[Bug]: Thought Signature not being returned

[rasswanth-lyzr]

What happened?

Gemini 3 pro is not returning the new field Thought Signature

My code -

response =  completion(
            model="gemini/gemini-3-pro-preview",
            messages=[{"role": "user", "content": "What is the capital of France?"}],
            tools=[{"type": "function", "function": {"name": "get_capital", "description": "Get the capital of a country", "parameters": {"type": "object", "properties": {"country": {"type": "string"}}}}}],
            tool_choice="auto",
            temperature=1,
            timeout=120,
            metadata={"run_id": "123"},
        )

My Output

ModelResponse(id='me8dadXQBIzxjuMPqJ7UwAM', created=1763569558, model='gemini-3-pro-preview', object='chat.completion', system_fingerprint=None, choices=[Choices(finish_reason='tool_calls', index=0, message=Message(content=None, role='assistant', tool_calls=[ChatCompletionMessageToolCall(index=0, function=Function(arguments='{"country": "France"}', name='get_capital'), id='call_2ebefc07c5584185a2b8080d3ef1', type='function')], function_call=None, provider_specific_fields=None))], usage=Usage(completion_tokens=82, prompt_tokens=45, total_tokens=127, completion_tokens_details=CompletionTokensDetailsWrapper(accepted_prediction_tokens=None, audio_tokens=None, reasoning_tokens=66, rejected_prediction_tokens=None, text_tokens=16), prompt_tokens_details=PromptTokensDetailsWrapper(audio_tokens=None, cached_tokens=None, text_tokens=45, image_tokens=None)), vertex_ai_grounding_metadata=[], vertex_ai_url_context_metadata=[], vertex_ai_safety_results=[], vertex_ai_citation_metadata=[])

I don't get the thought signature, therefore my system breaks. I need to make further tool calling and i'm currently unable to make it work

When I try to make another call, it throws me an error saying

ERROR:root:Error during process: litellm.BadRequestError: VertexAIException BadRequestError - {
  "error": {
    "code": 400,
    "message": "Function call is missing a thought_signature in functionCall parts. This is required for tools to work correctly, and missing thought_signature may lead to degraded model performance. Additional data, function call `default_api:agent_tool_67e4d6673e456a9f9291203d` , position 2. please refer to https://ai.google.dev/gemini-api/docs/thought-signatures for more details.",
    "status": "INVALID_ARGUMENT"
  }
}

Relevant log output

ModelResponse(id='me8dadXQBIzxjuMPqJ7UwAM', created=1763569558, model='gemini-3-pro-preview', object='chat.completion', system_fingerprint=None, choices=[Choices(finish_reason='tool_calls', index=0, message=Message(content=None, role='assistant', tool_calls=[ChatCompletionMessageToolCall(index=0, function=Function(arguments='{"country": "France"}', name='get_capital'), id='call_2ebefc07c5584185a2b8080d3ef1', type='function')], function_call=None, provider_specific_fields=None))], usage=Usage(completion_tokens=82, prompt_tokens=45, total_tokens=127, completion_tokens_details=CompletionTokensDetailsWrapper(accepted_prediction_tokens=None, audio_tokens=None, reasoning_tokens=66, rejected_prediction_tokens=None, text_tokens=16), prompt_tokens_details=PromptTokensDetailsWrapper(audio_tokens=None, cached_tokens=None, text_tokens=45, image_tokens=None)), vertex_ai_grounding_metadata=[], vertex_ai_url_context_metadata=[], vertex_ai_safety_results=[], vertex_ai_citation_metadata=[])

Are you a ML Ops Team?

No

What LiteLLM version are you on ?

1.75.2

Twitter / LinkedIn details

No response

[rasswanth-lyzr]

@Sameerlite I get a similar error when passing the tool response back

Error -

litellm.exceptions.BadRequestError: litellm.BadRequestError: GeminiException BadRequestError - {
  "error": {
    "code": 400,
    "message": "Function call is missing a thought_signature in functionCall parts. This is required for tools to work correctly, and missing thought_signature may lead to degraded model performance. Additional data, function call `default_api:get_capital` , position 2. Please refer to https://ai.google.dev/gemini-api/docs/thought-signatures for more details.",
    "status": "INVALID_ARGUMENT"
  }
}

Code to replicate -

from litellm import completion
import os
import json
os.environ['GEMINI_API_KEY'] = ""

tools = [{"type": "function", "function": {"name": "get_capital", "description": "Get the capital of a country", "parameters": {"type": "object", "properties": {"country": {"type": "string"}}}}}]

# First turn
response =  completion(
            model="gemini/gemini-3-pro-preview",
            messages=[{"role": "user", "content": "What is the capital of France?"}],
            tools=tools,
            tool_choice="auto",
            temperature=1,
            timeout=120,
            metadata={"run_id": "123"},
        )

# Extract the thought signature from the first response
thought_signature = None
if response.choices[0].message.thinking_blocks:
    thought_signature = response.choices[0].message.thinking_blocks[0].get('signature')

# Build the assistant message with tool calls including thought signature
assistant_message = {
    "role": "assistant",
    "content": None,
    "tool_calls": []
}

for tool_call in response.choices[0].message.tool_calls:
    tool_call_dict = {
        "id": tool_call.id,
        "type": "function",
        "function": {
            "name": tool_call.function.name,
            "arguments": tool_call.function.arguments
        },
        "index": tool_call.index
    }
    
    # Add thought signature if available
    if thought_signature:
        tool_call_dict["provider_specific_fields"] = {
            "thought_signature": thought_signature
        }
    
    assistant_message["tool_calls"].append(tool_call_dict)

# Create tool response message
tool_response_message = {
    "role": "tool",
    "name": "get_capital",
    "tool_call_id": assistant_message["tool_calls"][0]["id"],
    "content": "{\"capital\": \"Paris\"}"
}

print("--------------------------------")
print("MESSAGE REQUEST:")
print(json.dumps([
        {"role": "user", "content": "What is the capital of France?"},
        assistant_message,
        tool_response_message
    ], indent=4))

# Second turn with conversation history including tool response
response2 = completion(
    model="gemini/gemini-3-pro-preview",
    messages=[
        {"role": "user", "content": "What is the capital of France?"},
        assistant_message,
        tool_response_message
    ],
    tools=[{"type": "function", "function": {"name": "get_capital", "description": "Get the capital of a country", "parameters": {"type": "object", "properties": {"country": {"type": "string"}}}}}],
    tool_choice="auto",
    temperature=1,
    timeout=120,
    metadata={"run_id": "124"}
)

print("\n" + "="*80 + "\n")
print("Second response:")
print(response2)

Message Request structure -

MESSAGE REQUEST:
[
    {
        "role": "user",
        "content": "What is the capital of France?"
    },
    {
        "role": "assistant",
        "content": null,
        "tool_calls": [
            {
                "id": "call_2b2b40698e4146628d02821731d3",
                "type": "function",
                "function": {
                    "name": "get_capital",
                    "arguments": "{\"country\": \"France\"}"
                },
                "index": 0,
                "provider_specific_fields": {
                    "thought_signature": "EuMCCuACAdHtim8wS0W+P0ETtZZhVRo9zbRgDSt9TF0SSv7oroRDYWY1P0EEdA25m808oc6wwglJTdn5JyrxMgBNOrGIyqiWQh7P+0b8ZnsWiTuuwzHBvojd0dPXse5QTyJcaBW7Xn9dTfopuKcjv4P0CG/ew2h7qspuvNU06wCiF+1eRHaKRhFiLOby3HxwDKu3UJXrrW1obyl1hLq84+YExPTp8izypPcS8FVIdMqLigUgfwQ0pA8gbHmvKjbzjzpkft5R0hgAq8NYqw2Zv6GYrKSbjXhSNJ+6wdaszJev2KTi3Iyov9SbYxYnwl85+/ncvG78uEiB9+/74/Fw2EvjRHsQrI+X7EfZ0KDjIglUSbnjSyz3on4yt/eDpqsT0WuJpX65sKi2LqxoZh1yRj6Yj4HhZJfyDsEyaX6Wv459JJ4DwQf2OM+5fn/KuZpAxuunSgGiwKa3jbXCZ8tbb28YHRp8dQ=="
                }
            }
        ]
    },
    {
        "role": "tool",
        "name": "get_capital",
        "tool_call_id": "call_2b2b40698e4146628d02821731d3",
        "content": "{\"capital\": \"Paris\"}"
    }
]

[Sameerlite]

Hey @rasswanth-lyzr Can you try this with the latest code on main?

[rasswanth-lyzr]

Hey @Sameerlite the main branch is working well! When can I expect the pip package to be updated? Thanks for the help!

[Sameerlite]

@rasswanth-lyzr We will have a release this Saturday. Expect it in that release