Repository navigation
|
Interrupts shipped in What the spec does not say is which of these applies when a run ends normally — no interrupt — but a client-declared tool call has no result yet. That is not an edge case: as noted in #200, "front end tool calls are already implicitly long running... so this should likely continue to be the default", and #1505 describes the same shape from the LangGraph side, where "fire-and-forget placeholders leave the graph 'done' while the UI is still collecting input". Three shapes are in the wild
Why the ambiguity has a costClients have landed on different readings, and both are defensible under the current wording:
So the same well-formed event stream works against one client and silently dead-ends against another, and neither client is clearly violating the spec — because the spec does not address the case. QuestionIs If the former, it would help to say so explicitly, so clients stop inferring completion of client obligations from it. If the latter, something like:
would settle it for every producer and consumer at once, and would match what this repo's own Happy to open a docs PR for whichever reading is intended. Related: #200 (long-running tool calls), #1505 (LangGraph HITL vs AG-UI tools), #827 (interrupts draft questions). |
Replies: 1 comment 2 replies
|
1.0 settled this, and it chose your first reading: The rules are in the frontend tools section:
{ "type": "RUN_FINISHED", "threadId": "t1", "runId": "r1",
"outcome": { "type": "success", "pendingToolCallIds": ["call-1"] } }The same rule is in lifecycle I ran your shape (one frontend call, no result) through What this means for your table:
That puts a client that treats an explicit The concepts/interrupts page you cited still lists only omitted, success and interrupt. The docs PR you offered would be useful there. |
1.0 settled this, and it chose your first reading:
successmeans "the agent run completed". Any client-side work still outstanding is tracked separately. The "SHOULD NOT report success" wording was not adopted.The rules are in the frontend tools section:
interrupt. Interrupts are for a producer that explicitly stops to ask; a frontend call is answered by a tool message in the next run's history.pendingToolCallIds. When it does, the list MUST contain exactly the calls the run started that got noTOOL_CALL_RESULT.