fix(agents-server): runFinished child wakes delivered only for the last child — read shape delete row from value

[espaciofuturoio]

Symptom

A parent entity that spawns N children with wake: { on: 'runFinished' } is woken only by the last child spawned. The other children's run completions never reach the parent, and restarting the coordinator does not help. No warning is logged.

Reproduced on two 5-child swarms (electricax/agents-server:0.6.4, electricsql/electric:1.7.10): all 5 rows are present in wake_registrations, but only the last registration's source produces a wake event on the parent stream, before and after a coordinator restart. Sending a message to any of the first four children (so they complete a new run) never wakes the parent; the last child does.

It is easy to miss because a parent that runs the LLM on every wake stays 20 s in idle afterwards, and children finishing inside that window are consumed in-process without going through the coordinator. As soon as the parent closes quickly (sleep() or a state-only handler), fan-in silently breaks.

Root cause

WakeRegistry.applyShapeMessage handles shape delete messages by reading the deleted row id from message.old_value. Electric's shape stream with replica: full carries the deleted row in message.value — old_value is only set on updates. Measured against electricsql/electric:1.7.10:

{"value":{"id":"90","manifest_key":"child:…","source_url":"/survey_worker/…","subscriber_url":"/orchestrator/kit1"},
 "key":"\"public\".\"wake_registrations\"/\"90\"","headers":{"operation":"delete"}}

So Number(undefined) is NaN and the code takes its fail-closed branch, resetCachedRegistrations(), on every delete. syncManifestWakes does unregister + register per child manifest, so after N spawns the cache holds only the registration inserted after the last delete. The shape log replays those deletes on every restart, which is why a restart does not recover.

Fix

Read the id from value first and fall back to old_value; keep the full reset only when neither carries an id.

The existing sync test modelled the delete message without a row, which cannot distinguish "removed one registration" from "reset everything". The new test inserts two registrations, deletes one with the row in value, and asserts the other still evaluates.

Verified end to end with the image built from this branch: 5/5 children wake the parent (registrations 110–114 all deliver), and a child completing a run 28 s after the parent's session closed also wakes it.

Related

• Refactor wake registry sync and fix child wake delivery #4632 / Fix repeated wake delivery and desktop resume recovery #4726 replace this cache with a TanStack DB collection over the shape, which would remove the bug as a side effect; this is the minimal fix for the released 0.6.x line while that refactor is pending.
• fix(agents-server): dropped parent wake on parallel sub-agent spawn #4720 reports a dropped parent wake on parallel spawn — overlapping symptom, different mechanism; this fix may cover part of what it observes.

Changeset: @electric-ax/agents-server patch.