LangGraph

LangGraph owns graph checkpoints and paused-run state. ActionBox supplies the durable human request and typed response; your graph driver owns the side effect and its credentials.

Ownership boundary

ActionBox does not replace the LangGraph checkpointer and does not execute the operation for you. Use a durable checkpointer when a human response can outlive the worker that first surfaced the interrupt.

Supported flow

1. Run the graph with a stable thread_id and a checkpointer.
2. Call interrupt(...) inside the node with a bounded, JSON-serializable review payload. Do not create the ActionBox Action inside the node: LangGraph can re-execute the node when it resumes.
3. Let the graph driver observe result["__interrupt__"], then create one ActionBox Action per surfaced interrupt using a stable idempotency key.
4. Wait for the typed response. A boolean response is the envelope {"type":"boolean","value":true} or false.
5. Map the response to the value your node expects and resume the same checkpoint with Command(resume=...).
6. Run the side effect only after the resumed graph confirms approval. If you need an execution record, report its real result to ActionBox after it finishes.

The core resume helper looks like this after your driver has received a pending interrupt and resolved its ActionBox request:

from langgraph.types import Command

async def resume_interrupt(graph, pending, response, thread_id: str):
    approved = (
        isinstance(response, dict)
        and response.get("type") == "boolean"
        and response.get("value") is True
    )
    return await graph.ainvoke(
        Command(resume={pending.id: {"approved": approved}}),
        config={"configurable": {"thread_id": thread_id}},
    )

Install the public SDKs in the environment that runs the graph:

python -m pip install actionbox-sdk langgraph

Do not serialize framework-owned checkpoint state into an Action payload. Use a stable external ID, the interrupt ID, and bounded context blocks to correlate the Action with the graph run. Hash the stable thread_id and interrupt ID into the idempotency key so a retried driver recovers the same human request.

When multiple interrupts surface, resolve every item and pass an interrupt-ID map to Command(resume=...); do not assume the first item is the only one. Treat rejection, expiry, timeout, and a malformed response as approved=False.

For a complete walkthrough, see LangGraph human-in-the-loop approval.