A Claude Code + Codex skill that panels six independent review agents across three topics before you execute a plan or finish a spec. When model tiers disagree, an opus juror adjudicates. The result is a gated Blockers / Warnings / Observations verdict.
Designed to work with the Superpowers workflow:
brainstorming → spec → /multi-agent-review spec → writing-plans → /multi-agent-review plan → subagent-driven-development
/plugin marketplace add riekelt/multi-agent-review
/plugin install multi-agent-review@multi-agent-review
git clone https://github.com/riekelt/multi-agent-review \
~/.claude/plugins/multi-agent-reviewAdd to your Codex plugins via the .codex-plugin/plugin.json at this repo root.
/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 — haiku and sonnet tiers each review completeness, alignment, and risk. If the two tiers disagree on a finding, a single reasoning-model juror adjudicates. If they agree, no juror is invoked.
The verdict 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
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.
Copy skills/multi-agent-review/project-rules.example.md to your project root as project-rules.md. The skill injects it into alignment and risk reviewers. Without it, only generic checks run.
| Flag | Effect |
|---|---|
--fast |
Fast-tier only, skip standard tier, skip juror. ⚠ Do not use on plans that touch production-safety paths. |
| Failure | Behaviour |
|---|---|
| < 3 valid agent reports | Halt — no verdict produced |
| 3–5 valid reports | Continue with panel-health WARNING |
| Juror error/timeout | Conservative fallback — all contested findings → BLOCKER |
| All agents fail | Halt — empty panel never triggers clean verdict |
.claude-plugin/
plugin.json Claude Code plugin metadata
marketplace.json Claude Code marketplace entry
.codex-plugin/
plugin.json Codex plugin metadata (includes model tier config)
skills/multi-agent-review/
SKILL.md Coordinator
completeness-reviewer.md Completeness agent prompt
alignment-reviewer.md Alignment agent prompt
risk-reviewer.md Risk agent prompt
synthesis-agent.md Opus juror prompt
project-rules.example.md Project-specific rules template
Configured per platform in plugin.json. Claude Code defaults: haiku / sonnet / opus. Codex defaults: gpt-5.4-mini / gpt-5.4 / gpt-5.5.
MIT