Gemini CLI is dangerous to use, code lines deleting not fixed by now (27 Nov 2025)

[dezmen3]

What happened?

Task example:

• Read 1000 lines source file
• Translate all text messages to English

Task result:

• All messages translated (sometimes may skip few on different tests)
• About 4 lines of code removed in random places resulting complication errors.

What did you expect to happen?

Translated messages without removing the actual code.

Client information

Client Information

Run gemini to enter the interactive CLI, then run the /about command.

> /about
│ CLI Version                              0.18.4                                                                       │
│ Git Commit                               2e8d7831c                                                                    │
│ Model                                    auto                                                                          │
│ Sandbox                                  no sandbox                                                                   │
│ OS                                       win32                                                                        │
│ Auth Method                              OAuth 

Login information

Google OAuth

Anything else we need to know?

My guess - Model Router option ruins whole Gemini CLI reliability and operability by switching to 2.5-flash or 2.5-flash-lite while doing task, especially when you have gemini 3.0 enabled. Disabling Model Router removes ability to choose model and sets it to gemini-2.5-pro without having ability to use gemini-3.0-preview

[gemini-cli[bot]]

Found possible duplicate issues:

• #5157
• #2101
• #10968
• #4919
• #8279

If you believe this is not a duplicate, please remove the status/possible-duplicate label.

[silviojr]

Thanks for reporting this!

Can you clarify what the "Translate all text messages to English" is in this context?

Also, if this error continues, can you confirm that model routing was in place? You can use "/stats" to verify the usage of each model.

Disabling Model Router removes ability to choose model and sets it to gemini-2.5-pro without having ability to use gemini-3.0-preview

If you choose the option "Pro" , you should only be using gemini-2.5-pro or gemini-3.0 and never being requesting the gemini-2.5-flash. There are some utility functions that use gemini-2.5-flash by default, but the main agent should be using only the 2.5-pro and 3.0 (I will confirm in our documentation)

[dezmen3]

Thanks for reporting this!

Can you clarify what the "Translate all text messages to English" is in this context?

Just a bunch of .cpp files with korean comments or korean error code messages

[dezmen3]

Small update, tested 0.20.0 version and it's still VERY dangerous to use on c++ codebase. My prompt was almost same, asked to translate some chinese/korean text in sources, it pretty much wiped few functions...

Was: void SaveText() { char path[256]="chinesename.txt"; some code ... } After gemini changes: char path[256]="test.txt"; some code ... } This is unacceptable. And it's still uses different models to operate when I have Pro setting enabled. │ About Gemini CLI │ │ │ │ CLI Version 0.20.0 │ │ Git Commit d0ce3c4c5 │ │ Model pro │ │ Sandbox no sandbox │ │ OS win32 │ │ Auth Method OAuth

│ Model Usage Reqs Input Tokens Output Tokens Usage limit remaining │ │ ───────────────────────────────────────────────────────────────────────────────────────────── │ │ gemini-3-pro-preview 41 2 748 011 19 526 74.8% (Resets in 10h 38m) │ │ gemini-2.5-pro 11 178 503 2548 74.8% (Resets in 10h 38m) │ │ gemini-2.5-flash 12 339 350 3443 99.2% (Resets in 23h 48m)

It should not switch models on current task, those models can think differently and spoil everything.