Human‑in‑the‑loop (Interrupt / Resume) – LangGraph Demo

This repository contains a minimal example of a LangGraph that pauses execution to ask the user for confirmation via a custom interrupt.
The graph demonstrates how to:

Define a typed state with fields that will be filled by a human.
Trigger an interrupt from a node, passing a structured payload (type, question, allow_responds).
Handle the interrupt in a loop: display the question in the console (using questionary), capture the answer and resume the graph.

📦 Installation

# Create a virtual environment if you wish
python -m venv .venv
source .venv/bin/activate   # On Windows use `.venv\Scripts\activate`

# Install dependencies
pip install langgraph questionary

Tip

: If you plan to extend the graph with LLM calls, add langchain and an appropriate provider (e.g., langchain-openai) to your requirements.

📁 Project Structure

.
├── interrupt_graph.py   # The LangGraph implementation
└── README.md            # This file

🚀 Running the Example

The graph is executed in a simple loop that keeps running until the user chooses to stop.
Run it with:

python interrupt_graph.py

You should see something like:

[Graph] Starting new run...
[Interrupt] type: confirm
Question: Уверены, что хотите продолжить?
Options: approve / reject

Your choice: approve
[Graph] Resuming after interrupt...

[Graph] Run finished. Final state:
{'human_value': 'approve', 'foo': None}

If you answer reject, the graph will still resume but with a different value.

📄 Example Usage in Code

# interrupt_graph.py (excerpt)

from langgraph.types import interrupt

def node_with_interrupt(state: GraphState) -> GraphState:
    if state.get("human_value") is not None:
        return state

    payload = {
        "type": "confirm",
        "question": "Уверены, что хотите продолжить?",
        "allow_responds": ["approve", "reject"],
    }
    # Pause the graph until a human responds
    return interrupt(payload)

The main loop in interrupt_graph.py handles this payload:

while True:
    result = graph.run(state=state, config=config)
    if isinstance(result, Command):
        # Show question and get answer
        choice = questionary.select(
            result.payload["question"],
            choices=result.payload["allow_responds"]
        ).ask()
        state["human_value"] = choice
        continue  # resume graph
    else:
        break  # finished

🎯 What You Can Learn

How to pause a LangGraph node and wait for user input.
How to structure the interrupt payload so that it can be rendered in any UI (console, web, etc.).
The pattern of looping over graph.run() until no more Command objects are returned.

Feel free to fork this repository and experiment with different interrupt types or add LLM calls to make the graph more complex!

3.0 KiB Raw Permalink Blame History Unescape Escape