What problem does it solve? When running multi-stage AI development pipelines, problems like repeated rollbacks, thin stage documentation, and skipped intervention checks accumulate silently. This Skill audits in-flight or archived task folders and surfaces these anti-patterns with severity ratings so you know when to step in. ## Core Features & Use Cases - Anti-pattern detection: Detects five fixed-threshold anti-patterns (same-stage rollbacks, cross-stage rollback tallies, stage-doc thinness, missing intervention checks, missing archive calls) with deterministic INFO/WARN/ALERT severity. - Single-task and cross-task modes: Audit one task by slug, the last N archived tasks, or every archived task, with token-budgeted reads in cross-task mode. - Verdict emission: Writes a single SUPERVISION_REPORT.md ending in a HEALTHY, WATCH, or INTERVENE verdict that the verify_all passive guard can later observe. - Use Case: After a long autonomous pipeline run, invoke the supervisor on the archived task to check whether the AI looped on the same stage or skipped required human checkpoints before you trust the delivery. ## Quick Start Run /harness-supervise on my last 5 archived tasks and tell me which ones need intervention.
