and I see from your screenshot you are using the tool wrong.
it’s generally bad to stick to one chat, try to reimagine your workflow with no more than two messages.
so you prompt an agent to do the thing, and if it does it wrong - don’t ask it to fix it or change something, just edit the initial prompt addressing issues it faced and try again
every time you extend a chat the context becomes more and more compromised. LLMs are not humans, you can’t really have conversations with them. every back and forth will make the model a bit dumber, with all the things that went wrong still existing in context and all the critique from the user making it slowly “lose its mind”
have a look at this post: Complex Context - TIP!
and in general, try to one-two shot your request. every new message from the user dumbs models down. it’s not really on Cursor as it’s just how LLMs operate.