When Gemini 2.5 Pro uses the edit_file tool, it always seems to fail. The issue is that while Gemini 2.5 Pro can utilize edit_file to edit files, it fails to receive the correct response—for instance, the tool consistently indicates “no changes” even when edits should have been made. I used prompts to teach Gemini 2.5 Pro to bypass this issue. Although it seemed solvable on the surface, the experience was truly poor.
2 Likes
Tool calls are the biggest pain with this otherwise good model.
The other pain is the obfuscated thinking process, but that’s not something on Cursor’s end.
I think Cursor’s tool design could be better; it’s like building infrastructure for large language models.
1 Like
Try this custom model instruction and see if it improves, I’ve had good success with it: Gemini 2.5 pro 06-05 has issues with Edit & Reapply - #10 by yulin0629
It seems great! but now I’ve run into a new issue that’s directly preventing me from conversing with large language models!https://forum.cursor.com/t/service-unavailable-issue/107425
Sooo many issues in the past 2 weeks and not just in cursor, google messed up here
yeah gemini was unusably slow and dumb for half the day, which is fixed now.
now it constantly runs into the issue of not picking up after running a script file
yesterday gemini was in genius mode and today he is truly incapable.
is google running a quantized model to compensate for server space used on veo 3?
I have a similar experience with Gemini Pro. It claims it can’t use the file edit tool, and it won’t follow custom rules.
Also, it has a tamper! With the edit tool, it starts throwing a fit, like “I’ve had it with this tool… it’s so frustrating!” 
1 Like
I had gemini test his tools and he did find some ambiguity in how the diff is returned, which the model seems to misinterpret. Reapply also duplicates lines if it’s a prepending operation. Here’s the bug report he put together `edit_file` and `reapply` Tools Lead to Gemini-2.5-pro Misinterpretation and Task Failure
Yet another time I’m so disappointed. The code implementation is good but the tool calls… a true disaster.
For now, I’ve opted to write a full section in the project rules about dealing with stress caused by failed tool calls, from eating comfort food to practicing yoga.
Time to try out that yulin0629’s method.
Also, Cursor team’s decision to hide error messages from the user really doesn’t help. More like the opposite. 
I think found a fix. When it’s failing with the edit_tool, ask it to use sed to edit the files.
1 Like
Gemini was my go to but something happened with the edit tool lately.
I can make it read the file to check for success after a failed tool call but that is messy.
Forced to use o3 or Claude.
Gemini seemed the one with better balance. Hopefully they get it fixed.