Explorer subagents keep forgetting the read-only requirement and try to modify code
Resolved 💬 3 comments Opened Mar 9, 2026 by winoros Closed Mar 16, 2026
💡 Likely answer: A maintainer (github-actions[bot], contributor)
responded on this thread — see the highlighted reply below.
What version of Codex CLI is running?
0.112
What subscription do you have?
Pro
Which model were you using?
gpt-5.4
What platform is your computer?
Linux 6.17.0-14-generic x86_64 x86_64
What terminal emulator and version are you using (if applicable)?
ssh
What issue are you seeing?
I give a complex task to it. And it spawned explorers for some investigation.
After some time, the explorer forgot that we had granted read-only permission to it and began to edit.
I didn't meet such a case before when using gpt5.3codex.
What steps can reproduce the bug?
A complex task. Let it spawn agents to explore. The explore task is also hard.
What is the expected behavior?
Only give insights and findings.
Additional information
_No response_
3 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
It seems that the context is polluted.
The sub agents thought it was the main agent when the issue happened.
@jif-oai
Hi, I have more context about this issue, not sure whether it's an independent problem.
Based on the 0.115, I noticed a case and then read some code. When the
fork_contextis true, the spawned agent had a chance to recognize it as the main agent and then do the work that main agent should do.I think what I suffered in the previous case is highly related to this.