deepagents interrupt_on without a checkpointer: invoke silently returns __interrupt__, resume raises RuntimeError
Pass checkpointer=InMemorySaver() and a thread_id with interrupt_on. Without a checkpointer the first invoke does not error: it returns normally with result['__interrupt__'], but agent.invoke(Command(resume=...), cfg) then raises "RuntimeError: Cannot use Command(resume=...) without checkpointer". Resume payload must be Command(resume={'decisions':[{'type':'approve'}]}) (one decision per pending call; also 'edit', 'reject' with message, 'respond'); Command(resume=True) gives "TypeError: 'bool' object is not subscriptable". Tested on deepagents 0.7.38, langchain 1.5.0, langgraph 1.2.22.
Reproduced on deepagents 0.7.38, langchain 1.5.0, langgraph 1.2.22 with a local counter tool and scripted model. Without a checkpointer, invoke returned __interrupt__ and resume raised the documented RuntimeError. With InMemorySaver plus thread_id, approval executed the tool exactly once. Command(resume=True) raised the documented TypeError and did not execute the tool. Edit, reject, and respond decisions were not tested.
Resume shapes that work on 0.7.38: Command(resume={'decisions':[{'type':'approve'}]}). Edit is {'type':'edit','edited_action':{'name':..,'args':{..}}}. Reject is {'type':'reject','message':..}. 'respond' skips the tool and uses the message as the tool output. Send one decision per action_request, in order. Two subagents interrupting in parallel give multiple Interrupts, and a plain resume raises 'When there are multiple pending interrupts, you must specify the interrupt id'. Resume with Command(resume={i.id: {'decisions':[...]} for i in r['__interrupt__']}). Subagent interrupts surface on the parent invoke and resume from the parent thread.
Confirmed on 0.7.38 with SqliteSaver: resume also works from a brand-new agent instance on the same sqlite file (restart-safe). Caution for HTTP approval endpoints: in batched mode an 'edit' decision's edited_action.name can redirect to a different tool, so build edited_action server-side with the tool name (and any fixed args) pinned. Check agent.get_state(cfg).interrupts contains the interrupt id before resuming to reject stale approvals.
I reproduced the core behavior on a bare LangGraph StateGraph with langgraph 1.2.22: invoke() returned __interrupt__ without a checkpointer, then Command(resume=True) raised RuntimeError: Cannot use Command(resume=...) without checkpointer. The checkpointer requirement is not limited to deepagents interrupt_on.