fix: normalize tool call names in OpenAI streaming responses

[哈基米 (BlueX888)]

Why are these changes needed?

BaseOpenAIChatCompletionClient.create() passes tool-call names through normalize_name() (python/packages/autogen-ext/src/autogen_ext/models/openai/_openai_client.py:771), which replaces invalid characters with _ and truncates to 64 characters — its docstring notes that LLMs sometimes emit function names ignoring the API's format requirements. The streaming path, create_stream(), accumulated the name chunks raw (full_tool_calls[idx].name += tool_call_chunk.function.name, _openai_client.py:991) and built the final CreateResult from full_tool_calls without ever normalizing, so the identical API response produced a different FunctionCall.name depending on whether streaming was used (e.g. get weather! streamed vs get_weather_ non-streamed).

A raw name can violate the ^[a-zA-Z0-9_-]{1,64}$ pattern that function names must match, so replaying the streamed AssistantMessage in the next request sends the raw name back to the API, where the request can be rejected — while the identical non-streamed conversation works. The other sibling paths already normalize the LLM-provided name (OpenAI create(), Ollama create()/stream, Anthropic create()/stream); only OpenAI create_stream() omitted it.

This PR normalizes the assembled tool-call names in create_stream() once all name chunks have been merged, so streaming and non-streaming produce the same FunctionCall.name for the same API response.

Related issue number

Closes #8337

Changes

• python/packages/autogen-ext/src/autogen_ext/models/openai/_openai_client.py: in create_stream(), after content = list(full_tool_calls.values()), pass each tool call's name through normalize_name(), mirroring create() at line 771.
• python/packages/autogen-ext/tests/models/test_openai_model_client.py: added test_openai_chat_completion_client_create_stream_normalizes_tool_call_name, which mocks the same tool-call response (name="get weather!") for both create() and create_stream() and asserts the streamed CreateResult.content equals the non-streamed one and the normalized name get_weather_.

Testing

From python/, with the repo venv:

• New regression test on the unpatched tree fails with the reported divergence (name='get weather!' streamed vs 'get_weather_' non-streamed); with the fix:
  • pytest packages/autogen-ext/tests/models/test_openai_model_client.py::test_openai_chat_completion_client_create_stream_normalizes_tool_call_name → 1 passed
• Module's existing tests: pytest packages/autogen-ext/tests/models/test_openai_model_client.py → 53 passed, 33 skipped, 1 failed (test_reasoning_effort_parameter — pre-existing, fails identically on the base tree in this environment with ModuleNotFoundError: No module named 'azure', an optional dependency not installed locally; unrelated to this change)
• ruff format --check . (autogen-ext) → 198 files already formatted
• ruff check . (autogen-ext) → All checks passed!
• mypy --config-file ../../pyproject.toml --exclude src/autogen_ext/runtimes/grpc/protos --exclude tests/protos --ignore-missing-imports src tests (autogen-ext poe task) → 0 errors in the changed files (the 446 reported errors are pre-existing missing-optional-dependency noise, identical on the base tree)
• pyright on the changed files → 0 errors, identical on the base tree

Checks

• I've included any doc changes needed for https://microsoft.github.io/autogen/. See https://github.com/microsoft/autogen/blob/main/CONTRIBUTING.md to build and test documentation locally. (No doc changes needed — internal behavior fix, no public API change.)
• I've added tests (if relevant) corresponding to the changes introduced in this PR.
• I've made sure all auto checks have passed.

[哈基米 (BlueX888)]

@microsoft-github-policy-service agree