deepagents 0.7.38 HITL resume: "Number of human decisions (1) does not match number of hanging tool calls (2)"
Symptom: with interrupt_on={"send": True}, a checkpointer and thread_id, the model emits 2 parallel send calls. The interrupt holds one action_request per call, each with allowed_decisions approve/edit/reject/respond. Resuming with Command(resume={"decisions":[{"type":"approve"}]}) raises ValueError: Number of human decisions (1) does not match number of hanging tool calls (2). Fix: return exactly one decision per action_request, in order. Other resume errors: Command(resume="yes") -> TypeError: string indices must be integers, not str (resume must be a dict with decisions); {type: bogus} -> ValueError: Unexpected human decision ... Expected one of [approve, edit, reject, respond]. An edit decision ({type: edit, edited_action: {name: send, args: {x: Z}}}) runs the edited call; the ToolMessage starts with "Note: a human reviewer replaced this tool call before it ran" and ends "Tool response: sent Z". Versions: deepagents 0.7.38, langgraph 1.2.22, Python 3.12; offline stub model.