Skip to content

The agent engineering platform.

fix(langchain): need to inject all state for tools registered by middleware (#33087)

Sydney Runkle083bb3cdd7c7
diff --git a/libs/langchain_v1/langchain/agents/middleware_agent.py b/libs/langchain_v1/langchain/agents/middleware_agent.py
index 0f31993106..877c0e01f9 100644
--- a/libs/langchain_v1/langchain/agents/middleware_agent.py
+++ b/libs/langchain_v1/langchain/agents/middleware_agent.py
@@ -528,8 +528,8 @@ def _fetch_last_ai_and_tool_messages(
def _make_model_to_tools_edge(
first_node: str, structured_output_tools: dict[str, OutputToolBinding], tool_node: ToolNode
-) -> Callable[[AgentState], str | list[Send] | None]:
- def model_to_tools(state: AgentState) -> str | list[Send] | None:
+) -> Callable[[dict[str, Any]], str | list[Send] | None]:
+ def model_to_tools(state: dict[str, Any]) -> str | list[Send] | None:
if jump_to := state.get("jump_to"):
return _resolve_jump(jump_to, first_node)
@@ -548,8 +548,7 @@ def _make_model_to_tools_edge(
# of using Send w/ tool calls directly which allows more intuitive interrupt behavior
# largely internal so can be fixed later
pending_tool_calls = [
- tool_node.inject_tool_args(call, state, None) # type: ignore[arg-type]
- for call in pending_tool_calls
+ tool_node.inject_tool_args(call, state, None) for call in pending_tool_calls
]
return [Send("tools", [tool_call]) for tool_call in pending_tool_calls]
@@ -560,8 +559,8 @@ def _make_model_to_tools_edge(
def _make_tools_to_model_edge(
tool_node: ToolNode, next_node: str, structured_output_tools: dict[str, OutputToolBinding]
-) -> Callable[[AgentState], str | None]:
- def tools_to_model(state: AgentState) -> str | None:
+) -> Callable[[dict[str, Any]], str | None]:
+ def tools_to_model(state: dict[str, Any]) -> str | None:
last_ai_message, tool_messages = _fetch_last_ai_and_tool_messages(state["messages"])
if all(
@@ -598,7 +597,7 @@ def _add_middleware_edge(
"""
if jump_to:
- def jump_edge(state: AgentState) -> str:
+ def jump_edge(state: dict[str, Any]) -> str:
return _resolve_jump(state.get("jump_to"), model_destination) or default_destination
destinations = [default_destination]
diff --git a/libs/langchain_v1/tests/unit_tests/agents/test_middleware_agent.py b/libs/langchain_v1/tests/unit_tests/agents/test_middleware_agent.py
index 5953ba8d1a..996f3558dc 100644
--- a/libs/langchain_v1/tests/unit_tests/agents/test_middleware_agent.py
+++ b/libs/langchain_v1/tests/unit_tests/agents/test_middleware_agent.py
@@ -1,4 +1,3 @@
-from typing_extensions import TypedDict
import pytest
from typing import Any
from unittest.mock import patch
@@ -6,7 +5,7 @@ from unittest.mock import patch
from syrupy.assertion import SnapshotAssertion
from langgraph.runtime import Runtime
-from typing_extensions import Annotated, TypedDict
+from typing_extensions import Annotated
from pydantic import BaseModel, Field
from langchain_core.language_models import BaseChatModel
from langchain_core.language_models.chat_models import BaseChatModel
@@ -17,9 +16,10 @@ from langchain_core.messages import (
ToolCall,
ToolMessage,
)
-from langchain_core.tools import tool
+from langchain_core.tools import tool, InjectedToolCallId
from langchain.agents.middleware_agent import create_agent
+from langchain.agents.tool_node import InjectedState
from langchain.agents.middleware.human_in_the_loop import (
HumanInTheLoopMiddleware,
ActionRequest,
@@ -39,6 +39,7 @@ from langgraph.checkpoint.base import BaseCheckpointSaver
from langgraph.checkpoint.memory import InMemorySaver
from langgraph.constants import END
from langgraph.graph.message import REMOVE_ALL_MESSAGES
+from langgraph.types import Command
from langchain.agents.structured_output import ToolStrategy
from .messages import _AnyIdHumanMessage, _AnyIdToolMessage
@@ -1277,3 +1278,49 @@ def test_runtime_injected_into_middleware() -> None:
agent = create_agent(model=FakeToolCallingModel(), middleware=[CustomMiddleware()])
agent = agent.compile()
agent.invoke({"messages": [HumanMessage("Hello")]})
+
+
+def test_injected_state_in_middleware_agent() -> None:
+ """Test that custom state is properly injected into tools when using middleware."""
+
+ class TestState(AgentState):
+ test_state: str
+
+ @tool(description="Test the state")
+ def test_state(
+ state: Annotated[TestState, InjectedState], tool_call_id: Annotated[str, InjectedToolCallId]
+ ) -> str:
+ """Test tool that accesses injected state."""
+ assert "test_state" in state
+ return "success"
+
+ class TestMiddleware(AgentMiddleware):
+ state_schema = TestState
+
+ agent = create_agent(
+ model=FakeToolCallingModel(
+ tool_calls=[
+ [{"args": {}, "id": "test_call_1", "name": "test_state"}],
+ [],
+ ]
+ ),
+ tools=[test_state],
+ system_prompt="You are a helpful assistant.",
+ middleware=[TestMiddleware()],
+ ).compile()
+
+ result = agent.invoke(
+ {"test_state": "I love pizza", "messages": [HumanMessage("Call the test state tool")]}
+ )
+
+ messages = result["messages"]
+ assert len(messages) == 4 # Human message, AI message with tool call, tool message, AI message
+
+ # Find the tool message
+ tool_messages = [msg for msg in messages if isinstance(msg, ToolMessage)]
+ assert len(tool_messages) == 1
+
+ tool_message = tool_messages[0]
+ assert tool_message.name == "test_state"
+ assert "success" in tool_message.content
+ assert tool_message.tool_call_id == "test_call_1"