hermes-overnight-autonomy

Use Kanban as the durable execution kernel: SQLite-backed cards preserve task state, claims, and heartbeats across restarts, while the gateway-embedded dispatcher handles scheduling and stale-claim reclamation. Add an OS-supervised watchdog for liveness alerting only after a staged pilot verifies the lifecycle.

Deploy an alert-only collector that reads board state, heartbeats, claims, and dispatcher dry-run output, emitting `[SILENT]` when healthy and deduplicated JSON alerts on new issues. The collector must never kill, restart, reassign, or modify any task.

Automatic recovery is a separate, explicitly approved stage enabled only after synthetic failure tests prove the alert path. It uses atomic expiring claims, bounded stale sampling, and at most one automatic resume before circuit-breaking and alerting.

URI mode=ro can fail on WAL databases when SQLite cannot establish its shared-memory sidecar, and immutable=1 can hide committed WAL content. Use a normal connection followed by PRAGMA query_only=ON for safe read-only watchdog queries.

Kanban handles durable task lifecycle and dead-worker redispatch but cannot prove meaningful progress, resume in-memory reasoning, judge code quality, or survive total gateway loss alone. A replacement worker receives card context and filesystem checkpoints, not the failed worker's chain of thought.