Coordinator ends an unfinished multi-agent run after answering a side question

What issue are you seeing?

Feedback ID: 01a07f96-2529-7e33-9553-771efa707b4d

During an ongoing multi-agent repository task, the coordinator answered a side question with a final response and stopped coordinating even though the original authorized work was unfinished. Delegated agents continued for a while, then held waiting for decisions/actions reserved to the coordinator. Progress resumed only after the user returned and asked what was happening, approximately 37 minutes later.

The side question concerned an action-based reading map for a runbook; it did not cancel or replace the implementation run. The coordinator was instructed to answer side questions while preserving and continuing the original task. It nevertheless treated the answer as the end of its turn.

At the later check, the agents were idle. Pending coordinator work included resolving a missing directory and interpreting a diagnostic after a successful closeout operation. The coordinator inspected the native agent state and resumed handling the outstanding work. This was a coordination stall, not evidence that those worker operations had all failed.

A subsequent request to draft a message explaining the incident to the Codex team produced another final response while the original run was still active. The user immediately called out the same premature-ending behavior. No second 37-minute interval is claimed.

What steps can reproduce the bug?

This is the observed workflow, not a deterministic minimal reproduction or measured failure rate:

1. Start a sustained task with a coordinator and delegated implementation/review agents. Reserve integration, shared-state changes and some recovery decisions to the coordinator.
2. Instruct the coordinator to continue until completion or an explicit user stop, and to handle side questions without abandoning the original task.
3. While delegated work is ongoing, ask a related explanatory side question, such as how a runbook's reading map should work.
4. Observe whether the coordinator answers with a final response and ends active coordination despite unfinished delegation or pending coordinator authority.
5. Let workers reach a completion or blocker that needs the coordinator. In this incident, the run remained unattended until a later user message prompted inspection and continuation.

The supplied feedback ID identifies the session for investigation. No full transcript or private project files are attached.

What is the expected behavior?

Answering a side question should preserve the original authorized task. The coordinator should continue handling delegated work and pending decisions unless the user explicitly stops/replaces the task or a genuine blocker requires user input.

Please consider a guard or clear warning before a coordinator ends while it has ongoing delegation or unresolved coordinator-only actions. A supported completion/blocker event that wakes an otherwise eligible coordinator would also help prevent this class of stall. This should preserve normal user stop/cancellation and authorization boundaries, rather than automatically execute unapproved actions.

Additional information

• Observed on September 8, 2026, in a Codex multi-agent session.
• Read-only local checks during report preparation returned codex-cli 0.153.4 and Linux 7.0.0-31-generic x86_64 x86_64 (uname -mprs). The installed CLI version is not asserted to be the app/UI build or the exact component revision responsible for the incident. App build and subscription metadata were not collected.
• An existing workflow-level advisory watchdog did not restart the coordinator. This is an observation about this run, not a claim that such a timer has a documented product guarantee to restart ended turns.
• The directly observed failure is the agent not following task-continuity instructions. Whether a Codex lifecycle/event-delivery defect also contributed is unverified; no engine or model root cause is asserted. Please use the feedback record to distinguish instruction following from host/runtime behavior.
• Related reports: #15723 covers completion wake behavior, #42032 covers descendants continuing after a final response, and #43535 covers closure with a pending user-choice question. This report supplies the side-question trigger, pending coordinator authority and feedback ID; those reports may be relevant to triage.

Submitted at the affected user's explicit request, following the repository's Other Bug report fields.

Start with feedback ID 01a07f96-2529-7e33-9553-771efa707b4d and compare related issues #15723, #42032, and #43535. Trace the coordinator lifecycle and completion or blocker event handling for the side-question workflow described here. Done means an explanatory side response does not end unfinished delegation, while explicit stops and authorization boundaries remain intact.