Resume Multiple LangGraph Interrupts by ID
When parallel LangGraph branches are paused at interrupt(), one invoke can return multiple entries in __interrupt__. A scalar Command(resume=value) fails with RuntimeError because the pending interrupt is ambiguous. Build a mapping from each returned interrupt's id to its answer and pass Command(resume=answers); the graph then resumes each branch with its corresponding value. Reproduced with two parallel StateGraph nodes and SqliteSaver on LangGraph 1.2.22.
On LangGraph 1.2.22 with SqliteSaver, resuming only one interrupt ID in a two-branch parallel superstep completed that branch and left the other interrupt pending in get_state(config). A later invoke with Command(resume={remaining_id: answer}) completed the graph. This supports staged approvals without resending an answer for the branch that already finished.
On LangGraph 1.2.22 with SqliteSaver, after answering one of two parallel interrupt IDs, the checkpoint still carried both original __interrupt__ writes, but get_state({configurable: {thread_id}}).tasks contained only the unanswered branch. Reading pending_writes alone therefore reported a resolved approval as still waiting. For live task status, query the current thread snapshot and inspect its active tasks; a focused sample reproduction now reports both waits before resume and only the remaining wait afterward.
Follow-up reproduced three times on LangGraph 1.2.22 / checkpoint-sqlite 3.1.3. For a target thread with one active interrupt, resuming with another thread's real 32-character interrupt ID re-entered the target node and paused again without consuming that answer. A malformed key such as not-an-interrupt-id was instead treated as the resume value itself; the node completed with the whole mapping as its answer. Validate IDs against active get_state(config).tasks before resume: foreign IDs can still re-enter pre-interrupt code, and malformed keys can become user input.