MCP tools with a parameter named "config" silently broken by LangChain's RunnableConfig injection #532

Bug

When an MCP tool has an input parameter named config, convert_mcp_tool_to_langchain_tool produces a StructuredTool that silently replaces the user's config value with LangChain's internal RunnableConfig dict at invocation time.

Reproduction

from mcp.types import Tool as MCPTool
from langchain_mcp_adapters.tools import convert_mcp_tool_to_langchain_tool

tool = MCPTool(
    name="my_tool",
    description="A tool with a config param",
    inputSchema={
        "type": "object",
        "properties": {
            "config": {"type": "string", "description": "User config value"},
            "query": {"type": "string"},
        },
        "required": ["config", "query"],
    },
)

lc_tool = convert_mcp_tool_to_langchain_tool(session, tool)
# When invoked with {"config": "my_value", "query": "hello"},
# the coroutine receives config={"callbacks": ..., "tags": ..., ...} (a RunnableConfig)
# instead of "my_value"

Root cause

The execution path in BaseTool.arun (langchain-core):

1. _to_args_and_kwargs unpacks the tool input dict into tool_kwargs — so tool_kwargs["config"] = "my_value"
2. Then LangChain checks _get_runnable_config_param(func_to_check) against StructuredTool._arun, which has config: RunnableConfig in its signature
3. This returns "config", and the code does tool_kwargs["config"] = config — overwriting the user's value with the internal RunnableConfig

The same applies to run_manager and callbacks (via FILTERED_ARGS).

Version

• langchain-mcp-adapters: 0.2.2
• langchain-core: 0.3.x