google-gemini
Model outputting internal thought (which leads to bad reasoning)
What happened?
Description: The AI assistant's internal thought process is sometimes displayed to the user instead of the final, polished response. This can be confusing for the user and exposes the internal workings of the assistant.
Thank you for your help in improving my functionality.
What did you expect to happen?
Seems the gemini cli was getting into a weird state where it only showed me what it planned to tell me, but without actually telling me. (The above is what gemini suggested I write in the bug report.)
Client information
- CLI Version: 0.6.1
- Git Commit: f3078136
- Session ID: 7f708760-96de-4aee-a6a4-fe7d2d3162b7
- Operating System: darwin v24.8.0
- Sandbox Environment: no sandbox
- Model Version: gemini-2.5-pro
- Memory Usage: 516.3 MB
Login information
No response
Anything else we need to know?
The comment
Hi, we only display thoughts in the loading phrase, can you share a screenshot of what it looks like?
Here is a screen shot of this kind of behavior:
This is after updating (using brew). Current version info:
╭─────────────────────────────────────────────────────────────────╮
│ │
│ About Gemini CLI │
│ │
│ CLI Version 0.7.0 │
│ Git Commit f28ffaa3 │
│ Model gemini-2.5-pro │
│ Sandbox no sandbox │
│ OS darwin │
│ Auth Method OAuth │
│ │
╰─────────────────────────────────────────────────────────────────╯
Ah. Ok. So there is a separate "thoughts" feed that Gemini uses for thinking but you are looking at here is not that. This is the regular model output. Sometimes the model gets confused and starts talking in second person. Unfortunately, there's not much we, the CLI team, can do about it other than wait for the Google Deep Mind team to publish better models.
Just want to add an update that the latest version of gemini cli using "gemini-3-pro-preview" does this often. It will also stop mid chain of thought and I"ll have to say "continue" and it'll continue where it left iff. It will stay in this chain of thought verbose mode until I /clear it.
About Gemini CLI │ │ │ │ CLI Version 0.18.4 │ │ Git Commit 2e8d7831c │ │ Model pro │ │ Sandbox no sandbox │ │ OS linux │ │ Auth Method gemini-api-key
Deduped a couple issues to this. Once it starts spewing it's thought process to the output the reasoning is junk.
@mmaunder do you have examples of transcripts or screenshots that you can share to helps us reproduce/validate this bug?
@silviojr Here's an example: https://paste.googleplex.com/6584854227582976