skill-librarian

name	skill-librarian
description	Find, policy-check, and compose one or more specialized skills from the local agentic skill registry when a task needs domain guidance or spans multiple domains.

Skill Librarian

Use this skill when the right domain skill is unclear, specialized knowledge is useful, or a task may need several skills. Official Superpowers process skills take precedence over domain playbooks selected here.

Registry

Use the configured local registry:

REGISTRY_ROOT="${AGENTIC_SKILL_REGISTRY_ROOT:-$HOME/.agents/agentic-skill-registry}"

The registry CLI is the only discovery and loading runtime. Do not inspect or load catalog entries directly.

Workflow

1. Extract 2-5 keywords that describe the task, domain, output, and important constraints.

2. Search for at most ten candidates:

   skill-registry search --root "$REGISTRY_ROOT" --limit 10 --format json KEYWORDS...
   

3. If no useful candidate appears, retry exactly once with broader domain terms or synonyms. If the second search is also unhelpful, continue the task without a library skill.

4. Select 1-5 skills based on textual relevance. Mark each as primary or supporting.

5. Choose one composition:

   • single: one skill covers the task.
   • sequential: outputs or checks from one skill feed the next.
   • parallel: independent workstreams can use different skills.

6. Read every selected skill separately:

   skill-registry read --root "$REGISTRY_ROOT" --format json SKILL_ID_OR_LOAD_NAME
   

7. On exit code 3, show the user the candidate's risk, source, and reason, then request explicit confirmation. Only after confirmation, repeat that one read with --allow-unreviewed.

8. On exit code 1, discard the candidate. Never suggest or attempt a bypass.

9. Return a short composition plan to the main agent: selected skills, roles, ordering, and why each is needed.

Optional Librarian Subagent

For a simple or clearly matched request, perform the workflow directly. A Librarian subagent may be used only when the task spans multiple domains or several candidates are similarly relevant. The subagent may search and recommend; it must not execute the user's task, run bundled scripts, use secrets, or widen permissions.

Hard Rules

• Never select more than 5 skills.
• Never load the entire catalog or dump the whole discovery index.
• Do not execute bundled scripts automatically.
• Do not grant credentials or broad permissions to a selected skill.
• Active does not mean safe.
• Never bypass quarantine, path, symlink, or hash failures.
• Never treat Core or safe status as a substitute for textual relevance.
• Keep official Superpowers unmodified and use it only for process guidance.