Stats
Actions
Available In
Tags
Ralpr
Autonomous three-phase loop that takes GitHub issues to merge-ready PRs.
What is Ralpr?
Ralpr (Ralph Autonomous Loop for Pull-request Readiness) is a Claude Code plugin that autonomously implements GitHub issues and iteratively reviews the resulting PRs until they reach merge-ready confidence levels.
- Three-phase loop: Implementation, Review, Refactor — each phase runs as an independent Claude Code session
- Parallel AI reviewers: Three specialized agents (QA, Domain Expert, Codex) review code simultaneously
- Fix-by-default: Every issue found is fixed, not deferred — skipping requires structured justification
- Fresh context per iteration: Each loop iteration spawns a new Claude session to prevent context drift
- Cumulative confidence scoring: Confidence only goes up, measured against configurable thresholds
Architecture
┌─────────────────────────────────────────────────────────────────────────┐
│ Ralpr Pipeline │
│ │
│ GitHub ┌──────────┐ PR ┌──────────────┐ ┌──────────────┐ │
│ Issue ──▶ │ Phase 1 │ ──────▶ │ Phase 2 │ ─▶ │ Phase 3 │ │
│ │ Impl │ │ Review │ │ Refactor │ │
│ │ (1 pass) │ │ (loop → 90%) │ │ (loop → 95%) │ │
│ └──────────┘ └──────────────┘ └──────────────┘ │
│ │ │ │
│ ▼ ▼ │
│ Merge-ready PR │
└─────────────────────────────────────────────────────────────────────────┘
Phase 2 — Parallel Review Agents:
┌─────────────────┐
│ QA Reviewer │──┐
│ (test coverage) │ │
└─────────────────┘ │ ┌───────────┐ ┌───────────────┐
├──▶ │ Aggregate │──▶ │ Fix-by-Default│──▶ Confidence
┌─────────────────┐ │ │ & Score │ │ Decision │ Score
│ Domain Expert │──┤ └───────────┘ └───────────────┘
│ (arch/security) │ │
└─────────────────┘ │
│
┌─────────────────┐ │
│ Codex Reviewer │──┘
│ (3rd-party AI) │
└─────────────────┘
Prerequisites
System Tools
| Tool | Min Version | Check Command | Purpose |
|---|---|---|---|
| Claude Code | latest | claude --version | Runtime |
gh | any | gh auth status | GitHub API |
git | 2.20+ | git --version | Worktree support |
jq | any | jq --version | JSON parsing |
bc | any | echo "1+1" | bc | Confidence math |
Required Claude Code Plugins
| Plugin | Source | Install Command | Used In |
|---|---|---|---|
superpowers | Official marketplace | /plugin install superpowers | Phase 3 — superpowers:refactor skill |
pr-review-toolkit | Official marketplace | /plugin install pr-review-toolkit | Phase 3 — pr-review-toolkit:code-simplifier agent |
Required MCP Server
| Server | Tool Used | Setup |
|---|---|---|
| Codex CLI | mcp__codex__codex | Install Codex CLI, set OPENAI_API_KEY, configure as MCP server in Claude Code |
The Codex MCP server is used by the codex-reviewer agent in Phase 2 to get an independent third-party review from OpenAI's model.
Note: Ralpr itself has zero npm/node dependencies. Project-specific tools (npm, cargo, pytest, forge, etc.) are auto-detected by quality gates.
Installation
Option A: Marketplace (recommended)
# Inside Claude Code:
/plugin marketplace add Archethect/ralpr # Add marketplace
/plugin install ralpr # Install plugin
Option B: Local development
git clone https://github.com/Archethect/ralpr.git
claude --plugin-dir ./ralpr
Post-install verification
# Start Claude Code in any GitHub repo, then:
/ralpr --dry-run
You should see a dry-run summary confirming prerequisites are met.
Quick Start
# 1. Open Claude Code in a GitHub repo
cd your-github-repo && claude
# 2. Run Ralpr on an issue
/ralpr --issue 42
# 3. Watch it work:
# - Implements the issue (TDD: RED → GREEN → REFACTOR)
# - Creates a PR labeled ralpr:impl:done
# - Reviews with 3 parallel AI agents, fixes issues
# - Refactors for code quality
# 4. Result: PR with confidence labels
# ralpr:impl:done
# ralpr:review:90+
# ralpr:refactor:95+
