SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsskill-composition

skill-composition

General & other

View repository

Updated 10-15 22:12

About this skill

skill-composition

Objective

For one skill SKILL.md (catalog, .skill-proposals draft, or lifecycle fixture), extract invoke/forbid/require edges to catalog skill names and judge consistency. Return exactly one primary verdict: composition_ok, composition_fail with closed findings, or blocked. Ignore manner-only constraints. Do not edit files or build a full-repo graph.

Input

Exactly one of:

  • target — name listed in skills/INDEX.md with skills/<name>/SKILL.md
  • proposal_path — .skill-proposals/<name>/SKILL.md
  • fixture_path — skills/lifecycle-build/fixtures/**/SKILL.md

Empty, missing, or path outside those roots → blocked. If skills/INDEX.md is missing or unreadable → blocked.

Subject may be non-INDEX for proposal/fixture. Every edge target must resolve as INDEX ∩ skills/<name>/SKILL.md.

Extraction

Read only the subject SKILL.md. Extract only lowercase hyphenated catalog-shaped skill name tokens used as skill references. Ignore non-skill phrases.

Edge Include when text… Exclude
invokes tells the agent to call / run / invoke / use skill S as a step mere mentions in “do not use when” without a call instruction; parenthetical capability of another skill (“it may use S”); handoff stops (“stop for S”, “pending S”) without instructing this agent to run S
forbids must not call, run, invoke, absorb, or replace skill S manner-only constraints (e.g. “do not call bare grill-me”) — do not treat as forbids: grill-me
requires names skill S as a hard prerequisite skill this agent must have already completed or must run before proceeding soft “may” / “optional”; requiring an artifact/output shaped by another skill (“require skill-path-selector path”) without requiring that skill to be run here

Do not invent edges from WORKFLOW.md alone.

Consistency rules

Any break → composition_fail with closed findings (id, rule, quote, remedy hint):

  1. Every invokes / forbids / requires target exists in INDEX ∩ skills/<name>/SKILL.md
  2. No name appears in both invokes and forbids
  3. Subject must not invokes itself (same package name)
  4. Checker edits nothing

If all extracted edges pass → composition_ok (include the extracted edge lists as evidence).

Required deliverable

YAML with:

  • subject — the resolved SKILL.md path (for target: skills/<name>/SKILL.md; otherwise the supplied proposal_path or fixture_path)
  • verdict
  • extracted invokes / forbids / requires
  • findings on fail
  • missing on blocked (omit subject when no path was resolved)

When to use

  • Use when: a single skill package (catalog, draft proposal, or lifecycle fixture) needs an invoke/forbid/require consistency check against the INDEX catalog
  • Do not use when: Purpose review (skill-review); which-skill routing (skill-router); designing a new skill path; building a full-repo composition encyclopedia; editing skills to add dependency frontmatter

Invocation

  • skill-composition
  • composition check
  • check skill deps

Workflow

  1. Resolve input to a single subject SKILL.md path or blocked.
  2. Load catalog = INDEX-listed names that have skills/<name>/SKILL.md.
  3. Extract edges per the table (manner-only excluded from forbids).
  4. Apply consistency rules; emit composition_ok or composition_fail with subject set to the resolved path and F1.. findings on fail.
  5. Stop. Do not edit, ship, install, or claim purpose_pass.

Validation

  • Checks: Primary verdict is exactly one of composition_ok, composition_fail, blocked; no repository files edited by this skill
  • Evidence: Returned YAML including subject when a path was resolved; for fail, each finding cites a quote and rule number
  • Rules engine self-test: node skills/skill-composition/scripts/self-test.mjs covers consistency rules 1–3 on pre-extracted edge fixtures; rule 4 (checker edits nothing) is agent-enforced, not machine-checked; extraction from SKILL.md prose remains agent-judged

Boundaries

  • Do not edit SKILL.md, INDEX, README, WORKFLOW, or run npx skills install.
  • Do not invent edges from WORKFLOW alone.
  • Do not treat manner-only constraints as forbids.
  • Do not absorb skill-review or skill-router jobs.
  • Do not recommend or execute a skill chain.
  • Stop after one verdict.

Human review and stop

Human approval is required before: ship skill to skills/.

Stop after one primary verdict, or when blocked on missing input/catalog.