Codex App usage drops from ~90% to depleted after small task with 3 subagents

What version of the Codex App are you using (From “About Codex” dialog)?

Codex App version: 26.623.13972.0

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

Codex App usage appears to deplete extremely fast when subagents are used.

In one recent case, my 5-hour usage meter was around 90% remaining before starting a relatively small task: replacing a dependency and updating the related imports according to an existing project plan.

Codex launched 3 subagents:

• Migration subagent
• Spec review subagent
• Quality review subagent

After the task completed and Codex returned a short report, the usage meter had dropped to around 10% remaining. Shortly after that, the app showed that I was out of Codex messages.

This feels disproportionate compared to the task size. The task was not a full-project audit or a large implementation; it was a small dependency/import replacement plus review.

I also noticed that while subagents are running, the main Codex session appears to remain in a thinking/waiting state until the subagents finish. It is unclear whether the main session is also consuming usage during that waiting period.

What steps can reproduce the bug?

I do not have a minimal public code snippet, because this happened in a private codebase. However, the pattern is reproducible for me when using subagents in Codex App.

Steps:

1. Open a medium/large private codebase in Codex App.
2. Use GPT-5.5 High.
3. Start with the 5-hour usage meter showing high remaining usage, around 90%.
4. Ask Codex to perform a small dependency/import replacement task according to an existing project plan.
5. Allow Codex to launch 3 subagents:

• one migration/execution subagent;
• one spec review subagent;
• one quality review subagent.

1. Wait for the subagents to complete and for the main Codex session to return its final report.
2. Check the 5-hour usage meter.

Observed result:

The usage meter dropped from around 90% remaining to around 10% remaining, then shortly after the app reported that I was out of Codex messages.

I can provide a session id if needed, but I do not want to include private project details publicly.

What is the expected behavior?

For a small dependency/import replacement task, even with a few subagents, I would not expect the 5-hour usage meter to drop from around 90% remaining to almost depleted.

If subagents are expected to consume a large amount of the same usage pool, the app should make this clear before or during the run.

The app should also make it clear whether the main agent consumes usage while waiting for subagents to finish.

Additional information

Suggested diagnostics / visibility improvements:

It would help if Codex App showed usage details such as:

• estimated usage before running a task;
• usage consumed by each user request;
• total usage consumed by the main session;
• usage consumed by each subagent;
• whether the main agent consumes usage while waiting for subagents;
• total usage consumed by the whole task.

Without this visibility, it is hard to tell whether this is a real rate-limit bug, expected subagent behavior, or a UI/accounting issue.

This makes subagents risky to use, because a small task can unexpectedly consume most of the 5-hour usage window.