From ai-eng-devops
Coordinates multi-repo refactors by fanning out subagents per repo onto feature branches. Use for mechanical changes spanning 3+ repos.
How this skill is triggered — by the user, by Claude, or both
Slash command
/ai-eng-devops:cross-repo-refactorThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Coordinate a refactor that spans multiple repos. The orchestrating session reads the change description, plans the per-repo edits, fans out subagents to make them, collects the results onto a single feature branch in each repo, and pushes for review. The branch-guard hook ships alongside this skill — it physically prevents pushing to `main`/`master`/`develop` or any pre-existing branch.
Coordinate a refactor that spans multiple repos. The orchestrating session reads the change description, plans the per-repo edits, fans out subagents to make them, collects the results onto a single feature branch in each repo, and pushes for review. The branch-guard hook ships alongside this skill — it physically prevents pushing to main/master/develop or any pre-existing branch.
<repos>. Slug: <short-slug>. Description: <one-liner>."<short-sha> (typically from date +%s | sha1sum | cut -c1-7 or a stable hash of the slug+date).git checkout -b cook/refactor-<slug>-<short-sha> (the branch-guard hook validates the name is fresh and matches the pattern)refactor(<slug>): <one-line summary>git push -u origin cook/refactor-<slug>-<short-sha>cook/refactor-<slug>-<short-sha>
<slug> — kebab-case, alphanumeric + hyphens, ≤ 30 chars (e.g., bump-zod-3.23)<short-sha> — at least 7 hex characters (matches the regex ^[0-9a-f]{7,}$)The branch-guard hook regex is:
^cook/refactor-[a-z0-9-]+-[0-9a-f]{7,}$
Branches that don't match are rejected by the hook on git checkout -b. Pushes to main/master/develop are rejected outright. Pushes targeting HEAD:main (refspec form) are also rejected.
Stop when ALL of these are true:
origin successfullyWhen termination triggers:
rm -f ~/.claude/cooking/active-$PID ~/.claude/cooking/iter-$PID
If iteration 10 hits before completing the fan-out, write a partial-progress note (which repos shipped, which didn't, why) and exit. Do not skip the structured-message format to "make it work."
main. The hook will reject; if the agent retries with a different approach, retry will also be rejected. The fix is naming the branch correctly, not bypassing the hook.git checkout -b.hooks/cooking/branch-guard.sh to relax it during a session. The skill's safety boundary is the hook; relaxing it makes the skill silently dangerous.The branch-guard hook runs as a Claude Code PreToolUse Bash matcher (hooks.json PreToolUse matcher with Bash tool, command pattern). Cursor does not expose an equivalent PreToolUse Bash hook today, so under Cursor this enforcement is skill prose only — no runtime guard. Use this skill from Claude Code when correctness matters; from Cursor, audit the branch names before you push.
bash hooks/cooking/test/branch-guard.sh covers 11 cases including push-to-main, push-to-master, push-to-HEAD:main, the cook-pattern accept path, the pre-existing-branch reject, and pattern mismatch. Run it after any branch-guard.sh edit.
hooks/cooking/branch-guard.shhooks/cooking/stop-hook.sh/cook-statustest-fix-loop, investigation-loop, dreaming-consolidatornpx claudepluginhub p/v1truv1us-ai-eng-devops-plugins-ai-eng-devopsCreates structured, bite-sized implementation plans from specs or requirements before writing code. Useful for breaking down multi-step tasks into testable steps with file structure and task boundaries.
2plugins reuse this skill
First indexed Jul 8, 2026