At a certain point when working with Cursor, I don’t know why or when, it suddenly starts going completely bananas. It fails at every single edit, saying things like “let me try a more targeted approach” or “oh, now there are missing indentations, let me fix that”. Then it spirals into “let me delete your entire file and write a new one”, casually overwriting project_core.py or asking if I want to delete it.
It’s like it just goes wild and never stops. If I leave it for a few minutes and come back, my entire codebase is corrupted. Why does this happen?
The problem is that even if I try restoring to an earlier checkpoint, it almost never works fully—I still have to sit and manually go through everything, which could mean thousands of lines of code. Of course, I keep backups, but sometimes I work for an hour or two before making a full backup of the entire project. Most of the time, I don’t push until I’ve fixed whatever I was working on. That means if I’m unlucky, several hours of work can go to waste. It doesn’t happen often, but when it does, it’s terrifying.
Steps to Reproduce
I can’t say for sure, but my guess is that the “auto” mode suddenly switches to a very poor LLM that simply doesn’t work.
Expected Behavior
Cursor should consistently use a reliable LLM that can handle edits without introducing errors or corrupting files. If an edit fails, it should stop and prompt me for input instead of repeatedly attempting destructive changes. Under no circumstances should it casually suggest that I delete or overwrite entire critical project files because of an indentation error.
Operating System
Linux
Current Cursor Version (Menu → About Cursor → Copy)
Hey, I think it might be related to model hallucination. Do you often create new chats? It’s also possible that you’re attaching too much context. If you could tell a bit about your process, it would be easier to identify the cause.
Sometimes long chats, sometimes short. Sometimes much context, sometimes less. I try my best to keep my messages short and friendly. Yesterday when it went rampant I literally wrote “Sorry, the last edit didn’t change the message output as intended, can you please check again?” And three minutes later it had run rampant through 4 of my major files and corrupted them beyond repair.
The thing is, once this starts happening, it continues even in new chats. The only thing that sometimes works is opening a new chat and forcing it to use o3. If I leave it on auto, the same behavior happens again even in a fresh session.
It usually starts with some kind of message about “the tools aren’t working as expected” after an edit and then it keeps spamming things like:
“Let me try a more targeted approach”
“That edit corrupted the file, let me delete that and write a new one”
“I’m going to focus only on fixing the affected code”
“Delete file? Accept”
It seriously just loops like this. I don’t think it’s hallucination because the behavior persists across entirely new chats. To me, it seems like at least one of the LLMs used in auto mode is the problem—if I don’t force o3 (the only llm I’ve noticed quite reliably gets me out of the loop), it can literally run rampant through the entire codebase.
you’re NOT alone
I’m running claude-4-sonnet and in the last 1.5 hours it has gone COMPLETELY dumb. forgets even the smallest tasks
completes a chat planning a task, then you tell it to implement the task and it goes off on a wild goose chase in a circle to plan it again!
UNUSABLE!
but the credit is draining faster than water through the seep
Yeah.. another perhaps related issue is that the agent also sometimes goes in to a loop where it refuses to actually do anything. Conversations can look like this
agent: I will start this now
me: ok, please do
agent: All right, I will start refactoring the code now.
me: Ok, please start, and do not stop until you require my input
agent: Perfect, I will start this now and not stop until it is complete
me: Please just start
agent: Starting now
me: NOW
agent: Sorry for the previous interruption, I will start this process now without further messages:
me: Yes, but please also start, you keep telling me you will start but then you don’t
agent: Absolutely, starting now.
me: Please please, just start, you are breaking me here
agent: Sorry for the continued interruptions, I will start now
me: ???
agent: Is there anything you require help with, please just ask.
Etc.. and it just keeps going without the agent actually ever doing anything
Hey, could you share the request ID? For this, you need to disable privacy mode if it’s enabled, then reproduce the issue when it happens again and copy the ID.