Skip to content

aspiers/ai-config / .agents/skills/git-branch-management

git-branch-management

Verified

Structures git branches and worktrees: keeps unrelated changes on separate branches, isolates each in its own worktree created so the repository's setup runs, records dependencies with git-machete, and recombines them into a temporary mixdown branch for local testing. Use when starting any fix or change, diagnosis included, in a codebase that works on more than one branch or in a fork, clone or third-party checkout, before reading code for it; before creating any branch or worktree; when choosing what a new branch should be based on; when a branch is growing a second unrelated concern; when deciding whether to stack branches or keep them independent; when preparing a change for a pull request or patch submission; when handed an existing worktree or branch to implement and commit in a fork; or when a local build needs several in-progress branches combined. Not needed where the repository clearly commits everything to the trunk, or holds notes or personal data rather than code.

Git Branch Management

Prepare every change in the form that is cleanest for upstream submission, then recombine locally for testing and normal use.

First establish upstream ownership

These principles exist to satisfy a maintainer you do not control. Check who owns the upstream before applying any of them.

Inspect the remotes:

git remote -v

The usual convention names origin for the upstream and github for the user's own fork of an upstream owned by someone else. Where it holds, a github remote alongside origin suggests an upstream the user does not own.

A repository the user owns outright may have no origin at all — its only remote can be github, pointing at their own account. So the absence of a fork remote is not by itself a signal either way.

It is a convention rather than a rule, so read it as a hint about which URLs to check, not as the answer. Ownership comes from the URLs themselves — compare them against the user's own accounts and hosts. Ask if it stays unclear.

Where the user owns the upstream there is no external reviewer to satisfy, so clean separation stops paying for itself: combining concerns on one branch, or pushing straight to the trunk, is legitimate and often preferable. Follow the user's preference for that repository rather than imposing separation. Do not spend effort splitting branches that nobody else will review.

Why separation matters

Upstream reviewers accept or reject one concern at a time. When two unrelated changes share a branch, the contentious one blocks the uncontentious one: a maintainer who wants the bugfix must also litigate the feature. Keeping them apart means each can land on its own schedule, and a rejected change never strands an accepted one.

This is also why mixing beats arbitrary stacking. Stacking makes branch B depend on branch A even when the two changes are unrelated, so B cannot be submitted until A lands. Stack only where a real dependency exists; combine independent branches with a mixdown instead.

Isolate each branch in its own worktree

Set up the worktree as soon as the user asks for a fix or change, before diagnosing it or reading code for it. Diagnosis is part of the work, and a session that starts reading in the main checkout tends to stay there.

Develop each independent branch in its own worktree, so branches do not contend for a single checkout and each keeps its own build state.

Let the repository's own setup run

A new worktree is only usable once the repository's setup has run: dependency install, env files, generated code. Each tool runs only its own setup:

  • wt switch --create runs hooks from .config/wt.toml and the user's worktrunk config, nothing else;
  • Orca's worktree create runs the setup from the repository's orca.yaml (scripts.setup) or from Orca's settings for that repository;
  • raw git worktree add runs neither, so never use it to start work.

Orca can only create a worktree when you are running inside it, which its terminals signal by setting ORCA_PANE_KEY. So:

  • inside Orca, in a repository with an orca.yaml or registered in Orca (repo list --json), create the worktree with Orca and start an agent in it, as the orca-cli-local skill describes, even when Orca has no setup script for it, because only Orca gives the new agent its own workspace in Orca's sidebar. Do not work in it from the current session unless the user says so;
  • otherwise use wt, even when the repository has an orca.yaml:
wt switch --create <branch> --base <upstream-base>

--base matters here: without it wt uses the repository's default branch, which is usually right, but an explicitly named base makes the intended starting point unambiguous. See the worktrunk skill for configuration and hooks.

If setup did not run, run it before editing: for a wt worktree in a repository whose setup lives only in orca.yaml, run its scripts.setup from the new worktree, with ORCA_ROOT_PATH set to the main checkout. Never borrow another checkout's dependencies or build output. If the repository has setup for only one tool, or none, propose adding a shared script that both run: see the setting-up-worktree-hooks skill.

Reserve the main checkout for mixdowns and combined testing, matching the default that git-branch-mixer describes.

wt sync rebases worktree branches onto their parents in dependency order, which keeps a set of related branches current against a moving upstream without rebasing each by hand:

wt sync --fetch          # fetch, then sync the current stack
wt sync --all --fetch    # sync every worktree branch

Prefer it over manual per-branch rebases when several branches need bringing up to date. Combine with --prune to clear worktrees whose branches have landed upstream. Note that wt sync and git machete traverse overlap; use whichever the repository is already set up for rather than both at once.

To find which worktrees can go, start with wt list --format=json, whichever tool created them, rather than looping over git worktree list by hand. Each entry's main_state is integrated or empty once its work has landed, even through a squash or rebase, and working_tree flags uncommitted and untracked files. It knows nothing of agents still running there: inside Orca, check those with orca-ide terminal list.

Once a branch has merged into the trunk, remove its worktree and branch in the same step, after checking the worktree is clean, no agent is still working in it, and git clean -ndX lists no ignored file that is useful and not clearly scratch: wt remove <branch>, which keeps a branch holding unmerged work, or git worktree remove <path> then git branch -d <branch>. Inside Orca, remove it as orca-cli-local describes instead, so Orca's sidebar stays in step.

Choosing a base for new work

Base a new branch on the nearest upstream integration branch — typically origin/main or origin/master — not on whatever branch happens to be checked out. Deriving from a local feature branch silently makes that feature a prerequisite of the new work.

Base on another local branch only when the new work genuinely cannot compile, run, or make sense without it. Record that dependency in machete so the relationship is explicit rather than implied by history.

A local checkout's base branch is not always named main; confirm what the remote actually uses before branching.

Integration fixes are real dependencies

When combining two branches needs more than a textual conflict resolution (branch B moves code that branch A changes, so A's change must also be applied in B's new location, or tests from both need reconciling), that is a real dependency, not a mixdown concern. Stack the branch that needs adapting on the other and commit the fix on the child branch. If that branch is not in the machete layout yet, add it under its new parent without asking, unless choosing the parent is itself hard. If it already is, moving it changes where an existing branch sits in the graph, so propose the move and ask first. Rebasing or propagating a rebase along the existing layout needs no confirmation.

Never commit such a fix on working or hide it inside a merge commit. Rerere replays only conflict hunks, so edits made inside a merge, and commits made only on working, vanish at the next full rebuild. Before rebuilding, check for both: commits on working that are on no source branch, and differences between the old and rebuilt working.

The cost of separation is that no single branch contains everything, so a build from one branch lacks the others' features. That is what mixdowns resolve — they are the reason separation stays practical rather than merely principled.

Recording dependencies with git-machete

git machete maintains a branch layout file describing which branches descend from which. It makes the intended tree explicit, shows sync state against upstreams, and drives rebases that respect the structure.

git machete status          # tree plus sync state against parents and remotes
git machete discover        # infer an initial layout from existing history
git machete edit            # edit the layout file directly
git machete add <branch>    # place a branch in the tree
git machete traverse        # walk the tree, syncing each branch to its parent

Every worktree shares one layout file, by default in the main checkout's git directory (.git/machete), unless machete.worktree.useTopLevelMacheteFile is false. Get its path from git machete file rather than building it yourself. If git machete status prints a tree, a layout exists; with none it says "No branches listed" and exits non-zero. Its yellow-edge warning means only that a fork point was inferred, not the layout.

Roots sit unindented; each level of indentation marks a child of the branch above. Independent features are siblings under the same root, never nested inside one another:

master
    feat/one
    feat/two

Reserve nesting for genuine prerequisites. A sibling layout is the signal that two branches may be submitted in either order, and it is what makes them eligible to be mixed rather than stacked. Annotations after a branch name can record PR numbers and per-branch traverse qualifiers; see git machete help format.

Run discover with care in a fork carrying many fetched upstream branches — it proposes a layout from existing history, which may pull in branches you do not maintain. Review its proposal before accepting.

Recombining for testing and daily use

To build, run, or use the combined result, merge the independent branches into a throw-away integration branch with a mixdown. The git-branch-mixer skill owns the ggmx and ggmxd mechanics.

Source-branch scope and mixdown integration scope are independent. Keep each source branch narrowly scoped for submission. The working mixdown must include all configured source branches and always represent their latest best-of-breed combination for day-to-day use and testing.

Never develop or commit new source work on working. Make and commit every source change on its narrowly scoped source branch in that branch's worktree.

In a fork of an upstream the user does not own, with a working branch and mixdown.working config (git config --get-all mixdown.working), add each new source branch to the mix (git mix +<branch>) and to the machete layout (git machete add) on its first commit, without asking. This is only configuration: do it even when the main checkout is on another branch, so the next mixdown includes the branch.

One agent owns each mixdown: the one in the main checkout, which has working (or whichever mixdown branch applies) checked out and normally coordinates the project. It runs remixes one at a time, because parallel mixdowns conflict. An agent in a linked worktree never runs the mixdown and never just reports it as pending: once its work is ready to test, it asks that agent to remix, naming the branches and commits to pick up, and waits for the reply, treating the remix as pending until it arrives. Under Orca, use orchestration send --type question, then wake an idle coordinator as orca-cli-local describes. If no agent runs in the main checkout, ask the user.

Regenerate working only where the user actually runs it, for example checked out in the main checkout and served from there; mix config alone is not enough. Even then, wait until a logically complete piece of work is ready to test, such as when every commit a pull request needs has been made, or a rebase or review fix finishes one. Intermediate commits are not worth testing. Once it is ready, that agent, without asking:

  • regenerate working from the full configured set in the main checkout (git mixdown -c). If a merge conflicts and the fix is plain, such as keeping both of two adjacent edits, resolve it yourself and re-run the mixdown, as git-branch-mixer describes. Ask the user only when the fix needs a design decision, such as moving a branch already in the machete layout under a different parent, or choosing between incompatible behaviours;
  • only after a successful mixdown, rebuild, and only after a successful build, restart whatever serves the main checkout, so the user's own instance runs the new working. Look for it in user service units whose working directory is the main checkout, and in the project's README or agent instructions; ask if nothing turns up. A dev server you started in a feature worktree does not count.

That instance is the user's live one, so a rebuild that is not ready must never take it down: chain these steps with &&, never ;, and check each result before the next.

Never update working from only the branch that changed. Stop and ask only if a mixdown conflict needs a design decision or the mixdown would touch uncommitted changes in the main checkout. If the main checkout is on another branch, ask before switching it, and report the regeneration as pending. Never submit a branch upstream until the user has had the chance to test it in their rebuilt instance.

working is a local scratch branch. Never push it or open a PR from it; publish the independent source branches instead. This boundary keeps the upstream-clean structure intact while still giving you a single working build.

When a branch accumulates unrelated work

Splitting after the fact is more expensive than starting separately, so decide the shape when the work begins. If a branch has already grown a second concern, move the newer concern onto its own branch off the upstream base, record both in machete as siblings, and mix them for testing. The incremental-commits skill covers carving existing changes into reviewable commits.

Before submitting upstream

Confirm the branch contains only its own concern, sits on a current upstream base, and carries no mixdown merge commits. See the submitting-upstream skill for the submission channel and its expectations.

finishing-a-development-branchobra/superpowersUse when implementation is complete, all tests pass, and you need to decide how to integrate the workStars: 322kUpdated Claude CodeCodexVerifiedInstall
systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixesStars: 322kUpdated Claude CodeCodexVerifiedInstall
test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation codeStars: 322kUpdated Claude CodeCodexVerifiedInstall
skill-creatoranthropics/skillsCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.Stars: 197kUpdated Claude CodeCodexVerifiedInstall