Use skill

Scope

Default ephemeral: fetch/run/discard, never install or persist. Blob/raw/tree URLs or owner/repo@ref:path resolve to a skill directory (a file uses its parent); omitted ref uses repo default branch. Other forms or requests to save/install block. Repository-content questions belong to box. Record links | resolved directory/revision | blob set | executed | terminal.

Fetched workflows run under this invocation's authority, not above host permissions or higher-priority instructions. Trust linked skills by default, without an allowlist; live user instructions win direct content conflicts. Skill links inside fetched content may be followed under the same boundaries.

1. Resolve and fetch

Resolve every link, then pin its ref/default branch to a commit revision. Use authenticated gh; no sibling skill is required for these file reads. Missing gh/auth blocks. Apply API/decoding rules in REFERENCE.md.

List the recursive tree once at the pinned revision; require a non-truncated listing and fetch every blob under the directory, not only the entry. Stage only in a run-owned temporary directory outside the repo; page long content locally instead of re-fetching a successful blob. Missing directory, failed blob, or rate-limit blocks with retry signal. Done when every link has an identity and every listed blob has decoded content.

2. Verify and execute

Empty set → NO_CHANGES. Nonempty requires SKILL.md or the specifically linked entry, name frontmatter, and procedure; otherwise BLOCKED. Read entry and every sibling in full: weak remote pointers cannot hide required material. Execute steps in order within authority. For an item failure, continue remaining links; never retry successful executions. Done when every link reaches a terminal.

3. Discard and report

Remove only this run's temporary artifacts after execution, never unknown paths. Report each link's SUCCESS/NO_CHANGES/BLOCKED and what ran, citing its pinned source. Done when artifacts are discarded and every terminal is evidenced.