Quality problems

It seems this app is getting worse and worse. I am using Flutter

When I ask an LLM to add a new method or feature it will delete code. I never asked it to delete any code only add new code. This thing has suggested to delete entire blocks of code. Working code mind you. Code that has nothing to do with the subject in question.

Other times it will just repeat the entire file as its answer to the question that I asked it. It doesnt suggest any code changes, it just types out the entire file.

Then another issue I had is in the system settings I will tell it that it is an expert flutter developer, dont repeat code, dont apologize, pay attention to all instructions given,

It gave me python code, It always repeats the entire file, It apologizes profusely and like the example above it doesnt pay attention to instructions. It deletes working valid code when it was never told to do so.

Not really sure what is going on here but I am finding the main LLMs on their actual websites to be much more efficient than cursor these days.

I am paying for cursor pro, I should get better than what this is giving me.

4 Likes

I experience similar issues:
Occasionally it deletes, refactors unrelated code.
Sometimes it just repeats the code, without adding any changes.
Also it always removes the last empty line when doing ‘apply’, and I then have to readd it, to avoid polluting commit history.
Also I added a rule for AI to “never add comments to code” it generates. But it always, especially gpt, adds comments anyway, which I then have to remove each time.
Partial accepts for Copilot++ no longer work, despite working before and me never changing the settings for it…

1 Like

It is almost as if cursor has a cache that is too large and the entire app becomes unstable.

I literally asked it to fix a problem with a user profile screen loading infinitely and this was the response

" * One potential workaround is to ensure the isReady flag is properly set to true when the data loading is complete. This can be done using the onRender callback of the PDFView widget (assuming you're using the flutter_pdfview package). [Source]



PDFView(

onRender: (_pages) {

setState(() {

isReady = true;

});

},

// ...

)

* Another suggestion is to use the PDFViewController to control the loading state of the PDF. [Source]

PDFView(

onViewCreated: (PDFViewController pdfViewController) {

_controller.complete(pdfViewController);

},

// ...

)"

I dont have a single pdf file, or mention of any pdf document in the entire app. What is this response? This is absolutely left field and out of context 100%

2 Likes

I find that AI Rules do nothing. Maybe I don’t know how to use them. But in that case they need some documentation!

2 Likes

@rishabhy Fix this please. Cursor is giving awful responses when the normal Claude.AI or Chat.OpenAI.com work flawlessly

Are you all doing anything at all on the backend that would give bad results and not comparable results to the normal systems?

Can you double check that the partial accept setting in set to enabled in settings (cmd-shift-J > Features)?

Do you mind reporting the chat as bad? I’ll take a look to see what is going wrong!

After you generate the bad PDF chat response, hit cmd-shift-p and type “Report Latest Chat as Bad”.

1 Like

Yes, it was enabled.

I’ve found a fix though: I went to shortcuts and removed the Ctrl-Rigtharrow keybiding for CursorWordRight. It didn’t seem to have any impact on partial accepts before.

1 Like

Reported plenty of chats as bad.

The app is completely broken now. Now it gives code that has nothing to do with the current coding language.

Apply doesnt work any more.
It ignores my requests.

Looks like you have “Always search web” enabled. Can you try disabling that and trying a more few requests?

I’ve had the llm deleting code issue too. Sometimes it will just delete stuff and do nothing else (inline with ctrl+k). And telling the model to not delete stuff or anything like that doesn’t help. Switching the model does seem to help at this point, I’m always using sonnet 3.5 and switching to gpt4-o when this happens.

As someone used to working with llms this does look a bit like a repetition problem. But since this is the only place where this kind of thing happens my guess would be this is an issue with how the llm reacts to cursor’s prompt design.

could this be a caching issue? seems over time cursor ls llms starts hallucinating.

creating a new chat doesn’t help either

Each release this thing gets worse. Guys please stop updating the app. One fix breaks something else. At this point its time to start from scratch.

The apply button stopped working. I click apply and nothing happens. No code is changed nor updated.
Restarted the app and pc multiple times.

Please fix this what are we paying for when the product is broken and unreliable like this

1 Like

Hi all,

sorry we are trying our best to be stable and push the product forward. really sorry. A quick explanation:

  1. making cursor tab faster (prev. copilot++) meant doing a careful rewrite of a lot of the logic.
    feel the pain for partial accepts. we will fix it.
  2. the accept button not working is weird. we are scaling our usage really fast and it might be a model bug. i will investigate.

if you would like to use an older release for a bit, i can get you all an earlier release.

2 Likes

And anyone has any thoughs on debugging? The debugger tool doesn’t work correctly: controls not working, tab hiding when other are not, command palette for those debugging controls also not working nor shortcuts, etc.

In my project I have couple files that Cursor always fail to edit. Always same thing, adds new function fine or changes existing but deletes everything else from the file even asked not to delete anything.

1 Like

what’s the size of these projects? have you found any patterns on when it breaks, or is it all the time? we’re rolling out a fix for some of these issues soon, hope that will mitigate some of these problems