›_ Codex Issues
Codex Issues / #38533

[macOS][Codex Desktop] Completed sub-agents remain under Running until opened

Open 💬 3 comments Opened Aug 14, 2026 by gut-puncture
💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What version of the Codex App are you using?

ChatGPT/Codex Desktop 26.810.41047 (build 6570)
Codex app-server 0.148.0-alpha.9

Platform

macOS 26.6.1 (25G76), arm64

What issue are you seeing?

Completed sub-agents remain listed under Running.

Opening one of those sub-agents shows that it has already stopped/completed. After returning to the sub-agent list, that same agent finally moves into Done.

This makes the running-agent count unreliable and makes it difficult to tell whether work is genuinely still active.

Steps to reproduce

  1. Start several sub-agents in one chat.
  2. Allow one or more sub-agents to complete.
  3. Open the sub-agent list.
  4. Observe that completed agents are still under Running.
  5. Open one of those agents. Its detail view shows it has stopped/completed.
  6. Return to the list. It now moves to Done.

Expected behavior

The list should move a sub-agent to Done as soon as its turn completes, without requiring the user to open that agent.

Diagnostic clue

The local desktop log records:

  • collab_hydration initially failing because a newly created rollout is temporarily empty.
  • Later turn/started and turn/completed events being received for an unknown conversation.

This suggests a possible hydration/state-registration race: after the initial empty-rollout failure, the conversation may not be registered, so later lifecycle events are discarded. Opening the child appears to force a fresh read and reconcile the displayed state.

Potential handling options could be retrying hydration after an empty rollout or buffering lifecycle events until the conversation is registered.

A short screen recording showing the deterministic UI sequence will be added in a follow-up comment.

Related reports

View original on GitHub ↗

3 Comments

github-actions[bot] contributor · 14 days ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

  • #37729
  • #37916
  • #38478
  • #37222

Powered by Codex Action

gut-puncture · 14 days ago

Screen recording

Unrelated task content was redacted. The recording starts at Active 70 / Done 46. Opening the stale Workflow history audit entry shows Worked for 30s, and returning to the list immediately changes the counts to Active 69 / Done 47. No new agent completion occurred during that inspection; opening the already-completed child reconciled its status.

https://github.com/user-attachments/assets/c5c33c57-6d5e-49f0-853f-785a7c5b3626

This appears closely related to #37729, with additional desktop-log evidence described in the issue body.

runningPHPer · 7 days ago

I can reproduce this on macOS in a long-running Codex Desktop task.

Observed behavior:

  • The Subagents panel lists many historical children as 处理中 / Working for 9–11+ hours.
  • I sent close requests for every child agent ID available to the runtime. The runtime reported the agents as already closed or no longer found.
  • The Desktop UI continued to render the same agents as active and kept increasing their elapsed time.
  • Opening the panel and attempting to close them again did not reconcile the display.

This appears to be a stale UI lifecycle state rather than active work, but it makes the panel unreliable and obscures the genuinely active child. I have screenshots showing the entries and elapsed times, available if useful.