Fix: Resume graph on generic/custom interrupts instead of starting a new run

[Geoffrey Pruvost (geoffreyp)]

fix issue

#262

Problem

When a LangGraph backend uses interrupt() with a custom/generic payload (i.e. not the standard HumanInterrupt schema with action_request / config), the chat input always sends the user's response as a new run. This causes the graph to restart from START, re-triggering the interrupt() node in an infinite loop instead of resuming execution.

Only interrupts matching the isAgentInboxInterruptSchema format (with action_requests and review_configs) were properly handled via Command(resume=...) through the ThreadView / agent-inbox components. Any other interrupt format — like a simple {"type": "human_input_request", "message": "..."} — was displayed read-only via GenericInterruptView with no way to resume the graph from the chat input.

Solution

Modified handleSubmit in src/components/thread/index.tsx to detect when there is a pending interrupt that is not a standard HITL interrupt. In that case, the user's message is sent as Command(resume=input) instead of new input, allowing the graph to resume correctly.

Scenario	Before	After
Standard HITL interrupt (action_requests/review_configs)	✅ Handled by ThreadView	✅ No change
Generic/custom interrupt	❌ Chat input starts new run → infinite loop	✅ Chat input sends Command(resume=...)
No interrupt (normal chat)	✅ Normal submit	✅ No change

[vercel]

Geoffrey Pruvost (@geoffreyp) is attempting to deploy a commit to the LangChain Team on Vercel.

A member of the Team first needs to authorize it.