Pattern: Skill Catalogs Are Inverting — the Catalog Validates, the Coding Agent Chooses
AAS Core hit v15.6.0 on 2026-07-28 with 1,993 skills and 44,119 stars, and its architecture is a deliberate rejection of the recommendation-engine model: the catalog explicitly does not rank or recommend. Codex or Claude inspects the project, searches the complete local catalog over a read-only stdio MCP server, picks exact skill IDs, and a read-only `compose_stack` tool validates that selection in memory before the CLI persists it as `aas-stack.json` and emits an immutable plan for human review ahead of any target change. Apply and recovery remain experimental and outside the supported path. The generalizable shape for anyone building agent tooling: give the model complete catalog access plus a validator, not a ranked shortlist — then gate the mutation behind a reviewable plan artifact.
Source
↳ Follow the thread