Show per-subagent usage, model, and reasoning settings in Codex App

What variant of Codex are you using?

Codex App

What feature would you like to see?

I would like Codex App to show and optionally control per-subagent usage, model, and reasoning settings.

When a task launches subagents, the Subagents panel currently shows the subagent names and status, but it does not clearly show:

• which model each subagent used;
• which reasoning effort each subagent used;
• how much usage each subagent consumed;
• whether the main agent consumed usage while waiting for subagents;
• the total usage consumed by the whole task.

I would like the Subagents panel or final task summary to show a breakdown like:

| Agent | Model | Reasoning effort | Usage consumed | Status |
|---|---|---|---|---|
| Main session | GPT-5.5 | High | ... | Done |
| Migration subagent | GPT-5.5 | High | ... | Done |
| Spec review subagent | GPT-5.4 | Medium | ... | Done |
| Quality review subagent | GPT-5.4 | Low/Medium | ... | Done |

I would also like users to be able to choose subagent settings separately from the main session.

For example:

• run the main session on GPT-5.5 High;
• run migration/editing subagents on GPT-5.5 High;
• run lightweight review subagents on GPT-5.4 Medium;
• run simple validation subagents on GPT-5.4 Low;
• run security or architecture review subagents on GPT-5.5 Extra High when needed.

Possible controls:

1. Manual control:

• choose the model and reasoning effort for each subagent or subagent group.

1. Automatic control with visibility:

• Codex chooses the model/effort based on task complexity, but shows what it chose.

1. Policy/preset control:

• “Use lower-cost/lower-effort models for review subagents.”
• “Use the same model as the main session for implementation subagents.”
• “Ask before launching high-effort subagents.”
• “Show an estimated usage impact before launching multiple subagents.”

This would make subagents more predictable, easier to trust, and safer to use when usage limits are tight.

Additional information

This feature is related to usage transparency and rate-limit debugging.

When Codex launches multiple subagents, a large part of the 5-hour usage window can be consumed quickly. Without a breakdown, it is hard to tell whether:

• one subagent consumed most of the usage;
• all subagents used the same high-effort model as the main session;
• the main agent consumed usage while waiting for subagents;
• the usage meter is working as intended;
• or there is a UI/accounting bug.

Not every subagent needs the same model or reasoning effort as the main session. A simple import validation or style review may not need the highest reasoning setting, while a security or architecture review might.

Showing per-subagent model, reasoning effort, and usage would help users make better trade-offs between cost, speed, and quality. It would also make bug reports about unexpected usage depletion much more actionable.