The enterprise plan locks compute to single-instance because locking.py uses fcntl advisory file locks and the job queue is local JSON (one .agent-brain/ state dir). Replace with Postgres advisory locks (pg_advisory_lock/leader-election) and a Postgres queue table (SELECT ... FOR UPDATE SKIP LOCKED) so multiple workers coordinate safely. Keep the file-lock path for local/dev. This is the prerequisite for any autoscaling / HA deployment.
Follow-up from design doc docs/plans/2026-06-09-enterprise-hardening-and-cloud-deployment.md (enterprise hardening, GCP-first).
Local TODO: .planning/todos/pending/2026-06-09-enterprise-horizontal-scaling-lock-and-queue-into-postgres.md
The enterprise plan locks compute to single-instance because
locking.pyuses fcntl advisory file locks and the job queue is local JSON (one .agent-brain/ state dir). Replace with Postgres advisory locks (pg_advisory_lock/leader-election) and a Postgres queue table (SELECT ... FOR UPDATE SKIP LOCKED) so multiple workers coordinate safely. Keep the file-lock path for local/dev. This is the prerequisite for any autoscaling / HA deployment.Follow-up from design doc
docs/plans/2026-06-09-enterprise-hardening-and-cloud-deployment.md(enterprise hardening, GCP-first).Local TODO:
.planning/todos/pending/2026-06-09-enterprise-horizontal-scaling-lock-and-queue-into-postgres.md