Trustcall do not strip Injected arguments from BaseTool

[Dimitri-Accad]

Hello

I noticed that the Injected arguments do not get stripped before trustcall verification causing verification issues.

For example using an InjectedState, InjectedToolCallId or InjectedToolArgs in the tool arguments will trigger a verification of these fields even if parsed as Optional.

Here is the example of the handoff_tool from langgraph documentation:

def create_handoff_tool(*, agent_name: str, description: str | None = None):
    name = f"transfer_to_{agent_name}"
    description = description or f"Transfer to {agent_name}"

    @tool(name, description=description)
    def handoff_tool(
        state: Optional[Annotated[AgentState, InjectedState]] = None,
        tool_call_id: Optional[Annotated[str, InjectedToolCallId]] = None,
    ) -> Command:
        tool_message = {
            "role": "tool",
            "content": f"Successfully transferred to {agent_name}",
            "name": name,
            "tool_call_id": tool_call_id,
        }
        return Command(
            goto=agent_name,
            update={"messages": state["messages"] + [tool_message]},
            graph=Command.PARENT,
        )

    return handoff_tool

The (truncated) error it produces:

"Error:\n\n\nExpected either a dictionary with a 'type' key or an object with a 'type' attribute. Instead got type <class 'dict'>.\n\nExpected Parameter Schema:\n\n```json\n{'$defs': {'AIMessage': {'additionalProperties': True, 'description': 'Message from an AI.....

In other tools, injecting the state is also problematic but this time it can be avoided by marking it as Optional:

@tool("SomeTool")
def some_tool(
    some_argument: Type,
    state: Optional[
        Annotated[AgentState, InjectedState]
    ] = None,  # Optional or else trustcall will try to generate it itself
) -> Dict[str, Any]:

I have noticed that in _base.py this snippet of code that seems responsible for stripping the injected tool arguments is not called for both my tools mentioned above:

def csff_(function: Callable) -> Type[BaseModel]:
    fn = _strip_injected(function)
    schema = create_schema_from_function(function.__name__, fn)
    schema.__name__ = function.__name__
    return schema

used there:

def ensure_tools(
    tools: Sequence[TOOL_T],
) -> List[Union[BaseTool, Type[BaseModel], Callable]]:
    results: list = []
    for t in tools:
        if isinstance(t, dict):
            print(f"DEBUG: dict: {t}")
            if all(k in t for k in ("name", "description", "parameters")):
                schema = create_model_from_schema(
                    {"title": t["name"], **t["parameters"]}
                )
                schema.__doc__ = (getattr(schema, __doc__, "") or "") + (
                    t.get("description") or ""
                )
                schema.__name__ = t["name"]
                results.append(schema)
            elif all(k in t for k in ("type", "function")):
                # Already in openai format
                resolved = ensure_tools([t["function"]])
                results.extend(resolved)
            else:
                model = create_model_from_schema(t)
                if not model.__doc__:
                    model.__doc__ = t.get("description") or model.__name__
                results.append(model)
        elif is_typeddict(t):
            results.append(_convert_any_typed_dicts_to_pydantic(cast(type, t)))
        elif isinstance(t, (BaseTool, type)):
            # All my tools end up here and are not stripped of their injected args    <-----------------------
            results.append(t)
        elif callable(t):
            # This section calls the stripping function   <-----------------------
            results.append(csff_(t))
        else:
            raise ValueError(f"Invalid tool type: {type(t)}")
    return list(results)

I use trustcall using the create_extractor method:

bound_decision_llm = create_extractor(
    llm,
    tools=[some_tool, transfer_to_call_model],
    tool_choice="any",
)

Do you have any insights on this behavior? Is it a bug or am I missing something in my implementation?

Best regards,
Dimitri

[open-swe]

🤖 Open SWE has been triggered for this issue. Processing...

View run in Open SWE here (this URL will only work for @hinthornw)

Dev Metadata

{ "runId": "1f0634d5-ab66-6fb2-a2d4-eb625c5fb903", "threadId": "5e43f876-cfe6-42fc-bff8-b6b6ab1d285f" }