Codex doens't listen to anything it's told

Resolved 💬 2 comments Opened Mar 24, 2026 by uehlbran Closed Mar 24, 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?

codex-cli 0.116.0

What subscription do you have?

Max

Which model were you using?

GPT5.4

What platform is your computer?

Windows (WSL2)

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

WSL

What issue are you seeing?

Codex doesn't listen at all. I can't count how many times I will tell it to do something and it straight up will acknowledge my request and then do nothing.

What steps can reproduce the bug?

As an example, "Babysit pr #931 until it's merged". Codex: "Ok." and then doesn't do anything until prompted again. I have a skill, agents file and it knows all of this and yet, it still doesn't do a damn thing it's told and this applies to everything.

"I want you to do a full exhaustive audit of the core package. Look for bugs, code smells, architectural issues and for any findings open up a github issue for it". What does codex do? Takes a quick look and says done. When asked "Did you do a full audit?" Codex says, NO!

What is the expected behavior?

I expect it to follow instructions and display some level of "intelligence" instead of being an ignorant, gaslighting nuisance. It's honestly easier to get a toddler to do things they hate than it it is to get codex to do a damn thing. It's honestly like the 5.x models were trained exclusively on Reddit and 4chan. Helpful? Not even remotely. Intelligent? NOPE! Reading comprehension? Must of got lost.

Additional information

_No response_

View original on GitHub ↗

2 Comments

github-actions[bot] contributor · 5 months ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

  • #14414
  • #14543
  • #14228
  • #13950
  • #14341

Powered by Codex Action

etraut-openai contributor · 5 months ago

Thanks for the bug report. We'll share the feedback with the team responsible for training our codex models. If you see other model behavior that you'd like to report, please use the /feedback command. That will provide us with additional details including logs and conversation text.