test_create_agent_tool_validation.py — langchain Source File

Source Code

import sys
from typing import Annotated, Any

import pytest
from langchain_core.messages import HumanMessage
from langgraph.prebuilt import InjectedStore, ToolRuntime
from langgraph.store.base import BaseStore
from langgraph.store.memory import InMemoryStore

from langchain.agents import AgentState, create_agent
from langchain.tools import InjectedState
from langchain.tools import tool as dec_tool
from tests.unit_tests.agents.model import FakeToolCallingModel


@pytest.mark.skipif(
    sys.version_info >= (3, 14), reason="Pydantic model rebuild issue in Python 3.14"
)
def test_tool_invocation_error_excludes_injected_state() -> None:
    """Test that tool invocation errors only include LLM-controllable arguments.

    When a tool has InjectedState parameters and the LLM makes an incorrect
    invocation (e.g., missing required arguments), the error message should only
    contain the arguments from the tool call that the LLM controls. This ensures
    the LLM receives relevant context to correct its mistakes, without being
    distracted by system-injected parameters it has no control over.
    This test uses create_agent to ensure the behavior works in a full agent context.
    """

    # Define a custom state schema with injected data
    class TestState(AgentState[Any]):
        secret_data: str  # Example of state data not controlled by LLM

    @dec_tool
    def tool_with_injected_state(
        some_val: int,
        state: Annotated[TestState, InjectedState],
    ) -> str:
        """Tool that uses injected state."""
        return f"some_val: {some_val}"

    # Create a fake model that makes an incorrect tool call (missing 'some_val')
    # Then returns no tool calls on the second iteration to end the loop
    model = FakeToolCallingModel(
        tool_calls=[
            [
                {
                    "name": "tool_with_injected_state",
                    "args": {"wrong_arg": "value"},  # Missing required 'some_val'
                    "id": "call_1",
                }
            ],
            [],  # No tool calls on second iteration to end the loop
        ]
    )

    # Create an agent with the tool and custom state schema
    agent = create_agent(
        model=model,
        tools=[tool_with_injected_state],
// ... (319 more lines)