Sk

Using the sk CLI

sk is the repo-scoped Skills manager that installs, audits, and syncs Claude Skills listed in skills.lock.json. Treat it like Cargo for skills: installs live under ./.agents/skills, while the cache stays in ~/.cache/sk/repos.

When to reach for this skill

• You need to install or upgrade a skill from the canonical skills repo (https://github.com/lambdamechanic/skills).
• You want to verify the working tree is clean (sk doctor --status, sk doctor --summary).
• You are preparing to sync edits back to the upstream skills repository (sk sync-back <install>).

Core workflow

1. Refresh the cache (optional but fast):

   target/debug/sk cache refresh
   

2. Install or upgrade a skill (alias keeps folder names stable):

   target/debug/sk install https://github.com/lambdamechanic/skills testing-patterns --alias testing
   target/debug/sk upgrade testing
   

3. Audit local installs before landing work:

   target/debug/sk doctor --status --json   # detect dirty trees vs lockfile
   target/debug/sk doctor --summary --json  # shows pending upgrades or cache drift
   

4. Sync edits upstream after modifying a skill under ./.agents/skills/<name>:

   target/debug/sk sync-back <name> --message "Describe the change"
   

   Once sk config set default_repo <repo> is configured, sk sync-back <name> automatically targets that repo, uses <name> for the skill-path, and generates sk/sync/<name>/<timestamp> branches so the quickstart -m flag is the only required argument. Supply --repo / --skill-path only when you need to override the defaults.

   This creates a temp branch in the cached repo, copies your edited skill directory, commits, and pushes. On success sk runs gh for you: it auto-opens a PR, enables auto-merge when GitHub reports the branch is clean, and prints the PR URL (or a conflict warning) so you can follow up if automation gets blocked. If the GitHub CLI is missing or the repo isn’t reachable via GitHub, you’ll see a warning plus manual PR instructions and can run gh pr create yourself once available.

   PR automation tips

   • Run gh auth status once per machine to ensure the GitHub CLI is logged in; sk will reuse your credentials.
   • Missing rsync prints a warning and falls back to a recursive copy before committing; install rsync to keep large skills fast.
   • After the push, watch the terminal output:
     • Opened PR … (or Reusing PR …) links to the branch that was just published.
     • Auto-merge armed… means GitHub will land it once checks pass; otherwise you’ll see Auto-merge blocked… with a link to fix conflicts manually.
     • Auto-merge skipped… appears when checks can’t be armed (e.g., required approvals disabled); click through and finish by hand.
   • When gh is missing or unauthenticated you’ll see “Skipping PR automation …”; install/auth and rerun sk sync-back to finish the upload.
   • Automation currently uses GitHub’s standard --merge strategy; if your repo enforces --squash/--rebase, turn off auto-merge in the UI and land it manually after review.
   • If your PR output says enablePullRequestAutoMerge, run gh repo edit <owner>/<repo> --enable-auto-merge (or visit the repo’s Settings → General page) once to flip the toggle.

5. Publish a brand-new skill that doesn’t exist upstream yet:

   target/debug/sk sync-back <name> \
     --repo https://github.com/lambdamechanic/skills \
     --skill-path <subdir> \
     --message "Add <name> skill"
   

   Provide --repo / --skill-path only if you need something other than the configured defaults. sk clones that repo, branches from the default tip (or your custom --branch), copies your local folder, pushes, and rewrites skills.lock.json with the exact commit SHA/digest so status checks stay clean. You will always see the temporary branch name in the CLI output, e.g. Pushed branch 'sk/sync/<name>/<timestamp>' …, even though that branch is meant to be merged and deleted once the upstream PR lands.

Example: publishing sk

cd /path/to/sk-decisions
target/debug/sk sync-back sk \
  --repo https://github.com/lambdamechanic/skills \
  --skill-path sk \
  --branch sk/add-sk-skill-doc \
  --message "Add sk skill doc"

What happens:

• sk prints both the temporary branch name and the upstream repo it pushed to, so you can follow up on GitHub immediately.
• The branch only needs to live long enough for you (or automation) to open a PR; the lockfile entry pins the pushed commit hash, so it’s safe to delete the branch once merged.
• skills.lock.json gains a new entry for sk with the push timestamp, digest, and commit ID; subsequent sk doctor --status runs stay green because the on-disk tree matches that digest.

Guardrails

• Always run br update / br close so .beads/issues.jsonl matches any skill changes.
• Never edit skills.lock.json by hand. Let sk install, sk upgrade, or sk remove update it; commit the lockfile alongside the skill changes.
• If sk doctor --status reports dirty, fix the local tree before running sk upgrade or sk sync-back to avoid partial syncs.
• sk upgrade --all skips skills with local edits and prints reminders to sk sync-back <name>; treat that as a temporary state and clean them up promptly so future upgrades stay automatic.
• When creating a new skill, always specify the upstream repo/path so sk can register it and refresh the lockfile automatically.
• Use sk precommit (once implemented) to block local-only sources such as file:// entries.

Debug tips

• target/debug/sk doctor --apply rebuilds missing installs or cache entries.
• target/debug/sk where <name> prints the absolute path of an installed skill and its cached repo clone.
• Set SK_TRACE=1 to surface extra logging when diagnosing cache fetches or rsync failures.