code-review

Code Review Skill

Review changed code in this repo and report every issue with a severity and a confidence score.

Review Scope

By default, review all uncommitted work: git status --short --untracked-files=all for the file list, git diff HEAD for tracked changes (staged and unstaged), and each untracked (??) file read in full, since a new file has no diff. The user may specify different files or scope.

What to check in this repo

Beyond general correctness and security, the contracts most likely to break here (plugin paths are under plugins/harness-eval/):

• Scripts: target project root as $1 (aggregate.sh reads stdin instead), JSON on stdout / logs on stderr, exit codes 0/1/2, jq checked at startup (scripts/CLAUDE.md). The SKILL.md files duplicate this contract for plugin users, so a script change must update them.
• Prompt files are code. SKILL, agent, and command .md files carry contracts other code parses - handoff lines and markers, report headings, JSON field names, headings asserted by tests/structure/test-plugin-structure.sh. A changed string must match every consumer; grep for it.
• Auto-discovery: no non-component .md (including CLAUDE.md) under commands/ or agents/. Agents use tools; commands and skills use allowed-tools.
• The version must match in plugin.json version and marketplace.json metadata.version + plugins[].version.
• Plugin hooks must not modify user files without opt-in, and their failures stay visible (hooks/CLAUDE.md).
• The fixtures under tests/fixtures/ are frozen test inputs; new coverage goes in a new fixture directory.
• Bilingual output: English and Korean reports and doc sections stay in sync.

Confidence and severity

For each issue give a severity (critical / important / minor) and a confidence 0-100 that it is a real, newly introduced defect (90+: confirmed; 75-89: verified; 50-74: likely; below 50: speculative or pre-existing). Report every issue with its scores. The report shows items at 75+ first and groups the rest under "Lower confidence" - the reader filters, you don't drop.

Output Format

For each issue:

[CRITICAL|IMPORTANT|MINOR] <issue title> (confidence: XX)

File: path/to/file.ext:line Issue: Clear description of the problem Guideline: Reference to CLAUDE.md rule or security standard Fix: Concrete code suggestion

List items with confidence ≥ 75 first, then a "Lower confidence" section; if nothing reaches 75, say so and still list the rest.