Skip to main content

dual-review

A Claude Code slash command that sends your plan or diff to a second AI agent for independent, structured review before and after you implement.

What it does​

/dual-review Implement feature X
  1. Explore the codebase to understand the area being changed
  2. Plan a draft implementation
  3. Pre-review — send the plan to the review tool on the claude_reviewer (or codex) MCP server for a structured verdict
  4. Iterate — on NEEDS_REVISION, address critical/major issues and re-submit, up to 3 rounds
  5. Implement once the plan is APPROVED
  6. Post-review — send the final git diff to code_review for a post-implementation check
  7. Summarize with the full review history

The reviewer runs in a separate agent instance with read-only codebase access and returns:

{
  "verdict": "APPROVED | NEEDS_REVISION",
  "issues": [
    { "severity": "critical | major | minor", "description": "...", "recommendation": "..." }
  ],
  "suggestions": ["..."]
}

Iteration rules: critical issues must be addressed before re-submitting; major issues should be; minor issues are at your discretion. After 3 NEEDS_REVISION rounds, the issues are presented to you for a decision instead of looping forever.

Why it exists​

dual-review isn't a bolted-on extra — it's the workflow the MCP bridge itself was built to enable. The bridge's original design doc (SPEC.md) describes exactly this loop: give a second AI a plan or diff via the claude_reviewer/codex MCP tools, before and after implementing, so a review round doesn't mean manually copy-pasting a diff into another agent's chat window.

For most of the project's life it lived only as skills/dual-review/SKILL.md — usable if you had this repo cloned, but not installable on its own. That changed 8 days into the project's public life, via three commits that turned it into a real distributable skill:

  • 3c12bac — scaffolded multi-channel skill distribution infrastructure (pnpm workspace, package generation, marketplace generation, a skills-release Changesets pipeline) for the repo as a whole. At this point dual-review still had no metadata.json, so package generation was blocked for it specifically.
  • 0aac0a7 — authored dual-review's missing metadata.json, unblocking package generation. Fully generated, synced, and validated — but still private: true, nothing published yet.
  • bf7279b — flipped it to private: false and enabled it for npm publish.

The result: @mcp-agent-bridge/skill-dual-review is now a real, independently versioned package (currently 1.0.0 on npm) and a plugin in this repo's own Claude Code marketplace, installable without cloning the bridge itself.

Install​

Claude Code Plugin
/plugin marketplace add catesandrew/mcp-agent-bridge
/plugin install mab@mcp-agent-bridge
npm package

The skill is published as @mcp-agent-bridge/skill-dual-review. Inspect or fetch it directly:

npm view @mcp-agent-bridge/skill-dual-review

Its SKILL.md, metadata.json, LICENSE, and README.md are the package contents — copy SKILL.md into your agent's skills directory if your tooling consumes skill files directly rather than through a plugin manager.

Manual / clone
git clone https://github.com/catesandrew/mcp-agent-bridge.git
cp -r mcp-agent-bridge/skills/dual-review ~/.claude/skills/dual-review

Configuration​

The skill needs the claude_reviewer MCP server (this repo's Claude bridge) reachable. Default: http://localhost:8940/mcp, configured in .mcp.json:

{
  "mcpServers": {
    "claude_reviewer": {
      "type": "streamable-http",
      "url": "http://localhost:8940/mcp"
    }
  }
}

Environment variables (inherited from the bridge server):

  • CLAUDE_REVIEW_MODEL — model used by the reviewer (default: opus)
  • CLAUDE_REVIEW_MAX_TURNS — max agentic turns per review (default: 8)
  • CLAUDE_REVIEW_CWD — working directory for the reviewer's file access
  • CLAUDE_REVIEW_ALLOWED_TOOLS — comma-separated tools the reviewer can use