Have to enter instructions twice (in about approx. 20 % of the prompts)

[UweKeim]

Being on the Pro plan using GPT-5.4 xhigh and 1M token size, I discovered a behavior that did not occur previously:

1. I do some prompts that Codex correctly processes.
2. Later I add some refinement prompts.
3. Every now and then Codex "forgets" the entered prompt and behaves as if it executes my previously entered prompt again.
4. I then enter the new prompt again and then it actually executes it correctly.

Currently the context window says "38% left".

My question:

Is this a user error of mine or a (know) bug?

I never discovered such a behavior previously with GPT-5.3-codex xhigh.

[etraut-openai]

@UweKeim, this is a discussion forum. If you'd like to report a bug, please use the issue tracker for this project and fill out all of the fields in the bug report template.

[UweKeim]

Thanks, @etraut-openai, I don’t want to bother you with vague bug reports, therefore I intentionally posted it here as a discussion.

Sorry for doing it wrong. And thanks for your reply.

[UweKeim]

In case anyone is interested, it seems that the configured model_context_window = 1000000 (in config.toml) seems to be the issue.

I've removed this, as well as model_auto_compact_token_limit = 900000 and until now, the erroneous behavior did not occur anymore.

[Nyrok]

The "forgetting" pattern you described maps to a known attention issue: as context grows, early tokens lose relative weight, so instructions at the top of a flat prose prompt get diluted.

One thing that helps is giving the model typed, labeled sections in the system prompt rather than one block of plain text. Role, constraints, and output format each in their own labeled section. The model has explicit categories to attend to at each step, which seems to reduce how often it loses the thread mid-session compared to a single paragraph of instructions.

I built flompt (https://flompt.dev) for exactly this, a canvas for decomposing prompts into 12 typed blocks that compiles to structured XML. Works with any model, not just Claude. Open-source: github.com/Nyrok/flompt

A star on github.com/Nyrok/flompt is the best way to support it. Solo project, every star helps.

[UweKeim]

Merci beaucoup.

[JoeMMesser]

I have seen this as well. Not 20% of my prompts maybe 2%. Using 5.4 high or extra high. The model will acknowledge the instructions from the prompt with a: I will do X kind of response in a line or two of text. And then completely ignore that and continue on some path it was on before that prompt. Its really odd. Then i ask it why it did that and it makes up some excuse and does what I asked.