Tool Schemas And Invocation Diverge For Valid Pydantic And Python Signatures

The subset model creation for tool_call_schema uses a naive field extraction that drops Pydantic FieldInfo metadata. During invocation, the code calls model_dump() which applies serialization rules (e.g., Field(exclude=True)), stripping valid inputs. Parameter binding treats all validated fields as keyword arguments, breaking positional-only and *args signatures. Exception handling tests the exception instance's truthiness (`if exc:`) instead of checking for None, causing falsey exceptions to be ignored.

Run the provided Python snippet on langchain-core 1.5.6. It defines tools with aliased Pydantic fields, excluded fields, positional-only arguments, variadic *args, and a falsey exception. The observed output shows: alias schema lacking 'external' key, TypeError for excluded field, internal 'v__positional_only' leaking into schema, TypeError for positional-only as keyword, TypeError for variadic values, and None instead of raising FalseyError.

This fix may introduce compatibility issues with older Pydantic versions (below v2) or custom model configurations. Ensure that all installed packages are compatible with the updated internal API. In concurrent or multi-threaded environments, the new argument binding logic must be thoroughly tested with async and variadic functions to avoid race conditions. When rolling back, revert all related changes simultaneously; partial rollback could leave tools with mismatched schemas, causing model-facing errors.