Skill

review-code

Multi-stage CodeAgora-style code review with SARIF output

Install

npx claudepluginhub justn-hyeok/harness-for-yall --plugin review-pipeline

Tool Access

This skill uses the workspace's default tool permissions.

SKILL.md

Similar Skills

cache-components

Guides Next.js Cache Components and Partial Prerendering (PPR) with cacheComponents enabled. Implements 'use cache', cacheLife(), cacheTag(), revalidateTag(), static/dynamic optimization, and cache debugging.

cache-components

138.9k

claude-opus-4-5-migration

2 files

Migrates code, prompts, and API calls from Claude Sonnet 4.0/4.5 or Opus 4.1 to Opus 4.5, updating model strings on Anthropic, AWS, GCP, Azure platforms.

claude-opus-4-5-migration

83.2k

bmad-help

Analyzes BMad project state from catalog CSV, configs, artifacts, and query to recommend next skills or answer questions. Useful for help requests, 'what next', or starting BMad.

bmad-pro-skills

43.8k

Stats

Parent Repo Stars0

Parent Repo Forks0

Last CommitApr 7, 2026

Actions

View Source View Plugin View on GitHub View README

CodeAgora Multi-Stage Code Review

Orchestrates a full code review pipeline: fan-out to parallel L1 screeners, consolidation by moderator, final verdict by judge.

Workflow

Step 0 — Identify Target

Determine what to review:

If the user provided a file path or list of files, use those
If the user said "audit PR" or referenced a PR, get the diff with git diff against the base branch
If the user said "review code" without specifics, review staged changes (git diff --cached) or the most recent commit (git diff HEAD~1)

Read all target files/diffs so you have the full context.

Step 1 — Fan-Out: Parallel L1 Screening

Launch all three screener agents in parallel using the Agent tool:

review-screener-1 (correctness) — send the code/diff and ask for correctness findings
review-screener-2 (security) — send the code/diff and ask for security findings
review-screener-3 (performance/style) — send the code/diff and ask for performance and style findings

Each agent prompt must include:

The full code or diff to review
The file path(s) being reviewed
Instruction to return findings as a JSON array per their agent spec

Wait for all three to complete.

Step 2 — Fan-In: Moderator Consolidation

Launch review-moderator with:

The three JSON arrays from Step 1
Instruction to deduplicate, resolve conflicts, validate severity, and produce the consolidated report

Step 3 — Final Verdict: Judge

Launch review-judge with:

The moderator's consolidated report
Instruction to render a verdict and produce the final SARIF v2.1.0 report

Step 4 — Output

Save the SARIF JSON report to review-report.sarif.json in the project root
Present a human-readable summary to the user:

## Code Review Complete

**Verdict**: [APPROVE|APPROVE_WITH_NOTES|REQUEST_CHANGES|BLOCK]
**Risk Level**: [CRITICAL|HIGH|MODERATE|LOW|CLEAN]

### Findings Summary
- BLOCKER: X
- HIGH: X  
- MEDIUM: X
- LOW: X

### Top Issues
1. ...
2. ...
3. ...

### Action Items
- ...

Full SARIF report saved to: review-report.sarif.json

Important Notes

All agent launches must use the agents defined in .claude/agents/
Screener agents run in parallel (fan-out) — do not run them sequentially
Moderator and judge run sequentially (fan-in then producer-reviewer)
If any agent fails, report the failure and continue with available results
The SARIF report must be valid against the SARIF v2.1.0 schema