turn/completed can precede command completion and wedge later turns between active-writer and no-session errors

[pixexid]

Summary

A Codex-backed thread emitted turn/completed while a command-execution item from that same turn was still running. The command later emitted output and item/completed after the terminal turn event. Every subsequent new-turn request then failed in one of two contradictory states:

• thread <provider-thread> already has an active writer
• No active codex session for thread <bb-thread>

Stopping the broken runtime and retrying the exact input did not recover the thread. Provider retry refused because the input had not been accepted, and edit-message could not apply because the rejected request was agent-origin rather than an editable user turn.

Native ordering evidence

For one exact turn/provider-thread:

1. item/started for a long-running commandExecution.
2. turn/completed with status=completed.
3. item/commandExecution/outputDelta for the still-running command.
4. item/completed for that command, reporting success and a multi-minute duration.
5. A later client/turn/requested was rejected as already has an active writer.
6. A recovery new-turn request was rejected as No active codex session.
7. A further exact redelivery again rejected as already has an active writer.

This is a native event/state contradiction: the turn was publicly terminal before all of its writer-owned items terminalized, leaving BB and the provider bridge disagreeing about whether a session/writer existed.

Expected behavior

• Do not emit turn/completed until every command/tool/item belonging to that turn is terminal and no turn writer remains.
• If a provider sends premature completion, BB should delay/fail the terminal event or reconcile outstanding items deterministically.
• A terminal thread must not simultaneously report both active-writer and no-active-session states.
• Stop/restart should clear any orphaned writer ownership without deleting event history or replaying rejected input.
• Rejected agent-origin requests need a supported exact-redelivery recovery path; edit-message is not applicable.

Acceptance

1. A fixture delays command completion beyond provider turn completion; BB does not publish terminal completion first.
2. After true terminalization, one new-turn request is accepted exactly once.
3. Stop/restart clears stale writer ownership while preserving interruption/history.
4. Tests cover the contradictory already has an active writer / No active codex session sequence and prove it cannot persist.

No repository or canonical-state mutation is needed to reproduce; this report is based on bounded native event reads.