Bug report for "data:" being unrecognized by aider

[emincangencer]

Aider version: 0.73.0
Python version: 3.12.8
Platform: Windows-11-10.0.26100-SP0
Python implementation: CPython
Virtual environment: Yes
OS: Windows 11 (64bit)
Git version: git version 2.47.1.windows.2

Aider v0.73.0
Model: gemini/gemini-2.0-flash-thinking-exp with diff edit format
Git repo: .git with 47 files
Repo-map: using 4096 tokens, auto refresh

---

data: is not processed by aider.

diff> /ask What does "data: {" mean?

It looks like you are asking about the meaning of " {"  in the context of the prisma\seed.ts file, specifically within the prisma.novel.create  
call.

[paulsharratt0]

It looks like something is removing "data:" in gemini flash thinking responses. I don't see any indication of this in the Aider codebase

[paulbatum]

I think it must be something in the aider codebase (or dependencies), because the problem is not reproducible in any other environment.

Steps to reproduce:

1. Create an empty folder with just a .env file (configured for gemini-2.0-flash)
2. Navigate into the folder and run aider

Aider v0.74.2
Model: gemini/gemini-2.0-flash with diff edit format
Git repo: none
Repo-map: disabled

3. Send the command:

 /ask Please respond with just the string `entity_data: #asdf`

Aider responds with:
entity_ #asdf

The issue does not reproduce in the playground, nor via raw API:

POST https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:generateContent?
content-type: application/json

{
    "contents": [{
        "parts": [{"text": " Please respond with just the string `entity_data: #asdf`"}]
    }]    
}

Response:

{
      "content": {
        "parts": [
          {
            "text": "entity_data: #asdf\n"
          }
        ],
        "role": "model"
      },
      "finishReason": "STOP",
      "avgLogprobs": -0.00023910652713051865
    }

Does aider have settings for logging the raw API request/response? If yes, I would be happy to enable it to confirm that the raw model response is correct.

[paulbatum]

OK I debugged aider and noticed that I could not reproduce the issue when I turned streaming off (which I did to make debugging simpler).

So this means we at least have a workaround:
AIDER_STREAM=False

I dug a little more and it seems like its a liteLLM issue. See:
BerriAI/litellm#8143

The above issue points at the problematic code here:
https://github.com/BerriAI/litellm/blob/v1.59.10/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py#L1434