Adopt the Skill Mesh — a 7-tier (T0-T6) classification + composition-rule + self-evolution architecture for all 162 local skills + future plugin imports. Source-of-truth is the registry at /root/aj-wo

Decision

Adopt the Skill Mesh — a 7-tier (T0-T6) classification + composition-rule + self-evolution architecture for all 162 local skills + future plugin imports. Source-of-truth is the registry at /root/aj-workspace/state/skill-mesh/registry.yaml. Enforcement: setup-curator routing case 9b on every new skill + nightly audit cron at 03:30 IST. Anti-mega-skill threshold = 600 lines. New meta-skill skill-system-creator scaffolds future T1 orchestrators by composing existing skills.

Rationale

Prior state was a flat 162-skill catalog with no composition contract, no parent/child wiring, no anti-mega-skill enforcement, no orphan detection. Transcript-audited 2026-05-16: failure modes were (a) isolated skills with manual chaining, (b) mega-skills accreting end-to-end logic into one file (6 already exist: autoresearch 1393 lines, agents-architect 2011, mcp-server-dev 712, gws 686, snowflake 682, stitch 610), (c) no quality propagation when shared components improved. The mesh fixes all three: orchestrators wire children explicitly (no manual chaining), mega-skill detector flags >600-line SKILL.md files (decomposition path published via skill-system-creator), bidirectional parent/child registry on T3/T4 means quality/improvements propagate to every dependent orchestrator. Composition rules are mechanical (caller-tier → callee-tier matrix), enforced by audit script — cycles, tier violations, orphans, ghosts all caught. The mesh is additive (no skill lost, no protocol broken, no MCP touched) and integrates cleanly with existing setup-curator + skill-mirror-parity-check + claude-skills-sync infrastructure. Final audit: 0 ERROR, 9 informational WARN (all pre-existing). 162/162/162 parity verified.

Alternatives Rejected

Outcome

Pending