Intermittent session-wide exec failure after using gpt-5.6-luna sub-agents from gpt-5.6-sol on Windows

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

26.901.20858

What subscription do you have?

Plus

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

When using gpt-5.6-sol with medium reasoning effort as the parent agent in a project, sessions that use sub-agents explicitly configured with gpt-5.6-luna can sometimes enter a broken state where command execution fails for the entire session.

The failure is not 100% reproducible.

In at least one affected session, multiple Luna sub-agents successfully completed their tasks and returned results before the failure appeared. The issue therefore does not seem to occur simply at Luna sub-agent startup.

Once the issue occurs, command execution begins failing with:

helper_unknown_error: setup refresh had errors

At that point, the failure affects:

• the parent agent;

• Luna sub-agents;

• other existing sub-agents in the same session.

Starting a new session in the same project restores normal command execution.

Observed behavior

A typical affected session looks roughly like this:

gpt-5.6-sol (medium)
|
+-- spawn gpt-5.6-luna
| -> completes successfully
|
+-- spawn/use additional sub-agents
| -> may also complete successfully
|
+-- later command execution
-> helper_unknown_error: setup refresh had errors
-> parent exec fails
-> all sub-agent exec fails
-> current session remains broken

The exact action that causes the transition into the broken state is not yet clear.

Another session running Terra subagents are currently fine, uncertain on whether the bug won't happen with Terra, or I haven't go far enough.

What steps can reproduce the bug?

1. Open a project in the Codex App on Windows.

2. Start a new session using:

   • Parent model: gpt-5.6-sol

   • Reasoning effort: medium

3. Confirm that command execution works normally.

4. Ask the parent agent to spawn a sub-agent explicitly using:

   • Model: gpt-5.6-luna

5. Attempt to execute commands after the Luna sub-agent has been spawned.

6. Command execution begins failing with:

helper_unknown_error: setup refresh had errors

7. Attempt command execution from the parent agent and/or other existing sub-agents in the same session.

8. They also fail with the same setup-refresh error.

9. Start a new session in the same project.

10. Command execution works normally again.

What is the expected behavior?

No response

Additional information

/feedback report id: 01a056a8-f261-7713-a1a8-e3d671b89c2c

Start by reproducing on Windows with gpt-5.6-sol and explicitly spawned gpt-5.6-luna sub-agents, using the /feedback report 01a056a8-f261-7713-a1a8-e3d671b89c2c for context. Compare command execution before and after the setup-refresh failure and after starting a new session; done means the parent and all sub-agents continue executing commands without restarting the session.