Infinite response loop in Codex VS Code extension (chat & agent, all models)

Resolved 💬 3 comments Opened Dec 14, 2025 by MiSte-Git Closed Dec 14, 2025

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

What version of Codex is running?

0.4.51

What subscription do you have?

ChatGPT Plus

Which model were you using?

gpt-5 to gpt-5.2 - All available models affected

What platform is your computer?

Debian 12 Testing - Linux 6.17.9+deb14-amd64 x86_64 unknown - VS Code: Linux Codex VS Code extension (current stable)

What issue are you seeing?

Codex enters an infinite response loop even for trivial prompts.
The same output is generated repeatedly ("OK", "Thinking", "OK", …) until the user manually aborts.
This happens in both Chat and Agent modes and is independent of the selected model.

What steps can reproduce the bug?

Open VS Code
Open an empty workspace
Start Codex (Chat or Agent)
Select any model
Prompt: "Answer exactly once. No variations. No suggestions. Just the word: OK."

What is the expected behavior?

Single response:
OK

Additional information

Clearing ~/.config/Code/User/globalStorage and workspaceStorage does not help.
Appears to be a server-side execution/termination bug.

View original on GitHub ↗

3 Comments

etraut-openai contributor · 8 months ago

Are you seeing this repeatedly, or did this happen only once?

Please resume the session where it occurred and use the /feedback slash command to upload it so we can investigate.

MiSte-Git · 8 months ago

I do see it all the time since the codex update to gpt-5.2. Easy to reproduce for me.

It does't matter what prompt I put in.

Feedback sent -> 019b1c66-3abd-72b3-834a-c18a7b2604d5

Thank you.

etraut-openai contributor · 8 months ago

Thanks for uploading the session. I don't see any problem with the codex harness here — it looks like it's handling the model responses correctly. So this appears to be a model issue with gpt-5.2 itself. I'll forward the feedback on to the team responsible for training our models.

In the meantime, you can try modifying your prompt, switch to a different model, or try changing the reasoning effort.