Loop Engineering

The Loop Engineering ecosystem hub — index and entry point for all autonomous coding loop engines.

11 autonomous AI coding loop engines — from a single goal statement to production code, without human intervention.

11 个自主 AI 编程循环引擎 —— 从单一目标声明到生产级代码，无需人工干预。

English | 中文

English

🎯 What is Loop Engineering?

Loop Engineering is a methodology for autonomous AI coding. Instead of chatting with an AI assistant turn-by-turn, you give it a goal and it executes a structured multi-phase workflow:

Part 1 — Design Bubble (phases 1.1→1.2→1.3): Requirements → Direction → Solution
Routing Gate: Decide next action (implement / redesign / complete / pause / fail)
Part 2 — Implementation Chain (phases 2.1→2.8): Plan → Implement → Review → Test Strategy → Test Plan → Test Execution → Audit → Hard Verification
Convergence: Auto-detect completion and terminate

All 11 engines share this architecture, each wrapping a different AI tool.

🌐 Project Navigation

#	Project	Wrapped Tool	Language	Description
⭐	loop-everything	Meta-repo	—	Ecosystem hub — you are here
1	loop-aider	Aider CLI	Python	11-phase state machine, 10-point PhaseGuard, atomic write
2	loop-superpowers	Claude Code Skills	YAML/MD	7-skill autonomous mini-loop, Phase Contract DSL
3	loop-ollama	Ollama	Python	ReAct loop, 3-tier fault tolerance, fully local/air-gapped
4	loop-hermes	Hermes SDK	Python	24 modules, 6 gates, provider fallback, parallel delegation
5	loop-antigravity	Gemini API	Python	Circuit breaker, multimodal handler, billing tracker
6	loop-codex	Codex Desktop (CDP)	TypeScript	CDP+JSON-RPC dual-channel, CDPGuard L0/L1/L2
7	loop-copilot	GitHub Copilot SDK	TypeScript	Think-Act-Observe, watchdog, session hooks
8	loop-cursor	Cursor SDK	TypeScript	3-package monorepo, 22-step engine, 27-model registry
9	loop-opencode	OpenCode CLI	TypeScript	8 safety gates, 3-layer architecture
10	loop-openclaw	OpenClaw Gateway	Python	15 Jinja2 templates, dual-engine rendering, 3 topologies
11	loop-deepseek	DeepSeek API	TypeScript	ReAct loop, reasoning_content caching (40-60% token savings)
📋	loop-claudecode	Claude Code	Bash/Skill	Reference implementation, G1/G2/G3 OS-level gates

🤔 Which One Should I Use?

Your Scenario	Recommended Project
"I want the best overall quality"	loop-aider or loop-ollama
"I use Claude Code and want zero-dependency"	loop-superpowers
"I need fully offline / air-gapped operation"	loop-ollama
"I need production-grade safety with provider fallback"	loop-hermes
"I want Google Gemini with cost control"	loop-antigravity
"I need GUI automation (browser testing)"	loop-codex
"I use GitHub Copilot"	loop-copilot
"I use Cursor IDE"	loop-cursor
"I want DeepSeek's reasoning models"	loop-deepseek
"I need to generate agent workspace configs"	loop-openclaw

🙋 FAQ

Q: Isn't this just a wrapper around existing AI tools? A: No. Each engine adds a structured state machine, safety gates, convergence detection, and autonomous multi-turn execution. Without the loop wrapper, the underlying tools require manual turn-by-turn guidance. With it, you state a goal and the engine executes all phases autonomously.

Q: Why 11 different engines instead of one? A: Each engine wraps a different AI tool with unique strengths. Ollama for local/offline, Hermes for production safety, Codex for browser GUI, DeepSeek for reasoning models. Pick the one that fits your stack.

Q: Are these production-ready? A: Each engine implements a structured state machine with safety gates, convergence detection, and multi-phase execution. They are designed for autonomous operation, but we recommend testing in your own environment. See each project's README for detailed metrics and caveats.

Q: How do I contribute or build my own engine? A: Start with loop-claudecode as the reference implementation, then adapt the architecture to wrap your preferred AI tool. See CONTRIBUTING.md.

Q: What's the difference between Part 1 (Design) and Part 2 (Implementation)? A: Part 1 (1.1→1.3) runs in a single continuous context for creative design thinking. Part 2 (2.1→2.8) runs as discrete checkpointed steps for reliable implementation. A routing gate between them decides whether to implement, redesign, or terminate.

📄 License

All projects in the Loop Engineering ecosystem are licensed under Apache License 2.0.

中文