Skip to content

The-Rabak/compound-engineering-plugin

Repository files navigation

Compound Engineering Plugin

Development tools that compound. Every task you complete makes the next one faster -- not through magic, but through structured knowledge capture and reuse.

26 specialized agents. 22 commands. 18 skills. One workflow that actually works.

Quick Start

# 1. Add the marketplace
claude /plugin marketplace add https://github.com/The-Rabak/compound-engineering-plugin

# 2. Install the plugin
claude /plugin install compound-engineering

# 3. Navigate to your project and run setup
/compound-engineering:setup

Setup auto-detects your stack (Laravel, NestJS, Python, Vue, TypeScript, Rust, etc.) and configures the right review agents for your project.

The canonical source now lives in portable/compound-engineering/. This repository also ships generated Copilot assets under .github/, built from that same portable source.

Regenerate platform outputs

bun run build:platforms

That rebuilds:

  • plugins/compound-engineering/ for Claude Code
  • .github/ for Copilot

How It Works

The plugin is built around a plan → build → review → learn cycle. Each phase is a standalone command, and learnings from every cycle persist and feed into the next one. The learnings-researcher agent surfaces past solutions during planning and review, so you don't solve the same problem twice.

  ┌──────────┐      ┌───────────────┐
  │  PLAN    │─────▶│  DEEPEN PLAN  │
  └────┬─────┘      └───────┬───────┘
       │                    │
       ▼                    ▼
  ┌──────────────────────────────────────────┐
  │              WORK                        │
  │                                          │
  │  Orchestrator decomposes plan into tasks │
  │                                          │
  │    ┌────────────┐  ┌────────────┐        │
  │    │ subagent 1 │  │ subagent 2 │  ...   │──▶ commits
  │    └────────────┘  └────────────┘        │
  │           │              │               │
  │           ▼              ▼               │
  │    ┌─────────────────────────────┐       │
  │    │  learnings brief (shared)   │       │    ◀── Ralph Loop
  │    └─────────────────────────────┘       │    variation:
  │           │                              │    iterate until
  │           ▼                              │    all tasks pass
  │    regression guard ──▶ next task        │
  └──────────────────────────────────────────┘
       │
       ▼
  ┌──────────────────────┐
  │       REVIEW         │
  │  (parallel agents)   │──▶ findings
  └──────┬───────────────┘
         │
         ▼
  ┌──────────────────────┐
  │      COMPOUND        │
  │  (capture learnings) │──▶ docs/solutions/
  └──────────────────────┘

The work phase is where the heavy lifting happens. It implements a variation of the Ralph Loop -- an iterative execution engine that decomposes a plan into scoped chunks, delegates each to a focused subagent, accumulates learnings across tasks, and runs regression guards after every chunk. The orchestrator never writes code itself; it decomposes, delegates, records, and routes. Each subagent gets a scoped prompt with only the context it needs plus a learnings brief filtered by domain relevance.


The Ralph Loop

The Ralph Loop is a self-referential iteration mechanism built on Claude Code's hook system. A stop hook intercepts the session exit, checks whether a completion promise has been met, and if not, feeds the original task back into the conversation with updated iteration state. This creates persistent, goal-directed execution with a guaranteed termination condition.

The /workflows:work command uses a variation of this pattern for its internal task loop: the orchestrator iterates over execution chunks, spawning subagents, collecting learnings, running regression guards, and advancing to the next task -- repeating until every chunk in the plan passes or the pipeline is explicitly halted.

The loop terminates when:

  • All tasks complete successfully (work phase variation), or
  • The agent outputs the agreed-upon completion promise (raw Ralph Loop), or
  • The maximum iteration count is reached (default: 10)

No runaway processes. No infinite loops.

Commands that use the Ralph Loop:

Command What it does
/lrj "feature" Full autonomous cycle: plan → deepen → work → review → test → video
/slrj "feature" Same as /lrj but uses swarm mode for parallel execution
/ralph-loop "task" Raw loop -- wrap any prompt in the iteration engine
/cancel-ralph Emergency stop for an active loop

Core Workflow Commands

/workflows:brainstorm -- Explore before building

For when the problem space is fuzzy. Produces a structured brainstorm document.

/workflows:brainstorm "how should we handle user notifications?"

Output lands in docs/brainstorms/.

/workflows:plan -- Turn ideas into implementation plans

Takes a feature description and produces a structured plan with acceptance criteria and code examples. Checks for existing brainstorms, fetches any linked documents, runs local research against your repo patterns and past learnings.

/workflows:plan "add a search feature to the dashboard"

Output lands in docs/plans/. Follow up with /deepen-plan to enhance it or /workflows:work to start building.

/deepen-plan -- Add depth with parallel research

Takes an existing plan and runs per-section research agents in parallel. Re-fetches linked source docs, checks docs/solutions/ for relevant past learnings, runs every configured review agent against the plan.

/deepen-plan docs/plans/2026-02-17-feat-dashboard-search-plan.md

/workflows:work -- Build it

The workhorse, and where the Ralph Loop variation lives. Takes a plan file and executes it through orchestrated subagent delegation:

  1. Reads the plan, resolves ambiguity upfront
  2. Sets up a feature branch (or worktree for parallel dev)
  3. Decomposes into execution chunks with success criteria
  4. Iterates through the task loop:
    • Builds a scoped prompt for each chunk (task + relevant files + learnings brief filtered by domain)
    • Spawns a focused subagent that implements, tests, and retries independently
    • Processes results: writes session files, updates learnings brief, checks off plan items
    • Runs regression guards against all previously completed tasks
    • Commits incrementally when a logical unit passes
  5. Repeats until every chunk completes or the pipeline is explicitly halted
  6. Pushes the branch with a PR description template

/workflows:review -- Multi-agent code review

Reviews your branch diff against main using every configured review agent in parallel. If migrations are present, it also runs the data integrity, migration, and deployment verification agents.

/workflows:review                    # current branch vs main
/workflows:review feat/my-branch     # specific branch

Findings are categorized P1 (blocks merge), P2 (should fix), P3 (nice-to-have).

/workflows:compound -- Capture what you learned

After solving a hard problem, run this to create a solution document in docs/solutions/. These get automatically surfaced during future plans and reviews by the learnings-researcher agent. This is how knowledge compounds.


Utility Commands

Command Purpose
/deepen-plan Research-enhance an existing plan
/test-browser Browser tests on PR-affected pages (uses agent-browser)
/triage Interactive triage of pending todos
/resolve_todo_parallel Fix approved todos in parallel
/resolve_parallel Fix TODO comments in code
/changelog Generate a changelog from recent merges
/feature-video Record a video walkthrough and add to PR
/reproduce-bug 123 Investigate and reproduce a reported bug
/report-bug File a bug against the plugin itself

Agents

Review (15 agents)

The review agents are the core of the plugin. Each one is a deep specialist -- not a generalist with bullet points, but an agent that knows the current best practices, common pitfalls, and real-world tradeoffs in its domain.

Agent Focus
rabak-laravel-reviewer Laravel 11+, Eloquent, Pest testing, modern PHP 8.3+
rabak-vue-reviewer Vue 3 Composition API, Nuxt 3, Pinia, TypeScript
rabak-nest-reviewer NestJS architecture, validation, auth, performance, security
rabak-python-reviewer Python conventions, type hints, testing patterns
rabak-rust-reviewer Ownership, safety, idiomatic patterns, zero-cost abstractions
rabak-typescript-reviewer Strict TypeScript, type safety, module patterns
code-simplicity-reviewer Cognitive complexity, YAGNI, dead code, over-engineering
architecture-strategist SOLID, Clean Architecture, DDD, architectural anti-patterns
security-sentinel OWASP Top 10 2025, supply chain, secrets, API security
performance-oracle Big O, DB optimization, caching, Web Vitals, benchmarks
data-integrity-guardian Database migrations, referential integrity, rollback safety
data-migration-expert Data migration validation, ID mappings, transformation logic
deployment-verification-agent Go/No-Go deployment checklists for risky changes
agent-native-reviewer Agent-native architecture verification
pattern-recognition-specialist Code pattern and anti-pattern analysis

Research (5 agents)

Agent Focus
best-practices-researcher External best practices and examples
framework-docs-researcher Framework documentation deep dives
git-history-analyzer Git history and code evolution analysis
learnings-researcher Past solutions from docs/solutions/
repo-research-analyst Repository structure and convention analysis

Design (3 agents)

Agent Focus
design-implementation-reviewer UI vs. design verification
design-iterator Iterative UI refinement
figma-design-sync Figma-to-code synchronization

Workflow (3 agents)

Agent Focus
bug-reproduction-validator Systematic bug reproduction
pr-comment-resolver PR comment resolution and implementation
spec-flow-analyzer User flow analysis and gap identification

Skills

Skill What it does
setup Configure review agents for your project
laravel-conventions Modern Laravel coding standards reference
frontend-design Production-grade frontend interfaces
resolve-pr-parallel Resolve PR review comments in parallel
brainstorming Structured collaborative dialogue
compound-docs Capture solved problems as documentation
create-agent-skills Guide for creating Claude Code skills
skill-creator Skill creation templates and patterns
document-review Structured document review
file-todos File-based todo tracking
git-worktree Git worktree management for parallel dev
orchestrating-swarms Multi-agent swarm orchestration
agent-native-architecture Prompt-native architecture patterns
agent-browser CLI-based browser automation
gemini-imagegen Image generation via Google Gemini API
rclone Cloud storage uploads (S3, R2, B2)

Configuration

Run /compound-engineering:setup in your project root. It creates compound-engineering.local.md:

---
review_agents: [rabak-laravel-reviewer, code-simplicity-reviewer, security-sentinel, performance-oracle]
plan_review_agents: [code-simplicity-reviewer, architecture-strategist]
---

# Review Context

Project-specific notes for review agents:
- "We use Redis heavily -- check cache invalidation patterns"
- "Performance-critical: API serves 10k req/s"
- "All database IDs must use UUIDs"

Edit this file to tune which agents run and what context they receive.


Tips

  • Start with review. Run /workflows:review on your next PR before trying the full cycle.
  • Compound early. Run /workflows:compound after solving hard bugs. It pays dividends on future work.
  • Feed the planner. Paste ticket URLs, doc links, and design references into /workflows:plan. Better inputs, better plans.
  • Go autonomous. Use /lrj "feature" when you want the full cycle hands-free -- plan, build, review, test, all driven by the Ralph Loop.

Troubleshooting

Context7 MCP not loading -- Add to .claude/settings.json:

{ "mcpServers": { "context7": { "type": "http", "url": "https://mcp.context7.com/mcp" } } }

Review agents not running -- Make sure compound-engineering.local.md exists. Run /compound-engineering:setup.

agent-browser not installed -- npm install -g agent-browser && agent-browser install. Required for /test-browser and /feature-video.


License

MIT -- see LICENSE for details.

Originally forked from Every Inc's compound-engineering-plugin. Dual copyright (c) 2025 Every Inc, (c) 2025 The Rabak.

About

AI-powered compound engineering plugin for Claude Code. Specialized agents, commands, and skills for code review, research, and workflow automation.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors