Checked other resources

• I added a very descriptive title to this issue.
• I used the GitHub search to find a similar bug and didn't find it.
• I searched the LangChain documentation.
• I am sure that this is a bug in LangChain rather than my code.
• The bug is not resolved by updating to the latest stable version of LangChain (or the specific integration package).

Example Code

from langchain.agents import create_agent
from langchain.agents.middleware import TodoListMiddleware

agent = create_agent(
    model=llm,
    tools=tools,
    system_prompt=prompt,
    middleware=[TodoListMiddleware()],
)

Error Message and Stack Trace (if applicable)

No error — the agent produces duplicate output.

Description

When using TodoListMiddleware, the model sometimes generates both text content (the final answer) and a write_todos tool call (marking all todos as completed) in the same response.

Because the response contains a tool call, the agent loop continues:

1. Model generates text answer + write_todos(all completed)
2. Agent executes write_todos tool → gets ToolMessage back
3. Agent loops back to model (because there was a tool call)
4. Model regenerates the answer (slightly different)

The user sees the answer twice (or more with streaming).

Expected behavior

When the model's only tool call is write_todos and all todos are marked completed, and the response also contains text content (the final answer), the agent should terminate instead of looping back to the model.

Workaround

The TodoListMiddleware.after_model could detect this case and strip the write_todos tool call from the AI message, updating the todos directly in state instead. Since the AI message would then have no tool calls, the agent loop would terminate normally.

def after_model(self, state, runtime):
    # ... existing parallel write_todos check ...

    last_ai_msg = next(
        (msg for msg in reversed(messages) if isinstance(msg, AIMessage)), None
    )
    if not last_ai_msg or not last_ai_msg.tool_calls:
        return None

    tool_calls = last_ai_msg.tool_calls
    if len(tool_calls) != 1 or tool_calls[0]["name"] != "write_todos":
        return None

    # Check if message has text content alongside the tool call
    has_text = _has_text_content(last_ai_msg)
    if not has_text:
        return None

    # Check if all todos are completed
    todos = tool_calls[0].get("args", {}).get("todos", [])
    if not todos or not all(t.get("status") == "completed" for t in todos):
        return None

    # Strip write_todos tool call, update todos directly
    clean_msg = AIMessage(
        content=last_ai_msg.content,
        response_metadata=last_ai_msg.response_metadata,
        id=last_ai_msg.id,
    )
    return {"messages": [clean_msg], "todos": todos}

This uses the same-ID replacement pattern that add_messages supports to replace the original AI message with one that has no tool calls.

System Info

• langchain 1.2.10
• langchain-core 1.2.13
• Python 3.12
• Model: Gemini 2.5 Pro (via langchain-google-genai)