skill-stocktake

# skill-stocktake — Skill Quality Audit Evaluate installed skills and assign each a verdict: `Keep / Improve / Update / Retire / Merge`, plus `Out of scope` for skills this harness does not own (Phase 0). This skill does NOT do the improving — once it has a verdict, it **hands off to skill-creator (the improvement engine)**. That boundary is the point: stocktake is the quality gate, skill-creator is the fixer. > Design note (v3.0). v1 batched ~20 skills per subagent with no cross-batch view — > overlap was structurally invisible. v2.0 swung to everything-in-one-context on the > 1M-context premise. A 2026-07-13 controlled comparison falsified half of that premise: > the single-context pass had returned 84/84 Keep, while fresh-context batches with > **unconditional** reference verification surfaced 12/73 non-Keep (half with > deterministic evidence — 404 links, deleted files, retired CLI flags), and a dedicated > overlap probe reproduced the set-level dimension *without* needing all bodies in one > context (0 genuine duplications across 17 clusters, all documented layering). > Two mechanisms explained the gap: (1) per-item attention dilutes as the context fills, > and (2) a **conditional** verification trigger ("confirm if it looks stale") degrades > to "never verify" in a loaded context, because the trigger itself is a diluting > judgment. v3.0 is therefore a hybrid, split not by context length but by the property > being checked (see when-code-when-llm): deterministic check...