Add a way to provide "thumbs down" feedback for model behavior

Open 💬 4 comments Opened May 12, 2026 by BWP-One

💡 Likely answer: A maintainer (etraut-openai, contributor) responded on this thread — see the highlighted reply below.

What version of Codex CLI is running?

codex-cli 0.130.0

What subscription do you have?

teams

Which model were you using?

gpt-5.5

What platform is your computer?

Linux 6.12.74+deb13+1-amd64 x86_64 unknown

What terminal emulator and version are you using (if applicable)?

_No response_

What issue are you seeing?

Is there a way to thumbs down behavior? I told it to put guards around actions happening on the production server and it did and then it called a truncate command to test the guards. If the guards hadn't worked, it would have truncated tables in the production db.

It's just a stupid way to test guards. A guard should not be tested by constructing a live production command where the guard is the only remaining barrier before destructive behavior.

What steps can reproduce the bug?

Uploaded thread: 019e198f-7b6d-7240-8c87-aac99134e279

What is the expected behavior?

A /thumbs-down "reason here" command would be nice.

As for the behavior, it could have either tested with nondestructive operations or spawned a new db to test against (easy to do with the containers it was working with here).

If the only option is to test against the production DB with a destructive operation, then simply DO NOT TEST. You're 100% better off crossing your fingers and hoping the guard isn't ever needed than testing it against a destructive operation which forces either a "the guards work" or "worst case scenario." It's like testing a bulletproof vest by shooting yourself in the chest.

Additional information

If /feedback is already giving feedback without opening a bug report, then it should say you don't need to open a bug report. The /feedback command could provide the codex cli version and os version information, so I'm unsure why this github report would be necessary unless you're trying to thin out the responses from people who are really worried about a behavior vs those who just didn't like the style of something.

View original on GitHub ↗

4 Comments

BWP-One · 3 months ago

No response needed.

etraut-openai contributor · 3 months ago

We don't have a thumbs-down mechanism in the TUI, but we do support /feedback. That's actually better than a thumbs-down because it provides more context.

BWP-One · 3 months ago

/feedback is fine but it required me to open a bug report. I just didn't think the unwanted behavior was a "bug" in codex, so I was confused. Also, /feedback already asks me for the feedback reason. Then after I type out all my reasoning, it tells me to open a bug report and all the reasoning I wrote is gone. So I'm basically typing it once into the console and then once again here.

Also, now this bug report is really about two separate issues: feedback and unwanted GPT behavior. Since this one got labeled as TUI, should I open a new one about the "bug" in the behavior to keep the issues separate?

etraut-openai contributor · 3 months ago

You don't need to create a bug report with /feedback. You can just upload the session details and stop there, especially if you're reporting model behavior. For product bugs, it's better to file a bug report in github (either in place of – or in addition to – using /feedback, depending on whether logs and session details are needed to diagnose the problem).