macOS Desktop: shared threads can become uncontrollable, with no retry path after a network disconnect

Summary

In a shared Codex Desktop conversation, a failed/stalled model turn can leave participants without a usable recovery path:

• the model stops producing a response;
• the active turn cannot be paused/cancelled;
• the model picker cannot be used to switch models; and
• after a network disconnect, there is no effective user-initiated reconnect/retry action. Restarting the desktop app is the only way I have found to restore use of the conversation.

This is especially disruptive in shared threads because every participant is blocked by the same stuck state.

Environment

• Codex Desktop: 26.707.62119 (build 5211)
• Bundle identifier: com.openai.codex
• Platform: macOS 26.5.1, arm64

Steps to reproduce

1. Create or open a Codex Desktop conversation shared with multiple participants.
2. Start a model turn in the shared conversation.
3. While the turn is active, temporarily disconnect the local network, then restore it.
4. Return to the shared conversation after the connection is restored.

Actual behavior

• The model may no longer produce a reply in the shared conversation.
• The active turn cannot be paused or cancelled.
• Switching to another model is unavailable while the conversation is stuck.
• The client does not offer a reliable Retry connection / Reconnect path that reattaches to the existing turn and reconciles its state.
• Restarting Codex Desktop is required before the conversation becomes usable again.

Expected behavior

• Shared-conversation participants should retain a reliable way to cancel/pause a stuck active turn and select another model.
• When connectivity returns, Desktop should automatically reattach to the existing turn and reconcile whether it is running, completed, or interrupted.
• If automatic recovery fails, the UI should expose a visible Retry connection action that retries state reconciliation without requiring an app restart or rerunning work.
• The UI should show a clear terminal/recovery state rather than leaving all participants blocked.

Related issues

• #32466 — Desktop can miss live turn updates until the same thread is opened elsewhere.
• #32467 — Desktop may not surface a completed turn after a transient network disconnect.

The additional shared-thread regression here is the loss of user controls: participants cannot reply, pause/cancel the stuck turn, or switch models, and have no in-place retry path.