What problem does it solve?
AI agents increasingly propose consequential actions—deployments, purchases, credential operations, data mutations—and a plausible-looking approval screen is not proof that the action is safe. This Skill provides a rigorous, evidence-based review process that verifies whether an approval actually matches the exact proposed action before any human or enforcement point decides to proceed.
Core Features & Use Cases
- Six-Control Verification: Recomputes the assessment, matches the exact approved action, rejects replay and identity ambiguity, tests reviewer independence, preserves evidence and contradictions, and verifies lifecycle and monitoring freshness.
- Tamper Challenge Testing: Mentally mutates approvals (parameter swaps, nonce reuse, correlated reviewers, removed evidence, stopped heartbeats) to confirm the controls would actually catch forgery.
- Structured Verdict Output: Produces a standardized review report with a result of BLOCKED, INCONCLUSIVE, ELIGIBLE_WITH_CONTROLS, or ELIGIBLE_FOR_HUMAN_DECISION—while never executing or authorizing anything itself.
- Use Case: An agent requests approval to force-push to a production repository. Use this Skill to verify the approval digest matches the exact branch, commit, and scope, confirm the nonce is unused and unexpired, and check that monitoring telemetry is fresh before a human decides.
Quick Start
Ask the agent to review this proposed action and its approval packet using the verify-agent-action skill, and report whether it is blocked, inconclusive, or eligible for a human decision.
