multi-agent-review

A Claude Code, Codex, and Cursor skill that reviews specs and plans, not code. Six review agents cover three topics before you execute, and a reasoning-tier juror is invoked only when the model tiers disagree. The gate is fail-closed: a dead juror promotes contested findings to BLOCKER, a broken panel halts, and an empty panel never yields a clean verdict. Use it when you want a plan reviewed before execution, a spec checked before planning, or a second opinion before you build.

Install

Claude Code, plugin manager

/plugin marketplace add riekelt/multi-agent-review
/plugin install multi-agent-review@multi-agent-review

Claude Code, manual install

The plugin lives in the plugins/multi-agent-review/ subdirectory, not at the repo root:

git clone https://github.com/riekelt/multi-agent-review /tmp/multi-agent-review
cp -r /tmp/multi-agent-review/plugins/multi-agent-review ~/.claude/plugins/multi-agent-review

Codex

Clone the repo and copy the plugin directory into your Codex plugins directory:

git clone https://github.com/riekelt/multi-agent-review /tmp/multi-agent-review
cp -r /tmp/multi-agent-review/plugins/multi-agent-review <your-codex-plugins-dir>/multi-agent-review

Codex reads the plugin metadata and the model tiers from .codex-plugin/plugin.json inside the copied directory.

Cursor

Clone the repo and copy the plugin directory into your Cursor plugins directory:

git clone https://github.com/riekelt/multi-agent-review /tmp/multi-agent-review
cp -r /tmp/multi-agent-review/plugins/multi-agent-review <your-cursor-plugins-dir>/multi-agent-review

Cursor reads the plugin metadata and the model tiers from .cursor-plugin/plugin.json inside the copied directory.

Usage

/multi-agent-review spec    # after brainstorming, before writing-plans
/multi-agent-review plan    # after writing-plans, before subagent-driven-development

Six agents run in parallel: a fast and a standard tier each review completeness, alignment, and risk. If the two tiers disagree on a finding, a single reasoning-tier juror adjudicates. If they agree, no juror is invoked.

The verdict sorts findings into Blockers, Warnings, and Observations, then gates the next step:

• Blockers → stops execution, presents to operator (fix + re-run, or override with logged note)
• Warnings → presents to operator (fix or accept and continue)
• Clean → auto-proceeds to next skill

Superpowers integration

This skill is designed to slot into the Superpowers / Superpowers Extended CC workflow:

brainstorming skill
  → spec written
  → /multi-agent-review spec          ← this skill
  → writing-plans skill
  → /multi-agent-review plan          ← this skill
  → subagent-driven-development skill

Without Superpowers, invoke it directly before any plan execution: pass an explicit path argument, since the default artifact discovery assumes the Superpowers layout (docs/superpowers/specs/, docs/superpowers/plans/). That coupling is deliberate: this skill is a slot in that workflow, and the auto-invoked next skills (writing-plans, subagent-driven-development) are Superpowers skills by name. If Superpowers renames its paths or skills, Step 1 and Step 9 of plugins/multi-agent-review/skills/multi-agent-review/SKILL.md are the two places to update.

Project-specific rules

Copy plugins/multi-agent-review/skills/multi-agent-review/assets/project-rules.example.md to your project root as project-rules.md. The skill injects it into the alignment and risk agents. Without it, the skill falls back to the project's CLAUDE.md or AGENTS.md and applies only the parts that read as standing constraints on specs and plans.

Flags

Panel failure semantics

Both fail-closed paths have run in operation on Claude Code and Codex: the juror fallback promoting contested findings to BLOCKER, and the quorum halt on too few valid agent reports (operator-verified, 2026-08-19).

Repository layout

.claude-plugin/
  marketplace.json     Claude Code marketplace entry

plugins/multi-agent-review/
  .claude-plugin/plugin.json   Claude Code plugin metadata (model tier config)
  .codex-plugin/plugin.json    Codex plugin metadata (model tier config)
  .cursor-plugin/plugin.json   Cursor plugin metadata (model tier config)
  skills/multi-agent-review/
    SKILL.md                           Coordinator
    agents/completeness-reviewer.md    Completeness agent prompt
    agents/alignment-reviewer.md       Alignment agent prompt
    agents/risk-reviewer.md            Risk agent prompt
    agents/synthesis-agent.md          Juror prompt
    assets/project-rules.example.md    Project-specific rules template
  evals/evals.json                     Persisted review test prompts
  evals/trigger-evals.json             Skill-routing trigger measurements
  evals/fixtures/                      Planted-defect artifacts the evals review

Model tiers

Configured per platform in that platform's plugin.json under models. Claude Code: haiku / sonnet / opus. Codex: gpt-5.6-luna / gpt-5.6-terra / gpt-5.6-sol. Cursor: claude-haiku / claude-sonnet / claude-opus.

License

MIT