MultiReview Agent Skill

A reusable agent skill/runbook for running consolidated multi-model code reviews.

It supports two review modes:

1. PR review — multiple reviewer agents inspect a GitHub pull request; the orchestrator consolidates findings and posts one summary plus inline PR comments.
2. Scope review without PR — multiple reviewer agents inspect a folder/module/repo area; the orchestrator consolidates findings into a Markdown review file committed to the repo.

The workflow is designed for any capable coding agent that can:

• spawn independent reviewer agents/subagents,
• read local files and run shell commands,
• use GitHub CLI/API for repository access and PR comments,
• consolidate structured findings.

It has been tested in production-style workflows with OpenClaw and Hermes.

Files

• SKILL.md — the complete skill/runbook.

Default model pattern

The skill documents a five-reviewer setup:

• GPT-5.5
• DeepSeek V4 Pro
• Kimi K2.6
• Qwen 3.6 Plus
• GLM-5.1

The orchestrator is GPT-5.5 in the tested setup, but the method is agent-agnostic.

Why multi-model review?

Different models have different failure modes, strengths, and blind spots. Running several independent reviewers increases the chance that at least one model catches a real bug that others miss.

The orchestrator then consolidates the results. Findings reported by multiple models carry more confidence because independent reviewers converged on the same issue. Single-model findings are still useful, but they are filtered more aggressively for evidence, traces, and false positives.

In practice, this gives two benefits:

• Broader bug discovery: diversity catches more edge cases.
• Stronger signal: consensus makes important findings easier to trust and prioritize.

The goal is not to produce five separate reviews; it is to produce one high-signal review backed by independent analysis.

Core ideas

• Reviewers are analysis-only; they do not post to GitHub or edit files.
• The orchestrator prepares local snapshots/diffs, spawns reviewers, deduplicates findings, verifies obvious false positives, and publishes one consolidated result.
• PR reviews avoid GitHub pending-review collisions by using standalone inline PR comments instead of parallel draft reviews.
• Non-PR reviews are delivered as Markdown files in the reviewed repo.

Usage

Copy SKILL.md into your agent's skills/runbooks directory, or adapt its workflow into your agent instructions.

For OpenClaw-style skills, place it as:

skills/multireview/SKILL.md

Then ask your agent for a multi-review of a PR or a repo scope.

Status

This skill is evolving from real multi-review workflows. Update it when you discover better reviewer prompts, model sets, false-positive filters, snapshot patterns, or publishing conventions.