code-review-report - SKILL.md Agent Skill

name: code-review-report version: 0.1.0 category: development description: Turn code changes into a single-file HTML code review report that pairs the diff with context a diff can't show — design decisions, rejected alternatives, tradeoffs, unfinished work — plus prioritized review points. Use when preparing for review after finishing a sizable change, refactor, or feature, or whenever the user asks for a review report, change report, or an HTML summary of their changes.

Code Review Report

Build a single-file HTML report a reviewer reads before digging through the raw diff. A diff shows what changed; it doesn't show why, what was rejected, or what's still missing. Filling that gap with the context accumulated in this conversation is what separates this from a diff viewer.

1. Collect the change

Check the change with git status and git diff. Default scope: current branch vs the default branch, plus the uncommitted working tree.

git diff <base>...HEAD (committed), git diff HEAD (uncommitted)
Untracked files don't appear in diffs — find them with git status and read them directly.

If the scope is mixed or contains unrelated changes, ask the user. Exclude unrelated changes or mark them separately.

2. Collect the working context

Gather from this conversation the things a diff never records:

What was changed and why
Design decisions and their reasons, rejected alternatives, tradeoffs
Unfinished work and next steps
Concerns raised and reversals made along the way

If the conversation lacks this (e.g., the skill was invoked fresh), reconstruct from git log, commit messages, PRs, and related docs. If still empty, ask the user.

3. Extract review points

Inject the changed code plus the working context into subagents and collect suspicious spots as P1–P3 findings. Read references/review-perspectives.md for the perspective list and the subagent prompt skeleton. The context injection is the point — without it you only get surface-level remarks, not findings like spec mismatches or contradictions between decisions. For large changes, run one subagent per perspective in parallel.

4. Write the HTML

Start from assets/template.html. Follow references/report-structure.md for sections and references/style-guide.md for visual rules. Two things are non-negotiable:

Syntax highlighting on code blocks (add language-* classes)
Bidirectional anchors between change details and the review checklist

Write the report in the language the user works in: translate the template's headings and set <html lang>. Keep code as-is. Before finalizing, run the self-check at the end of references/report-structure.md.

5. Finalize

Save to the project root as <task-slug>-review-report.html (kebab-case task name).
Inline the vendored highlight.js — replaces the template's placeholder, keeps the report self-contained and offline-friendly: python3 <skill-dir>/scripts/inline_hljs.py <report-path>
Open it for the user: open <report-path> (macOS), xdg-open (Linux), start (Windows).

Principles

State tradeoffs, limitations, and unfinished work plainly.
Every change gets a "why", not just a "what".
Quote code; show before→after where possible.
Hierarchy through typography and spacing, not color.