review-changes · git:20261117.4a52dff · 2026-11-17 · sha256 93d214a31144e46d

review-changes git:20261117.4a52dffA

Immutable. This exact content never changes and is served at /api/v1/blob/93d214a31144e46d.

No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_laxezz7mapsg7erw/93d214a31144e46db644ff2d7de44ebfc78a80a0c626f81927184384a7abafe1.json

Then /plugin install review-changes@markdownregistry-laxezz7mapsg7erw in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.

---
name: review-changes
description: Review working-tree, branch, or PR changes on separate axes — Standards (repo conventions + a smell baseline), Spec (does it do what the active feature spec asked), and Security when the diff touches a trust boundary — with blast-radius and test-coverage checks. Graph tools accelerate when present; never required.
argument-hint: "[file | function | branch | PR number]"
---

# Review Changes

Separate axes, never merged. Code can pass one and fail another, and merging lets one mask the other.

## Inputs

- The diff: no argument -> working tree vs the default diff base; file/function -> that target and its impact; branch -> vs `main`/`master`; PR number -> the checked-out PR branch or local diff.
- `.safe-code/context/code-standards.md` (documented repo standards).
- The active spec, if any: `.safe-code/context/feature-specs/<NN>-*.md` with `status: in-progress` (or the one the user names).
- Graph accelerators when a codegraph index is ready (`codegraph sync` first): `codegraph impact <symbol>` for high-risk changes, `codegraph callers <symbol>` for changed signatures, `codegraph affected <changed files>` for the tests the diff reaches (over-reports — the safe side). Without them: `git diff`, `rg`, direct reads. Read full source only for changed or high-impact files.

## Axis 1 — Standards

Does the change follow the documented standards, and does it avoid the smell baseline?

Smell baseline (Fowler), always carried against the diff: Mysterious Name · Duplicated Code · Feature Envy · Data Clumps · Primitive Obsession · Repeated Switches · Shotgun Surgery · Divergent Change · Speculative Generality · Message Chains · Middle Man · Refused Bequest. Plus two structural smells: a pure helper living in a heavy module drags that module into every importer (move it to a leaf module); two branches that check the same state with near-identical conditions leave a gap between them.

Binding rules: a documented repo standard always overrides the baseline; every smell is a labelled judgement call ("possible Feature Envy"), never a hard violation; skip anything tooling (lint, type-check, formatter) already enforces.

Test shapes to reject on this axis: **implementation-coupled** (mocks internal collaborators or asserts through a side channel — the tell is that it breaks on refactor while behaviour is unchanged); **tautological** (the assertion recomputes the expected value the way the code does, so it passes by construction — expected values must come from an independent source: a known-good literal, a worked example, the spec); **horizontally sliced** (all tests first, then all implementation — it verifies imagined behaviour); **substring-matched** (`contains("valid")` also passes on "invalid", or fails on a formatting change — assert the exact value or structure); **simulated-only** (fake timers or jsdom can hide bugs that exist only in the real runtime — ask for one real-runtime probe).

Evidence checks:

- A performance claim ("faster", "lighter") needs a before-baseline measured under the same conditions; without one it is labelled `hypothesis`. "Fewer lines" is not performance evidence.
- A changed setting that several paths share (timeout, pool size, retry count) is checked against each path's time budget using real latency, not just the path that motivated the change.
- Test depth by change size is allowed only when the project declares tiers in `code-standards.md`; never pick a lighter tier yourself — a short review compresses the report, never the verification.

## Axis 2 — Spec

Only when an active spec exists; otherwise report `Spec: no active spec`. Does the change implement what the spec asked?

- **Missing**: a requirement in the spec with no corresponding change.
- **Scope creep**: a change no requirement asked for (note it; it may be prefactoring, which is fine if labelled).
- **Wrongly implemented**: the requirement is addressed but the behaviour differs from the spec's Design / Verify When Done.
- **Reinvention**: the spec or change adds a new module, table, or harness where something existing already does most of the job — name the existing one.

## Axis 3 — Security

Only when the diff touches a trust boundary (input handling, queries, file paths, shell, outbound requests, auth, permissions, secrets, tenant scope, dependencies or CI, LLM or agent tool input); otherwise report `Security: no boundary touched`. Under safe-code, follow `references/security.md` (Part 2). Standalone, for each touched boundary name the lower-trust party, the control meant to stop misuse and where it is enforced, and the concrete result if it fails.

- `confirmed` — the path from input to result is traced in source; severity by demonstrated impact.
- `needs-validation` — depends on a named fact outside the repo; no severity.
- `hardening` — a missing extra layer with no reachable violation; never called a vulnerability.

No named party, boundary, and result -> not a finding. Never probe a deployed service or real accounts.

## Output

```text
## Standards
- [severity] path:line - issue, impact, fix
Worst issue: <one line, or none>

## Spec
- [missing | creep | wrong | reinvention] <requirement or change> - detail
Worst issue: <one line, or "no active spec">

## Security
- [confirmed <severity> | needs-validation | hardening] path:line - party, boundary, result, fix
Worst issue: <one line, or "no boundary touched">

## Summary
- Risk: <low|medium|high>
- Blast radius: <files/functions/flows>
- Tests checked: <commands or graph coverage>
- Open questions: <or none>
```

Findings first, ordered by severity **within** each axis; never re-rank across axes. Report only actionable findings.