runtime-verification-for-agents

What problem does it solve?

Many agent coordination daemons implement complex concurrency and recovery logic where subtle invariant violations (data loss, ordering errors, claim conflicts) can corrupt state or trigger cascading failures; this Skill provides a practical pathway to convert formal invariants into lightweight runtime monitors that continuously audit live sessions and trigger safe remediation when properties fail.

Core Features & Use Cases

• Invariant compilation: Patterns and guidance to compile safety (TLA+, Alloy) and bounded liveness properties into per-action or periodic checks.
• Checking strategies: Decision trees for choosing synchronous, sampled, or hybrid monitoring based on criticality and latency budgets.
• Remediation and safety: Safe remediation flows (alert, salvage, halt), double-check gates to avoid false positives, clock-skew handling with HLC/Lamport recommendations, and injection test guidance.
• Operational guidance: Metrics to track monitor overhead and health (heartbeat canaries, false-positive rates, crash-loop policies) and examples for wiring monitors into daemon hot paths or background sweeps.
• Use case examples: NoteMonotonicity as a synchronous monitor, CrashRecovery as a sampled/ bounded-liveness monitor, and Arbiter deployment patterns for autonomous monitoring agents.

Quick Start

Compile your TLA+ safety invariant into a (previous,current)->Violation monitor, wire cheap precondition checks into the hot path and schedule full sweeps in a background Arbiter agent to begin continuous auditing.