3 comments

[ 0.22 ms ] story [ 8.7 ms ] thread
(comment deleted)
(comment deleted)
I built this to fix a pattern I kept running into with gpt-5.6-sol: after you correct the model for over-featuring, the correction itself starts leaking into later output. It shows up in READMEs, comments, and follow-up replies as defensive wording like “the product does not contain redundant X since…” or “this is Y instead of Z.” That’s noise for readers.

So I had Codex keep the internal correction notes out of user-facing output. The hook just shows the actual result; any real safety or limitation notes are appended once at the end.

Curious if others find this useful, and what’s still missing. Thanks.