Skillroster

SkillRoster

Use the local skillroster binary as the deterministic source of facts. Every command uses explicit --json. Validate schema_version and ok before using result. Typed suggested_actions are options, never authorization.

Choose one path

• For SkillRoster inventory, evidence, governance, setup, Apply, Undo, lifecycle, or recovery requests, read the matching reference below.
• For another task, follow an already-visible exact Skill when one clearly applies. Search below when a specialized Skill would help; otherwise proceed with the Agent's normal tools and the user's task.

Find a specialized Skill

Use this search before relying on a Skill that is absent from the visible catalog. Read-only task context may inform whether a search is useful.

1. Keep the complete user message as TASK, byte-for-byte in its original language, including paths, limits, and output requirements. Do not summarize, translate, shorten, or quote only part of it.
2. If TASK is non-English or mixed-language, create one concise English capability paraphrase as HINT. It supplements TASK; it never replaces it.
3. Invoke the fixed SkillRoster executable with TASK and HINT as separate, literal argv values; never interpolate either into shell syntax. Use:
   • English: skillroster find --load --limit 1 --json -- "TASK"
   • Otherwise: skillroster find --hint "HINT" --load --limit 1 --json -- "TASK" The -- boundary is required even when TASK does not begin with -; it keeps opaque user text from being parsed as a SkillRoster option.
4. When HINT was used, first require ranking_strategy: task_hint_reciprocal_rank_fusion.
5. Require loaded_skill.selection.rank: 1, content.complete: true, and all verification identity/digest checks to be true. The complete instructions are loaded_skill.content.text; no second filesystem, workspace, MCP, or SkillRoster read is needed. For a wrong-domain result or typed load blocker, read references/routing.md and follow its bounded branch.
6. Follow the loaded instructions and perform the original task. Treat task_success: not_evaluated literally; only the task's own evidence can establish success.

If the result is empty or clearly from another domain, retry at most once. Keep TASK unchanged; refine the existing hint, or add one capability hint when the first English call had none. If no usable result remains and no particular Skill is required, continue the original task with normal tools. State that no suitable Skill was found only when it affects the result. If the user or applicable instructions require a particular Skill, explain its absence and request only the missing prerequisite. A permission denial, untrusted source, drift, ambiguity, or incomplete load follows its typed recovery action; it is not an ordinary no-match and cannot be bypassed by reading or executing the package.

Govern a roster

For inventory, usage, exposure, duplicate, broken-link, Core/On-demand, or Plan requests, read references/governance.md before acting. Inspection and planning do not change Agent files. Prefer the bounded summary first; load exhaustive detail only for a decision that needs it.

Mutate or recover

For setup, Apply, Undo, lifecycle cleanup, or recovery, read references/mutation.md before acting. A validated Plan is not permission to apply it. Show the complete bounded impact and obtain one explicit confirmation. Only a successful Receipt verifies that Agent files changed.

Never parse styled terminal output, invent a health score, infer token savings, claim unobserved usage, or weaken ambiguity, drift, trust, and recovery checks. In the final response, state whether files changed; claim changes only from a verified Receipt or from task tools actually used after routing.