branch-agent

name	branch-agent
description	Creates a git branch from origin/<default> with no upstream tracking. Auto-names the branch from staged or unstaged changes in the working tree. Use when the user asks to create or start a new branch.
allowed-tools	Bash(git *), Bash(date *), ToolSearch, AskUserQuestion, ExitPlanMode
argument-hint	[branch-name] (optional) — omit to auto-generate from uncommitted changes using <type>/<scope>-<description>
disable-model-invocation	true
model	haiku

Create a new branch from the latest origin/<default> with no upstream tracking ref. When called with no argument and the working tree has uncommitted changes, the branch name is auto-generated from those changes. A -YYYY-MM-DD date suffix is always appended to the final branch name so branches sort and group chronologically. Follow these steps in strict order. STOP immediately after step 6.

When not to use

Does not commit, push, or create PRs — use commit-agent or pr-agent for that.

Step 0: Exit Plan Mode

If currently in plan mode, call ExitPlanMode first and silently before any other action — Branch creation is a git mutation and cannot proceed inside plan mode. Skip this step entirely when not in plan mode. ExitPlanMode is a deferred tool — use ToolSearch with select:ExitPlanMode first, then call it silently.

Step 1: Guards

Run all checks before proceeding. Stop on the first failure.

Not a git repository: Run git rev-parse --is-inside-work-tree. If it fails, output: "Not a git repository." and STOP.

Detached HEAD: Run git branch --show-current. If the output is empty, output: "Cannot create branch: repository is in detached HEAD state. Checkout a named branch first." and STOP.

No origin remote: Run git remote get-url origin. If it fails, output: "branch-agent requires a remote named 'origin'." and STOP.

Step 2: Resolve Branch Name

Read $ARGUMENTS.

Case A — $ARGUMENTS is empty or whitespace-only: Run git status --porcelain=v1.

• If output is empty (clean working tree): output "Provide a branch name. Example: branch-agent feat/login-fix" and STOP.
• If output is non-empty (working tree has changes): auto-generate the branch name as described in Step 2a, then proceed to Step 2b.

Case B — $ARGUMENTS contains spaces or reads as a descriptive phrase: Convert it to a human-readable slug — lowercase, replace spaces and special characters with -, collapse consecutive dashes, strip leading/trailing dashes. Keep the user's words whole; if the slug exceeds 60 characters, drop trailing words (never chop mid-word). If a single word alone exceeds 60 characters (so no word boundary exists to trim at), hard-truncate that word at 60 characters as a last resort. Example: "add allowed tools to skills" → "add-allowed-tools-to-skills". Use the slug as the branch name and proceed to Step 2b.

Case C — $ARGUMENTS is already a valid branch name (no spaces): Use it verbatim as the branch name. Do not slugify, abbreviate, or transform it. Proceed to Step 2b.

Step 2a: Auto-Generate Branch Name from Changes

Use this format: <type>/<scope>-<description> (or <type>/<description> if the scope is omitted). Total length ≤ 60 characters — this reserves 11 chars for the -YYYY-MM-DD suffix appended in Step 2b so the final branch name stays within 72 chars.

Type inference (first match wins):

1. Only markdown / docs/** / README* changed → docs
2. Only test files (**/test/**, **/tests/**, *.test.*, *_test.*, tests/fixtures/**) → test
3. Only CI configs (.github/workflows/**, .gitlab-ci.yml, .circleci/**) → ci
4. Only build/dependency manifests (package.json, pnpm-lock.yaml, Cargo.toml, pyproject.toml, etc.) → build
5. Diff is pure renames/moves with no logic delta → refactor
6. New files added under source dirs → feat
7. Existing source files modified, diff < 20 lines → fix
8. Existing source files modified, diff ≥ 20 lines → feat
9. Otherwise → chore

Scope inference:

Run git status --porcelain=v1 and group changed paths by their first path segment. Pick the group with the most files and use that segment as <scope>. If the top group contains ≤50% of changed files, OR more than 2 groups contain files, omit the scope entirely (use <type>/<description>).

Description inference:

From the changed file basenames and git diff --stat, write the description as a short verb-led phrase (3–7 words) that reads like a commit subject a human would write: start with an imperative verb (add, fix, update, remove, rename, improve, …) followed by what changed. Lowercase, hyphen-separated, alphanumeric only. Strip non-alphanumeric characters; collapse repeated hyphens; trim leading/trailing hyphens.

Readability rules:

• Use whole dictionary words — never abbreviate a word to save space (validation, not valid or val; config, not cfg)
• Prefer describing what the change does over listing filenames
• If the name runs long, drop the least important trailing words instead of shortening individual words

Examples:

Change	Good	Bad
New login validation in src/auth/	feat/src-add-login-form-validation	feat/src-login-form-valid
Typo fixes across README files	docs/fix-readme-typos	docs/rdme-typo-fx
CI workflow updated to Node 22	ci/update-workflow-to-node-22	ci/wf-node22

Validation:

• Lowercase only; characters in [a-z0-9/-]; no leading/trailing hyphens
• Total length ≤ 60 chars (drop trailing description words to fit; never chop mid-word; never truncate the type or scope)
• Must contain a / separator after the type

If validation fails, regenerate once with chore as the type and a shortened description. If it still fails, fall back to chore/auto-branch and proceed.

Output one line before continuing:

Auto-generated branch name from working tree changes: <branch>

Then proceed to Step 2b.

Step 2b: Append Date Suffix

Run:

date +%Y-%m-%d

Append the result to the resolved branch name with a - separator, producing the final branch name: <branch>-<YYYY-MM-DD> (e.g. feat/login-fix → feat/login-fix-2026-04-17). This always runs, regardless of whether the name came from Case A, B, or C.

If the final name exceeds 72 characters, drop trailing words from the description portion until it fits. Never chop mid-word, and never truncate the date suffix, the type prefix, or the scope segment.

Use this date-suffixed name as <branch> for the rest of the flow. Proceed to Step 3.

Step 3: Detect Default Branch

Run:

git symbolic-ref refs/remotes/origin/HEAD --short 2>/dev/null

Strip the origin/ prefix to get the default branch name.

If that fails, run:

git remote show origin | grep 'HEAD branch'

Extract the branch name after HEAD branch:.

If both fail, try git rev-parse --verify --quiet main, then git rev-parse --verify --quiet master. Use the first that succeeds.

If none resolve, report the git error verbatim and STOP.

Step 4: Fetch Latest from Origin

Run:

git fetch origin <default>

On failure (offline, network error, auth required): report the git error verbatim and STOP. Do not proceed with a stale ref.

Step 4.5: Detect Checkout Conflicts

Compute the intersection of (a) tracked files modified in the working tree and (b) files that differ between HEAD and origin/<default>:

bash -c 'comm -12 <(git diff --name-only HEAD | sort -u) <(git diff --name-only HEAD origin/<default> | sort -u)'

(The bash -c wrapper ensures process substitution works regardless of the invoking shell.)

• Empty output → no stash needed; uncommitted changes will carry forward to the new branch automatically. Proceed directly to Step 5.
• Non-empty output → the listed files would be overwritten by git checkout; a stash is required. Record the conflicting paths and set "stash needed". Proceed to Step 5.

Note: untracked files (?? in git status) never appear in git diff --name-only HEAD and never conflict — they always carry forward untouched.

Step 5: Create Branch with No Upstream Tracking

The --no-track flag prevents git from setting origin/<default> as the upstream. Without it, any future git push would target the wrong remote ref.

When stash is not needed (Step 4.5 returned empty):

git checkout -b <branch> --no-track origin/<default>

On failure (branch already exists or other error): report the git error verbatim and STOP. Do not retry. Do not force.

When stash is needed (Step 4.5 returned conflicting paths):

Run each command separately so failure is detectable at each step:

git stash push -m "branch-agent auto-stash <branch>"

On failure: report the error verbatim and STOP.

git checkout -b <branch> --no-track origin/<default>

On failure: report the error verbatim. Note that the stash is preserved — run git stash pop to restore uncommitted changes. STOP. Do not retry. Do not force.

git stash pop

On failure (pop conflict): report the git error verbatim, then output:

Your uncommitted changes are safe in the stash. To restore manually:

1. git stash list — confirm the stash entry is present
2. Resolve any conflicts in the listed files
3. git add <resolved-files> — stage the resolved files so the index reflects the resolved state
4. git stash drop to remove the stash entry Do not retry. Do not drop the stash before staging resolved files.

STOP.

Step 6: Confirm and STOP

Run git rev-parse --short HEAD to get the short SHA.

Output one line:

Created and checked out <branch> from origin/<default> @ <sha> (no upstream tracking)

---

STOP here. Do not stage, commit, push, create PRs, or take any further action.