branch-review

Branch Review Skill

Pre-push sanity check. Runs correctness gates first (Pass 0), then invokes the three review skills against the branch diff (origin/main...HEAD), collects their native outputs verbatim, appends a merge-by-file index. Use before git push / PR.

Do not auto-fix. Each sibling refuses to auto-edit; this skill preserves that.

Sibling to /launch-prep — that skill runs the full release pipeline (CHANGELOG, version bump, PR to main); this one is the lightweight pre-push counterpart.

Commands

• /branch-review — default: Pass 0 + /code-cleanup + /test-review (docs skipped)
• /branch-review --full — includes /docs-review
• /branch-review --fast — skips Pass 0's test run; uses --recent on code-cleanup where useful
• /branch-review --code-only / --tests-only / --docs-only — single-skill shortcut
• /branch-review --base=<ref> — non-default base (default: origin/main)
• /branch-review --include-dirty — add staged+unstaged to the diff scope

Preflight

1. Refresh base. git fetch origin main --quiet (or <base>). Stale local main re-surfaces upstream-merged commits.
2. Git state check. Refuse on detached HEAD or active rebase/merge (git status --porcelain=2 --branch → look for in progress). Warn if working tree is dirty; offer --include-dirty to add staged+unstaged to scope, else diff is branch-only.
3. Compute diff. git diff --name-only origin/main...HEAD (+ git diff --name-only HEAD if --include-dirty).
4. Scale gate.
   • Empty diff → exit with "no changes vs base".
   • ≤30 files → proceed as normal.
   • 31–60 files → proceed but force --fast and print a "large branch" warning.
   • >60 files → refuse. Print the partition counts and suggest: narrow base (--base=<recent-commit>), scope to subtree (--paths=src/benchflow/agents), or invoke the siblings directly.

Partition

Route changed files by path:

• tests → /test-review: tests/**/test_*.py, tests/**/*_test.py
• src → /code-cleanup: src/benchflow/**/*.py
• docs → /docs-review: *.md, docs/**/*, .claude/dev-docs/**/*, src/benchflow/**/*.md

Skip silently (no routing, no findings, no warning): uv.lock, .venv/**, __pycache__/**, *.egg-info/**, dist/**, build/**, .pytest_cache/**, generated files.

Flag in rollup but do not review — config changes the author should eyeball themselves: pyproject.toml, .pre-commit-config.yaml, .github/**, .claude/**, ruff.toml, pytest.ini.

Renames: if git diff -M --name-status shows R100 (pure rename, no content delta) → surface as rename, route nothing. R<100 → route the new path normally.

Test-orphan detection: if src/benchflow/X.py changed but tests/test_X.py (or tests/**/test_X.py) did not, include the matching test file in the test partition. Benchflow's test layout is flat under tests/ with some subdirs — glob for test_<basename>.py rather than assuming co-location. Do not extrapolate beyond direct name match.

Pass 0 — correctness gates (unless --fast)

Run in parallel Bash calls. CI gates all four per AGENTS.md, so this skill mirrors CI:

• .venv/bin/ruff format --check src/ tests/ — formatting.
• .venv/bin/ruff check src/ tests/ — lint.
• .venv/bin/ty check src/ — typecheck (whole-src, cheap).
• .venv/bin/python -m pytest <changed test files + orphan-detected siblings> — scoped to the test partition, not full suite. Unit only; do not pass -m live (requires Docker + API key, not appropriate for a pre-push gate).

With --fast: skip the pytest run only. Keep ruff + ty — they're fast enough to always run.

If any gate fails, surface at the top of the report as BLOCKERS and short-circuit the review: there's no point polishing a branch that doesn't lint, compile, or pass tests. Author fixes those, then re-runs.

Pass 1 — review skills

Invoke the enabled siblings via the Skill tool in the main thread, one after another. Pass each the space-separated partition as the argument. Each sibling already runs its internal subagent fan-out in parallel, so within-skill concurrency is preserved; the orchestrator just waits on each skill's native report.

Example invocations:

• Skill("code-cleanup", "<src paths>")
• Skill("test-review", "<test paths>")
• Skill("docs-review", "<doc paths>") (only if --full)

With --fast, prefer Skill("code-cleanup", "--recent") when the branch touches >10 src files — it caps subagent work. test-review and docs-review take explicit path lists; pass the partition directly.

Collect the three native outputs verbatim. Do not summarize, rewrite, or translate their vocabularies — each skill's bucketed structure (cleanup's 7 categories, test-review's delete/collapse/loosen/add, docs-review's drift/stale/duplicate) encodes its calibration. Preserving it lets the author re-use their mental model from direct invocations.

Output

[If BLOCKERS from Pass 0]
## Blockers
- ruff format: <files needing reformat>
- ruff check: <rule violations, file:line>
- ty: <type errors, file:line>
- pytest: <failing test names>

[Else continue to reviews]

<verbatim code-cleanup report>

<verbatim test-review report>

<verbatim docs-review report, if --full>

## Merge-by-file index
Author-oriented cross-reference. Does NOT re-rank; lists which skills flagged which files so co-located fixes surface together.

### src/benchflow/agents/registry.py
- [cleanup] 2 findings (1 verified, 1 needs-judgment)
- [test] tests/test_registry.py: 1 delete, 1 loosen

### docs/architecture.md
- [docs] 1 blocker, 2 stale

### Flagged config (not reviewed)
- pyproject.toml, .github/workflows/ci.yml

...

End with a one-line rollup: N files reviewed · <cleanup-count> cleanup · <test-count> test · <docs-count> docs · <blocker-count> blockers.

Anti-patterns

• Don't re-verify findings. Each sibling runs its own verification pass. Re-auditing here is duplication.
• Don't re-rank or translate vocabularies. The skills calibrate on their native buckets; flattening to a common severity destroys that signal.
• Don't expand scope beyond test-orphan detection. If the author wanted arch-level review, they'd invoke /arch-audit directly.
• Don't run at repo scope. That's the scale gate's job — enforce it, don't silently accept 76-file branches.
• Don't skip Pass 0 outside --fast. A clean review of uncompileable code is worse than useless.
• Don't run pytest -m live. Live tests need Docker + an API key; pre-push gates should be hermetic and fast.
• Don't re-run immediately after fixes. Use the sibling skills' scoped modes (<path> argument, --recent) to re-verify just the touched files.
• Don't duplicate /launch-prep. If the user is cutting a release (version bump, CHANGELOG, PR to main), route them there instead.