Branch: QA Review

Assess branch readiness for PR submission: full review methodology plus the checks only a local checkout allows

SKILL.md

GitHub

Assess whether this branch is ready to be submitted as a reviewable pull request. Three layers: the mechanical facts, the gate run and the same review methodology pr-review applies after the PR exists. Anything that would surface in the PR review should surface here, one step earlier.

Step 0: Facts

Gather the exact numbers first: "$HOME"/.claude/library/scripts/branch-facts.sh $base (blank $base = the script's default, main) emits JSON (ahead/behind, conventional-commit and branch-name compliance, WIP commits, diff size, conflict markers, TODOs and console.logs added, test files touched, svu bump). Judge from these facts; the steps below are the judgement layer, not fact-gathering to repeat by hand.

Step 1: Run the gate

Discover the project's test, typecheck and lint commands from package.json (or the ecosystem equivalent) and run all three. This is the one check only a local checkout can perform, the whole reason to review before the PR instead of after, so never skip it or substitute a static skim of test files. Distinguish failures the branch caused from pre-existing breakage (run the same check on the base branch when unclear); pre-existing breakage is reported, never blamed on this branch.

Step 2: Review the code with pr-review's methodology

Load the pr-review-dry_run skill and run its full methodology against this branch: foci (correctness, security, conventions, reinforcement), taxonomy (🔴/🟠/🟡/🟣), scope classification and its conventions-research step (CLAUDE.md, .claude/**/*, docs/). The local diff git diff <base>...HEAD (the same $base, default main) substitutes for its gh pr diff step; everything else applies unchanged.

Do not re-derive or abbreviate that methodology here. A thinner duplicate of it is exactly what this skill used to carry, and it routinely missed what the PR review then caught: structured per-focus passes find what "check for obvious bugs" skims past.

One addition the diff alone can't show: for every changed export, function signature or component prop, Grep for its callers and confirm they still hold. Changed-contract breakage lives outside the diff.

Step 3: Branch mechanics

From the Step 0 facts, briefly:

  • Conventional-commit compliance, atomicity, no WIP/fixup/junk commits
  • Branch name: <prefix>/<short-description>, prefix from the canonical set (matches CLAUDE.md §8.6): feat, fix, enhance, refactor, test, docs, config, chore, ci, deps, hotfix, spike, agents; breaking-change branches use <prefix>/breaking-<description>

Step 4: Breaking changes

  • Removed or renamed exports, functions, types or components
  • Changed function signatures (required parameters added/reordered)
  • Modified return types in a widening direction
  • Database schema changes (columns removed/renamed, constraints changed)
  • HTTP API changes (routes, methods, request/response shapes)
  • New required environment variables
  • Changed component props (removed, renamed, type-changed, newly required)

Verdict: defined, not vibed

Verdict Criteria
Blocked Unresolved conflict markers, or a red gate that predates the branch (needs a human call before any PR makes sense)
Needs Work Any 🔴 finding (per pr-review-dry_run's <blocking-gate/> closed trigger list, not a severity judgement made here), a red gate this branch caused or WIP/fixup commits
Ready Green gate and nothing above 🟠/🟡; remaining findings are listed as improvements, not blockers

Output Format

Verdict

Ready | Needs Work | Blocked, with the one-line reason from the mapping above.

Gate

Test / typecheck / lint results, and whether any failure predates the branch.

Findings

The Step 2 findings in taxonomy order, file and line references included. Omit if none.

Branch Health

Branch name compliance, commit message quality, commit atomicity.

Breaking Changes

Any detected, with context. Flag format:

⚠️ Breaking change: consider feat!: or BREAKING CHANGE: footer

Omit if none.

Recommended Next Steps

Ordered list: blockers first, then improvements, then nice-to-haves.


If the verdict is Ready, offer to run the pr-create skill immediately, forwarding base <branch> when $base isn't main so the PR opens as a stacked layer on the same base this review judged, rather than silently targeting main.