Skill · v1.0

PR Code Review

PR Code Review is a skill for catching real issues before a pull request merges, without adding a human reviewer to the critical path. Five agents read the diff from five independent angles (convention compliance, bugs, git history, prior PR comments, code comments), every finding gets a 0-100 confidence score, and only scores of 80 or above post as a GitHub comment. Quiet on clean PRs. Precise on the ones that need a second look.

Install

Claude Code (CLI / WSL / Git Bash)

/plugin marketplace add https://www.infinitegameos.io/marketplace.json
/plugin install pr-code-review@igos-library

Claude Code (VS Code)

Install in VS Code

Opens the Claude Code plugins dialog with the marketplace and skill prefilled. Requires the Claude Code VS Code extension installed and signed in. Or paste the snippet below into .claude/settings.json for VS Code, JetBrains or any setup that prefers manual config.

{
  "extraKnownMarketplaces": {
    "igos-library": {
      "source": {
        "source": "url",
        "url": "https://www.infinitegameos.io/marketplace.json"
      }
    }
  },
  "enabledPlugins": {
    "pr-code-review@igos-library": true
  }
}

Direct markdown URL (Claude Code, Cursor, Codex CLI)

https://www.infinitegameos.io/markdown/skills/pr-code-review

This URL returns the narrative concept page. The plugin install path above delivers the operational SKILL.md instruction file.

Cursor (.mdc rules file)

curl -O https://www.infinitegameos.io/install/cursor/pr-code-review.mdc

Aider, Cline, any agent with --read

curl -O https://www.infinitegameos.io/markdown/skills/pr-code-review
aider --read pr-code-review.md

This skill began inside the Sovereign Ecosystem Foundation and graduated to the public library in April 2026.

Definition

PR Code Review runs an automated review on any open pull request before it merges. Five specialized agents analyze the diff from independent angles: convention-file compliance (your CLAUDE.md, README or equivalent), a shallow bug scan scoped to the diff, git history and blame context, prior PR comments on the same files, and inline code comment compliance. Each finding from those five agents is scored for confidence against an explicit 0-100 rubric by a dedicated scoring pass, and only findings at 80 or above survive the filter. Surviving issues post as a single GitHub PR comment with exact file links and full commit SHAs, using the gh CLI end to end. A false positive taxonomy of eleven categories (pre-existing issues, linter-catchable issues, pedantic nitpicks and more) rides along with every scoring pass to keep the signal-to-noise ratio high on a public surface.

Five lenses, one confidence gate

An eligibility check runs first and last: skip draft PRs, closed PRs, automated PRs and anything already reviewed. In between, five agents work the diff in parallel, each blind to the others: what does the convention file require, what looks like an outright bug, what does the git history say about intentional patterns, what did earlier PR comments on these files already resolve, and what do the code comments themselves promise.

Every issue those five agents surface gets its own confidence-scoring pass against a verbatim rubric, no exceptions. Below 80 the issue is a note that never leaves the session. At or above 80 it earns a spot in the single comment that posts to GitHub, complete with the full SHA and a direct line link. One comment per run, or a clean no-issues confirmation. Nothing in between.

Use Cases

Merge-gate review without a second human in the loop

A solo builder or small team ships PRs faster than a human reviewer can keep pace with. PR Code Review runs automatically on each open PR, catches the convention violations and real bugs that matter, and stays silent on the PRs that are already clean. The merge gate holds without adding review latency.

Enforcing a convention file that keeps drifting in practice

A CLAUDE.md or CONTRIBUTING.md exists but nobody consistently checks new PRs against it by hand. Agent 1 reads the convention file fresh on every run and flags only the violations the file actually calls out, so the standard stays live instead of decaying into a document nobody rereads.

Multi-repo consistency across a small fleet of sites or services

A team runs several repos with a shared review bar but no shared reviewer bandwidth. The same skill runs identically on each repo, reading whatever convention file lives there, so the bar holds without a person context-switching between codebases all day.

FAQ

Does this replace human code review?

No. It is a merge gate, not a reviewer. It catches convention violations and obvious bugs scoped to the diff, at a high confidence threshold, so a human reviewer spends their attention on judgment calls rather than mechanical compliance checks.

Why the 80 confidence threshold specifically?

A false positive on a public PR costs credibility, not just review time. The threshold defaults high on purpose, and the skill treats lowering it as a deliberate operator decision rather than a tuning knob to reach for casually.

What if my repo does not have a CLAUDE.md?

Point Agent 1 at whatever convention file you do have: README, CONTRIBUTING.md, a .cursorrules file, a style guide. The lens generalizes to any document that states how your team wants code written.

Related

PR Code Review is one entry point. The wider operating system that uses it is Infinite Game OS. The full life-design frame that motivates clean architecture in the first place is the Sovereign Life Playbook.

See the Sovereign Life Playbook