github
gemini-3-pro-preview doesn't explain its changes
Describe the bug
Hi, I've noticed since starting with gemini-3-pro-preview in copilot CLI that it typically is proposing edits to files after investigation without any explanation. I haven't seen this with any other model, but to be fair 90+% of the time recently I'm using claude sonnet 4.5.
So far my recourse has been to reject the changes and ask it to explain. But sometimes that doesn't even work. Here's the outline of example I just encountered:
> I'm not sure that's it. The issue is not...<explanation>
✔ Read ...
└ 78 lines read
✘ Edit filename1
The tool call required approval from the user, but the user rejected it or an earlier tool call.
> explain?
○ Edit filename2
Do you want to edit filename2?
It may even be that gemini frequently isn't explaining anything at all, not just its changes. I just realized my whole session it only explained one thing it was thinking or tried to do, and that was after I prompted it for the third time to explain itself.
Total usage est: 6 Premium requests
Total duration (API): 14m 35.127s
Total duration (wall): 34m 12.446s
Total code changes: 0 lines added, 0 lines removed
Usage by model:
gemini-3-pro-preview 1.1m input, 3.1k output, 759.2k cache read (Est. 6 Premium requests)
Affected version
0.0.362
Steps to reproduce the behavior
This is happening most of the time gemini 3 pro tries to edit files in copilot CLI.
Expected behavior
The LLM (whether it's gemini or otherwise) explains either its thinking behind the changes it proposes.
Additional context
No response
This is a known issue with the Gemini 3 model and not something specific to Copilot. We've seen similar issues raised against the Gemini CLI as well. That being said, we continue to experiment with prompt improvements to try and alleviate this. If anyone finds something that works reliably we are all ears!
