What problem does it solve? Sub-agents in multi-agent pipelines can silently contradict earlier architectural decisions, and without an independent check these contradictions propagate downstream undetected. This Skill provides a 5-step protocol for an external consistency-checker agent to detect such contradictions before they cause damage. ## Core Features & Use Cases - Anchor integrity validation: Verifies the SHA-256 checksum of the decision_anchor before any comparison, aborting with an [anchor-tampered-pre-check] error on mismatch. - Contradiction detection with severity grading: Compares each anchored decision against the output artifact and classifies findings as critical, major, or minor with a confidence score. - Deterministic verdict matrix: Computes pass/warn/fail verdicts from findings and confidence thresholds, with degraded mode when the token budget (default 5000) is exceeded. - Fail-closed escalation: Fail verdicts log an EP-013 consistency_decision event, write an ERROR to wiki/log.md, and escalate to a human gate without auto-rollback. - Use Case: After a dev-agent generates an implementation plan, the consistency-checker agent runs this protocol to confirm the plan does not contradict the locked stack and security decisions recorded in the decision anchor. ## Quick Start Ask the consistency-checker agent to run the consistency check protocol on the sub-agent output artifact against the current decision anchor and report the verdict with any findings.
