›_ Codex Issues
Codex Issues / #9446

[FEATURE] Change 'stop and ask' behavior to 'ignore' for unexpected changes

Resolved 💬 2 comments Opened Jan 18, 2026 by PaulRBerg Closed Jan 18, 2026
💡 Likely answer: A maintainer (etraut-openai, contributor) responded on this thread — see the highlighted reply below.

What feature would you like to see?

The Developer Instructions currently include this guidance:

if I notice unexpected changes I didn't make, I must stop and ask how to proceed

I'd prefer Codex to simply ignore unexpected changes rather than stopping to ask. When I'm working on a task, I don't want to be interrupted with questions about unrelated modifications in the repo—I just want Codex to proceed with what I asked it to do.

Proposed change: Update the Developer Instructions to say something like:

if I notice unexpected changes I didn't make, I should ignore them and proceed with the requested task

This matches the workflow expectation that Codex should stay focused on the task at hand without derailing for unrelated repo state.

Additional information

  • Codex version: codex-cli 0.87.0
  • Platform: Darwin 25.2.0 arm64 arm

View original on GitHub ↗

2 Comments

etraut-openai contributor · 7 months ago

Thanks for the suggestion.

We have to be very careful about updating system prompts because there can be unexpected interactions between the model behavior (that is, the behavior that's trained in to the model). Changes to system prompts can have unintended side effects and introduce regressions. We're more likely to change system prompts for new models only.

Have you observed the behavior where the model actually stops when it detects an unexpected change? I haven't personally seen this, and I make unrelated changes quite often when the model is working. If you see such a behavior, it would be helpful if you'd use the /feedback command to upload the session details. The folks who train our models use that feedback.

PaulRBerg · 7 months ago

I understand where you're coming from.

My experience stems from >6 months with Claude Code, where Claude has absolutely never stopped because of unrelated changes. My expectation is that Codex CLI works the same way.

Have you observed the behavior where the model actually stops when it detects an unexpected change?

No, but I do see it often in its reasoning logs. And this, at a minimum, slows Codex down.