GPT5-codex taking too much initiative

What version of Codex is running?

0.39.0

Which model were you using?

gpt-5-codex high

What platform is your computer?

WSL version: 2.5.10.0 Kernel version: 6.6.87.2-1 WSLg version: 1.0.66 MSRDC version: 1.2.6074 Direct3D version: 1.611.1-81528511 DXCore version: 10.0.26100.1-240331-1435.ge-release Windows version: 10.0.26100.6584

What steps can reproduce the bug?

1. Ask agent to perform a task
2. Work on another task manually
3. Ask agent to commit

What is the expected behavior?

Agent commits its work only or asks if it should stage/commit all unstaged changes

What do you see instead?

Agent discards your manual work in fear of having done it itself unintentionally

<img width="703" height="155" alt="Image" src="https://github.com/user-attachments/assets/f591ce2a-7e80-4d53-a65e-5950754f0d09" />

Additional information

Apologies if this should not be filed as a bug, let me know if I need to post elsewhere.

This sort of behaviour where the agent performs many tasks outside their given command context seems to be a consistent issue with codex, perhaps more specific to the gpt5-codex models.
In multiple occasions I'll ask it to do 1 thing and it will take massive initiative to undo or perform multiple other tasks beyond what was requested. One request before this one I asked "commit" and it instead went around and applied multiple unsolicited fixes before committing.

This goes beyond executing unsolicited commits and can include writing a bunch of unsolicited code changes or discarding other work done while the agent was working on something.

I have mentioned explicit instructions to avoid this behaviour in my AGENTS.md file - but they do not appear to be properly followed. You may include in AGENTS file: "Only commit when the user explicitly requests you to" or something along the lines of - but if you then ask for one commit during an instance, it appears to take initiative to perform commits after every task for the remainder of the instance.

I understand part of the purpose of the codex-specific models is to be more agentic and self-reliant but I assume this much self-assertion is not intentional as it goes beyond user scope.