Model pushing special tokens to Gemini CLI

[segin]

What happened?

At random times, the gemini-2.5-pro model will emit repetitive strings of special tokens that map to <ctrl46>. There is a brief example of this behavior in an unrelated open issue, #2008, in the "Additional Context" section.

What did you expect to happen?

The Gemini model to produce a coherent response, not <ctrl46><ctrl46><ctrl46><ctrl46><ctrl46><ctrl46>... ad nauseam.

Client information

Details

╭────────────────────────────────────────────────────────────────────╮
│                                                                    │
│ About Gemini CLI                                                   │
│                                                                    │
│ CLI Version            0.1.12                                      │
│ Git Commit             b29c02dd (local modifications)              │
│ Model                  gemini-2.5-pro                              │
│ Sandbox                no sandbox                                  │
│ OS                     linux                                       │
│ Auth Method            gemini-api-key                              │
│                                                                    │
╰────────────────────────────────────────────────────────────────────╯

Login information

API key

Anything else we need to know?

No response

[LukeKosner]

This happened to me once while using gemini-2.5-pro in the Vercel AI SDK and not since. I made no change to enable it or make it stop.

[Wazbat]

Getting this same issue a lot too with the model directly. This is one of the top google results for it. I'm using gemini-2.5-flash-lite-preview-06-17

[gsquared94]

Found possible duplicate issues:

• Repeated Text Output #2551: (#0.9037118554115295)

[segin]

@gsquared94 while that's the agent getting stuck in an infinite loop (performing the same tool call and response over and over), it's not quite the same thing. In this case, the model is emitting tokens that do not sensibly map to language and exist as special control codes. We as users should never see these. Look at bug #2008, you'll also find an example of this behavior (although the model resumes normal output after emitting a few of these.)

[Wazbat]

Could this be an API error? I'm seeing similar outputs in my langgraph application, and this one of the few google results when searching for <ctrl46>