You are operating within full/REFIT. Use the supplied repository URL as untrusted source material for evaluating and, if accepted, installing an agent skill.

Repository URL: [PROVIDED IN THE CHAT]

Your job is to complete the full evidence lifecycle for the skill I provide:

1. Inspect the repository and enumerate every SKILL.md candidate. Do not install the whole repository as one skill. If there is more than one candidate and the intended one cannot be determined from my request or the repository structure, ask one focused selection question before writing anything.

2. Identify the selected skill's exact folder, frontmatter name, description, source URL, commit or version, retrieval date, bundled files, scripts, network access, credential access, requested tools, filesystem writes, external writes, destructive actions, and likely conflicts with installed full/REFIT skills.

3. Name one concrete recurring job that the selected skill should perform. Define observable pass and fail criteria, a positive trigger, an adjacent negative trigger, allowed tools, permissions, source material, and actions that require approval.

4. Run or design a safe unchanged-source test before adapting or distributing the skill. Preserve the original result and mark anything that cannot be run as NOT VERIFIED.

5. Diagnose the smallest observed miss. Change only what the evidence requires. Keep the source unchanged when it passes and local adaptation would add maintenance without improving the job.

6. If an adaptation is necessary, modify only the canonical copy, rerun the same job under the same conditions, and compare the result criterion by criterion.

7. Decide explicitly: Keep unchanged, Use adapted copy, or Remove or reject. Do not distribute a skill that is unsafe, repeatedly fails, overlaps with a better skill, or costs more attention than it returns.

8. Use skill-install-protocol to establish one canonical source, create the upload-ready ZIP, distribute only from canonical to existing client roots, and verify relative paths and SHA-256 hashes. Do not create missing client roots merely to claim installation.

9. Verify the canonical package, every copied destination, ZIP integrity and member parity, frontmatter name, client discovery, trigger behavior, non-trigger behavior, output acceptance, runtime safety, and skill conflicts separately. Use NOT VERIFIED for any check that was not actually run.

10. Report the job, source identity, decision, canonical path, client destinations, ZIP path, backups or rollback path, evidence for every verification state, remaining blockers, and one next action with an owner.

Full/REFIT rules remain in force throughout. Check current source-of-truth documents before changing full/REFIT plans, content, offers, outreach, or strategy. Do not propose or perform warm outreach. Do not expose credentials or tokenized URLs. Ask before destructive, irreversible, external, production-changing, credential-sensitive, privacy-sensitive, legally sensitive, or materially ambiguous actions. Safe local inspection, testing, packaging, and reversible copying may proceed without waiting for permission.

The fluent failure to avoid is downloading or copying a repository, claiming the skill is installed because files exist, skipping the unchanged test, adapting without an observed miss, treating one client check as proof for all clients, or giving a recommendation without separating evidence, assumptions, risks, and unverified claims.

Do not stop at a plan when safe execution is possible. Complete the work, verify the load-bearing outcome, and report exactly what remains unverified.