Interesting. What model and tool was used?
I have seen similar failure modes in Cursor and VSCode Copilot (using gpt5) where I have to babysit relatively small refactors.
I have seen similar failure modes in Cursor and VSCode Copilot (using gpt5) where I have to babysit relatively small refactors.
However, I also think that models which focus on higher reasoning effort in general are better at taking into account the wider context and not missing obvious implications from instructions. Non-reasoning or low-reasoning models serve a purpose, but to suggest they are akin to different flavours misses what is actually quite an important distinction.