Skip to content
Agent ConfigAgent Config

AI code review that demands evidence

Review your local changes before you open the pull request:

/review:changes

What it does

Section titled “What it does”

/review:changes collects everything on your branch that is not yet on main and hands it to focused reviewers — bugs, security, tests, quality, architecture, spec and overbuild — then merges their findings into one list. Every finding names a file and line, so you can check it before you act on it.

It first brings your branch up to date with its base. If the base cannot be found it stops and says so, rather than quietly reviewing a different diff.

Check that done means done

Section titled “Check that done means done”
@production-validator is this branch really done?

production-validator is a read-only checker for Claude Code. It reads the code that would ship — not the tests — and looks for mocks, stubs, fakes, TODOs and placeholders on the real path and for evidence that the change ran against real systems. It answers READY or NOT READY with the exact file and line. It changes nothing.

We measured it on two planted cases: on a hollow implementation it answered NOT READY with the file and line; on a clean one it answered READY with no spurious findings. Two cases on Claude Code, not a hit rate.

When to use it

Section titled “When to use it”
  • Before every pull request — /review:changes is cheap compared to a reviewer finding the same bug.
  • Before you trust an agent’s “done” — @production-validator.
  • When you want people, not just models, to look: /review:routing suggests reviewer roles from the paths a change touches.

What it needs

Section titled “What it needs”
  • /review:changes: any supported AI coding tool, git and a main branch on origin.
  • @production-validator: Claude Code, which runs it as a subagent.

Limits

Section titled “Limits”
  • The reviewers are passes of the model you already use. They catch what a careful read catches; they do not run your application.
  • Treat findings as claims to check. The file and line are there so you can.

Read the review skill without installing

Section titled “Read the review skill without installing”

The review procedure is also available as a skill your AI tool can read over MCP, without the full install. agent-config mcp-server is a read-only server: it can suggest a skill for a task and return a skill’s text. It runs nothing and writes nothing.

{
  "mcpServers": {
    "agent-config": {
      "command": "npx",
      "args": ["-y", "@event4u/agent-config", "mcp-server"]
    }
  }
}

The server offers two tools: suggest_skill_for_task ranks skills for a task you describe, and read_skill returns one skill by name — here code-review. Your agent follows the skill’s procedure with its own tools.

Next

Section titled “Next”