ECC: Turning AI Programming into an Auditable, Evolvable Engineering System

{
  "repoFullName": "affaan-m/everything-claude-code",
  "repoUrl": "https://github.com/affaan-m/everything-claude-code",
  "repoName": "everything-claude-code",
  "language": "javascript",
  "stars": 83098,
  "analysisContent": "Hi, I'm Zhou Xiaoma — an 8-year Java veteran who's been dizzy from Spring Boot auto-configuration, cried over Hibernate's second-level cache pitfalls, and woken up in cold sweat after seeing Claude Code's `/model opus` bill. Today, we won't talk about JVM tuning or dig into MyBatis internals. Instead, let's dissect this GitHub sensation — now at 83K stars — whose README ships in five languages (Chinese, Japanese, Korean, Traditional Chinese, English): **everything-claude-code** (henceforth ECC).\n\nLet’s start with blunt honesty: This isn’t a library, not a framework, nor an SDK — it’s an \"AI programmer’s field manual + tactical gear box + military-grade logistics hub.\" Plug it into Claude Code, Cursor, Codex, or OpenCode, and it’s like equipping a special forces operator with night vision, a tactical tablet, real-time intel feeds, and programmable munitions — it won’t pull the trigger for you, but ensures every shot is precise, ammo-efficient, reviewable, and evolutionary.\n\n### 🧩 What problem does it *actually* solve?\n\nDon’t be fooled by the \"Claude Code\" name — ECC targets the *industrial pain points of AI programming agents*:\n- **Writes but doesn’t review**: Agents generate code, yet lack senior architects’ instinct to ask, \"Should this API be idempotent?\"\n- **Runs but doesn’t remember**: Every conversation is \"amnesiac development\" — last week’s optimized SQL index? Rewritten from scratch today.\n- **Has skills but no system**: Scattered prompts, isolated hooks, hand-crafted rules — like LEGO bricks dumped unsorted on the floor.\n- **Cross-platform via voodoo**: A hook that works flawlessly in Cursor crashes in Codex; an agent running smoothly in Claude Code goes silent in OpenCode.\n\nECC’s answer is hardcore: **Turn AI programming into a configurable, testable, auditable, and evolvable engineering system.** It decomposes the mission of \"making AI smarter\" into five pluggable, composable, and canary-deployable subsystems: Skills (skill pipelines), Instincts (intuition memory), Memory (context lifecycle management), Security (runtime defense layer), and Research-first (pre-research workflow). This isn’t magic — it’s 10 months and 2,000+ real-project iterations, compiled into YAML, Markdown, and Node.js scripts.\n\n### ⚙️ What does its architecture resemble? — An autonomous \"AI OS kernel\"\n\nAt first glance at its directory structure, my brain instantly flashed Linux kernel metaphors:\n- `agents/` = **Process scheduler** (planner, architect, security-reviewer… each a lightweight coroutine focused on a single responsibility)\n- `skills/` = **System call library** (tdd-workflow, strategic-compact, cost-aware-llm-pipeline… each SKILL.md encapsulates an atomic capability)\n- `hooks/` = **Interrupt handler** (`beforeFileEdit` automatically scans for `console.log`; `beforeSubmitPrompt` blocks API key leaks — stricter than frontend XSS protection)\n- `rules/` = **Kernel-enforced policy** (common/ says \"all PRs must have ≥80% coverage\"; typescript/ mandates \"TS interfaces require JSDoc\" — violation? Hook fails fast and exits)\n- `scripts/` = **User-space toolchain** (fully rewritten in Node.js, cross-platform across Windows/macOS/Linux — even PowerShell-compatible — yes, it’s *that* obsessive)\n\nMost brilliant is its **cross-platform adapter layer**: Cursor exposes 15 hook events; Claude Code only 8. How does ECC run one codebase everywhere? The secret lives in `.cursor/hooks/adapter.js` — it translates Cursor’s stdin JSON stream into Claude Code’s hook format in real time, then feeds it into the shared `scripts/hooks/session-start.js`. This isn’t adaptation — it’s building an HTTP/2 tunnel *between different AI IDEs!*\n\n### 💡 Code examples: Not Hello World — \"Hello Production\"\n\nECC has no traditional `npm install && node index.js`. Its \"installation\" *is* an engineering exercise. Here are three most representative snippets:\n\n#### 【Installation】— Not dependency, but \"deployment\"\n```bash\n# Step 1: Plugin-based integration (like installing a VS Code extension)\n/plugin marketplace add affaan-m/everything-claude-code\n/plugin install everything-claude-code@everything-claude-code\n\n# Step 2: Manual rule injection (⚠️ Critical! Claude Code doesn’t support automatic rule distribution)\ngit clone https://github.com/affaan-m/everything-claude-code.git\ncd everything-claude-code\n./install.sh typescript python golang  # Install language stacks on demand\n```\n\nBehind these commands lies ECC’s profound insight into AI programming paradigms: **Rules aren’t optional config — they’re the constitution of production environments.** It prefers users typing two extra lines over compromising on \"rules silently disabled post-installation\" pseudo-convenience.\n\n#### 【Quick Start】— Your first AI-powered collaboration command\n```bash\n# Launch a feature plan with security review\n/everything-claude-code:plan \"Add user authentication with OAuth\"\n\n# Trigger TDD workflow immediately (no tab switching, no terminal opening)\n/tdd\n\n# Let AI conduct real code review (not ceremonial — OWASP Top 10 verified)\n/code-review\n```\n\nSee it? This isn’t `console.log('Hello')`. It’s **initiating a cross-role collaboration meeting inside your IDE**: planner sets the route, tdd-guide controls cadence, code-reviewer guards the baseline. Three slash-commands orchestrate a full R&D pipeline.\n\n#### 【Advanced usage】— Runtime dynamic tuning (this is where it gets hardcore)\n```bash\n# Dynamically switch hook strictness level (akin to Kubernetes PodSecurityPolicy)\nexport ECC_HOOK_PROFILE=strict\n\n# Temporarily disable error-prone hooks (faster than deleting code, faster than restarting)\nexport ECC_DISABLED_HOOKS=\"pre:bash:tmux-reminder,post:edit:typecheck\"\n\n# Save LLM tokens: default to Sonnet, escalate to Opus only for deep reasoning\n# ~/.claude/settings.json\n{\n  \"model\": \"sonnet\",\n  \"env\": {\n    \"MAX_THINKING_TOKENS\": \"10000\",\n    \"CLAUDE_AUTOCOMPACT_PCT_OVERRIDE\": \"50\"\n  }\n}\n```\n\nThis config reminds me of tuning JVM flags — not \"bigger is better\", but finding the **golden balance among performance, cost, and stability**. ECC slashes `MAX_THINKING_TOKENS` from the default 32K down to 10K — saving 70% hidden token costs. It drops auto-compaction threshold from 95% to 50%, preventing late-conversation \"eureka moments turning into eureka blackouts\". This is no longer prompt engineering — it’s **LLM resource orchestration engineering**.\n\n### 🔍 As a Java veteran, here’s my honest experience\n\nFrankly, my hand shook when I first saw `skills/springboot-security/` and `skills/java-coding-standards/` — finally, someone turned Spring Security’s CSRF protections, JWT refresh logic, and Actuator exposure risks into executable SKILL.md files! But I also spotted two \"Java-exclusive minefields\":\n1. **MCP server conflict**: ECC enables 14 MCPs by default (Supabase/Vercel/Railway…), while our Java stack relies on Spring Cloud Gateway + Nacos + Seata. Blindly enabling all shrinks the 200K context window down to 70K — recommend precise disabling via `\"disabledMcpServers\": [\"supabase\", \"vercel\"]` in `settings.json`.\n2. **Hook vs. Spring Boot DevTools philosophical clash**: ECC’s `post:edit:typecheck` hook triggers compilation checks *immediately* upon saving a Java file, while DevTools hot-reload waits for class file generation. Fix? Refine the hook matcher to `\"tool == \\\"Edit\\\" && tool_input.file_path matches \\\"\\\\.(java)$\\\"\"`, then invoke `mvn compile -q` instead of raw `javac` — which perfectly embodies ECC’s design philosophy: **It provides the armory — but the hand pulling the trigger remains yours.**\n\n### ✅ Is it worth deep diving? My verdict is crystal clear:\n- **If you’re an AI-native developer**: ECC is mandatory. Its `skills/continuous-learning-v2/` auto-extracts coding patterns from your Git history — more accurate than *my own* \"Spring Boot Best Practices\" book.\n- **If you’re a Java tech lead**: Don’t roll out enterprise-wide yet. Pilot `skills/springboot-tdd/` and `skills/jpa-patterns/`: enforce TestNG via `/tdd`, detect N+1 queries via `/jpa-review` — *that’s* how AI truly empowers teams.\n- **If you’re still using Copilot for CRUD**: ECC may be overkill. It’s built for teams already tasting AI’s sweetness — now crushed under the triple burden of \"unstable output quality, lost context, and unverifiable security.\"\n\nFinally, a golden line from ECC’s README: **\"Star this repo if it helps. Read both guides. Build something great.\"** — I’ve starred it, read both guides, and next? Time to write `/skill-create --instincts`, feeding ECC’s continuous learning engine with my company’s 3 years of Spring Cloud Alibaba battle-tested experience. After all, the real productivity revolution was never about replacing humans with AI — it’s about evolving *how humans wield AI*, transforming us from spellcasters into engineers.",  "codeExamples": [\n    {\n      \"type\": \"installation\",\n      \"description\": \"Plugin-based installation + manual rule injection (Claude Code official restriction)\",\n      \"code\": \"# Step 1: Plugin-based integration\n/plugin marketplace add affaan-m/everything-claude-code\n/plugin install everything-claude-code@everything-claude-code\n\n# Step 2: Manual rule injection (mandatory!)\ngit clone https://github.com/affaan-m/everything-claude-code.git\ncd everything-claude-code\n./install.sh typescript python golang\"\n    },\n    {\n      \"type\": \"quickstart\",\n      \"description\": \"Three-step launch of AI-powered collaborative R&D pipeline\",\n      \"code\": \"# 1. Plan feature (with security review)\n/everything-claude-code:plan \\\"Add user authentication with OAuth\\\"\n\n# 2. Enforce TDD workflow\n/tdd\n\n# 3. Auto code review\n/code-review\"\n    },\n    {\n      \"type\": \"advanced\",\n      \"description\": \"Runtime dynamic tuning: hook policy + token cost control\",\n      \"code\": \"# Dynamic hook profile\nexport ECC_HOOK_PROFILE=strict\nexport ECC_DISABLED_HOOKS=\\\"pre:bash:tmux-reminder,post:edit:typecheck\\\"\n\n# Token optimization in ~/.claude/settings.json\n{\n  \\\"model\\\": \\\"sonnet\\\",\n  \\\"env\\\": {\n    \\\"MAX_THINKING_TOKENS\\\": \\\"10000\\\",\n    \\\"CLAUDE_AUTOCOMPACT_PCT_OVERRIDE\\\": \\\"50\\\"\n  }\n}\"\n    }\n  ],\n  \"keyFeatures\": [\"Cross-platform AI agent system (full support for Claude Code/Cursor/Codex/OpenCode)\", \"Four-layer architecture: Skills–Instincts–Memory–Security\", \"Runtime hook policy engine & token cost optimization\"],\n  \"techStack\": [\"TypeScript\", \"Node.js\", \"Markdown\", \"YAML\", \"Shell\"],\n  \"suggestedTags\": \"AI programming, intelligent agents, LLM engineering, developer tools, ClaudeCode\"\n}

{\n  \"title\": \"English title (highlighting technical value)\",\n  \"summary\": \"English summary (emphasizing technical highlights)\",\n  \"content\": \"Full English content (preserving all code)\",\n  \"category\": \"Open Source\",\n  \"tags\": \"GitHub,OpenSource,corresponding technical tags\",\n  \"zhBlogId\": \"Extract blogId from chinese_article response\",\n  \"repoUrl\": \"From github_analysis\",\n  \"repoName\": \"From github_analysis\"\n}