buffer_streamed_tool_calls=True crashes when a Chat Completions provider omits the tool call index

[rioyu123]

Summary

The buffered tool-call path assumes every streamed tool_calls[] delta carries an index. The OpenAI SDK's lenient construct_type parsing leaves index=None when a provider omits it; the buffered replay then raises pydantic_core.ValidationError (ChoiceDeltaToolCall.index must be an int) and, when indexed and index-less calls coexist, TypeError: '<' not supported between instances of 'NoneType' and 'int'. The unbuffered path handles the same stream correctly, and the option's documented audience (src/agents/models/openai_provider.py:85-88) is exactly OpenAI-compatible providers whose chunk semantics are not reliable.

Affected code (main @ 89c02c8), src/agents/models/chatcmpl_stream_handler.py

• _accumulate_tool_call_delta (:309-316) keys buffered calls on tool_call_delta.index, which may be None.
• _buffered_tool_call_delta (:355-356): ChoiceDeltaToolCall(index=buffered_call.index, ...) raises ValidationError for None.
• _buffered_tool_calls_chunk (:379-383): sorted(buffered_calls.items()) raises TypeError with mixed None/int keys.

Reproduction

from openai._models import construct_type
from openai.types.chat.chat_completion_chunk import ChatCompletionChunk
from agents.models.chatcmpl_stream_handler import ChatCmplStreamHandler

def chunk(delta, finish_reason=None):
    return construct_type(
        type_=ChatCompletionChunk,
        value={"id": "c", "object": "chat.completion.chunk", "created": 1, "model": "m",
               "choices": [{"index": 0, "delta": delta, "finish_reason": finish_reason}]},
    )

chunks = [
    chunk({"role": "assistant", "tool_calls": [{"id": "call_1", "type": "function",
                                                "function": {"name": "f", "arguments": '{"a":'}}]}),
    chunk({"tool_calls": [{"function": {"arguments": "1}"}}]}),
    chunk({}, "tool_calls"),
]

async def gen():
    for c in chunks:
        yield c

events = [e async for e in ChatCmplStreamHandler.handle_stream(
    response, ChatCmplStreamHandler.buffer_tool_call_stream(gen()))]
# pydantic_core.ValidationError: 1 validation error for ChoiceDeltaToolCall ... index ... int_type

Observed: ValidationError, and TypeError when one delta has index: 0 and another has none.
Expected: the same response.completed output as the unbuffered path, [('function_call', 'call_1', '{"a":1}')], and for the mixed stream both calls in a deterministic order.

Root cause

The buffered replay hard-codes buffered_call.index and sorts on it without accounting for None.

[Hughhhhcoder]

I prepared a focused fix for this issue on my fork:

• Branch: codex/openai-agents-python-buffer-tool-index
• Commit: af8265fdc981
• The original missing-index reproduction now yields a valid index=0 tool call with the complete arguments.
• The fix also keeps indexed and index-less calls in deterministic order.

Validation on the current main baseline: mypy, Ruff, and format checks pass; the complete tests/models/test_openai_chatcompletions_stream.py file passes (110 tests). The branch is ready for an upstream PR once the contribution slot is available.

[Hughhhhcoder]

Update: the issue author has now opened #4824, which covers the missing-index path described here and links this issue. To avoid duplicate work, I will not open a second PR from my reference branch and will track #4824 instead. The branch remains available as an independent reproduction/reference if useful.