fix: make ContentToolResult.error JSON serializable

[cpsievert]

Summary

• Fix serialization of ContentToolResult.error field when calling .get_turns() followed by .model_dump_json()
• Add test covering tool call failure serialization

Problem

When a tool call fails and the error is stored in ContentToolResult.error, attempting to serialize the chat turns would fail because Python Exception objects are not JSON serializable by default.

Reprex

from chatlas import ChatOpenAI, UserTurn, AssistantTurn
from chatlas.types import ContentToolRequest, ContentToolResult

# Create a tool request and a failed result
request = ContentToolRequest(
    id="test-123",
    name="my_tool", 
    arguments={"x": 1},
)
result = ContentToolResult(
    value=None,
    error=Exception("Something went wrong"),  # <-- Exception not serializable
    request=request,
)

chat = ChatOpenAI(api_key="fake_key")
chat.set_turns([
    UserTurn("Call the tool"),
    AssistantTurn([request]),
    UserTurn([result]),
])

# This would fail with:
# pydantic_core._pydantic_core.PydanticSerializationError: 
#   Unable to serialize unknown type: <class 'Exception'>
turns = chat.get_turns()
turns_json = [x.model_dump_json() for x in turns]

Solution

Add a field_serializer and field_validator to ContentToolResult to handle the error field:

• serialize_error: converts Exception to string for JSON output
• validate_error: accepts string or Exception during deserialization, reconstructing a generic Exception from the string

This allows round-trip serialization/deserialization of turns containing failed tool results, though the specific exception type is not preserved (only the error message).

Test plan

• Added test_json_serialize_with_tool_failure test that verifies serialization and deserialization of turns with tool errors
• All existing tests pass (458 passed)
• Type checking passes

🤖 Generated with Claude Code