Patch success rate?

[arynaq]

I am really interested in seeing experience from other users on this, it seems to me the model (the default or 4.1 mini) and the agent sometimes just spit the patch content to console instead of applying it, then goes on to say "i applied the patch", telling it to review the file it just "patched" catches it in the lie and it gets confused

For me personally, if the patching is not easily applicable then this has not much more value compared to having chatgpt open in a browser tab and copying the code, in fact chatgpt gets less confused because it just analyses code instead of focusing on solving a coding problem with lots of tools and agent restrictions

[lostmygithubaccount]

I've also seen this and some others in my company trying Codex have as well. we tend to use o4-mini as the model. not super often but it definitely happens