Skill Inventory

Skill Inventory — Unified Audit, Atlas, Mesh Health & Unification

DITEMPA BUKAN DIBERI — Intelligence is forged, not given. The mesa above the terrain. See the whole. Find the missing. Route the right.

What This Skill Is

This is the meta-skill — a skill about skills. It provides:

1. Inventory — live multi-surface counts (not a frozen number)
2. Gap detection — missing, thin, dual-named, harness-divergent
3. Routing — given a task and harness, which skill(s) to load
4. Health scoring — freshness, collision, catalog vs view drift
5. Cross-harness unification — AAA catalog ↔ CLI agent views
6. Mesh sync — per-skill SYNC / DIVERGED / MISSING_FROM_X status
7. Cross-surface contrast — 10-agent audit with rot classification

It does NOT execute, judge, or seal. It classifies, routes, and illuminates.

Iron rule: AAA is the catalog. Harnesses are views. Do not invent parallel catalogs.

When to Use

• Reviewing the overall capabilities, naming conventions, and performance of the active skill portfolio
• A new skill is drafted or proposed for installation
• An existing skill's SKILL.md is updated or modified
• Executing portfolio maintenance checks to clean up outdated APIs, dead links, or legacy documentation
• Tuning the triggering accuracy of skills when experiencing trigger drift
• Answering "are my agents in sync right now?"
• Auditing skill mesh, resolving dual names, rebinding harness skills
• Before claiming skill inventory is complete

When NOT to Use

• Creating a single new skill from scratch (use skill-creator)
• Linting individual skill trigger statements (use skill-creator linter mode)
• The task is a general system performance audit unrelated to skills

§0. MULTI-HARNESS UNIFICATION

Architecture

BOOTSTRAP_MANIFEST (9 universals)     ← always first
        ↓
AAA/skills  (catalog + V3 registry)   ← sole named truth
   +  .agents/skills (doctrine/stage) ← shared federation core
        ↓ symlink mesh
~/.grok | ~/.claude | ~/.codex        ← views (+ harness-native)
        ↓ separate trees
Hermes categories | Kimi roles | OpenClaw owned

Live inventory (OBSERVED)

Canonical artifacts

Resolve a V3 short name

python3 -c "import json;d=json.load(open('${AAA_HOME:-/root/AAA}/skills/SKILL_ALIAS_TABLE.json'));
print([a for a in d['aliases'] if a['v3_name']=='meta-atlas'][0])"

Mesh hygiene

# dry-run
bash ${AAA_HOME:-/root/AAA}/skills/scripts/skill-mesh-sync.sh
# apply missing links
bash ${AAA_HOME:-/root/AAA}/skills/scripts/skill-mesh-sync.sh --apply
# CI / pre-seal
bash ${AAA_HOME:-/root/AAA}/skills/scripts/skill-mesh-sync.sh --check

BOOT gate

BIND → GROUND → ROUTE → RECALL → VERIFY → SEAL → KNOW → READY

Only after READY may domain skills load.

§1. CROSS-SURFACE INVENTORY (10 Surfaces)

Cross-surface compare

# Dump each surface
comm -23 <(ls /root/AAA/skills/ | sort) <(ls /root/.kimi-code/skills/ | sort)   # orphans: AAA has, kimi lacks
comm -13 <(ls /root/AAA/skills/ | sort) <(ls /root/.kimi-code/skills/ | sort)   # drift: kimi has, AAA lacks

Classification Matrix

§2. MESH SYNC PROTOCOL

Compare command

find /root/AAA/skills /root/.hermes/skills /root/.kimi-code/skills -maxdepth 2 -name 'SKILL.md' | xargs grep '^version:' | sort

PITFALL: -maxdepth 1 returns empty on this mesh — skill dirs nest one level deep. Use -maxdepth 2.

Health report format

skill_name | AAA_version | hermes_version | kimi_version | status

Status values

Escalation protocol

§3. ROT CLASSIFICATION

Age thresholds

§4. ROUTING TABLE

§5. HEALTH SCORING

Per-skill health check

skill_health = {
    "name": str, "version": str, "forged_date": str,
    "days_since_forge": int, "references_count": int,
    "phantom_refs": int, "has_prompt": bool, "has_test": bool,
    "organ_coverage": float,
}

§6. ANTI-PATTERNS

§7. PRE-SEAL CHECKLIST

1. skill-mesh-sync.sh --check exits 0
2. V3 logical count is 64; alias rows separately classified
3. No live primary resolves through a tombstone row
4. No source-less broken alias remains active
5. Harness-native keepers remain real directories
6. This skill points to a dated live receipt

§8. INDEPENDENT VERIFICATION LANE (WAJIB 2)

A-FORGE planning, execution, AND verification in the same trust chain is a primary substrate defect. The required separation:

A-FORGE executes mutation
    ↓
Independent observe-lane verifier reads resulting reality
    ↓
Kernel checks evidence against original success criteria
    ↓
Only then may completion be recorded to VAULT999

Verifier hard rules

The verifier MUST:

• NOT have performed the mutation
• Use independently obtained state
• Receive original success criteria, NOT executor's rewritten summary
• Be unable to modify the state it is checking
• Return: VERIFIED, MISMATCH, INCONCLUSIVE, or STALE
• NEVER issue constitutional approval — only the kernel may SEAL

Kernel rejection rules

The kernel MUST reject "completion" when:

• Verifier identity == executor identity
• Evidence originated ONLY from the executor
• Verifier had mutation permission over the target
• Evidence is older than freshness_requirement
• Original success criteria are missing
• Results cannot be independently reproduced

§9. BOUNDED INDEPENDENT AUDIT PROTOCOL

Subagents are optional, scope-bounded evidence collectors. They do not inherit authority, cannot validate one another recursively, and never replace a live source-of-truth probe.

Rules

1. Route uncertainty to the owning evidence source first
2. Spawn only when scopes are independent and handoff cost < direct inspection
3. Do not ask one model to recursively agree with copies of itself
4. The root agent compares evidence, labels contradictions/UNKNOWNs
5. Irreversible SEAL still requires real human/external witness path

References

• contracts/AAA_SKILL.md — full orthogonal + subagent contract spec
• contracts/HERMES_ROLE.md — polymorphic runtime
• BOOTSTRAP_MANIFEST.json — signed manifest with 9 skills
• FEDERATED_SKILLS_REGISTRY_V3.yaml — 64 canonical skills
• SKILL_ALIAS_TABLE.json — 133 rows (104 active + 29 tombstone)