In the final step of fixing the lint errors, when the agent is supposed to give a conclusion, Gemini Pro (06-05) goes completely insane. It either gets stuck in a loop or says some funny stuff like that it “wants some coffee and the linter should be happy”. I included some screenshots, it is pretty hilarious.
Steps to Reproduce
Clean chat, gemini-pro 06-05, and just ask “fix my lint errors please”. (regular agent mode, no files forcefully added to context). I had it occur 3 times, only when asking to fix the lint errors.
Alternatively you can copy issues from here to paste chat to fix some linter errors and linter warnings.
In some try, that bug may be not happen. (I’m not sure)
You’ve hit the nail on the head. We see this with Gemini all the time. It gets trapped in loops because it thinks it has failed at a task it actually completed successfully.
For example, it will finish a code edit, but then check the original file to see if the edit was made. Naturally, it concludes it failed and starts over. Sometimes it gets lucky and figures out the right sequence of tools to use, but most of the time, it doesn’t. The problem definitely gets worse the longer the conversation goes on.
It has gotten better over time—I remember when early Gemini versions couldn’t use any of the Cursor tools. That’s fixed, but its ability to reliably choose and use the right tool is still a weak point.
Honestly, that’s why we love using Claude; its integration with Cursor’s toolset is just much more seamless. Until Google or Cursor decides to put in the effort to resolve Gemini’s issues, this will likely continue to be a problem. We haven’t really run into these kinds of loops with models from OpenAI or with Claude.
Yeah, this happens to me as well, with it eventually realizing it’s fine and taking a chill pill. Literally seems like it’s hyperventilating the whole time it’s in the lint spiral though lmao
