Fix bug where empty messages might be created in the Anthropic model

[oscar-broman]

This PR fixes an issue where empty messages are created.

The error happens when the function _map_message runs with this input:

[
    ModelRequest(parts=[SystemPromptPart(content='<my system prompt>', dynamic_ref=None, part_kind='system-prompt')], kind='request'),
    ModelResponse(parts=[TextPart(content='<model response>', part_kind='text')], model_name=None, timestamp=datetime.datetime(2025, 3, 2, 7, 10, 12, 560121, tzinfo=datetime.timezone.utc), kind='response'),
    ModelRequest(parts=[UserPromptPart(content="<user message>", timestamp=datetime.datetime(2025, 3, 2, 7, 10, 12, 566706, tzinfo=datetime.timezone.utc), part_kind='user-prompt')], kind='request'),
]

The following error is eventually returned from the Anthropic API:

{'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'messages.0: all messages must have non-empty content except for the optional final assistant message'}

[Kludex]

Can you provide an MRE, please?

[oscar-broman]

from anthropic import AsyncAnthropic
from openai import AsyncOpenAI
from pydantic_ai import Agent
from pydantic_ai.messages import ModelRequest, SystemPromptPart, TextPart, ModelResponse, UserPromptPart
from pydantic_ai.models.anthropic import AnthropicModel
from pydantic_ai.models.openai import OpenAIModel

async def test_anthropic_issue() -> None:
    openai_key = ""
    anthropic_key = ""

    models = [
        ("OpenAI", OpenAIModel(model_name="gpt-4o-mini", openai_client=AsyncOpenAI(api_key=openai_key))),
        ("Anthropic", AnthropicModel(model_name="claude-3-7-sonnet-latest", anthropic_client=AsyncAnthropic(api_key=anthropic_key))),
    ]


    for name, model in models:
        print(f"\n\nModel: {name}")

        agent = Agent(
            model=model,
        )

        messages = [
            ModelRequest(parts=[SystemPromptPart("Respond with the user's last message in upper-case")]),
            # ModelRequest(parts=[UserPromptPart(content="Hello")]), <-- Issue does not happen if uncommented
            ModelResponse(parts=[TextPart(content="HELLO")]),
        ]

        await agent.run(
            user_prompt="Test",
            message_history=messages,
            result_type=str,
        )

[alexmojaki]

Please convert the MRE into a test

[oscar-broman]

I've tried to do this without any other changes in the codebase -- it's quite hard. The function is an internal function not accessible without some indirections.

One way would be to create a live test, but this seems a bit extreme perhaps.

@alexmojaki What would you suggest I do?

[alexmojaki]

OK then I think it's fine unless @Kludex has some ideas.

[Kludex]

I'll create a test with VCR.

[mike-luabase]

@Kludex I'm hitting this error on most of our attempts to use Claude 3.7