But WHY?

@Dave Aitel i'm revisiting this patch for BaseTools -- I think this is a langchain issue with how the subclass initialization is created. If you annotate the

args_schema

as

args_schema: type[BaseModel] = CommandRunnerInput

instead of leaving it unannotated I think it will work. This is how LangChain's docs show BaseTool use as well: https://python.langchain.com/v0.2/docs/how_to/custom_tools/#subclass-basetool. LangChain appears to strip the unannotated args_schema. I'm not sure why it's working for you in LC, though they seem to dive into the raw Pydantic annotations at one point during subclass initialization so maybe they grab it there. I think that's why you setting it forcefully in your init worked as well. This check runs without error for me, asserting the args schema is not present unless annotated in pure langchain and pydantic v1:

from pydantic.v1 import BaseModel
from langchain_core.tools import BaseTool


class LCBaseToolInput(BaseModel):
    x: int


class LCBaseTool(BaseTool):
    name = "TestTool"
    description = "A test tool"
    args_schema = LCBaseToolInput

    def _run(self, x: int) -> str:
        return str(x)


class LCBaseToolWithAnnotation(LCBaseTool):

    # Note the annotation here
    args_schema: type[BaseModel] = LCBaseToolInput


assert not LCBaseTool().args_schema
assert LCBaseToolWithAnnotation().args_schema

Ok fair I definitely could see me doing it wrong ! (Or following older docs from LC)

Yeah entirely possible something changed there, just wanted to make you aware in case it unblocked you!

I unblocked at the moment btw - but my control flow code does have a bunch of weird patches from my work :)