[Windows Desktop] Parent task completes while delegated child continues GPT-5.6-Sol Max requests

Component

• Codex Desktop on Windows
• Codex CLI/app version recorded in the session: 0.153.4
• Multi-agent v2 workflow
• Observed on 2026-09-07 (Asia/Hong_Kong)

Summary

A parent task returned its final answer while a delegated child task remained active and continued making model requests for almost an hour. The child was still visible in the desktop UI as a separate subagent, while the parent conversation had already completed.

Observed behavior

The parent task was a GPT-6 Astra task titled 查询GPT Pro Token额度. It spawned a child task titled Csv analysis.

• Parent task returned a final answer at approximately 02:32 Asia/Hong_Kong.
• Child task started at approximately 02:33.
• The child continued sending requests until approximately 03:30.
• The usage export contains 447 token-usage records for the child session during this period.
• The child session initially inherited a GPT-6 Astra context, then ran with gpt-5.6-sol at ultra effort. The usage export labels these requests as gpt-5.6-sol / Max.
• Later child operations repeatedly failed with Script error: exec cell unused not found, but the child continued retrying instead of terminating.
• The child was linked to the parent by forked_from_id and parent_thread_id, so this was not an unrelated standalone task.

The parent task's final response was already visible in the UI while the child continued consuming quota. The desktop panel showed the child as an active subagent after the parent had completed.

Expected behavior

When a parent task reaches finalization, Codex should either:

1. wait for all descendants to complete,
2. cancel the descendant tree and verify that it is quiescent, or
3. explicitly detach the descendants and clearly warn the user that they will continue running.

A child that has lost its tool execution context should stop or enter a bounded failure state. It should not continue retrying indefinitely after the parent has completed.

Impact

This caused hundreds of unexpected gpt-5.6-sol / Max requests and unnecessary usage cost. The user selected GPT-6 in the parent conversation and did not intentionally start a long-running GPT-5.6-Sol child task.

Reproduction pattern

1. Start a Codex Desktop task using GPT-6 Astra.
2. Delegate CSV or document analysis to a child agent.
3. Allow the parent to produce its final answer while the child is still working.
4. Observe that the child remains active and continues making requests after the parent is complete.
5. If the child's tool context becomes invalid, observe repeated exec cell unused not found failures and continued retries.

Requested investigation

Please investigate parent finalization versus descendant lifecycle management, cancellation propagation, detached-child handling, model/effort inheritance for child tasks, and retry termination after invalid tool context.

Related reports: #42032, #36141, #38478.

Local logs and the usage export can be provided privately if needed. No account credentials or API keys are included here.

Start by reproducing the delegated CSV or document workflow on Windows with the parent allowed to finalize, then observe the child UI state, usage, and retry behavior. Trace parent finalization, descendant cancellation or detachment, child model and effort inheritance, and invalid-tool retry handling; done means descendants are awaited, cancelled and quiescent, or explicitly detached, while invalid contexts reach a bounded failure state without unexpected requests.