Codex has stopped reading full files and refers to I think other older conversations since the updates >= 0.4

What version of Codex is running?

0.50.0

What subscription do you have?

Pro

Which model were you using?

GPT-5

What platform is your computer?

WSL on Windows

What issue are you seeing?

I don't know how to explain it, it feels like Codex is using cached conversations and file-reads selectively.

It starts to forget things (which is probably missing context) after ~5-10% context (i.e. >= 90% left). it seems to refer to old conversations, it forgets rules and it really has a hard time following files that are > 300 lines. It preferes snippets.

I added hard rules to use the file edit and read tool but it still gets around it.

The only thing that actually works is copy pasting the whole file, but then I could just use the web interface.

I believe there have been several optimizations, system prompts and tooling to reduce usage. I do understand that but it's why I came from Claude and what I believe was their August issue, please be more careful.

Context is really important if I can't rely on this I can't rely on the tool, GPT-5 is otherwise amazing.

What steps can reproduce the bug?

Chat with Codex for a while, reference 2 or 3 decently sized (300-700 lines of code) files and ask it for specific changes adaptations or reviews.

See if it's correct especially if it has to follow the whole thread of files and dependencies.

It's not an exaggeration to say that maybe 30% of the time or more it'll make assumptions and not read the whole file and will thus be incorrect, interleave dependencies in blocks and it's more likely to occur because it won't read the whole file and miss the other context.

What is the expected behavior?

_No response_

Additional information

_No response_