/prune

Inventory, Retention and Overlap

Use before SCORE/CLASSIFY. Prune remains read-only/propose-only; SKILL.md owns approval, protected-skill and archive requirements.

Inventory

Resolve the actual host's installed skill root through _common/CLI_COMPATIBILITY.md; include project-local sources and distinguish intentional mirrors via _common/PROJECT_LOCAL_SKILLS.md. Scan the full roster for hidden dependencies even for TARGETED classification; cache within the session and re-scan for FOLLOWUP. An unavailable log/profile is unknown, never zero use/dependencies.

Read in order: roster; every frontmatter and CAPABILITIES_SUMMARY/COLLABORATION_PATTERNS/BIDIRECTIONAL_PARTNERS/PROJECT_AFFINITY; reference count/size and actual consumers; .agents/PROJECT.md activity; per-skill journal; project CLAUDE.md/AGENTS.md and _common/ dependencies; current _common/SKILL_PACKS.md; all actual host profiles; Nexus skill, signal keywords and routing consumers. Use 90 days for usage scoring but inspect the full ≥6-month horizon for sunset evidence. Validate empty/missing fields before scoring.

inventory:
  - skill: <name>
    frontmatter:
      description_chars: <int>
      description_has_when_clause: <bool>
    capabilities:
      count: <int>
      collaboration_partners_in: [<list>]
      collaboration_partners_out: [<list>]
    files:
      skill_md_lines: <int>
      references_count: <int>
      references_total_chars: <int>
    activity:
      project_md_entries_90d: <int>
      journal_last_modified: <YYYY-MM-DD or null>
      journal_entry_count: <int>
    dependencies:
      claude_md_mentions: <int>
      common_md_mentions: <int>
      pack_memberships: [<pack-name>]
      profile_coverage: [<profile-name>]
      nexus_routing_mentions: <int>

Scoring Axes

Each axis is 0–5; retain the observed inputs and maximum 25-point sum. Size/reference count are maintenance signals, not targets: optional references need no minimum count. Never create references to improve a score.

1. Usage (0-5)

Activity in .agents/PROJECT.md over the last 90 days.

2. Overlap (0-5, inverted)

Maximum overlap percentage against any other skill, from the overlap matrix.

3. Uniqueness (0-5)

Number of CAPABILITIES_SUMMARY entries that are NOT covered by any other single skill.

4. Coverage (0-5)

Number of PROJECT_AFFINITY rows with H or M relevance, plus Pack membership count.

5. Maintenance Cost (0-5, inverted)

SKILL.md size, reference count, journal freshness.

Ambiguous bands: the original Usage table overlaps at one entry, Coverage has overlapping Pack/domain bands, and Maintenance conditions may match more than one row. Expose the eligible score range; do not silently invent a tie policy. Seek a decision if that range changes the verdict. Missing evidence is not a zero score.

Pairwise Overlap

Normalize declared capabilities to (verb, object, qualifier), then use semantic equivalence checked against description/Recipes/Trigger Guidance. For every pair in the audit scope:

raw_overlap_pct = 100 × shared_capabilities / max(capabilities_A, capabilities_B)
boundary_adjusted = raw_overlap_pct − min(20, 5 × explicit_mutual_exclusions)
pack_adjusted = boundary_adjusted × 1.2 if the pair shares a Pack, otherwise boundary_adjusted

Keep raw and adjusted results. Empty capability sets require missing-data review, not division by zero. Explicit mutual-exclusion clauses are intentional boundaries; verify each counted exclusion. Use adjusted overlap in axis 2's bands. The written cross-Pack modifier subtracts 0.1 from axis 2 for cross-Pack overlap ≥30%; report it separately rather than silently reversing its sign. Adjusted overlap is a policy index and can exceed 100, not a literal shared-capability percentage.

Report only pairs with adjusted overlap ≥30% in the triangular matrix, with matching capabilities, boundary deductions, Pack adjustment and canonical-owner evidence. Do not use invented below-threshold matrix values as examples. ≥50% requires merge investigation; ≥70% requires absorb/sunset investigation, not permission to bypass retention/protection gates. Usage and Coverage select the owner through reference/merge-protocol.md.

For _common/reference near-duplicates, the SKILL's HD-ENTROPY test is a lossless merge trial with fewer net lines; heading/token similarity alone is not proof.

Registry Drag

For below-KEEP scores and new proposals, inspect selection ambiguity (HD-OPAQUE), unnecessary external/fan-out latency, and excess permission surface (HD-PERM). Use observed routing/rework/authorization evidence. Drag supports a MERGE argument, never a separate SUNSET path. Every addition must declare what would make it unnecessary.

Classification Thresholds

Sum the 5 axes (max 25):

3-Condition Sunset Gate

All three are mandatory, including DEPRECATE candidates: ≥6 months without evidenced activity; a clear capability-covering alternative; and no active project dependency (project CLAUDE.md/AGENTS.md, _common and active profiles). Missing evidence or a failed condition means DEPRECATE-WATCH, not approval. Scores cannot override this gate.

Apply the protected set from SKILL.md and current mandatory Pack membership; project-CLAUDE.md dependencies are protected too. Score for completeness but classify protected skills KEEP. Use reference/pack-impact.md before any removal proposal, then reference/merge-protocol.md or reference/sunset-protocol.md for delivery.