GPT-5.1 responds with reasoning message as part of final response

[shankar-dh]

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

When using GPT-5.1 with reasoning_effort(low, medium or high) via OpenAI Agents SDK (Python), internal reasoning summaries leak into message content as msg_{uuid} instead of being isolated in separate ReasoningItem objects rs_{uuid}.

This causes reasoning traces to be indistinguishable from actual assistant responses

Debug information

• Agents SDK version: (0.5.1)
• Python version (Python 3.11)

Repro steps

from agents import Agent, ModelSettings, function_tool, Runner
from openai.types.shared import Reasoning

@function_tool()
async def get_weather(city: str, unit: str = "C"):
    mock = {
        "Austin": {"temp": 32, "condition": "Sunny"},
        "Seattle": {"temp": 18, "condition": "Rain"},
        "New York": {"temp": 25, "condition": "Cloudy"},
    }
    if city in mock:
        return {
            "city": city,
            "temp": mock[city]["temp"],
            "unit": unit,
            "condition": mock[city]["condition"],
        }
    return {"city": city, "error": "No mock data found"}

async def main():
    agent = Agent(
        name="Weather Agent",
        model="gpt-5.1",
        instructions="Always call the weather tool.",
        tools=[get_weather],
        model_settings=ModelSettings(
            tool_choice="required",
            reasoning=Reasoning(effort="low", summary = "auto"),
            parallel_tool_calls=True,
        ),
    )
    
    result = Runner.run_streamed(
        agent,
        input="What is the weather like in Austin?",
    )
    
    async for event in result.stream_events():
        if event.type != "run_item_stream_event":
            continue
        
        item = event.item
        if item.type == "message_output_item":
            msg = item.raw_item
            print(msg.id)
            if msg.content and hasattr(msg.content[0], "text"):
                response = msg.content[0].text
                return response

the model emits reasoning trace messages wrapped as message_output_item events and assigns them IDs with the msg_ prefix.

The response will look something like this

I should call the tool `get_weather` to fetch the weather status.

Message ID:

msg_05b63e7e027f96e700691d02a67b4c819183984d35170f1ddc

This makes reasoning items appear as normal assistant messages.

Note: This happens occasionally, not on all the turns. If you run this experiment 10 times it may appear once or twice

Expected Behavior

• Reasoning traces should be emitted as ReasoningItem objects with rs_ ID prefix (type: 'reasoning')
• Only the final assistant output should be emitted as a message_output_item with a msg_ ID

[seratch]

Thanks for writing in. I've checked the behavior with your code example (I adjusted some details to make it functional), but I see reasoning event items like RawResponsesStreamEvent(data=ResponseReasoningSummaryPartAddedEvent(item_id='rs_0a628200ac576ad600691d56893f9481968a8b886a89d7e47e', output_index=0, part=Part(text='', type='summary_text'), sequence_number=3, summary_index=0, type='response.reasoning_summary_part.added'), type='raw_response_event').

There may be some issues with your use case. If that's the case, it should be related to the model's behavior rather than this client-side SDK implementation. For those issues, please contact our technical support team to investigate with request IDs (we're unable to do the same in this public issue tracker). It'd be appreciated if you could understand this.

[ihower]

@shankar-dh I believe this is a feature, not a bug. The output you see comes from the preamble messages behavior in the GPT-5 series models. Please check the docs:

• https://cookbook.openai.com/examples/gpt-5/gpt-5_prompting_guide
• https://cookbook.openai.com/examples/gpt-5/gpt-5-1_prompting_guide

Before calling a tool, the model may send an assistant message to explain why it will call that tool. This message is meant to help the user understand what is happening, and is not part of the reasoning process.

In one turn, it can output:

1. reasoning items
2. preamble assistants message (this is what you're seeing)
3. tool calls
4. final assistant message (the last assistant message is the real final output)

You can adjust your prompt if you want to reduce or avoid this preamble message.