Support reasoning from LM Studio OpenAI Responses API with gpt-oss

[umut-firat]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

I utilizing LM Studio with openai/gpt-oss-20b model to develop my applications. Before LM Studio got OpenAI Response API support, I was usingOpenAIChatModel which was just working fine. After the update and switching to it with new Pydantic AI v1.1.0, I'm getting following warning and ThinkingPart stopped working.

warnings.warn(
/Users/umutfirat/myproject/.venv/lib/python3.13/site-packages/pydantic_ai/models/openai.py:1904: UserWarning: Handling of this event type is not yet implemented. Please report on our GitHub: ResponseReasoningTextDeltaEvent(content_index=0, delta=' bur', item_id='rs_w2j8k5cvclmzxum5sgkhr', output_index=0, sequence_number=34, type='response.reasoning_text.delta')

I'm also sharing my function to understand what I'm doing wrong. My theory is maybe pydantic_ai.models.openai.OpenAIResponsesStreamedResponse missing Responses API parts? Thanks in advance.

Example Code

import json
from fastapi.responses import StreamingResponse
from pydantic_ai import (
    Agent,
    FinalResultEvent,
    PartDeltaEvent,
    PartStartEvent,
    TextPartDelta,
    ThinkingPartDelta,
)
from pydantic_ai.messages import TextPart, ThinkingPart
from pydantic_ai.models.openai import OpenAIResponsesModel, OpenAIResponsesModelSettings
from pydantic_ai.providers.openai import OpenAIProvider


async def create_ai_response(id: str, text: str):
    """
    Create an AI response using an agent with streaming output.
    """
    agent = Agent(
        model=OpenAIResponsesModel(
            "openai/gpt-oss-20b",
            provider=OpenAIProvider(base_url="http://localhost:1234/v1"),
        ),
        model_settings=OpenAIResponsesModelSettings(
            openai_reasoning_effort="medium", openai_reasoning_summary="detailed"
        ),
    )

    async def create_stream_output():
        """
        Stream the agent's response as Server-Sent Events (SSE).
        """
        async with agent.iter(text) as agent_run:
            async for node in agent_run:
                if Agent.is_user_prompt_node(node):
                    print(f"=== UserPromptNode: {node.user_prompt} ===")
                elif Agent.is_model_request_node(node):
                    async with node.stream(agent_run.ctx) as stream:
                        async for event in stream:
                            print(event)
                            if isinstance(event, PartStartEvent):
                                if isinstance(event.part, ThinkingPart):
                                    yield f"data: {json.dumps({'type': 'reasoning-start', 'id': id})}\n\n"
                                elif isinstance(event.part, TextPart):
                                    yield f"data: {json.dumps({'type': 'text-start', 'id': id})}\n\n"
                            elif isinstance(event, PartDeltaEvent):
                                if isinstance(event.delta, ThinkingPartDelta):
                                    yield f"data: {json.dumps({'type': 'reasoning-delta', 'id': id, 'delta': event.delta.content_delta})}\n\n"
                                elif isinstance(event.delta, TextPartDelta):
                                    yield f"data: {json.dumps({'type': 'text-delta', 'id': id, 'delta': event.delta.content_delta})}\n\n"
                            elif isinstance(event, FinalResultEvent):
                                print(
                                    "[Result] The model started producing a final result"
                                )

                elif Agent.is_call_tools_node(node):
                    ...
                elif Agent.is_end_node(node):
                    print(f"=== Final Agent Output: {agent_run.result.output} ===")
                    yield "data: [DONE]\n\n"

    return StreamingResponse(create_stream_output(), media_type="text/event-stream")

Python, Pydantic AI & LLM client version

My `pyproject.toml`:

.toml
[project]
name = "myproject"
version = "0.1.0"
description = "My Project"
readme = "README.md"
requires-python = ">=3.13"
dependencies = [
    "fastapi[all]>=0.119.0",
    "langchain-text-splitters>=0.3.11",
    "lmstudio>=1.5.0",
    "psycopg[binary]>=3.2.10",
    "pydantic-ai>=1.1.0",
    "pypdf>=6.1.1",
    "qdrant-client>=1.15.1",
    "sqlalchemy>=2.0.44",
]

[DouweM]

@umut-firat Thanks for the report. This event is not currently supported because the official OpenAI Responses API never emits it, but now that alternative implementations are starting to use it we should support it as well. The same goes for the ResponseReasoningItem.content field that we're not currently handling here:

pydantic-ai/pydantic_ai_slim/pydantic_ai/models/openai.py

Lines 980 to 1003 in c5b1495

	if isinstance(item, responses.ResponseReasoningItem):
	signature = item.encrypted_content
	if item.summary:
	for summary in item.summary:
	# We use the same id for all summaries so that we can merge them on the round trip.
	items.append(
	ThinkingPart(
	content=summary.text,
	id=item.id,
	signature=signature,
	provider_name=self.system if signature else None,
	)
	)
	# We only need to store the signature once.
	signature = None
	elif signature:
	items.append(
	ThinkingPart(
	content='',
	id=item.id,
	signature=signature,
	provider_name=self.system,
	)
	)

I've put it in the November milestone but hope to get to it earlier.