{
  "markdown": "# multi-agent-review\n\n[![skills.sh](https://skills.sh/b/riekelt/multi-agent-review)](https://skills.sh/riekelt/multi-agent-review)\n\nA 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.\n\n## Install\n\n### Claude Code, plugin manager\n\n```\n/plugin marketplace add riekelt/multi-agent-review\n/plugin install multi-agent-review@multi-agent-review\n```\n\n### Claude Code, manual install\n\nThe plugin lives in the `plugins/multi-agent-review/` subdirectory, not at the repo root:\n\n```bash\ngit clone https://github.com/riekelt/multi-agent-review /tmp/multi-agent-review\ncp -r /tmp/multi-agent-review/plugins/multi-agent-review ~/.claude/plugins/multi-agent-review\n```\n\n### Codex\n\nClone the repo and copy the plugin directory into your Codex plugins directory:\n\n```bash\ngit clone https://github.com/riekelt/multi-agent-review /tmp/multi-agent-review\ncp -r /tmp/multi-agent-review/plugins/multi-agent-review <your-codex-plugins-dir>/multi-agent-review\n```\n\nCodex reads the plugin metadata and the model tiers from `.codex-plugin/plugin.json` inside the copied directory.\n\n### Cursor\n\nClone the repo and copy the plugin directory into your Cursor plugins directory:\n\n```bash\ngit clone https://github.com/riekelt/multi-agent-review /tmp/multi-agent-review\ncp -r /tmp/multi-agent-review/plugins/multi-agent-review <your-cursor-plugins-dir>/multi-agent-review\n```\n\nCursor reads the plugin metadata and the model tiers from `.cursor-plugin/plugin.json` inside the copied directory.\n\n## Usage\n\n```\n/multi-agent-review spec    # after brainstorming, before writing-plans\n/multi-agent-review plan    # after writing-plans, before subagent-driven-development\n```\n\nSix 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.\n\nThe verdict sorts findings into **Blockers**, **Warnings**, and **Observations**, then gates the next step:\n- **Blockers** → stops execution, presents to operator (fix + re-run, or override with logged note)\n- **Warnings** → presents to operator (fix or accept and continue)\n- **Clean** → auto-proceeds to next skill\n\n## Superpowers integration\n\nThis skill is designed to slot into the [Superpowers](https://github.com/obra/superpowers) / [Superpowers Extended CC](https://github.com/pcvelz/superpowers) workflow:\n\n```\nbrainstorming skill\n  → spec written\n  → /multi-agent-review spec          ← this skill\n  → writing-plans skill\n  → /multi-agent-review plan          ← this skill\n  → subagent-driven-development skill\n```\n\nWithout 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.\n\n## Project-specific rules\n\nCopy `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.\n\n## Flags\n\n| Flag | Effect |\n|---|---|\n| `--fast` | Fast-tier only: three agents, no cross-model check, no juror; every finding is accepted at its emitted severity. Safety limit: the skill refuses `--fast` when the artifact mentions authentication, authorization, security, secrets, payment, billing, migration, data deletion, or production infrastructure. It runs the full panel instead. |\n\n## Panel failure semantics\n\n| Failure | Behaviour |\n|---|---|\n| < 3 valid agent reports | Halt: no verdict produced |\n| 3 to 5 valid reports | Continue with panel-health WARNING |\n| Juror error/timeout | Conservative fallback: all contested findings become BLOCKER |\n| All agents fail | Halt: an empty panel never triggers a clean verdict |\n\nBoth 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).\n\n## Repository layout\n\n```\n.claude-plugin/\n  marketplace.json     Claude Code marketplace entry\n\nplugins/multi-agent-review/\n  .claude-plugin/plugin.json   Claude Code plugin metadata (model tier config)\n  .codex-plugin/plugin.json    Codex plugin metadata (model tier config)\n  .cursor-plugin/plugin.json   Cursor plugin metadata (model tier config)\n  skills/multi-agent-review/\n    SKILL.md                           Coordinator\n    agents/completeness-reviewer.md    Completeness agent prompt\n    agents/alignment-reviewer.md       Alignment agent prompt\n    agents/risk-reviewer.md            Risk agent prompt\n    agents/synthesis-agent.md          Juror prompt\n    assets/project-rules.example.md    Project-specific rules template\n  evals/evals.json                     Persisted review test prompts\n  evals/trigger-evals.json             Skill-routing trigger measurements\n  evals/fixtures/                      Planted-defect artifacts the evals review\n```\n\n## Model tiers\n\nConfigured 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`.\n\n## License\n\nMIT\n",
  "bytes": 6160,
  "sha": "f93f16c69681c04799b9e8630ab773869d0f92fdf6dcf7f214337a00228f99a8",
  "repo_slug": "riekelt/multi-agent-review",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/skl_riekelt_multi_agent_review_multi_agent_r_2d425de2/readme"
}