Missing thought_signature when using Gemini 3 pro preview

[silvasw]

Please read this first

• Have you read the docs?Agents SDK docs
• Have you searched for related issues? Others may have faced similar issues.

Describe the bug

Gemini 3 Pro Preview (google/gemini-3-pro-preview) model requires thought_signature similar to response_id for subsequent calls. Using the Agent and Runner the thought_signature is not inject for next loop during call. I have tried to use RunHooks to receive full response from the model but unsuccessfully.

File "C:\Users\dom\anaconda3\envs\genai\Lib\site-packages\openai\_base_client.py", line 1594, in request
    raise self._make_status_error_from_response(err.response) from None
openai.BadRequestError: Error code: 400 - [{'error': {'code': 400, 'message': 'Unable to submit request because function call `default_api:get_current_temperature` in the 2. content block is missing a `thought_signature`. Learn more: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/thought-signatures', 'status': 'INVALID_ARGUMENT'}}]

Debug information

• Agents SDK version: 0.6.1
• Python version: 3.11.13
• OpenAI version: 2.8.1

Repro steps

This snippet using plain OpenAI client returns the thouch_signarute but fails when use Agent and Runner

• OpenAI library:

# Create an async-compatible OpenAI client pointing to GOOGLE's endpoint
# Set up the OpenAI client pointing to GOOGLE endpoint
client = OpenAI(
    api_key=creds,
    base_url=f"https://aiplatform.googleapis.com/v1/projects/{PROJECT_ID}/locations/global/endpoints/openapi",
)

prompt = """
If it's too hot or too cold in London, set the temperature to a comfortable level.
Make your own reasonable assumption for what 'comfortable' means and do not ask for clarification.
"""

messages = [
    {"role": "user", "content": prompt},
]

response1 = client.chat.completions.create(
    model=MODEL_NAME,
    messages=messages,
    tools=tools,
)

print(response1)

OUTPUT:

ChatCompletion(id='3Kstacq6MerChcIPi9qngAE', choices=[Choice(finish_reason='tool_calls', index=0, logprobs=None, message=ChatCompletionMessage(content=None, refusal=None, role='assistant', annotations=None, audio=None, function_call=None, tool_calls=[ChatCompletionMessageFunctionToolCall(id='function-call-e37e486d-d21e-4203-a7d5-220fa5c27dac', function=Function(arguments='{"location":"London"}', name='get_current_temperature'), type='function', extra_content={'google': {'thought_signature': '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'}})]))], created=1764600796, model='google/gemini-3-pro-preview', object='chat.completion', service_tier=None, system_fingerprint='', usage=CompletionUsage(completion_tokens=11, prompt_tokens=70, total_tokens=400, completion_tokens_details=CompletionTokensDetails(accepted_prediction_tokens=None, audio_tokens=None, reasoning_tokens=319, rejected_prediction_tokens=None), prompt_tokens_details=None, extra_properties={'google': {'traffic_type': 'ON_DEMAND'}}))

• Using OpenAI Agents SDK:

# Create an async-compatible OpenAI client pointing to GOOGLE's endpoint
# Set up the OpenAI client pointing to GOOGLE endpoint
client = AsyncOpenAI(
    api_key=creds,
    base_url=f"https://aiplatform.googleapis.com/v1/projects/{PROJECT_ID}/locations/global/endpoints/openapi",
)

# Define the orchestrator agent that routes tasks
climatizer_agent = Agent[SharedContext](
    name="climatizer_agent",
    instructions=(
        "You are a smart home climatization agent. "
        "Your task is to monitor the temperature in a given city and adjust the thermostat to maintain comfort. "
        "Use the provided tools to read the current temperature and set a new thermostat value. "
        "If the temperature is too high or too low, adjust it to a reasonable 'comfortable' value, "
        "which you must assume without asking for clarification. "
        "Consider comfortable indoor temperature to be around 20–22°C (68–72°F) unless otherwise specified. "
        "Always perform actions using the tools and provide concise confirmation of the action taken."
    ),
    model=OpenAIChatCompletionsModel(model=MODEL_NAME, openai_client=client),
    model_settings=ModelSettings(parallel_tool_calls=False, tool_choice='required', temperature=0, top_p=1.0),
    tools=[function_tool(get_current_temperature), function_tool(set_thermostat_temperature),]
)

prompt = """
If it's too hot or too cold in London, set the temperature to a comfortable level.
Make your own reasonable assumption for what 'comfortable' means and do not ask for clarification.
"""

invoke_inputs = [{"role": "user", "content": prompt}]


async def main():
    while True:
        try:
            with trace("Climatization"):
                run_result = await Runner.run(climatizer_agent, input=invoke_inputs)
                print(run_result)
        except KeyboardInterrupt:
            break

OUTPUT:

File "C:\Users\dom\anaconda3\envs\genai\Lib\site-packages\openai\_base_client.py", line 1594, in request
    raise self._make_status_error_from_response(err.response) from None
openai.BadRequestError: Error code: 400 - [{'error': {'code': 400, 'message': 'Unable to submit request because function call `default_api:get_current_temperature` in the 2. content block is missing a `thought_signature`. Learn more: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/thought-signatures', 'status': 'INVALID_ARGUMENT'}}]

Expected behavior

A clear and concise description of what you expected to happen.

[ihower]

FYI, this issue is about Gemini 3 Pro support in OpenAIChatCompletionsModel. It is similar to the LiteLLM issue here:
#2109

[silvasw]

Not reflecting when using Agent.

[silvasw]

The problem is that if u are using orchestration for multi agents the initial call may not have tools calling so your implementation to fails to extract signature.

ChatCompletionMessage(content='Hello! How can I help you with your trading analysis or visualization today?', refusal=None, role='assistant', annotations=None, audio=None, function_call=None, tool_calls=None, extra_content={'google': {'thought_signature': 'CrYCAePx/177L7ysX5ant4Z3QWssMk9VJVCVOYEoYJX3bDkezrYUcV9fNkBCsNSzmzjHMNuinA9bYobyI+DEv7AdVtRKku1UQx5cf0H955AoTzSYkatK4pA4BQUYfhiOqGQ8QXdvkfs+y+2np+RFjH98X59qX6bm+u59KwkdNHISAX+mF2Ql31je+dt1D/69G3Oy/01V2tgTtcthYfpL9K94kFLTyafyLASxqfj4nKm0/eXMqwGCTf8KhCkVkTYMEodUTEo2YUL1iEiLOwFGYm9remHN/TmvH+w9lJvT5NRxNytQLFilsCXNvvQtECbhTddL3V+GlDsALt54k5iBLDuT0YYeW4QAYijyCKoxPtnblZeUmZauKcxTQQtf7S/+L1gpxoC2ZDy1cDXLMl5Rtp/loFjCRf4XCg=='}})