[Bug] spawn_agent hangs indefinitely without returning control

[jimicze]

Summary

A spawn_agent call remained blocked for approximately 18,901 seconds (5 hours 15 minutes) instead of returning promptly with the spawned agent ID or an actionable error.

Uploaded thread: 019f78cf-a24c-7f30-80ac-c32b3ccfc887

Steps to Reproduce

1. Start a Codex session with existing completed or idle subagents.
2. Call spawn_agent to create another reviewer subagent.
3. Observe that the tool call does not return, time out, or report capacity/status information.
4. Manually interrupt the turn.

Actual Result

The call remained pending for over five hours and finally returned:

aborted by user after 18901.0s

No agent identifier, timeout, capacity error, or recovery guidance was provided.

Expected Result

spawn_agent should:

• Return the new agent identifier within seconds; or
• Fail quickly with a clear capacity or scheduling error; or
• Apply a bounded timeout and return control to the root agent.

A stalled spawn must not block the entire orchestration turn indefinitely.

Impact

• Stops all implementation and review progress.
• Prevents the root agent from sending user updates.
• Requires manual interruption.
• Creates uncertainty about whether the requested agent was partially created.
• Wastes substantial execution time.

Environment

• Date: 2026-07-22
• Collaboration mode: Default
• Repository: selectorChromePlugin
• Existing subagents: p1_t2_implement, p1_t2_spec_review
• Requested agent: p1_t2_quality_review

Suggested Fix

Add a bounded scheduling timeout, return explicit agent-capacity state, and make agent creation idempotent so retrying cannot create duplicate agents.

[github-actions]

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

• MultiAgentV2 spawn_agent can hang indefinitely while evicting a terminal resident thread #33777

Powered by Codex Action

[jimicze]

Scope clarification after reviewing the suggested issues:

This report is not a duplicate of the planning, persistent-goal, worktree, or performance reports filed from the same uploaded thread. Its sole failure boundary is a synchronous spawn_agent call that never returned, never produced an agent ID, and monopolized the parent turn for 18,901 seconds.

I also reviewed #33777. This is an independent reproduction with a different uploaded thread, exact 5h15 duration, CLI 0.144.6, and the concrete resident-agent state at failure. It may share the same underlying V2 pre-create eviction defect as #33777; I am not claiming a proven different root cause. The value of keeping this report linked is the directly uploaded reproduction and exact lifecycle evidence. Maintainers can decide whether to retain it as a separate reproduction or consolidate it into #33777.