fix(core): support tool parameters named like BaseModel members

[abhay-codes07]

Summary

A function tool whose parameter name shadows a Pydantic BaseModel member crashes at tool-definition time, deep inside function_schema's create_model call, with an error that gives no hint about the real cause:

from agents import function_tool

@function_tool
def export_row(model_dump: bool) -> str:   # ValueError: Field 'model_dump' conflicts with member ...
    ...

@function_tool
def configure(model_config: str) -> str:    # TypeError: 'FieldInfo' object is not iterable
    ...

model_dump, model_validate, model_dump_json, model_json_schema and the other model_* methods raise a ValueError from Pydantic's protected-namespace guard; model_config raises a TypeError because create_model treats a model_config keyword as configuration.

This completes #4627, which taught the runtime call layer to read argument values from instance __dict__ (rather than by attribute access) precisely so parameters can shadow BaseModel members like model_extra / model_fields_set — but those names still had to be definable first, and the model_* methods crash before that support is ever reached.

Change: build the generated args model on a small base whose model_config sets protected_namespaces=(), so method-shadowing parameter names no longer trip Pydantic's guard. Because the call layer already resolves values via __dict__, a field that shadows a model member is still passed to the tool correctly. model_config itself cannot be a Pydantic field name, so it is rejected with an actionable UserError ("Rename the parameter…") instead of an opaque TypeError.

Test plan

• test_parameter_named_like_a_basemodel_method_is_supported (parametrized over model_dump, model_dump_json, model_validate, model_validate_json, model_json_schema): the tool is defined, validates input, and round-trips through to_call_args to the original function.
• test_parameter_named_model_config_raises_actionable_error: model_config raises a UserError mentioning the parameter, not a TypeError.
• 5 of the 6 new cases fail on main without the source change.
• The existing fix(core): read tool args without Pydantic property shadowing #4627 model_extra / model_fields_set tests still pass unchanged.
• uv run pytest tests/test_function_schema.py tests/test_function_tool_decorator.py tests/test_tool_converter.py tests/test_released_api_contract.py → 231 passed. ruff format --check, ruff check, and mypy on the changed file pass.

Issue number

Follows up #4627 (closed #3547 / #3549 covered the same crash; the runtime shadowing was fixed there, the definition-time crash remained).

Checks

• I've added new tests, if relevant
• I've run make lint, make format and make tests (ran targeted lint/format/typecheck and the tool suites locally)
• I've confirmed all verification steps pass
• If using Codex, I've run /review before submitting this PR

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[seratch]

Thanks for investigating these Pydantic name collisions. The previously closed reports asked for an actionable failure, not for every BaseModel member name to become a supported tool parameter.

Disabling protected_namespaces for every generated argument model creates a broad new contract and leaves model instances with fields shadowing methods such as model_dump and model_validate. Renaming the parameter or using a wrapper function remains a reasonable supported alternative.

The narrower silent-value corruption for already constructible models was fixed separately in #4627. I am closing this broader support change.