Hermes Overnight Autonomy

Use this skill when the user asks for Hermes projects to continue across the night, gateway restarts, worker failures, or 24/7 operation. The target is durable progress with bounded risk—not a tower of supervisors.

Architecture decision

Use Kanban as the durable execution kernel:

• SQLite-backed cards preserve task bodies, dependencies, claims, heartbeats, run history, comments, retries, and circuit-breaker state.
• The gateway-embedded dispatcher is the normal scheduler and stale-claim reclaimer.
• Each unattended card has an explicit runtime cap, retry budget, named assignee, workspace, and acceptance contract.
• A separate OS-supervised watchdog provides liveness and alerting when the gateway or dispatcher is unavailable.
• Recovery remains disabled during calibration. Do not add a standalone Kanban daemon against the same board database or multiple competing relaunchers.

Kanban solves durable task lifecycle and dead-worker redispatch; it does not prove meaningful progress, resume in-memory reasoning, judge code quality, substitute models automatically, or survive total gateway loss by itself.

Rollout stages

Stage 0 — Read-only baseline