mcpField
C
code-review

code-review

github.com/sjuneparkTesting & quality171 installsUpdated 3w ago
Repository

About code-review

code-review is a skill for bounded, evidence-based reviews of code changes and governed implementation. It reports actionable findings, applies safe in-scope fixes by default, and distinguishes issues requiring a decision from speculative concerns.

Key features

  1. 01Review a named target or infer local changes
  2. 02Trace requirements and consumer paths
  3. 03Report evidence-backed actionable findings
  4. 04Apply narrow, safe fixes by default
  5. 05Recheck fixes and affected validation

Use cases

  1. 01Review a diff, commit, branch, or pull request
  2. 02Assess implementation against a specification
  3. 03Find contract issues across changed interfaces
  4. 04Check a safe fix and its validation

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/sjunepark/agent-scripts --skill code-review

After installing, type/code-reviewin Claude Code to confirm it loaded.

Bundled files

References and scripts installed along with SKILL.md.

agents/
openai.yaml
evals/
evaluation-2026-08-23.md
lenses/
design.md
diet.md
implementation-review.md
system.md

Details

From the project README

Code Review

Deliver one bounded review with evidenced, actionable findings or a defensible clean result. Apply the implementation, system, design, and diet perspectives proportionately; no finding is preferable to a speculative one.

Scope and guidance

Use the named files, subsystem, branch, commit, PR, or diff. A roadmap, specification, issue, ADR, or migration review covers its governed implementation, not just the document. Establish the intended outcome, claimed milestone, implementation range, and edit policy. Without a named target, use uncommitted changes when present; ask only if no target can be inferred.

Read the implementation lens for the core review and safe-fix policy. Load the other detailed lenses when their dimensions are material to the target:

  • System: initiative delivery, technical choices, or cross-cutting consequences. Check that the requested outcome is actually delivered.
  • Design: changed responsibilities, interfaces, dependencies, or consumers. Check that contracts and invariants have clear owners.
  • Diet: added concepts, states, wrappers, compatibility paths, or other complexity. Check that the machinery earns its maintenance cost.

For a routine local edit, these can be brief checks without loading every lens or expanding into a repository-wide audit. Follow affected requirements and consumer paths; separate introduced issues from pre-existing debt.

Findings and fixes

Record each issue once, where its remedy belongs, with a trigger scenario, affected path or governing requirement, consequence, evidence, and likely action. Architectural taste and hypothetical flexibility are not findings.

  • Bucket I — Safe fixes: narrow, in-scope, mechanically verifiable changes without unresolved product, design, rollout, compatibility, or risk judgment. Apply these by default unless the user requested review-only; then mark proposed.
  • Bucket II — Needs decision: real issues with unresolved tradeoffs or authority. Explain the smallest viable alternative and decision needed. Apply an already supplied decision within the authorized scope; ask only for what remains unresolved.

Recheck fixes and affected validation until the evidenced problems are resolved or a concrete limitation prevents continuation. Do not turn the pass into an autonomous review loop or repeat passing checks without new evidence. Do not stage, commit, or push unless asked.

Report

Lead with consequential findings and their applied/proposed status. Include delivery coverage for initiative reviews; validation and material residual risk; and meaningful keep-as-is decisions when useful. A clean review needs only the result and verification limits, not empty buckets or a checklist transcript.

Comments

Markdown supported