AGENT SKILL

create-skill

Add a repo-owned runtime-kit skill with source, manifests, product render surfaces, acceptance coverage, and governance validation.

License 许可
MIT
最近更新
2026-08-29
Stars
★ 0
Forks
0

01 前言 Front Matter

name
create-skill
description
Add a repo-owned runtime-kit skill with source, manifests, product render surfaces, acceptance coverage, and governance validation.

03 SKILL.md 正文

Create Skill

Contract

Prereqs:

  • Run from the agent-runtime-kit repository root.
  • The request is for a repo-owned managed skill, not a project-local overlay.
  • agent-runtime is installed from the released nils-cli package and available on PATH.
  • The caller has supplied or accepted a canonical skill ID in the form <domain>.<skill>.

Inputs:

  • Skill ID, domain, skill name, supported products, description, required_clis floors, parent intents, example user request, and admission rationale.
  • Optional skill-owned scripts/, bin/, fixtures/, or docs support folders when the workflow needs them.
  • Explicit approval for any new plugin domain.

Outputs:

  • core/skills/<domain>/<skill>/SKILL.md.tera.
  • manifests/skills.yaml v2 entry with complete invocation, honest exposure.profile: default, and concrete required_clis values or an explicit empty map for prose-only skills.
  • manifests/plugins.yaml containment update, and product plugin metadata only when a new domain was approved.
  • Rendered product output, golden snapshots, sandbox expected skill list updates, runtime-smoke matrix/case coverage, and reminder metadata when the skill should be agent-invoked.

Failure modes:

  • The skill ID is malformed or collides with an existing source or manifest entry.
  • A new domain was requested without explicit user approval.
  • required_clis contains placeholders or references an unreleased binary.
  • The capability is agent bookkeeping, a one-command wrapper, execution mode, or child lifecycle phase without a reviewed direct-user outcome.
  • The entry claims advanced, internal, or opt-in exposure that the three product installers cannot implement.
  • The workflow would leave a skill without render, sandbox, or runtime-smoke acceptance coverage.

Entrypoint

Use the repository shape directly:

skill_id="<domain>.<skill>"
domain="${skill_id%%.*}"
skill="${skill_id#*.}"

mkdir -p "core/skills/$domain/$skill"
$EDITOR "core/skills/$domain/$skill/SKILL.md.tera"
$EDITOR manifests/skills.yaml
$EDITOR manifests/plugins.yaml

After editing, run governance and render validation:

bash scripts/ci/skill-governance-audit.sh --update-counts
bash scripts/ci/skill-governance-audit.sh
agent-runtime render --product codex --update-golden
agent-runtime render --product claude --update-golden
agent-runtime render --product hermes --update-golden
bash scripts/ci/sandbox-install-rehearsal.sh
bash tests/runtime-smoke/run.sh --mode deterministic --domain "$domain"

Use the full gate before committing or opening a PR:

bash scripts/ci/all.sh

Workflow

  1. Resolve the canonical ID and reject project-local overlay requests.
  2. Inspect existing core/skills/<domain>/, manifests/skills.yaml, and manifests/plugins.yaml before editing.
  3. If the domain is new, stop and get explicit approval because it changes both product surfaces.
  4. Create SKILL.md.tera with the standard front matter, H1, Contract, Entrypoint, Workflow, and Boundary sections.
  5. Add the v2 skill manifest entry for every supported product. Name the parent intent, a direct example request, and why this is a user outcome. New skills must never enter migration.pending_disposition or the #562 ledger.
  6. Add plugin containment for the skill. For existing domains, do not touch unrelated plugin metadata.
  7. Add reminder metadata only when agents should invoke the skill as a workflow.
  8. Add runtime-smoke matrix coverage and any case-runner probe needed to prove the skill surface.
  9. Run bash scripts/ci/skill-governance-audit.sh --update-counts after the source, manifest, sandbox, and runtime-smoke surfaces are in place.
  10. Render both products, refresh golden snapshots, run governance audit, and run the relevant smoke checks.
  11. Leave staging, commit, PR creation, and issue state updates to the caller's delivery workflow.

Boundary

This skill owns scope judgment and the ordered lifecycle checklist for adding a repo-owned skill. It does not hide broad YAML rewrites inside prose. If repeated adds require structured mutation, dry-run/apply plans, or reference graph output, extract that behavior to released nils-cli first and then call it from this workflow.

04 同仓库更多 Skills