One package.
Six trust decisions.
Work through admission, discovery, selection, loading, execution, and evolution. This is a review aid, not an automated security assessment.
Who gets into the library?
A package is accepted without reproducible identity, scope, or ownership.
Explore the relevant guide ↗Source identity, exact revision, file inventory, assigned owner, admission decision.
Keep discovered packages separate from approved packages. Review the complete artifact.
What becomes visible?
Metadata claims exaggerated relevance or hides the package’s actual purpose.
Explore the relevant guide ↗Name, description, registry entry, approved task scope, discovery candidates.
Compare metadata with implementation and preserve the reviewed discovery state.
Why is this skill chosen?
An unrelated or unsafe skill is selected for a task outside its reviewed scope.
Explore the relevant guide ↗Task, candidates, selected package, approval requirement, host configuration.
Test in-scope and out-of-scope tasks; constrain selection to reviewed packages.
What enters context?
Supporting files or mutable references add unreviewed instructions.
Explore the relevant guide ↗Loaded file versions, referenced URLs, resource map, task context.
Follow the full loading chain and review resources according to how they are consumed.
What can it do?
A skill-induced action reaches unrelated data, destinations, or permissions.
Explore the relevant guide ↗Tool calls, targets, arguments, approval records, runtime and application effects.
Apply least privilege and independent authorization to consequential actions.
What changed after review?
An update, dependency, or host change expands behavior or authority.
Explore the relevant guide ↗Diff, dependency state, permission changes, new test results, rollback artifact.
Re-review material changes and promote the exact assessed artifact.