lsm / neokai / 24756060735

82%

push

github

fix(space): auto-resume idle node-agent session when a message is queued for it (#1583)

Previously, when a Task Agent called send_message targeting a node agent
whose session had completed/gone idle, the message was persisted in the
pending queue (PendingAgentMessageRepository) with status "pending
activation" but never delivered — it would sit in the queue until the
10-minute TTL expired.

Three changes close the gap:

1. spawnWorkflowNodeAgentForExecution now passes `agentName` and `nodeId`
   into `createSubSession`, enabling session reuse (each named agent lives
   in one session per task lifetime) and automatic queue flush when the
   session is created/reused.

2. A new public method `tryResumeNodeAgentSession(workflowRunId, agentName)`
   is added to TaskAgentManager. It looks up the most recent NodeExecution
   with an agentSessionId, then either flushes the queue directly (fast
   path when the session is in memory) or calls rehydrateSubSession (slow
   path when the session must be restored from DB). Both paths drain the
   pending queue into the live session.

3. send_message in task-agent-tools and deliverMessage in AgentMessageRouter
   now call tryResumeNodeAgentSession (best-effort, non-fatal) after
   enqueueing a non-deduped message, so the send → queue → resume → drain
   cycle completes in a single flow without requiring an external activation
   trigger. A matching onMessageQueued callback is wired into the
   AgentMessageRouter constructor from buildNodeAgentMcpServerForSession.

9071 of 12358 branches covered (73.4%)

Branch coverage included in aggregate %.

22 of 32 new or added lines in 3 files covered. (68.75%)

66866 of 75709 relevant lines covered (88.32%)

301.22 hits per line