Unexpected Behavior: Interrupt() when invoked for the second time, failed to wait for the user input

[shtanko-michael]

Description
I am using interrupt function inside a subgraph, the first time it works as expected, it terminates the graph execution and I can type text in the console. But after I resume the graph and call interrupt for the second (or more) time, it returns the same previously cached value. So it does not work as expected.
When I remove the subgraph and impletemt the same logic inside parent graph, it works fine, so I think it is a bug, and also I found some discussions here on Github explaining this behavior and a way to fix it, hope this helps!

langchain-ai/langgraph#3072
langchain-ai/langgraph@main...vigneshmj1997:langgraph:interrupt_patch_for_subgraph

As far as I understand it fixed in 0.2.63 in Python Langgraph repo

Code

import { StateGraph, Annotation, START, MessagesAnnotation, Command, interrupt, MemorySaver, END } from "@langchain/langgraph";
import dotenv from "dotenv";
import { HumanMessage, AIMessage } from "@langchain/core/messages";
import readline from 'readline';
import { v4 } from 'uuid';

let rl = readline.createInterface({
    input: process.stdin,
    output: process.stdout
});

dotenv.config();

const SugGraphStateAnnotation = Annotation.Root({
    ...MessagesAnnotation.spec,
    counter: Annotation<number>,
    caller: Annotation<string | null>
});

const askHuman = async (state: typeof SugGraphStateAnnotation.State) => {
    const humanMessage = interrupt({
        'id': v4(),
        'request': state.messages[state.messages.length - 1].content,
    }) as string;

    console.log('Execution=', state.counter, 'Human message=', humanMessage);

    return new Command({
        goto: state.caller ?? END,
        update: {
            messages: [new HumanMessage({ content: humanMessage, name: 'askHuman' })],
            caller: null
        },
    });
}

const initial = async (state: typeof SugGraphStateAnnotation.State) => {
    return new Command({
        goto: 'askHuman',
        update: {
            caller: 'initial',
            counter: state.counter + 1,
            messages: [new AIMessage({ content: 'Question to ask user', name: 'initial' })]
        }
    });
};

const subGraphWorkflow = new StateGraph(SugGraphStateAnnotation)
    .addNode("askHuman", askHuman)
    .addNode("initial", initial, { ends: ['askHuman', END] })
    .addEdge(START, "initial");

const subGraph = subGraphWorkflow.compile();


const ParentStateAnnotation = MessagesAnnotation;
const router = (state: typeof ParentStateAnnotation.State) => {
    return new Command({
        goto: 'subGraph',
        update: {
            caller: 'router',
            messages: [new AIMessage({ content: 'Go to subgraph', name: 'router' })]
        }
    });
}

const parentGraphWorkflow = new StateGraph(ParentStateAnnotation)
    .addNode("router", router, { ends: ['subGraph'] })
    .addNode("subGraph", subGraph)
    .addEdge(START, "router");

const checkpointer = new MemorySaver();

const parentGraph = parentGraphWorkflow.compile({ checkpointer });

const config = { configurable: { thread_id: "console_app" } };

async function main() {
    try {
        await parentGraph.invoke({
            messages: [
                new HumanMessage("Inital message")
            ]
        }, config);

        const state = await parentGraph.getState(config);
        console.log((JSON.stringify(state.values, null, 2)));

        while (true) {
            const answer = await askQuestion();

            await resumeGraph(answer);
        }
    } catch (error) {
        console.error("An error occurred:", error);
    } finally {
        rl.close();
    }
}

async function askQuestion(): Promise<string> {
    return new Promise(function (resolve, reject) {
        rl.question("\User answer: ", async function (answer) {
            resolve(answer);
            rl.close();
        });
    });
}

async function resumeGraph(answer: string) {
    await parentGraph.invoke(new Command({ resume: answer }), config)
}

main();

Output

User answer: test answer
Execution= 1 Human message= test answer
Execution= 2 Human message= test answer
Execution= 3 Human message= test answer
Execution= 4 Human message= test answer
Execution= 5 Human message= test answer
Execution= 6 Human message= test answer
Execution= 7 Human message= test answer
Execution= 8 Human message= test answer
Execution= 9 Human message= test answer
Execution= 10 Human message= test answer
Execution= 11 Human message= test answer
Execution= 12 Human message= test answer
Execution= 13 Human message= test answer
Execution= 14 Human message= test answer

An error occurred: GraphRecursionError: Recursion limit of 25 reached without hitting a stop condition. You can increase the limit by setting the "recursionLimit" config key.

Troubleshooting URL: https://js.langchain.com/docs/troubleshooting/errors/GRAPH_RECURSION_LIMIT/

    at CompiledStateGraph._runLoop (file:///C:/projects/talkapp-ai/langchain/node_modules/.pnpm/@[email protected]_@[email protected][email protected][email protected]__/node_modules/@langchain/langgraph/dist/pregel/index.js:887:23)
    at process.processTicksAndRejections (node:internal/process/task_queues:105:5)
    at async createAndRunLoop (file:///C:/projects/talkapp-ai/langchain/node_modules/.pnpm/@[email protected]_@[email protected][email protected][email protected]__/node_modules/@langchain/langgraph/dist/pregel/index.js:980:17) {
  lc_error_code: 'GRAPH_RECURSION_LIMIT',
  pregelTaskId: 'd90e7254-46a7-582c-90d7-1fea82c89f52'
}