/work

/work [slug] is a declaration for this task: "spin up a dedicated worktree and do the work there." It applies to what the user is starting right now — it is not a standing rule that all work always happens in worktrees.

When invoked:

1. Look at what the repo already does, and copy it.
2. Create the worktree + branch and cd into it.
3. Do the requested work there.

Work aimed at an issue ("if it's open and reproducible")

Do these in order before writing any fix:

0. Load the repo's own gate skill first. If the repo has a deployment or brain skill (a <prefix>-repo skill, a triage/claim system), load it before searching. Its per-issue gates (claim, overlap check, side-effect screen, risk title, auto-merge classifier) apply to a single /work fix as much as to a sweep. Skipping them because the request said "/work" rather than "triage" is the failure: the PR ships without the risk and side-effect sections she expects, and can duplicate her own open PR.
1. Find the issue. Search open and closed issues with several phrasings (gh issue list --state all --search "<q>"), plus in:title variants. A closed issue with a close-by-PR may already cover it. Check that before you call a newer report a duplicate.
2. Look for a PR that already fixes it. Search with gh pr list --state all --search "<issue#>" and with keywords. Drafts count: a contributor can park a finished fix as a draft because of an open-PR cap. If a complete fix exists, salvage it (cherry-pick the commit so authorship is kept) instead of writing a parallel one. Check her own open PRs too (gh pr list --author @me --state open --json number,title,files,mergeable,autoMergeRequest), not only contributors': an armed-but-CONFLICTING PR of hers on the same function is live work. Fold both into ONE combined PR that supersedes them (cherry-pick what applies, port the rest by hand onto main's current file layout, keep every Co-authored-by), then disarm (gh pr merge N --disable-auto) and close the sources with a comment linking the replacement.
3. Prove it red on the base. Apply only the tests first (gh pr diff <N> > f.diff && git apply --include='tests/*' f.diff), run them, and see them fail. Then remove the file and git cherry-pick the real commit, so the test file isn't left as an untracked conflict.
4. Reproduce the user-visible path end to end. Run it on a detached base worktree and on the branch, and keep both outputs as before/after evidence for the PR body.
5. Re-run any failures on the base worktree before calling them unrelated. "Fails identically on untouched base" is the claim to report.

Answer the "is it open / can it be repro'd" part first in the handoff, then the fix and the PR link.

Read the repo's conventions

git worktree list                                   # sibling dirs + branches
git branch -a --sort=-committerdate | head -20      # branch prefix/slug pattern
git remote show origin | sed -n 's/.*HEAD branch: //p'   # default branch

• Branch prefix — mirror the dominant one (e.g. <initials>/<slug>). None obvious → ask.
• Worktree dir — mirror sibling naming/location (commonly <repo>-<short> next to the primary checkout). Short suffix ≠ full branch is fine.
• Base — current default branch unless they name another.
• Slug — from the /work argument, else the task description.

Create it

repo="$(git rev-parse --show-toplevel)"
parent="$(dirname "$repo")"
default="$(git -C "$repo" remote show origin | sed -n 's/.*HEAD branch: //p')"
git -C "$repo" fetch origin "$default"
git -C "$repo" worktree add -b <prefix>/<slug> "$parent/$(basename "$repo")-<short>" "origin/$default"
cd "$parent/$(basename "$repo")-<short>"

Reuse a matching worktree/branch if one already exists — don't duplicate. Match how siblings handle node_modules / .venv / symlinks.

• Never git stash in a worktree while other agents run: the stash list is shared repo-wide, so a pop can grab another worker's WIP. Commit WIP instead.

Then

• Do the task in the worktree; never in the primary checkout, and never in another agent's worktree.
• Never git stash in a worktree: refs/stash is shared by every worktree of the repo, so a parallel agent can pop your entry (or you pop theirs). To test against base, copy the file aside (git show HEAD:<path> > …) or use a throwaway worktree instead.
• Every worker and subagent gets its own worktree, and the worktree path is in its instructions. Fixes use a bb/* branch. Verification, typecheck, and tests use a detached worktree at origin/<default>. When its lockfile matches the primary's, symlink every node_modules (root and workspace packages) to the primary's; install only on a lock change, after removing those links. Never edit, install, build, typecheck, or test in the primary checkout.
• Prefer direct execution for focused fixes. Do not delegate routine sibling-path audits or cleanup; reserve subagents for substantial independent work that clearly saves time.
• Default to pnpm for JavaScript package commands and install examples when the repository supports it. Respect an existing different package-manager/lockfile contract rather than introducing a second lockfile.
• UI → ui-only + ui-system. Tests → whatever the repo uses.
• When handing off, tell the user the worktree path and how to run it.
• Ship → commit, push the branch, open a PR/MR to the default branch.
• Cleaning up → offer to remove worktrees whose branch is already merged.
• Reviewing an existing PR instead of new work → pr-triage.