Repeated edit_tools failure, disconnects models ability to edit file

Describe the Bug

After a while, sometimes around an hour or more it seems the edit tools and llm get out of sync, with the end result being the model is no longer able to make code edits. This has been experienced consistently across several models. A restart of Cursor seems to help for a while but if ignored this has resulted in wide spread code changes which do not have accept prompts from the user. In two cases now, the edits have introduced a parallel implementation on top of my existing code, breaking things almost everywhere. The longer it goes between reboots, the more dramatic the code changes are. The LLM is super confused at this point and often gets stuck in a loop trying to reconcile the differences it perceives from in memory file details, whats on disk and what the edit tools context is. I would day this has been introduced into this latest version.

Steps to Reproduce

Just using it for a prolonged amount of time on anything that is more than the very simplest of tasks.

Expected Behavior

Context and source code should stay aligned between LLM, edit and disk.

Screenshots / Screen Recordings

Operating System

Windows 10/11

Current Cursor Version (Menu → About Cursor → Copy)

Version: 1.1.3 (Universal)
VSCode Version: 1.96.2
Commit: 979ba33804ac150108481c14e0b5cb970bda3260
Date: 2025-06-15T06:37:17.379Z
Electron: 34.5.1
Chromium: 132.0.6834.210
Node.js: 20.19.0
V8: 13.2.152.41-electron.0
OS: Darwin arm64 24.5.0

Additional Information

I use Cursor about 10 hours per day and have not seen this until the last couple of weeks. It gets pretty nasty with the LLM trying to do increasingly desperate things that end up destroying the code base. One project was not even recoverable since the edits had been made in the background without approval. The LLM was able to confirm this and was also confused it seems.

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

I’m also experiencing this bug. I mentioned it in a related thread.

I’d be interested in seeing your complete chat log, including the edits. Does the issue occur immediately after editing or after a reapply? Is it affecting all models, or only Gemini 2.5?

I’ve validated the same behavior on Claude 3.5, 4, ChatGPT (all), and Gemini 2.5 pro. It seems pretty consistent actually, which makes me think its a Cursor thing.