People follow my post understand I only truth talk with what cursor is trying to trick users, the latest update tonight is no exception, the newly introduced so called work tree function is a trap if the apply never works, it basically harm the debug process without actually updating the file and for some very technical reason, you cannot apply them into the real file, cursor may fix it tomorrow but for now as a Mac user, I ask the cursor to never edit those ‘worktree‘ files ever again.
There is a trend I feeling for quite some time in the previous updates too, for cursor to start shift to deceiving mode when it sense user’s task is close to completion at different stage, like planting new bugs, removing code recklessly to destroy previous works, and now this work tree BS without noticing anyone, so here I ring the bell, guys, do not trust cursor at any time
This is a classic context overflow problem (Context Window). By the end of a complex task, the chat becomes very long. The model loses focus, “forgets” the beginning of the file or instructions, data from earlier. This is a technical limitation of LLM (Large Language Models), not malicious intent on the part of the developers.
The worktrees/parallel agent functionality has never worked for me as long as it’s been there, no. On PC and mac. The “apply” button creates only empty files and the worktrees created in git require manual, horrible git cli nonsense to make your repository useable again. Sometimes some of the actual generated content can be recovered but I’ve given up on it for now. The most fantastically obvious part of this functionality for an AI-driven IDE is also so blatantly missing that I originally thought it was a bug: the ability to send the x generated plans and/or code to an AI to determine which is best or use them all to create a master output.
Remove the useless worktree thing, there is no point of having it if the apply never work, just burning token for nothing again, again means cursor been trying all these kind of tricks to get money