Pydantic exceptions when using gemini-3-flash

[kevincojean]

Good afternoon,

I'm using (main): 20260110-1-db797fb and am seeing many exceptions relative to pydantic when using gemini-3-flash.

Here is the logs:

llm-proxy  | Attempting stream with credential antigravity_oauth_1.json (Attempt 1/2)
llm-proxy  | Stream connection established for credential antigravity_oauth_1.json. Processing response.
llm-proxy  | /usr/local/lib/python3.11/site-packages/pydantic/main.py:464: UserWarning: Pydantic serializer warnings:
llm-proxy  |   PydanticSerializationUnexpectedValue(Expected 10 fields but got 5: Expected `Message` - serialized value may not be as expected [field_name='message', input_value=Message(content=None, rol...er_specific_fields=None), input_type=Message])
llm-proxy  |   PydanticSerializationUnexpectedValue(Expected `StreamingChoices` - serialized value may not be as expected [field_name='choices', input_value=Choices(finish_reason='st...form structure.\n\n\n'}), input_type=Choices])
llm-proxy  |   return self.__pydantic_serializer__.to_python(
llm-proxy  | /usr/local/lib/python3.11/site-packages/pydantic/main.py:464: UserWarning: Pydantic serializer warnings:
llm-proxy  |   PydanticSerializationUnexpectedValue(Expected 10 fields but got 5: Expected `Message` - serialized value may not be as expected [field_name='message', input_value=Message(content=None, rol...er_specific_fields=None), input_type=Message])
llm-proxy  |   PydanticSerializationUnexpectedValue(Expected `StreamingChoices` - serialized value may not be as expected [field_name='choices', input_value=Choices(finish_reason='st... visual design.\n\n\n"}), input_type=Choices])
llm-proxy  |   return self.__pydantic_serializer__.to_python(

Here is my provider configuration:

        "google.antigravity.gemini-3-flash": {
          "id": "antigravity/gemini-3-flash",
          "limit": { "context": 1048576, "output": 65636 },
          "options": { "thinking_level": "high" },
          "variants": {
            "minimal": { "thinking_level": "minimal" },
            "low": { "thinking_level": "low" },
            "medium": { "thinking_level": "medium" },
            "high": { "thinking_level": "high" }
          }
        },

Cheers and thank you for the great project,

[mirrobot-agent]

@kevincojean Thank you for submitting this issue. I am now beginning my analysis and will report back shortly.

[mirrobot-agent]

Initial Analysis Report

Summary: The Pydantic serialization warnings are a cosmetic issue caused by using litellm.ModelResponse (designed for non-streaming responses) to construct streaming chunks, resulting in type mismatches during serialization.

Issue Validation: Confirmed - The warning messages are reproducible and stem from a known pattern in the codebase.

Reproducibility Assessment: Reproducible - This issue affects all streaming responses from the Antigravity provider, not just gemini-3-flash.

Root Cause Analysis:

The warnings originate from a type mismatch in how streaming responses are constructed:

1. Location: In src/rotator_library/providers/antigravity_provider.py:4032, streaming chunks are created using:

   yield litellm.ModelResponse(**openai_chunk)

2. Problem: The openai_chunk dict contains streaming-format data with delta objects (OpenAI's streaming format):

   {"choices": [{"index": 0, "delta": {"content": "..."}}]}

3. Mismatch: However, litellm.ModelResponse expects non-streaming format with message objects:

   {"choices": [{"index": 0, "message": {"role": "assistant", "content": "..."}}]}

4. Result: When client.py:700-702 calls chunk.model_dump() to serialize these objects, Pydantic detects the structural mismatch and issues warnings:

   • "Expected 10 fields but got 5: Expected 'Message'" - Delta objects have fewer fields than full Message objects
   • "Expected 'StreamingChoices' - serialized value may not be as expected" - The internal structure doesn't match

Important Note: These are warnings only, not errors. The proxy continues to function correctly and delivers valid streaming responses to clients.

Suggested Labels: bug, low-priority - This is a cosmetic issue that generates log noise but doesn't affect functionality.

Proposed Next Steps:

Option 1 (Quick fix - suppress warnings):

import warnings
warnings.filterwarnings('ignore', category=UserWarning, module='pydantic')

Option 2 (Proper fix - yield plain dicts instead of ModelResponse for streaming):
In antigravity_provider.py, replace:

yield litellm.ModelResponse(**openai_chunk)

with:

yield openai_chunk  # Return plain dict for streaming chunks

Then update client.py to handle plain dicts without calling .model_dump().

Option 3 (Alternative - use a streaming-specific wrapper class):
Create a simple wrapper that doesn't trigger Pydantic validation for streaming chunks.

Missing Information (if any): None - the issue is reproducible from the provided logs.

---

Additional Context:

• The unpinned litellm dependency in requirements.txt means this warning may have appeared after a litellm update with stricter Pydantic v2 serialization
• The same pattern exists in other providers (gemini_cli_provider.py, iflow_provider.py, qwen_code_provider.py) which may also produce similar warnings

This analysis was generated by an AI assistant.

[Mirrowel]

This is a litellm issue and everyone has it. It is just warnings though - litellm needs to fix this in their library. There are plenty of issues on their tracker.

I'll look into it however from the bot's explanation - might be worth it.

[Mirrowel]

Fixed in b505886