner; primary changed owner to follower.
• 21:07:15: thread/realtime/stop succeeded.
• No later ownership transition released the overlay in that desktop run.
• 21:09 onward: hidden avatarOverlay sent turn/steer requests that returned "no active turn to steer"; built-in fallback immediately started new turns successfully.
• Persisted session confirms completed final answers at 21:07:20, 21:09:15, 21:10:04, 21:10:26, and 21:11:12.
• Navigating away/back at 21:10 did not transfer ownership: primary re-entered as follower of the same hidden overlay.
• After relaunch, primary became owner at 21:11:34. Subsequent turns completed normally; no no-active-turn failures appeared in the new run.

Important: the no-active-turn fallback itself works. Backend computation is not hung in these examples. The evidence points to thread ownership/state synchronization between the hidden avatar overlay and main window after realtime stop. Exact causal code defect is not established.

Suggested investigation: ownership lifecycle and final turn-state propagation after realtime stop, including owner release/transfer and primary rehydration from persisted completed turns.

Acceptance test: start voice from an existing thread, end it, send text, confirm exactly one completed answer appears and persists, with no permanent thinking indicator or force-quit. Repeat across several voice-end cycles. Include the floating overlay path.

This report contains no raw conversation, credentials, private thread identifiers, usernames, or full logs.