codex - 💡(How to fix) Fix Codex Outputs jumbled output, and that contagiously errors ChatGPT (web)

What version of Codex CLI is running?

codex-cli 0.129.0

What subscription do you have?

Plus

Which model were you using?

gpt-5.3-codex

What platform is your computer?

Linux 6.18.26-gentoo-dist x86_64 Intel(R) Core(TM) i7-8850H CPU

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

No response

What issue are you seeing?

I asked for a code review, but got scratchpad-like caveman-talk mixed with tool use json, and Chinese+Armenian+Gujarati. That output, if copy-pasted into ChatGPT causes the model to loop and error consistently. That is, it becomes contagious to ChatGPT with GPT 5.5 instant and thinking: With Instant: Error in message stream. (80% of the time, 20% it gives responds, but gives a strictly innaccurate response, see additional info) With Reasoning: Hangs on "Thinking" -- without showing any trace at all. That is, If I include the review output, and ask "What is the chinese text saying?" for example, or even "What languages are these"? .

What steps can reproduce the bug?

Uploaded thread: 019e0974-efcb-7481-8f7a-004ba776dcae

(I ran /review -> selected 2. Review uncommitted changes )

What is the expected behavior?

No response

Additional information

I did some experimenting (kinda fun) on what reproducibly causes it to error chatGPT: It cannot tell me there is both Gujarati and chinese at the same time, otherwise it errors. it will either error, or say "Gujarati+Armenian" OR "Chinese+Armenian". If it says "Gujarati+Armenian", and I ask "what does the chinese output say?" -- it says "there is no chinese text present in what you gave me" (but there is) , and also vice versa, which is very interesting.

It seems to be the mixture of chinese+armenian+gujarati, that is all. It cannot tell me that all three languages exist at the same time. It may classify the chinese as "weird unicode" -- but not say its chinese. If i copy paste it, and give examples of all three explicitly, error in stream instantly, and not a single token gets through.

In conclusion, the output contains Chinese+Armenian+Gujarati(characters, not actual Gujarati words/language), but it causes GPT 5.5 to hallucinate with 100% reproducibility, saying there is only one or two non-english languages present, exclusive-or causes an error in the stream/failure to complete reasoning.