Find interface regressions in a change

interface-review reviews a change and the surfaces it affects. It resolves the target, reads both sides of the full diff, expands changed files to their consumers and classifies each finding by cause. It owns the change scope and report structure, while domain rules come from the better-* skills and severity, coverage, consolidation, caps and verdicts come from better-interface. General correctness, tests, security and performance remain with the project's code review.

An explicit invocation supplies the target, such as a pull request, branch or range. Without one, the skill first checks whether HEAD is ahead of the default branch's merge base and includes uncommitted changes alongside that range. Otherwise it reviews dirty working-tree changes. It records the base commit and distinguishes committed work from uncommitted files. Lockfiles, snapshots, generated output, vendored code and binaries are excluded and named.

A clean tree with no unmerged work is not silently replaced by the last commit. The workflow states the facts and offers an open pull request, an explicitly named last commit, another target or a separate repository-wide interface audit. An empty scope never receives approval.

Changed files expand into the surfaces they render. Direct importers and callers form the first hop; shared tokens, themes and primitives may require another. Expansion is limited to five consumers across both hops, with skipped consumers disclosed. Removed lines receive particular attention because regressions may disappear from the post-change view. A removal becomes a regression only when the change does not replace it and the owning domain confirms the loss.

Findings are Introduced when created by the change, Regression when previously correct behavior is weakened, or Pre-existing when the change did not worsen it. Classification follows causation rather than proximity to a changed line. Blame can help establish the origin:

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

For uncommitted work, blame omits the range; Not Committed Yet or a commit within the reviewed range identifies changed lines.

The review also checks stated intent from the pull request, linked issue and commits. It looks for incomplete state coverage, missing maintained translations, absent empty or loading states and controls omitted from sibling surfaces. Pre-existing issues remain a limited courtesy rather than becoming a new codebase audit.

Use it to ask whether a specific interface change introduced problems or weakened existing behavior. Use better-interface directly for repository-wide review. The working tree remains read-only: fetching refs is permitted, but switching branches, checking out a pull request or stashing the author's files is outside the review workflow.

Rendered verification is opt-in unless a cheap project preview exists. Working-tree and branch targets can render in place; other refs use an isolated worktree. Visual and runtime claims remain Not verified when they were not exercised.

Inputs include the target diff, relevant intent and affected consumers. The report opens with target, base and head refs, commit and file counts, exclusions and expanded surfaces. Coverage follows, then findings with severity, domain, status, location, before, after and impact. Up to three highest-severity pre-existing findings are explicitly separated from the change's responsibility. Domains with no evidence are marked Not reviewed: no evidence in the change scope. With no Introduced or Regression findings, the report states that no actionable interface findings were found and omits the empty table. Verification and verdict follow better-interface's format.

better-interface supplies the final review framework, and better-* domain skills judge domain-specific issues. If better-interface is unavailable, the skill reports the resolved scope and file inventory and stops instead of inventing a severity scale or verdict. The workflow is adapted from jakubkrehel/skills under the MIT license.