Gitkit

gitkit

One place where the git facts live that every other skill in a toolchain keeps re-deriving: where a worktree goes and what it's named, how it is created, adopted, and removed, which branch is the base, and when to rebase versus merge.

The core here is not clever. All of it is native git worktree, with no vendor CLI, no GUI dependency, and no recorded state anywhere but git's own. That is the point: a laptop with a desktop dev tool and a headless Linux box run identical commands, so there is no second code path to keep in sync and no "degraded mode" to reason about.

One branch of this skill sits outside that core, and says so. Stacked branches drive GitHub's gh stack extension, so that branch alone needs gh and a GitHub remote. It degrades to plain git rather than failing, and nothing else in this file depends on it, so a box with no gh runs everything above untouched.

gitkit is a primitives skill. Most of its runs come from another skill calling it, not from a human saying its name.

When this fires

• worktree. "Spin up a worktree for this branch", "make me a worktree for #42", "where is the worktree for issue-42-…", "tear down this worktree", "list my worktrees".
• base ref. "What's the base branch here", "diff me against the base".
• sync. "Should I rebase or merge here", "sync this branch with main", "this branch is behind, bring it up to date", "make this PR review-ready on the latest base".
• clean. "Clean up my merged worktrees", "which branches can I delete", "this repo is full of dead branches".
• rescue. "I lost a commit", "my rebase ate my work", "I reset the wrong branch", "what's in this stash".
• stack. "Stack this on #43", "start #44 on top of #43's branch", "restack this chain", "show me the stack".
• called by another skill. Any skill that needs a worktree, a base ref, a branch name, or a sync decision delegates here and uses what comes back.

Not this skill: committing, opening a PR, reviewing a diff, judging code. gitkit answers where and how, never what the change should be.

Modes

worktree and sync sit inline because nearly every run and every calling skill reaches them. The other three load only when their mode fires.

When the mode is not clear from the ask, pick it from the verb and say which one you picked. "Tidy this up" after a merge is clean; "tidy this up" on a branch behind its base is sync. Naming the mode in the first line is what lets the human correct a wrong pick before anything runs.

The convention

Path

$WORKTREE_ROOT/<repo-basename>/<local-branch-name>

$WORKTREE_ROOT defaults to ~/worktrees; honor the environment variable when it is set. <repo-basename> is the main checkout's directory name.

Slashes in a branch name flatten to dashes. A branch mimukit/6-marketing-site gets the directory mimukit-6-marketing-site, not a nested mimukit/ parent. This keeps one flat level per repo, so ls "$WORKTREE_ROOT/<repo>" is a complete inventory rather than a tree to walk. The trade is that the directory name no longer round-trips back to the branch name, which costs nothing, because lookup is always through git, never by reading a path. On the rare collision (branches a/b and a-b both flatten to a-b), git worktree add fails on the existing path: report it and let the human pick a name rather than silently appending a suffix.

Worktrees live outside the repository, never in a .worktrees/ directory inside it. An in-repo worktree gets swept into docker build contexts, bind mounts, file watchers, and test globs, and every one of those failures shows up far from its cause.

The invariant: one branch, one worktree

A branch has at most one worktree, and the directory is named after it. A worktree is keyed to a branch, not to a workflow stage, so implementing a feature and later reviewing its pull request use the same worktree, because they are the same branch.

This is not a style preference. Git enforces it:

$ git worktree add ../wt-b feature-x
fatal: 'feature-x' is already used by worktree at '.../wt-a'

Any scheme that names worktrees by stage (review-…, qa-…) hard-fails the moment two stages touch one branch, which is the normal case, not the edge case.

Consequences:

• Lookup is by branch, always, through git, never by reading or guessing at a directory path. This is what makes the flattening above free, and what lets a worktree sitting in some older root keep working untouched.
• The "kind" of a worktree rides along for free, because the branch name already carries it. issue-42-auth-capability-module is named that because the branch was named that when the work started.

Branch naming

• issue-<n>-<slug> for work that starts from a tracker issue. Slug from the issue title: if it follows a conventional type(scope): summary form, strip the prefix; kebab-case the rest, cap at roughly 50 characters on a word boundary, drop a trailing hyphen. An empty slug degrades to a bare issue-<n>. The number guarantees uniqueness, so no tie-break is ever needed.
• hotfix-<slug> for an urgent fix that patches the base branch directly. Slug from the symptom, kebab-cased, capped at roughly 40 characters. The number is absent because a hotfix starts from the symptom rather than from a tracker issue; when one does trace to an issue, the branch still takes this shape and the pull request carries the Closes #<n> link. On a collision with an existing branch, add a short distinguishing word rather than a counter. The separator is a hyphen: write hotfix-<slug>, never hotfix/<slug>. A slash makes git treat hotfix as a directory in the ref namespace, so hotfix-restore-session and hotfix/restore-session cannot both exist, and every skill that reads the type off the branch name (commitkit, prkit) matches the hyphen form only.
• pr-<n>-<slug> only for a fork pull request, where no local branch exists yet and one must be invented to hold the fetched head. Slug from the head branch name, kebab-cased, capped at roughly 40 characters.
• Anything else takes a branch name the human or the repo's convention supplies. gitkit does not rename it.

For a same-repo pull request there is already a local branch name: use it. Inventing a pr-* name for a branch that exists is how you land on the two-worktrees-one-branch failure above.

worktree

All four are native git, and all four are idempotent, so running one twice is a normal thing to do and must never error or destroy work.

Look up a branch's worktree

The one lookup primitive everything else is built on:

git -C "$REPO" worktree list --porcelain

Each record carries a worktree <path> line and, when a branch is checked out, a branch refs/heads/<name> line. Match on the branch to get the path. If the porcelain shape ever shifts, check git worktree list --help rather than guessing, but treat the worktree/branch pairing as stable; that is what porcelain output is for.

Create, or adopt

Always look first. If a worktree already exists for the branch, adopt it: report the path and stop. Never create a second one, never error.

When there is none:

git -C "$REPO" fetch origin --prune
git -C "$REPO" worktree add -b "$BRANCH" "$WORKTREE_ROOT/$(basename "$REPO")/$BRANCH" "$BASE"

• Fetch first, always. Branching off a stale base is silent and only surfaces as conflicts later.
• $BASE comes from The base ref, never a hard-coded origin/main, and never a sibling feature branch. The one exception is a stack layer, whose base is the branch below it, named deliberately by the caller.
• Drop -b when the branch already exists locally (a same-repo PR you have fetched, a branch you made earlier); git worktree add <path> <branch> checks it out.
• For a fork pull request, fetch the head into a local branch first with git fetch origin "pull/<n>/head:pr-<n>-<slug>", then add the worktree on that branch, no -b.

Report the path and the branch. Creating a worktree does not imply doing anything in it.

Remove

git -C "$REPO" worktree remove "$WT"
git -C "$REPO" branch -d "$BRANCH"     # only a branch you created, and only if merged

Three rules, each one guarding against a real way to lose work:

• A dirty worktree stops teardown. Show exactly what would be lost (uncommitted changes, untracked files, unpushed commits) and let the human decide. Never reach for --force on their behalf.
• Never remove a worktree you adopted rather than created. If it was already there when you arrived, it is someone else's context; you have no idea what is open in it.
• Never delete a branch you did not create. Use -d (not -D) so git itself refuses an unmerged branch.

Already gone? Report "already gone" and succeed. Teardown is idempotent in the same spirit as adopt.

List

git -C "$REPO" worktree list

A plain inventory. When the user asks to tidy up rather than to look, that is the clean mode, which classifies every row before it offers anything.

If paths look wrong after a move or a restore, git worktree repair <path> rewrites the back-pointers. Git stores absolute paths in .git/worktrees/<name>/gitdir and in each worktree's .git file, so moving a worktree by hand always needs a repair.

The base ref

Never assume main. Repos default to develop, trunk, and master in the wild, and getting this wrong silently produces an empty or enormous diff. The ladder, in order:

1. gh repo view --json defaultBranchRef --jq .defaultBranchRef.name, authoritative when the GitHub CLI is available and authenticated.
2. git symbolic-ref --short refs/remotes/origin/HEAD, the local record, if it has been set.
3. git remote set-head origin --auto to repair an unset origin/HEAD, then retry the previous step. This is the step most implementations skip, and it is the one that turns a failure into an answer.
4. Whichever of origin/main / origin/master actually exists.
5. Ask the human. Do not guess past this point.

Everything downstream uses the answer: git diff <base>...HEAD, ahead/behind counts, and the branch a worktree is cut from.

A stack layer is the one deliberate exception, and the ladder still runs. A layer's base is the branch below it rather than the repo default, so the caller names it and this ladder is not consulted for that branch. The ladder is still what resolves the trunk the stack's bottom layer sits on. The ban on a sibling-branch base keeps its teeth everywhere else, because what it catches is the accidental case: a worktree cut from whatever happened to be checked out. A layer base is legal because somebody said which branch and why. See Stacked branches.

Rebase or merge

Syncing a feature branch with its base has one default here, and it does not depend on whether a pull request is open:

Rebase onto the base to sync a feature branch, published or not. Merging the base in is an exception that needs a stated reason and the user's consent.

The default is rebase because that is the history worth having: a feature branch that reads as a linear sequence of its own commits, with the base's work underneath rather than braided through it. A repo that merges the base in on every sync accumulates one merge commit per sync, and the branch's own story becomes unreadable long before it lands.

Rebasing does have a real cost, and it is mechanical rather than aesthetic: it rewrites commit SHAs. Once a pull request is open, its review threads are anchored to those commits, so a force-push marks every one of them outdated, human and bot findings alike. That is the cost to disclose and weigh, not a bar. It is also the one pre-baked reason to propose the merge exception.

In practice:

git fetch origin
git rev-list --left-right --count "origin/$BASE"...HEAD   # "<behind>\t<ahead>"; left > 0 means behind
git rebase "origin/$BASE"

• Rebase is the answer unless someone has said otherwise for this branch. No pull request, an open one, a shared branch: the recommendation is the same.
• A branch with no remote counterpart rebases straight through. Nothing points at those SHAs, so nothing can break; this is the cheap-and-reversible class, and it needs no confirmation.
• A published branch previews once and waits. One confirmation covers the rebase and the --force-with-lease push that follows, because they are a single decision, and asking twice asks the same question twice. Name what it costs: the base being synced, the commits being rewritten, and the number of unresolved review threads that will go outdated.
• Count the threads before you ask. On a branch with an open PR, query the GraphQL reviewThreads connection for unresolved threads and put the number in the preview. "3 review threads will be marked outdated" is a fact the user can decide on; "this may outdate review comments" is not.
• Unresolved threads are the reason to offer the merge exception, not to take it. Still recommend the rebase; name merge as the available alternative alongside the count. A preference that folds in the case that argues against it was never a preference.
• --force-with-lease, never bare --force. The lease is what stops you overwriting a commit someone else pushed while you were rebasing.
• On conflict, stop and surface it (git diff --name-only --diff-filter=U). Propose a resolution per file and confirm before writing. A conflict resolution is a code change: run the repo's test gate afterward.

sync

The runnable form of the rule above. A sync has two halves, in this order: first bring the local branch level with its own counterpart on origin, then bring it up to date with the base. Skipping the first half is what makes a later --force-with-lease reject, or silently drop a commit pushed from another machine. Run it in the branch's worktree. Each step names the condition that ends it.

1. Fix the ground. Resolve the base with the base ref ladder. Run git -C "$WT" fetch origin --prune. Confirm the worktree is on the feature branch and that git status --porcelain is empty. A dirty tree stops the sync: report the files and let the human stash or commit. Ends when the base name, the branch name, and a clean tree are all known.
2. Sync with origin/$BRANCH. Run git rev-list --left-right --count "origin/$BRANCH"...HEAD when the remote branch exists. Behind count 0 means nothing to take. Otherwise run git pull --rebase origin "$BRANCH" to replay your local commits on the remote ones, and resolve every conflict by the rule in Resolve every conflict below. A branch with no remote counterpart skips this step. Ends when the local branch contains every commit on origin/$BRANCH.
3. Measure the gap to the base. Run git rev-list --left-right --count "origin/$BASE"...HEAD. Behind count 0 means the branch is already on the latest base: push whatever the previous step took, report that, and stop. Ends with a behind count and an ahead count.
4. Preview and confirm, once. An unpushed branch skips this step and goes straight through. A published branch gets one preview that covers the rebase and the force-push: the base, the behind and ahead counts, and the unresolved review thread count from the GraphQL reviewThreads connection when a pull request is open. Name merge as the alternative, and still recommend the rebase. Ends on the user's answer.
5. Rebase onto the base. Run git rebase "origin/$BASE". Ends when the rebase reports success or stops on a conflict.
6. Resolve every conflict. For each stop, list the files with git diff --name-only --diff-filter=U. Read each conflicted file, propose a resolution that keeps the branch's intent and the base's new code, and confirm before writing. Stage the file, then git rebase --continue. Repeat for every remaining stop. git rebase --abort restores the pre-rebase state, and it is the answer when a conflict is not yours to settle. Ends when no conflict marker remains and the rebase is complete.
7. Prove the branch still works. Run the repository's own test and build gate. Report a failure with its output and stop before the push. Ends with a pass, or a stop.
8. Push with a lease. Run git push --force-with-lease origin "$BRANCH". A rejected lease means somebody pushed while you rebased: fetch, show the new commits, and ask before any retry. Ends when the remote branch matches the local one.

Hand off. Report the commits taken from origin/$BRANCH, the base, the behind and ahead counts, the files whose conflicts you resolved, the gate result, and the pushed branch. Say when the branch was already current and nothing changed. Next, review the pull request diff on the new base, and re-request review when threads went outdated.

The merge exception

When the user takes the merge, or asks for it outright, it is still a sync, and it says so:

git merge "origin/$BASE" -m "chore(repo): sync with origin $BASE"

• Never git's default subject. Merge branch 'main' into issue-42-tailwind-4-foundation is what git writes when nobody chose a message, and it reads that way forever in git log. The fixed subject is the message.
• Interpolate the base, don't hardcode main. The base ref ladder exists because develop and trunk repos are real; a subject that says main in one of them is simply false.
• No --no-ff. If the sync can fast-forward, the branch had nothing of its own to preserve and git writes no commit at all, so there is no subject to fix, and forcing one manufactures an empty merge commit.
• State the reason in the same breath as the ask. "Merging instead of rebasing because this PR has 3 unresolved review threads" is a reason. "Merging to be safe" is not; if you cannot name what rebasing would break, rebase.

Never bare git pull

git pull on a branch behind its remote merges by default, and writes Merge branch 'main' of github.com:owner/repo without asking anyone. Use git pull --rebase, or git pull --ff-only when you want the sync to fail loudly rather than resolve itself. The same holds for git pull invoked to refresh the base branch itself: configure it or pass the flag, but never let the default run.

clean

Sweep away the worktrees and branches whose work has landed, on your machine and on origin. It classifies every worktree, local branch, and remote branch into one bucket — active, adopted, dirty, reapable, orphan — and removes only the reapable ones, one confirmation at a time. A remote delete takes its own confirmation and its own proof that the work landed.

Two things make it more than a git branch --merged loop, and both live in clean.md: a squash-merged branch is invisible to ancestry, so "merged" needs three detections rather than one; and the per-item confirmation is deliberate, because a sweep's rows are not equally safe to delete. Read that file when a run actually sweeps.

The Remove rules govern every removal the sweep makes. It never deletes a branch with -D, never touches a worktree it adopted, and stops on a dirty one. On origin it never deletes the base branch and never deletes the head of an open pull request.

rescue

Find work that looks lost and put it back. A bad rebase, a hard reset, a deleted branch, a stash nobody can find: the commits usually still exist, and git's own logs say where they went.

The procedure is in rescue.md. Two rules shape it. It is read-only until the restore, and the restore adds a branch rather than moving one, because a reset is the operation that lost the work in the first place. It also states the bound rather than implying safety: an unreachable commit lives until git prunes it.

Never run git gc, git prune, or git reflog expire during a rescue, in any mode.

Stacked branches

A stack is a chain of branches in one repository where each branch is cut from the one below it and the bottom sits on trunk. It is what you reach for when work depends on work that is built but not yet merged: instead of waiting for the prerequisite's pull request to land, you branch from it and keep going.

GitHub supports this natively, and the gh stack extension drives it. This is the one part of gitkit that needs gh and a GitHub remote. The full command surface, the preflight, and the degradation path live in stacks.md; read it when a run actually touches a stack. What follows is the part every run needs.

One worktree per layer

Each layer keeps its own worktree, cut from the layer below. The one branch, one worktree invariant is unchanged: a layer is a branch, so it gets a directory named after it, and lookup is still by branch.

The alternative, one worktree for the whole stack with gh stack up and down moving between layers inside it, is what the extension's own navigation assumes. It is not what this collection does, because every caller here keys a worktree to the issue being worked, and a stack is several issues. So gh stack's navigation commands are out of scope, and up, down, switch, top, and bottom are not gitkit's to run: they would move a checkout that another worktree already holds.

What gitkit owns here, and what it does not

gitkit takes the plumbing and nothing else: create a stack, add a layer, restack, sync, inspect, and adopt existing branches into one. Two commands are deliberately excluded, because they cross boundaries this skill holds elsewhere:

• gh stack submit opens pull requests. gitkit never opens anything. That belongs to whichever skill authors pull requests.
• gh stack merge merges them. gitkit never merges, and a stack merge is a cascade that lands several pull requests at once, which needs a human confirmation naming every one of them. That belongs to whichever skill owns merging.

Naming the owner rather than the command is the point: gitkit prepares ground, and publishing is somebody else's job whether the branch is stacked or not.

The cost of depth

Every layer added to a stack is one more pull request to review and one more branch to rebase each time anything below it changes. Restacking is cascading, so a change to the bottom layer rewrites every layer above it.

No number is stated here on purpose. How deep a stack should go depends on how big each layer is and how fast they get reviewed, which is the human's call. State the mechanism when they ask, so they can size it themselves.

Containers and remote boxes

Git worktrees store absolute paths. If a repo is bind-mounted into a container at a path different from the host's, every worktree inside it breaks.

• Mount both $WORKTREE_ROOT and the main checkout. A worktree's real git directory lives under the main repo's .git/worktrees/; mounting only the worktree yields a broken checkout.
• Mount both at the same absolute path inside and outside the container. This is the robust answer.
• Second belt, on git 2.48+: git config worktree.useRelativePaths true and git worktree repair --relative-paths. Relative pointers survive remapping as long as the repo and the worktree root keep their relative positions, which argues for siding them, e.g. ~/code/<repo> and ~/worktrees/<repo>.

A GUI dev tool that manages worktrees is welcome to observe these. Orca and its peers read git worktree list and typically have a setting to surface externally-created worktrees. That is a one-time UI configuration on whichever machine runs the GUI, and it is the only asymmetry between machines.

No skill may create or remove a worktree through a vendor CLI: the moment one does, the two machines diverge. A vendor's own metadata about a worktree, such as the issue its card links to, a status, or a comment, is a different layer and is fair game, because it exists only inside that tool and has no git equivalent to diverge from. A companion skill may reconcile that layer (orcakit does, for Orca) as long as the worktree itself is still created and destroyed here. gitkit never calls in that direction: it must keep working on a machine where no such tool is installed.

Notes

• gitkit is the single source of truth for what it owns. A calling skill may hold policy about when to ask, such as that a PR is worth pulling down or that a branch is ready to publish, but never about how the git operation is done.
• Restating a conclusion is allowed; restating the derivation is the bug. Every public skill is installed on its own, into repos where gitkit may not exist, so a caller that needs a git fact has to carry enough of one to keep working, and that requirement outranks tidiness. The line runs between the two: a caller may state gitkit's answer together with its degradation fallback ("gitkit owns the sync rule, and it resolves to rebase; without gitkit, rebase onto the base"), and may not reproduce the reasoning that produces the answer, meaning the full base-ref ladder, the worktree path formula, or the rebase-versus-merge argument. The answer is one line that a rename cannot silently invalidate; the derivation is the thing that drifts. Compliant examples worth reading: prkit and mergekit each name the rebase conclusion and its one-confirmation rule without re-deriving them, and wikikit and designkit each name gh repo view --json defaultBranchRef as a base-ref fallback without carrying the five-rung ladder.
• gitkit never implements, commits, or opens anything. It prepares the ground and tears it down. Writing code in the worktree, committing, and publishing belong to other skills.
• Destructive steps preview and confirm. Removing a worktree, deleting a branch, and rebasing a published branch each show what is about to happen and wait for an OK. Creating, adopting, and rebasing an unpushed branch are cheap and reversible, so those run straight through. The line is whether anything outside this machine points at what you are about to rewrite, not which command you typed.
• No shell available (e.g. a browser-based agent)? Then you cannot run git. Reason from what the user provides and print the exact commands as a codeblock for them to run, and never report a worktree created or removed that you could not perform.