Using `event_stream_handler` causes `text` to appear before `thinking` for `gpt-oss` models on ollama/vllm

[daryllimyt]

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

Issue

When using gpt-oss family models through OpenAIProvider, we see text appear before thinking parts. This only happens when event_stream_handler is present and invoking Agent.run().

When running using Agent.run_sync() or Agent.run() without event_stream_handler (comment out), text correctly appears after thinking.

Example Code

#!/usr/bin/env python3
# /// script
# requires-python = ">=3.12"
# dependencies = [
#     "pydantic-ai-slim[openai,anthropic,bedrock]==1.2.0",
#     "python-dotenv==1.1.1",
# ]
# ///
import asyncio
from collections.abc import AsyncIterable

from dotenv import load_dotenv
from pydantic_ai import Agent, RunContext
from pydantic_ai.messages import AgentStreamEvent
from pydantic_ai.models.openai import OpenAIChatModel
from pydantic_ai.providers.openai import OpenAIProvider
from pydantic_core import to_json

load_dotenv()


model = OpenAIChatModel(
    model_name="gpt-oss:20b",
    provider=OpenAIProvider(base_url="http://localhost:11434/v1/", api_key="ollama"), # Running on ollama
)


async def noop(context: RunContext[None], events: AsyncIterable[AgentStreamEvent]):
    pass


async def main():
    agent = Agent(
        model=model,
        instructions="You are a helpful assistant.",
        event_stream_handler=noop,
    )

    result = await agent.run("Hi")

    print(to_json(result.new_messages(), indent=2).decode())


if __name__ == "__main__":
    asyncio.run(main())

Outputs

> uv run scripts/repro/async_gpt_oss_vllm_ollama.py
[
  {
    "parts": [
      {
        "content": "Hi",
        "timestamp": "2025-10-21T15:47:00.949069Z",
        "part_kind": "user-prompt"
      }
    ],
    "instructions": "You are a helpful assistant.",
    "kind": "request"
  },
  {
    "parts": [
      {
        "content": "Hello! How can I help you today?",
        "id": null,
        "part_kind": "text"
      },
      {
        "content": "User says \"Hi\". Need respond politely.",
        "id": "reasoning",
        "signature": null,
        "provider_name": "openai",
        "part_kind": "thinking"
      }
    ],
    "usage": {
      "input_tokens": 82,
      "cache_write_tokens": 0,
      "cache_read_tokens": 0,
      "output_tokens": 28,
      "input_audio_tokens": 0,
      "cache_audio_read_tokens": 0,
      "output_audio_tokens": 0,
      "details": {}
    },
    "model_name": "gpt-oss:20b",
    "timestamp": "2025-10-21T15:47:11Z",
    "kind": "response",
    "provider_name": "openai",
    "provider_details": {
      "finish_reason": "stop"
    },
    "provider_response_id": "chatcmpl-758",
    "finish_reason": "stop"
  }
]

Python, Pydantic AI & LLM client version

#!/usr/bin/env python3
# /// script
# requires-python = ">=3.12"
# dependencies = [
#     "pydantic-ai-slim[openai,anthropic,bedrock]==1.2.0",
#     "python-dotenv==1.1.1",
# ]
# ///

[dsfaccini]

You're using the "OpenAIProvider" but the docs show you should use the OllamaProvider

Can you try that and see if it fixes the issue?

[daryllimyt]

Using OllamaProvider emits thinking -> text correctly with Ollama, but when serving gpt-oss-120b on vllm (using the same code with different endpoint) I get the same sequence. What provider should I use for a vllm backend in this case?

[DouweM]

This looks like a bug in this code that incorrectly handles streamed text before thinking:

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

Lines 1619 to 1653 in c5b1495

	# Handle the text part of the response
	content = choice.delta.content
	if content is not None:
	maybe_event = self._parts_manager.handle_text_delta(
	vendor_part_id='content',
	content=content,
	thinking_tags=self._model_profile.thinking_tags,
	ignore_leading_whitespace=self._model_profile.ignore_streamed_leading_whitespace,
	)
	if maybe_event is not None: # pragma: no branch
	if isinstance(maybe_event, PartStartEvent) and isinstance(maybe_event.part, ThinkingPart):
	maybe_event.part.id = 'content'
	maybe_event.part.provider_name = self.provider_name
	yield maybe_event
	# The `reasoning_content` field is only present in DeepSeek models.
	# https://api-docs.deepseek.com/guides/reasoning_model
	if reasoning_content := getattr(choice.delta, 'reasoning_content', None):
	yield self._parts_manager.handle_thinking_delta(
	vendor_part_id='reasoning_content',
	id='reasoning_content',
	content=reasoning_content,
	provider_name=self.provider_name,
	)
	# The `reasoning` field is only present in gpt-oss via Ollama and OpenRouter.
	# - https://cookbook.openai.com/articles/gpt-oss/handle-raw-cot#chat-completions-api
	# - https://openrouter.ai/docs/use-cases/reasoning-tokens#basic-usage-with-reasoning-tokens
	if reasoning := getattr(choice.delta, 'reasoning', None): # pragma: no cover
	yield self._parts_manager.handle_thinking_delta(
	vendor_part_id='reasoning',
	id='reasoning',
	content=reasoning,
	provider_name=self.provider_name,
	)

Would you be up for filing a PR that switches that around?

Edit: Actually, the fix is different, working on it