New Branch

Creating or switching branches never needs permission. In a checkout other sessions are using, branch in a new worktree instead of switching that checkout's branch under them. Platform-assigned Builder.io and Fusion branches stay where they are. Never move a branch that another worktree has checked out (git worktree list --porcelain), and never stash, reset, or force anything.

Name

<login>/changes-N, for example steve8708/changes-545.

login=$(gh api user --jq .login)
git fetch origin --quiet || { echo "fetch failed: not branching from a stale origin/main" >&2; exit 1; }
last=$(git for-each-ref --format='%(refname:short)' refs/heads refs/remotes/origin \
  | sed -nE "s#^(origin/)?$login/changes-([0-9]{2,6})\$#\2#p" | sort -n | tail -1)
branch="$login/changes-$(( ${last:-49} + 1 ))"

Only your own prefixed branches count, and six digits at most, which skips legacy and timestamp-named branches.

Create

Look at what would come along first:

git status --short --untracked-files=all
git log --oneline origin/main..HEAD

Then pick the base. The git switch commands below are for a checkout only this task uses; in a checkout other sessions are using, create the branch in a new worktree instead (git worktree add -b "$branch" <path> <base>).

• Nothing in the log, or every commit is from a PR that already merged (squash merges leave the source commits off main's history; check with gh pr view <number> --json state): git switch -c "$branch" origin/main.
• Commits that aren't merged anywhere yet, all from this task: git switch -c "$branch" to keep them at HEAD.
• A mix of merged and newer commits, or commits or dirty paths from another task: don't switch. Leave the checkout where it is and report what you found.

Never branch from a local main, which can be stale. If the branch is for an existing PR update, base it on that PR's freshly fetched live head instead of origin/main (see babysit-pr for the push target). Don't push the new branch until it has changes to ship.

After a merge

Rotate only when the PR merged, the tree is clean, and nothing is unpushed.

git fetch origin --quiet || { echo "fetch failed: keeping this branch" >&2; exit 1; }
git merge-base --is-ancestor <mergeCommit.oid> origin/main
git status --porcelain --untracked-files=all -- ':/' ':(top,exclude)learnings.md' ':(top,exclude)bridge/**' ':(top,exclude)data/**'
git log --oneline <ship_merge_head_oid>..HEAD

• mergeCommit.oid comes from gh pr view <number> --json mergeCommit. If it isn't in origin/main yet, wait a few seconds and fetch again.
• ship_merge_head_oid is the head /ship saved before merging. The git log must print nothing, which means no commits landed after the merged head. Also confirm the PR's final head is the one that merged: gh pr view <number> --json headRefOid --jq .headRefOid must print ship_merge_head_oid. This works even when the source branch was deleted or lives on a fork. If the command fails, keep the branch.
• git cat-file -e "<ship_merge_head_oid>^{commit}" and git merge-base --is-ancestor <ship_merge_head_oid> HEAD must both succeed first, which proves this checkout holds the merged head. If the saved OID is missing or invalid, an empty git log proves nothing, so keep the branch.
• If any of these commands fails, or anything is dirty or unpushed, keep the branch and list what you found.

When it's all clear, compute $branch as above and run git switch -c "$branch" origin/main. Always use origin/main here, never the HEAD-keeping branch from "Create": a squash merge leaves the source commits off main's history, so HEAD would carry already-merged work. In a shared checkout, don't rotate; leave its branch as it is.