Anthropic streaming returns in single chunk when agent result_type is specified

[seunggs]

I just tried Anthropic streaming (pydantic-ai v0.0.20) and the stream comes in as one big chunk, not as an actual stream (although there is no longer an error about anthropic streaming not being supported). I can confirm the code is correct since a simple switch to "openai:gpt-4o" model creates an actual stream.

...
async with agent.run_stream(
    user_prompt=prompt,
    deps=agent_deps,
    message_history=message_history
) as result:
    # Stream the model's response
    async for chunk in result.stream():
        print("\nStreaming chunk:", chunk)
        yield f"event: document_messages\ndata: {json.dumps(chunk)}\n\n"
...

I discovered that this only happens if I specify the result_type (streaming works fine without):

class DocumentModelResponse(TypedDict, total=False):
    model_response: str
    document_md: str

agent = Agent(
    agent_data["model"],
    deps_type=AgentDeps,
    result_type=DocumentModelResponse,
    tools=tools,
    system_prompt=system_prompt
)

[samuelcolvin]

My guess is this will be fixed by #582.

[seunggs]

@samuelcolvin thanks for the response!

Hmm, I don't have a deep understanding of that discussion but isn't that ticket related to structured outputs? Claude works fine with result_type for non-streaming use cases so why would that ticket resolve the streaming issue?

[sydney-runkle]

I can confirm that I also experienced this issue with anthropic.

[seunggs]

Thanks for confirming @sydney-runkle - let me know if you have an ETA on this. Would really love to get streaming work for my app. Thanks!