Model issue, 98 issue, or personality issue? Off the rails strange behavior - 4 minutes of reasoning for "STOP!" prompt

Resolved 💬 2 comments Opened Feb 6, 2026 by MisterRound Closed Feb 6, 2026
💡 Likely answer: A maintainer (etraut-openai, contributor) responded on this thread — see the highlighted reply below.

What version of the Codex CLI are you using (From “About Codex” dialog)?

v0.98.0

What subscription do you have?

Pro

What issue are you seeing?

Just generally bizarre behavior on newest version and model, suuuuper long fever dream reasoning loops, prompt steering is ignored even though mentioned in the CoT, not sure what the deal is. It reasoned for almost FOUR MINUTES after telling it to STOP... repeatedly? 5.2 on previous versions were near instant. Mind you this was from an ESC interrupt message. This may be related to switching models in the middle of a long compacted 5.2 thread, which was working really well. I've read that switching the model in the middle of the chat can cause issues... but I don't see how this would be context related? I'm literally hitting escape, and telling it to stop. And it doesn't. Beyond that it just goes into long thought loops, like very very long reasoning loops that are much a do about nothing, not related in any way to recent user messages. Just tangents, feels super regressive, at least in this model + version combo. Also, it's slow. This is not a speed up. Is this a personality layer? I've noticed a huge increase in reasoning tokens even on fresh threads.

<img width="797" height="1725" alt="Image" src="https://github.com/user-attachments/assets/693fd336-e291-4bde-a46f-7c34ef40788d" />

<img width="642" height="1655" alt="Image" src="https://github.com/user-attachments/assets/90511ce9-ddd3-4eb6-a280-6bb4a6c488bd" />

<img width="344" height="848" alt="Image" src="https://github.com/user-attachments/assets/fb686918-3ef9-484d-b4f8-9435970a2747" />

<img width="627" height="1514" alt="Image" src="https://github.com/user-attachments/assets/70521dcc-e1df-46f3-bc66-b3bbcbbdeeed" />

What steps can reproduce the bug?

Use 98 and/or 5.3

What is the expected behavior?

Model/version should essentially "act the same" more or less with existing 5.2 workflows.

Additional information

thread ID 019c151f-d443-7812-8e8a-b7fcb914d60e

View original on GitHub ↗

2 Comments

MisterRound · 6 months ago

This is CLI, not sure how it got labeled APP

etraut-openai contributor · 6 months ago

Thanks for the bug report. We'll share the feedback with the team responsible for training our codex models. We're tracking this feedback and similar reports so we can improve future models. And thanks for using the /feedback command. That additional information is useful.