skills/worktrees/SKILL.md
Git worktree hygiene — when to create, how to clean up, and what fails silently. Captures three classes of bug that recur with worktree-based agent isolation: subagent dangling commits, orphan branches with unique work, and time-sensitive dangling-commit garbage collection. Includes detection commands, the salvage decision flow, and the trust-but-verify rule for subagent commit reports. Use when: a subagent reports a commit SHA, after a worktree-based agent finishes, cleaning up `.claude/worktrees/`, salvaging old crew worktrees, "is this branch in main?", verifying agent-claimed commits actually landed, planning multi-agent parallel work in worktrees.
npx skillsauth add mikeparcewski/wicked-garden worktreesInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Git worktrees let multiple branches be checked out simultaneously and let
agent runtimes (Claude Code's Agent({isolation: 'worktree'}), crew runners,
etc.) make changes in isolation without disturbing the main checkout. Useful —
and a quiet source of three repeating bugs.
Subagent dispatched into an isolated worktree commits inside that worktree. Branch ref lives only in the worktree; when the worktree is removed, the commit is dangling — reachable only via git fsck --no-reflogs --lost-found until git gc collects it (14-day default via gc.pruneExpire). The subagent's "Commit SHA: abc1234" report becomes a lie. Code may also be sitting in the main worktree as unstaged changes.
Trust-but-verify rule — every time a subagent reports a commit, run the ancestry check from refs/recipes.md §1.
Worktree dir is gone (auto-cleaned, deleted weeks ago); branch (local or origin/crew/<name>) still points at commits NOT in main; git worktree list says nothing about it. Bulk-deleting branches because "no worktree = no work" silently loses everything between fork-point and tip.
Detection commands in refs/recipes.md §2. Each branch needs an explicit salvage decision (see Salvage flow below).
Commit object exists but no branch points at it. Time sensitive — once git gc runs (default grace 14 days), the object is unreachable forever. List + preserve commands in refs/recipes.md §3.
Decision must be prompt — there is no third state. Either preserve or accept the loss.
Branch + remote ref state, not on-disk worktree presence. Worktree-dir absence is the LEAST reliable indicator of merge state. Safe-delete recipe in refs/recipes.md §4.
For each candidate branch:
.claude/worktrees/agent-* and project-specific dirs (e.g. ~/.command_iq/worktrees/crew-*). Run git worktree list.refs/recipes.md §5.Manual file-by-file extraction often beats git cherry-pick on noisy branches — see refs/recipes.md §7.
Dangling commit SHA + unstaged changes in main worktree shows up when the subagent edited shared files (main worktree sees as unstaged) before its own commit (lands in worktree branch). Both halves need attention. Re-landing recipe in refs/recipes.md §6.
Whenever a subagent's summary says "commit SHA: XYZ":
git merge-base --is-ancestor XYZ origin/main — must return true (exit 0). Use origin/main (not HEAD); the latter false-positives if you're on a feature branch. git fetch origin main first if origin is stale.git status -s — should be empty if the commit is real and completegit show XYZ --stat | head — confirm the change set matches what the subagent describedSkipping this check has cost time before. The closing comment on a GitHub issue citing a dangling SHA is misleading future archaeology.
wicked-garden:engineering:review — that
reviews code changes; this reviews whether the changes actually
landed where they should.development
--- name: large-scale-migration description: How to execute a LARGE MECHANICAL change across any codebase with LEVERAGE instead of an agent-grind or hand-edits — a cross-cutting migration, refactor, rename, dialect/framework/DB port, library adoption, or bulk transform. The map→transform→gate pattern: a deterministic transform driven by a source-of-truth map, proven by a differential-equivalence gate. Use when the work is "migrate all X to Y", "rename Z everywhere", "port to a new DB/dialect/fra
testing
v11 LLM-based work-shape classifier. Replaces the regex archetype detector with the model's own reasoning. Reads the user's prompt, picks the right archetype(s) from the catalog, identifies signals (blast_radius, novelty, reversibility, etc.), and persists to SessionState so subsequent turns steer correctly. Use when: the prompt_submit hook emitted a `<wg classify-due />` directive, OR explicitly invoked at session start, OR when re-classifying after the user changes scope mid-session.
tools
v11 work-shape archetype runner. When a prompt has been routed to one of the 9 archetypes (triage, explore, specify, decide, ship, review, incident, build, migrate), this skill is the entry point. It picks the right per-archetype playbook from refs/ and executes the phase shape declared in `.claude-plugin/archetypes.json`. Use when: a `<wg archetype="X">` or `<wg archetypes>` system-reminder tag appears, an explicit "let's run the X archetype" request, or when one of the per-archetype slash commands resolves to this skill.
development
Show or set the session intent variable. Intent gates how loud the framework is — simple-edit (silent), feature/research (synthesis directive), rigor (full crew context). Auto-detected on turn 1; this skill overrides explicitly. Sticky for the session. Use when: "set intent", "intent override", "/wicked-garden:intent", "make the framework quiet", "force rigor", "what's my intent".