Ended a turn whose task panicked, instead of hanging.

[moedash]

What changed?

A panic in a turn task no longer escapes the spawn site in codex-rs/core/src/tasks/mod.rs. The task future runs under catch_unwind, and a panic becomes a Fatal task error, so on_task_finished emits the turn's terminal event like it does for any other failure.

Why?

That spawn site owns turn lifecycle, as its own comment says: "Finish uniformly from the spawn site so all tasks share the same lifecycle." A panic was the one path that escaped it, unwinding past both on_task_finished and done.notify_waiters().

The result is worse than a crash. A client waiting on the turn waits for a terminal event that never arrives, and the process stays alive holding the thread's writer lock, so every later resume of that thread fails with already has an active writer. codex exec hangs rather than exiting, because its wait loop only ends on an event or a closed stream.

This is the general guard. It is not hypothetical: it happened, and the specific panic that triggered it is fixed separately.

How did you test it?

• Unit Tests
• Staging
• End to End Tests

describe_panic has a unit test for the payload shapes it can read.

End to end: put the old panic back, resumed a thread with a dangling tool call, and watched it log turn task panicked, emit {"type":"turn.failed",...}, and exit. The same command hung indefinitely before this change.