Python: [Bug]: _build_pydantic_model_from_json_schema does not handle anyOf — all optional/union MCP tool parameters collapse to str

[lukasz-cledar]

Description

What happened:
When using MCPStreamableHTTPTool (or any MCP tool integration), any MCP tool parameter whose JSON Schema uses anyOf — which is how Pydantic represents Optional[X], X | None, or any union type — is incorrectly resolved to str by _build_pydantic_model_from_json_schema().

What I expected to happen:
Parameters with anyOf schemas (e.g. list[str] | None, int | None) should be resolved to the correct Python union type

Steps to reproduce:

1. Create an MCP server with a tool that has an optional list parameter (e.g. columns: list[str] | None = None)
2. Connect to it using MCPStreamableHTTPTool
3. Have the LLM call the tool with that parameter
4. Observe that the LLM sends a string instead of an array, and the MCP server rejects it

Code Sample

# Reproducing the bug directly with the internal function
from agent_framework._tools import _build_pydantic_model_from_json_schema
import json

schema = {
    "properties": {
        "required_id": {"type": "string", "description": "Some required ID."},
        "start": {
            "anyOf": [{"type": "integer"}, {"type": "null"}],
            "default": None,
            "description": "Start index."
        },
        "columns": {
            "anyOf": [
                {"items": {"type": "string"}, "type": "array"},
                {"type": "null"}
            ],
            "default": None,
            "description": "List of columns."
        },
    },
    "required": ["required_id"],
}

Model = _build_pydantic_model_from_json_schema("my_tool", schema)
print(json.dumps(Model.model_json_schema(), indent=2))
# All anyOf fields become "type": "string" — start and columns are both str

# String is accepted (WRONG — should require list):
m = Model.model_validate({"required_id": "abc", "columns": "name"})
print(type(m.columns))  # <class 'str'>

# Actual list is REJECTED:
m = Model.model_validate({"required_id": "abc", "columns": ["name"]})
# ValidationError: Input should be a valid string

Error Messages / Stack Traces

Error: Argument parsing failed.

Package Versions

agent-framework: 1.0.0b260130

Python Version

Python: 3.13.2

Additional Context

No response

[eavanvalkenburg]

Please update to rc1 (and we will release rc2 later today) as we have this fixed there, because we pass the importschema directly through to the agents, instead of converting to a pydantic model