bug: model occasionally does not read enough context to correctly answer

Resolved 💬 0 comments Opened Sep 25, 2025 by techjoec Closed Nov 5, 2025
TL;DR: Codex doesn't seem to properly use surrounding context especially in doc review tasks (markdown in my case) to understand file system paths (and likely more). Forcing it to consume a few lines around its target appears to fix it.

What version of Codex is running?

Happens on old and new

Which model were you using?

gpt-5-codex high reasoning but existed prior

What platform is your computer?

Linux 6.8.0-79-generic #79-Ubuntu

What steps can reproduce the bug?

When reviewing documentation more-so than user prompts Codex CLI does not consider surrounding context at all. For example, have it scan a repo for file system path updates. It goes totally dumb. If it matches "\S/\S" (anything slash anything) it has to be fought about it not being a path. I've had it create bonkers directory structures a few times but if you can't repro it I have an easy way. Tell it to scan all docs and scripts to move a project to a new folder. Then go scan the docs and scripts. Then tell it to go scan the scripts, check all the paths for sanity, and output the result to a JSON file to review.

What is the expected behavior?

Actual contextual awareness to complete the task, especially after telling the model _multiple times_ that the output was trash and that it was catching non-file paths in its work.

What do you see instead?

No sanity checking and no contextual awareness. Returns foo/bar as a file path liberally and defends it.

Potential Fix / Current Workaround (Additional information)

I feel this is also hampering some of the "review this repo for XYZ" tasks. As a test I forced it to consume context by telling it to go to each file it found paths in, and extract +/- 3 lines (6 lines each hit). Boom, instant smartness upgrade. Totally tackled the task and actually used reasoning as was very apparent in a 'better' response, and the expected level of output quality.

Perhaps something in tooling is negatively impacting the natural ability of the models to consume and consider surrounding context? Just giving it a few lines did wonders for me, and upon reflection, I think I've been fighting this more than I previously thought globally in my work since day #1.

View original on GitHub ↗