LangGraph human-in-the-loop (interrupt approvals) is implemented incorrectly

[n-sviridenko]

https://www.loom.com/share/f188635c9fea433c87115c04de52bb9a

How it's done now?

https://github.com/Yonom/assistant-ui/blob/main/packages/react-langgraph/src/useLangGraphRuntime.ts#L168-L176

How it should be done?

https://langchain-ai.github.io/langgraph/how-tos/human_in_the_loop/review-tool-calls/#example-with-no-review

(REST-level analogue: https://langchain-ai.github.io/langgraph/cloud/reference/api/api_ref.html#tag/thread-runs/POST/threads/{thread_id}/runs/stream (command param))

[n-sviridenko]

And one more thing, out of topic, but putting here to not forget.

When calling runtime.thread.subscribe() in combination with useThreadRuntime().getState(), it only has messages in the passed state, and I don't see a way to get the whole langgraph state values w/o manually calling langgraph APIs directly, which adds a redundant API call.

[n-sviridenko]

Also, somehow thread id is not stored in the thread state:

[Yonom]

Thank you for flagging these!

I believe the new interrupt API was added after the stockbroker example was created. I still need to add support for it :)

Three issues:

1. feat: support new interrupt API
2. feat: add a standard way of streaming and syncing state (outside of messages)
3. feat: pass threadId via ThreadListItem / thread.metadata state

For 2.: for the time being you can add additional streaming options when calling LangGraph APIs, then intercept langgraph's response stream (in onNew) to synchronize, say, a zustand store

[Yonom]

#1299

here's the proposed API for interrupt + Command flow

const sendCommand = useLangGraphRuntimeSendCommand();
sendCommand({
  resume: "your value"
});