A production-grade starter for orchestrating several AI coding agents on one host — with resource governance, admission gates, lane supervision, and honest status reporting. Extracted from a live fleet that has been running continuously for months.
A published JSON snapshot of CPU, memory pressure, disk, load, and worker counts — with UNKNOWN semantics that fail closed.
Deterministic can-start decisions with hysteresis, cooldowns, and retry hints. Refuses work before the host is on fire.
Per-provider supervisor units that own queue depth, cooldown, and health — not a competing scheduler.
Projects, owners, worktrees, and health in one queryable store, with a read-only CLI surface.
Honest band reporting (HEALTHY / PRESSED / RECOVERING / CRITICAL) instead of invented green lights.
Bounded per-task scratch registration with retention policy and reap on completion.
UNKNOWN, never a fabricated sample.$49 one-time, self-hosted, MIT-licensed, free updates for a year.
You'll get the source, the registry schema, the admission policy, and a setup walkthrough. Runs on any Linux host with Python 3.10+.