Execution protocol for AI coding agents. Routes tasks through complexity tiers, enforces phase gates, TDD, evidence grading, and verification gates.
Apex Forge is a protocol orchestrator, not a framework. It provides execution discipline (complexity routing, phase gates, evidence grading) while delegating domain knowledge to independent companion skills. The pipeline runs: brainstorm, plan, execute, review, ship, compound.
Seven companion skills handle specialized work:
systematic-debugging— root-cause debugging with Iron Law disciplinethorough-code-review— pre-merge review + incoming feedback evaluationsecurity-audit— infrastructure-first CWE-tagged auditbrowser-qa-testing— tiered QA with headless browser automationtasteful-frontend— opinionated frontend design guidancedesign-to-code-runner— spec-first Figma/screenshot to codeproduct-review— product experience evaluation
All companion skills install automatically. Each works standalone or orchestrated through the AF pipeline.
# Clone and install (installs AF + all 7 companion skills)
git clone https://github.com/d-wwei/apex-forge
cd apex-forge
bash skill/install.sh
# Initialize a project
apex init
# Run the pipeline
/apex-forge brainstorm # requirements exploration
/apex-forge plan # implementation planning
/apex-forge execute # TDD-first implementation
/apex-forge review # multi-persona quality gate
/apex-forge ship # commit, push, PR| Mechanism | What it does |
|---|---|
| Complexity Router | Tier 1 (single pass), Tier 2 (round-based PDCA), Tier 3 (wave-based cross-session) |
| Phase Gates | Brainstorm (WHAT), Plan (HOW), Execute (DO). No leaking between phases. |
| TDD Iron Law | Write test, RED, implement, GREEN, refactor. No exceptions. |
| Evidence Grading | E0 (guess) through E4 (validated). Minimum E2 for action, E3 for success claims. |
| Verification Gate | 5-step gate before any success claim. Skip a step = lying. |
| Escalation Ladder | L0-L4. After 5 failures, escalate to human with full context. |
| Skill Dispatch | bindings.yaml maps pipeline stages to external skills with version constraints. |
apex-forge (this repo)
├── skill/SKILL.md Core protocol
├── skill/bindings.yaml Stage → external skill mapping
├── skill/gates/ AF-owned quality gates
├── skill/stages/ Pipeline stage definitions
├── skill/roles/ Orchestration + utility roles
├── skill/aliases/ Backward-compatible command aliases
├── src/ CLI + MCP server + state management
└── dist/ Compiled binaries
│
│ Hard dependencies (install.sh auto-installs)
▼
7 companion skill repos (independent, each works standalone)
Works with any AI agent that can execute shell commands: Claude Code, Codex, Gemini CLI, Cursor, OpenCode, Windsurf.
# Detected automatically during install:
~/.claude/skills/apex-forge # Claude Code
~/.codex/skills/apex-forge # Codex
~/.gemini/skills/apex-forge # Gemini
~/.opencode/skills/apex-forge # OpenCodebash skill/install.sh updatePulls latest for all companion skills. AF's session-start hook also checks for missing skills on every session and installs them automatically.
MIT
AI 编程 Agent 的执行协议。通过复杂度分级路由任务,强制执行阶段门控、TDD、证据分级和验证门。
Apex Forge 是协议编排器,不是框架。核心价值是执行纪律(复杂度路由、阶段门控、证据分级),领域能力由独立的 companion skill 提供。Pipeline 流程:brainstorm, plan, execute, review, ship, compound。
7 个 companion skill 处理专业工作:
systematic-debugging— Iron Law 根因调试thorough-code-review— 代码审查(发起 + 接收反馈)security-audit— 基础设施优先的 CWE 安全审计browser-qa-testing— 分级 QA 测试 + 无头浏览器自动化tasteful-frontend— 有态度的前端设计指导design-to-code-runner— 设计稿还原代码product-review— 产品体验评审
所有 companion skill 自动安装。每个既可独立使用,也可通过 AF pipeline 编排。
# 克隆并安装(自动安装 AF + 全部 7 个 companion skill)
git clone https://github.com/d-wwei/apex-forge
cd apex-forge
bash skill/install.sh
# 初始化项目
apex init
# 执行 pipeline
/apex-forge brainstorm # 需求探索
/apex-forge plan # 实现规划
/apex-forge execute # TDD 实现
/apex-forge review # 多角色质量门
/apex-forge ship # 提交、推送、PR| 机制 | 作用 |
|---|---|
| 复杂度路由 | Tier 1(单次通过)、Tier 2(PDCA 轮次)、Tier 3(跨会话波次) |
| 阶段门控 | Brainstorm(WHAT)、Plan(HOW)、Execute(DO),阶段间不可泄漏 |
| TDD 铁律 | 写测试, RED, 实现, GREEN, 重构,无例外 |
| 证据分级 | E0(猜测)到 E4(多源验证),行动最低 E2,成功声明最低 E3 |
| 验证门 | 5 步门控,任何成功声明前必须通过。跳步 = 撒谎 |
| 升级阶梯 | L0-L4,5 次失败后带完整上下文升级给人类 |
| Skill 派发 | bindings.yaml 将 pipeline 阶段映射到外部 skill,含版本约束 |
适用于任何能执行 shell 命令的 AI Agent:Claude Code、Codex、Gemini CLI、Cursor、OpenCode、Windsurf。
bash skill/install.sh updateSession-start hook 每次启动时自动检查并安装缺失的 companion skill。
MIT