Using Mistral Agents with tool and message_history leads to "Unexpected role 'user' after role 'tool'"

[TKaluza]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI

Description

When using the Mistral model with pydantic-ai agents, passing message history from a previous result to a new run causes an error. The API returns a 400 error with the message "Unexpected role 'user' after role 'tool'".

Example Code

from pydantic_ai.models.mistral import MistralModel
from pydantic_ai.agent import Agent
from pydantic import BaseModel, Field

class PoemInfo(BaseModel):
    title: str = Field(..., description="Create a fitting titel for your poem")
    creator: str = Field(..., description="the creator of the poem")
    text: str = Field(..., description="The complete poem")

model = MistralModel('mistral-small-latest', api_key="API_KEY_HERE")
agent = Agent(model=model, result_type=PoemInfo)

# This works fine
result0 = agent.run_sync("Make a Poem, about singing like a bird on a motorbike")

# This also works fine
result1 = agent.run_sync("repeat the last answer")

# This causes the error
result2 = agent.run_sync("repeat the last answer", message_history=result0.new_messages())

### LOG Error Message
mistralai.models.sdkerror.SDKError: API error occurred: Status 400
{"object":"error","message":"Unexpected role 'user' after role 'tool'","type":"invalid_request_error","param":null,"code":null}

Python, Pydantic AI & LLM client version

Python 3.12.3 on ubuntu 24.04 amd64
mistralai==1.6.0
pydantic-ai==0.0.43

[TKaluza]

I have tested it with OpenAIModel('gpt-4o-mini', ...) and AnthropicModel('claude-3-5-haiku-latest', ... both work as expected.

[TKaluza]

if someone needs a hotfix:

def new_mixtral_messages(result: AgentRunResult) -> list[ModelMessage]:
    """
    Converts messages to a format compatible with Mistral:
    - Converts ToolCallPart to TextPart containing the args as text
    - Skips ToolReturnPart entirely
    - Preserves message metadata and other part types
    """
    new_messages = []

    for msg in result.new_messages():
        new_parts = []
        for part in msg.parts:
            if isinstance(part, ToolCallPart):
                # Convert tool call to text, handling both dict and string args
                content = part.args_as_json_str() if isinstance(part.args, dict) else str(part.args)
                new_parts.append(TextPart(content=content))
            elif isinstance(part, ToolReturnPart):
                # Skip tool returns
                continue
            else:
                # Keep all other parts unchanged
                new_parts.append(part)

        # Create the appropriate message type
        if isinstance(msg, ModelResponse):
            new_msg = ModelResponse(
                parts=new_parts,
                model_name=msg.model_name,
                timestamp=msg.timestamp
            )
        elif isinstance(msg, ModelRequest):
            new_msg = ModelRequest(parts=new_parts)
        else:
            raise NotImplemented("There should not be another type of message!")
        new_messages.append(new_msg)
    return new_messages

[tamir-alltrue-ai]

@TKaluza How did you incorporate this method into your agent?