Self-Healing Multi-Agent Runtime Kernel

Reference implementation of a supervised runtime that executes agents in isolated sandboxes with:

• lifecycle state machine
• heartbeat monitoring
• transactional snapshots
• rollback/replay recovery
• deterministic event logs
• resource quotas
• policy-driven escalation
• external integration hooks/event bus

Operability

• Register external hooks with kernel.subscribe(hook).
• Unregister with kernel.unsubscribe(hook).
• Consume bus events from runtime/logs/event-bus.jsonl.

Runbook

Operational procedure for ESCALATED incidents:

• docs/RUNBOOK.md

See docs/ARCHITECTURE.md and tests/ for executable behavior.