Skill Library Cleaner Dedupe Cluster

Skill Library Cleaner / Dedupe & Cluster

Role

Group the messy pile into useful decision clusters.

Input:

machine/INVENTORY.json
machine/RISK_BLOCKLIST.json

Output:

reports/DEDUP_GROUPS.md
machine/DEDUP_GROUPS.json

Cluster Types

Detect:

• exact duplicate folder or file;
• same name in SKILL.md;
• same behavior with different folder names;
• runtime copies of one skill;
• old draft vs newer variant;
• broad overlapping skills;
• source dump copies;
• private/local variants.

Keep exact duplicates separate from overlaps. Do not recommend physical archive for an overlap cluster unless the evidence shows the members are safe replacements for each other.

Required Group Fields

Each group must have:

• group id;
• group label;
• member paths;
• likely canonical candidate;
• archive candidates;
• blocked members;
• reason;
• confidence;
• next action.

Decision Routes

Allowed routes:

• archive-exact-duplicate
• merge-near-duplicate
• keep-separate
• manual-dedupe-review
• split-active-vs-blocked
• reference-only
• backlog

Use manual-dedupe-review when names or categories overlap but responsibilities differ. Use archive-exact-duplicate only when the same behavior/content is duplicated with high confidence.

Confidence Rules

• high: same name/content/purpose and one clear canonical candidate.
• medium: strong overlap but possible role differences.
• low: weak keyword similarity only.

Low-confidence groups must not produce archive operations.

Pass Condition

The cleaner knows what should become one canonical item and what should leave the active shelf.