200. Supervision repairs only what is idempotent, owned and verified, and never restarts blindly
Kontext
Owner 2026-10-17 asked for a watchdog that keeps the mesh up, drains tasks and kills zombies. serve ensure exits idle servers by design, claims expire on heartbeat, some processes hold leases, and a zombie is reaped only by its parent.
Rozhodnutí
Supervision is observe, classify, decide, recover, verify inside the existing runtime. Automatic repair is limited to idempotent safe actions; resuming executions or requeueing work is verified first; live sessions, name-based kills, taking valid leases, deleting worktrees, bypassing trust and retrying uncertain merges are never automatic. Each repair has a reason, one owner, an idempotent action, a verification and a circuit breaker.
Zamítnuté alternativy
- A new autonomous daemon that restarts everything
- Restart=always for every process
- Forcing nodes to cooperate
Důsledky
I3182 then I3183; extends ADR 0039's liveness rules from advice to a read capability.
Sources
Promoted on 2026-10-18 from the owner's ChatGPT project (workspace chatgpt-majordomus, issue I3100). The owner has not accepted it; it is proposed for the owner's decision.
chatgpt-majordomus:c288#16