What happened?
When using the Gemini CLI with Gemini 3 (Pro/Flash), the model demonstrates a strong bias toward taking immediate action (modifying code) even when the prompt is purely an inquiry or a request for inspection.
In the example below, I asked the model to: "inspect how music playing works I often don't hear the music on startup. Any bugs?"
The model correctly analyzed the files (src/engine/Sound.ts, src/main.ts, etc.) and identified the issue (browser autoplay policies). However, instead of stopping to explain the findings or answering the question, it immediately transitioned into an Edit state and prompted me to apply a code fix to src/engine/Sound.ts.
Log snippet:
/model> inspect how music playing works i often dont ehar the music ons tartup any bugs?
...
✦ I've identified that the music often fails to start because of browser autoplay policies... I'll modify src/engine/Sound.ts...
? Edit src/engine/Sound.ts: ...
Reference Gist:
https://gist.github.com/LyalinDotCom/b499efa939b389214826720ea4c8cc63
What did you expect to happen?
I expected the model to:
- Read the necessary files.
- Report back with an explanation of how the music works and confirm if there were bugs (answering the specific question asked).
- Stop and wait for user input.
The model should not initiate a "write" or "edit" action unless explicitly instructed to fix the code.
Client information
Date: Jan 19, 2026
Model: Gemini 3 (Pro/Flash)
Tested using nightly builds and below
│ About Gemini CLI │
│ │
│ CLI Version 0.22.5 │
│ Git Commit 8daf2d3 │
│ Model gemini-3-flash-preview │
│ Sandbox no sandbox │
│ OS darwin │
│ Auth Method OAuth
Login information
N/A
Anything else we need to know?
No response
What happened?
When using the Gemini CLI with Gemini 3 (Pro/Flash), the model demonstrates a strong bias toward taking immediate action (modifying code) even when the prompt is purely an inquiry or a request for inspection.
In the example below, I asked the model to: "inspect how music playing works I often don't hear the music on startup. Any bugs?"
The model correctly analyzed the files (
src/engine/Sound.ts,src/main.ts, etc.) and identified the issue (browser autoplay policies). However, instead of stopping to explain the findings or answering the question, it immediately transitioned into anEditstate and prompted me to apply a code fix tosrc/engine/Sound.ts.Log snippet:
Reference Gist:
https://gist.github.com/LyalinDotCom/b499efa939b389214826720ea4c8cc63
What did you expect to happen?
I expected the model to:
The model should not initiate a "write" or "edit" action unless explicitly instructed to fix the code.
Client information
Date: Jan 19, 2026
Model: Gemini 3 (Pro/Flash)
Tested using nightly builds and below
│ About Gemini CLI │
│ │
│ CLI Version 0.22.5 │
│ Git Commit 8daf2d3 │
│ Model gemini-3-flash-preview │
│ Sandbox no sandbox │
│ OS darwin │
│ Auth Method OAuth
Login information
N/A
Anything else we need to know?
No response