Change review

This skill reviews a change rather than a screen. It resolves the scope, expands the changed files to the surfaces they affect, reads both sides of the diff and classifies every finding.

Scope is all it owns. Domain rules belong to the better-* skills. Severity, consolidation, coverage, the cap and the verdict belong to better-interface, which this skill hands the review to.

Correctness, tests, security and performance belong to the project's general code review. Name the concern once and move on.

The change, not the codebase

The author is asking "did I make this worse?". Report what the change caused and stay mostly quiet about what it merely touched. Three pre-existing findings is a courtesy; thirty is a different review and one nobody asked for.

Read the change before forming an opinion of it. The stated intent decides what counts as incomplete, and a skimmed diff produces findings about code the next hunk already fixed.

Core principles

1. Resolve the change scope first

The whole invocation is the target, so /interface-review pr 482 reviews pull request 482. Scope resolution holds the accepted targets and how each resolves.

With no target supplied, resolve in this order and stop at the first match:

1. HEAD is ahead of git merge-base origin/<default-branch> HEAD: that range plus any uncommitted changes, with the commit count and the uncommitted file count stated separately.
2. The working tree is dirty: the uncommitted changes.
3. Neither: there is no change to review. Stop and ask, per With no change, ask rather than invent one.

Order matters. Check the working tree first and one stray formatting edit shadows a twelve-commit branch, with the report still claiming full coverage.

Exclude lockfiles, snapshots, generated output, vendored code and binaries, and name what you excluded. An empty scope after exclusions reaches the same place by a different route.

2. With no change, ask rather than invent one

A clean tree with nothing ahead of the merge base means the user asked to review a change that does not exist. Never fall back to HEAD~1..HEAD on your own. The last commit is whatever happened to land, often a merge, often someone else's work, and a report on it is indistinguishable from a report on what the user meant.

State the repository facts you found, then offer the routes and wait. Nothing to review holds the facts to gather:

• The last commit, HEAD~1..HEAD, named by short SHA and subject, so the user sees what they would get before choosing it.
• A target they name: pr <n>, a branch, a ref, or a range, resolved per Resolve the change scope first.
• A whole-repository interface audit, which is not a change review. Hand it to better-interface as a repository-scope review, without this skill's scope block, statuses, or pre-existing section. With no change, every finding is pre-existing and the classification says nothing.

Check for an open pull request on the current branch before asking, and offer it first. A branch whose commits already landed resolves to no change, while its pull request is still exactly what the user meant.

Where the scope emptied out after exclusions, say which files were excluded and ask the same way. Never report a review of nothing as Approve.

3. A diff is not a surface

A changed file is evidence, not the review subject. Its blast radius is the set of surfaces it renders in; review those.

Expand the blast radius one hop by default: the direct importers and callers. Expand a second hop only for design tokens, theme values and shared primitives, where one line reaches the whole product.

Review at most five consumers, ordered by the rule in Scope resolution, then state how many you did not expand. A sweep with no bound cannot support the coverage it claims, and an unstated cutoff reads as completeness.

4. Read the removed lines

Regressions are invisible in the post-change state. Read the - side of every hunk against Removed signals.

A signal is a lead, not a finding. A removal is only a regression when nothing in the change replaces it, and the domain skill owns that judgement. Route each unmatched removal to its owner, report only what that skill confirms and status it Regression. That tells the author they broke something that worked rather than made a new mistake.

5. Classify every finding

Give every finding one status:

• Introduced: the change created it.
• Regression: the change weakened something previously correct.
• Pre-existing: present in the touched code but not caused by this change.

Status by what the diff touched, not by which file it sits in: a line the change never touched is Pre-existing even three lines from a hunk. Confirm against the base ref when it matters:

git blame -L <line>,<line> "$BASE" -- path/to/file

Hand every finding up with its status attached and let better-interface apply its cap and verdict rules.

6. Hold the change to its stated intent

Read the pull request title and body, the linked issue and the commit messages, then review whether the interface delivers what they claim.

This is what surfaces the incomplete change. A surface review cannot see it, because it inspects the states that are present, and here the point is the ones that are absent:

• A new variant, size, or theme applied to some states but not all: hover, focus, active, disabled, loading, selected.
• A new user-facing string with no entry in the translation catalogue the project maintains.
• A new component with no empty, loading, error, disabled, or narrow-width state.
• A control added to one surface but not to the siblings that already carry its peers.

Do not report scope creep. Whether a change does too much is a process question, not an interface one.

7. Hand the review to better-interface

Hand better-interface the scope block, the affected surfaces and a status on every finding. It routes to the domain skills, applies severity, consolidates, enforces the cap and issues the verdict.

If better-interface is unavailable, report the resolved scope and the file inventory, name it as the missing skill and stop. Do not invent a severity scale, a cap, or a verdict.

8. Never mutate the working tree

A change review is read-only, including the checkout. Fetch pull request refs; never check them out. git fetch writes only to .git and is permitted. gh pr checkout, git checkout, git switch and git stash rewrite the files the author has open. They fail against local edits or discard them, so they are never permitted.

Rendered verification is opt-in. Mark visual and runtime claims Not verified unless the project exposes a cheap preview or the user asks for a rendered review. When they do, use an isolated worktree (git worktree add /tmp/review-<n> refs/remotes/pr/<n>) and remove it when done.

Before you finish

Review output format

Open with the scope block:

The coverage table follows it unchanged. A domain with no evidence in the change scope is Not reviewed: no evidence in the change scope, which is a coverage statement rather than a gap.

Then the findings, with a Status column per Classify every finding:

With no Introduced or Regression findings, omit the table and state "No actionable interface findings in this change."

Then Pre-existing findings, at most three, highest severity first, stated plainly as not this change's responsibility. Omit the section when there are none.

The cap and the verdict cover Introduced and Regression only. Pre-existing findings sit outside the cap, so touching a legacy file cannot turn into a full-file audit. They sit outside the verdict too, so a change whose only findings are pre-existing is an Approve.

End with Block when any HIGH remains and Approve otherwise, leaving the remaining findings in the table as work to do. When better-interface is available, the severity scale and the cap come from it.