lsm / neokai / 28134666630

82%

push

github

fix(space-runtime): guard activation flush against superseded worker sessions (#2177)

The pending external-event activation flush (flushPendingNodeQueue) drains
queued deliveries through the rate-limited digest path via a setTimeout(0)
macrotask. By the time the digest fires, the worker session captured at
spawn time may have been superseded — it crashed and was respawned (new
agentSessionId) or its execution was reset to pending (agentSessionId
cleared). resolveDigestDeliveryTarget's allowTargetSessionFallback then
fell back to the stale/dead captured sessionId, injecting the digest into
a superseded conversation.

Add resolveLiveDeliveryTarget, which re-resolves the authoritative current
session from the node execution record immediately before injection. The
digest target now retargets onto the current live session, and only falls
back to the captured activation session when it is still alive; otherwise
it returns a sessionless target so deliverDigestToSession requeues the
items as pending for the next activation rather than injecting into a
dead session.

The direct (non-digest) dispatch path has no async window between
resolveSubscriptionTarget (at flush time) and commandBus.dispatch, so it
is already protected and left untouched.

Tests:
- does not deliver activation-flushed digest to a superseded session
  (requeues instead) — fails against the unfixed code.
- retargets activation flush to the current session when the activation
  session is superseded (direct-path re-resolve invariant).
- update the overflow-cap test to mirror realistic activation (spawn sets
  agentSessionId before the flush), which the new guard requires.

11362 of 16393 branches covered (69.31%)

Branch coverage included in aggregate %.

21 of 21 new or added lines in 1 file covered. (100.0%)

95295 of 113549 relevant lines covered (83.92%)

300.4 hits per line