Compound Engineering Plugin

2026-09-08 · source · source ↗ · raw capture · wiki page

by Kieran Klaassen and Trevin Chow (Every, Inc.), MIT-licensed, v3.24.0.

What it is

A plugin of 33 skills for AI coding agents that operationalizes the Compound Engineering philosophy as a loop: brainstorm → plan → work → simplify → review → compound. The compound step is the differentiator — each cycle writes a durable "learning" (a solution doc under docs/solutions/) that the next brainstorm and plan read as grounding, so the toolset gets smarter with every use. "Run one teaches it. Run two remembers."

Scope

Runs on 14 agent hosts: Claude Code, Cursor, Codex (app and CLI), Kimi Code CLI, Cline, Grok Build CLI, Devin CLI, GitHub Copilot, Factory Droid, Qwen Code, OpenCode, Pi, oh-my-pi (omp), and Antigravity CLI. The repo root is itself a valid plugin package (plugin.json + skills/), and ships native manifests for each host (.claude-plugin/, .codex-plugin/, .kimi-plugin/, .cursor-plugin/, .grok-plugin/, .devin-plugin/, .omp-plugin/, .pi/, .opencode/, .agy/).

Two parts

  1. The skills (skills/<skill>/SKILL.md + references/ + scripts/): the actual engineering methodology, written as progressively-disclosed markdown. Each skill has an outcome, a done bar, and loads its references/ files only at the acting step. Heavy orchestration (ce-code-review, ce-plan, ce-compound) dispatches generic subagents seeded with skill-local "specialist prompt assets" rather than exposing standalone agents.
  2. The converter (src/, ~8K lines of TypeScript on Bun): parses the Claude-native skill format (frontmatter + body) and converts it to each host's native plugin format — Codex TOML agents, Pi extensions, OpenCode config, Antigravity, Copilot, Kiro, Droid. Uses explicit converters + writers (one per target) and an install-manifest ledger so a writer never overwrites a path it didn't write.

The loop, concretely

/ce-brainstorm (WHAT) → /ce-plan (HOW, a unified plan artifact) → /ce-work (execute, optionally via a cross-model author) → /ce-simplify-code → /ce-code-review (report-only multi-agent review) → /ce-compound (write the learning). /lfg runs the whole pipeline hands-off: plan, work, simplify, review-and-apply-fixes, browser tests, commit, push, PR, CI watch with a bounded repair loop.