fix(python): serialize datetime/UUID/Decimal/Enum/set/dataclass in plain-dict tool results

[rinceyuan]

Summary

Fix silent tool failures when a @define_tool handler returns a plain dict (non-Pydantic) containing datetime, UUID, Decimal, Enum, set, or dataclass values.

Problem

_normalize_result uses json.dumps with a default hook that only handles BaseModel. Any other non-JSON-native type raises TypeError, which wrapped_handler catches and reports to the model as:

Invoking this tool produced an error. Detailed information is not available.

This is easy to hit (returning a dict with a timestamp or UUID) and hard to diagnose (the error is intentionally redacted).

Fix

Add type-specific fallbacks in default():

Type	Serialization
datetime / date / time	.isoformat()
UUID / Decimal	str()
Enum	.value
set / frozenset	list()
dataclass	dataclasses.asdict()

Tests

New test test_dict_with_non_primitive_values_is_serialized covers all 9 types in a single plain-dict return. Existing test_pydantic_model_with_non_primitive_fields_is_serialized continues to pass (Pydantic path uses model_dump(mode=json)).

Partially addresses #2203.

[SteveSandersonMS]

Thanks for investigating this. I don’t think broadening Python’s serializer is the right direction.

The legitimate Pydantic case is already handled by model_dump(mode="json"). Plain dictionaries containing non-JSON-native values should fail rather than gain Python-specific implicit conversions.

Tool calls can fail for many reasons—handler exceptions, invalid arguments, permission failures, timeouts, and serialization errors. Applications already need to display or log these failures. The current behavior fits that model: the precise reason is available in the persisted and live tool.execution_complete.error.message, while the model-facing message remains intentionally redacted.

This PR instead turns an intentional failure into success and expands the API contract without a product requirement. I’m closing it; we can separately improve documentation if the standard failure-observation path is unclear.