Fix ResponseTextDeltaEvent validation error for non-OpenAI model providers

[vmalyi]

The ChatCmplStreamHandler.handle_stream() method fails with a Pydantic validation error when creating ResponseTextDeltaEvent objects during streaming operations with non-OpenAI model providers (e.g., Gemini via Google's OpenAI-compatible API).

Error:

pydantic_core._pydantic_core.ValidationError: 1 validation error for ResponseTextDeltaEvent
logprobs
  Field required [type=missing, input_value={'content_index': 0, 'del...', 'sequence_number': 3}, input_type=dict]

Root Cause:
The ResponseTextDeltaEvent constructor in chatcmpl_stream_handler.py:193 is missing the required logprobs parameter. The OpenAI SDK type definition for ResponseTextDeltaEvent defines logprobs: List[Logprob] as a required field, but the Agents SDK doesn't provide it when creating these events.

Type Inconsistency:

• ResponseTextDeltaEvent.logprobs: List[Logprob] (required)
• ResponseOutputText.logprobs: Optional[List[Logprob]] = None (optional)

This inconsistency causes issues when using model providers that don't return logprobs data.

Debug information

• Agents SDK version: v0.2.3
• Python version: Python 3.10
• OpenAI SDK version: 1.97.1
• Pydantic version: 2.11.7

Repro steps

import asyncio
from agents import Agent, Runner
from agents.models.openai_chatcompletions import OpenAIChatCompletionsModel
from openai import AsyncOpenAI

async def reproduce_bug():
    # Using Gemini via OpenAI-compatible API (doesn't provide logprobs)
    gemini_client = AsyncOpenAI(
        base_url="https://generativelanguage.googleapis.com/v1beta/openai/",
        api_key="your-gemini-api-key",
    )
    
    agent = Agent(
        name="Test Agent",
        instructions="You are a helpful assistant.",
        model=OpenAIChatCompletionsModel(
            model="gemini-2.5-pro",
            openai_client=gemini_client,
        ),
    )
    
    # This will fail with the logprobs validation error
    stream_result = Runner.run_streamed(agent, "Hello, world!")
    
    async for event in stream_result.stream_events():
        if event.type == "raw_response_event":
            continue  # Error occurs here during internal stream processing

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

The error occurs in agents/models/chatcmpl_stream_handler.py at line 193 where ResponseTextDeltaEvent is constructed without the required logprobs field.

Expected behavior

Streaming should work seamlessly with all OpenAI-compatible model providers, including those that don't provide logprobs data (like Gemini). The ResponseTextDeltaEvent should be created with proper logprobs handling that:

1. Uses actual logprobs data when available (OpenAI models)
2. Falls back to empty list when not available (non-OpenAI providers)

Solution

The fix uses getattr() to gracefully handle the logprobs field, following the existing pattern used elsewhere in the codebase (like line 91 with chunk.usage):

Before:

yield ResponseTextDeltaEvent(
    content_index=state.text_content_index_and_output[0],
    delta=delta.content,
    item_id=FAKE_RESPONSES_ID,
    # logprobs field missing - causes validation error
    output_index=state.reasoning_content_index_and_output
    is not None,
    type="response.output_text.delta",
    sequence_number=sequence_number.get_and_increment(),
)

After:

yield ResponseTextDeltaEvent(
    content_index=state.text_content_index_and_output[0],
    delta=delta.content,
    item_id=FAKE_RESPONSES_ID,
    logprobs=getattr(delta, "logprobs", []) or [],
    output_index=state.reasoning_content_index_and_output
    is not None,
    type="response.output_text.delta",
    sequence_number=sequence_number.get_and_increment(),
)

Impact

This bug affects:

• All users of non-OpenAI model providers (e.g. Gemini)
• Any streaming operations with models that don't return logprobs
• All official SDK examples when used with non-OpenAI providers
• Integration with LiteLLM and other OpenAI-compatible proxy services

Testing

I've tested the fix with:

• ✅ Gemini 2.5 Pro via Google's OpenAI-compatible API
• ✅ Streaming operations with Runner.run_streamed()
• ✅ Both raw_response_event and run_item_stream_event handling
• ✅ Backward compatibility with OpenAI models (when logprobs are provided)

[vmalyi]

Most likely caused by #715

[parthushah8]

an easy fix for me was to downgrade to OpenAI SDK version: 1.97.0

[handrew]

Also getting this issue. Switching to openai==1.96.1 works but unideal - hope this gets merged soon. Hopefully a small change, 🙏🏼 @rm-openai