AutoDev — 让 AI 真正能交付大型项目的开发引擎。

需求分析→系统设计→代码执行→质量审计→交付验收，一键完成大型项目全自动开发。 Requirements Analysis → System Design → Code Execution → Quality Audit → Delivery Acceptance. One command to autonomously develop large-scale projects.

#生产力#产品

评分

需要更多评价

订单数

如何使用

在 Capafy 上运行

也可用于外部应用

创作者提供

GLM 5.1

AutoDev · Full-Stack Autonomous Development Engine

From requirements to delivery in one command. It doesn't change scope, it doesn't quit — and when it hits an error, it fixes itself. If it can't, it asks for help.

🎯 One-Liner

AutoDev is an AI-powered autonomous development engine for large-scale projects. Describe what you need — it handles everything: requirements analysis → system design → code implementation → quality audit → delivery. Supports greenfield development, checkpoint resume, and mid-flight requirement changes.

✨ Key Capabilities

Capability	Description
Full-Pipeline Automation	Requirements → Design → Execute → Audit → Deliver, zero manual intervention
Checkpoint Resume	Auto-validates state after interruption, picks up where it left off — no lost progress
Mid-Flight Pivot	Requirements changed mid-project? `pivot` command seamlessly adapts
Overnight Mode	Crash recovery + module-level isolated processes + DAG parallelism — ship while you sleep
Strict TDD	Tests first, code second. Full test suite runs after every task
API Contract Verification	Auto-compares frontend/backend field names — kills the classic "interface mismatch" bug
Self-Learning Loop	Mistakes are recorded automatically and prevented in future projects
Quota Management	Tracks token usage in real-time, intelligently downgrades when limits approach

🚀 Usage

Full Development (Greenfield)

/autodev "Build an HR management system with employee management, attendance tracking, and payroll calculation"

Other Commands

/autodev run            # Execute from existing design docs
/autodev resume         # Resume from interruption (with auto state validation)
/autodev pivot "add export feature"  # Mid-flight requirement change
/autodev status         # Check current progress
/autodev doctor         # Self-diagnostic
/autodev overnight      # Overnight mode: crash recovery + DAG parallelism

Lightweight Entry Points (Existing Projects)

/autodev review         # Full code review
/autodev audit          # Security audit
/autodev optimize       # Performance/quality optimization
/autodev fix "login returns 500"  # Root cause → fix → regression test
/autodev iterate "add search"     # Small feature iteration

Watch Mode (Periodic Tasks)

/autodev watch "check PR status every 5 minutes"
/autodev watch "keep fixing failing tests until all pass"

🔄 Execution Pipeline (5 Phases)

Phase 0  Requirements Clarification
  │  Deep interview → Tech stack selection → Architecture design → API contracts → Task decomposition
  ▼
Phase 1  Autonomous Execution
  │  TDD coding → Incremental testing → Smoke verification → Contract verification → Auto commit
  ▼
Phase 2  Quality Audit
  │  Static analysis → Security scan → Dead code detection → Coverage check → Adversarial review
  ▼
Phase 3  Delivery & Acceptance
  │  Documentation generation → Git wrap-up → Environment cleanup → Delivery report
  ▼
Phase 3.5  Retrospective
     Lesson extraction → Self-learning loop → Knowledge retention

🛡️ Nine Execution Disciplines

#	Discipline	Why It Matters
1	Incremental Validation — Run full test suite after every task	Batched testing causes error accumulation; debugging cost grows exponentially
2	Module Interface Protocol — Read and confirm before execution	Implicit dependencies are the #1 source of cross-module bugs
3	Review Closure — Re-verify each module against checklist	CRITICAL issues will slip through without explicit verification
4	E2E Smoke Tests — Run minimal E2E after every module	Data flow breaks between modules get harder to fix the later they're found
5	Self-Learning Loop — Never make the same mistake twice	Recurring errors indicate a broken feedback loop
6	Execution Continuity — Stop Guard + Auto Mode triple guarantee	Slow is fast — every step must be solid before moving on
7	Impact Pre-Check — Grep for references before changing enums/models	Enum changes trigger test avalanches in ~40% of failures
8	Mock Transparency — Mock tests must verify real signatures	270 mock tests all passed, but production crashed
9	API Contract Consistency — Auto-compare frontend/backend fields	Dashboard showed 0 people — root cause was a field name mismatch

🔧 Technical Architecture

Error Handling (Auto-Classification + Circuit Breaker)

Error Type	Action
Network glitch / SDK crash	Auto-backoff and retry
Validation failure	Auto-repair
Semantic error	Re-plan
Quota exhausted	Pause and wait

Circuit breaker: Same task fails 3× → upgrade to stronger model; 5× → mark as blocked, write report.

Context Management

Each phase loaded on demand, no pre-loading
Task execution runs in isolated sub-agents, main context keeps only summaries
Auto-restarts new session every 3 modules to prevent context bloat

Quota Management

Usage	Level	Action
50-69%	SLOW	Route simple tasks to lighter models
70-84%	THROTTLE	Auto-downgrade
85-94%	WARN	Pause after current module completes
≥95%	CRITICAL	Immediate pause, save state

📊 Use Cases

✅ Best For	❌ Not For
Mid-to-large full-stack projects	One-off scripts / small tools
Projects requiring strict quality assurance	Quick prototypes without tests
Long-running projects needing checkpoint resume	Simple Q&A scenarios
Team projects with automated dev pipelines	Conservative teams that don't trust AI coding

📁 File Structure

autodev/
├── SKILL.md              ← Core engine (routing + rules + disciplines)
├── phase0.md             ← Phase 0: Requirements → Architecture
├── phase1.md             ← Phase 1: Autonomous execution
├── phase2.md             ← Phase 2: Quality audit
├── phase3.md             ← Phase 3: Delivery & acceptance
├── phase3.5.md           ← Phase 3.5: Retrospective
├── phase-lite.md         ← Lightweight mode (review/audit/fix etc.)
├── error-handling.md     ← Error handling protocol
├── hooks/                ← 15+ automation hook scripts
├── templates/            ← Template files
└── learnings.json        ← Self-learning experience database

🏆 Version History

Version	Highlights
v1.20	Auto Mode integration, dual-track permission system, DAG parallelism boost
v1.19	Stop Guard for execution continuity
v1.18	API contract verification, SPA route smoke tests, Mock transparency enforcement
v1.17	Overnight mode, crash recovery, module-level isolated processes
v1.16	Module-boundary segmented restarts, eliminates context bloat
v1.13	Auto error classification, per-module circuit breaker, auto-commit
v1.12	DAG parallel execution, context budget, impact pre-check

AutoDev — The development engine that actually ships large projects.

AutoDev · 大型项目全自动开发引擎

从需求到交付，一键完成。不改需求不走人，遇到错误自己修，修不好会喊你。

🎯 一句话介绍

AutoDev 是一个 AI 驱动的大型项目全自动开发引擎。你只需描述需求，它自动完成：需求分析 → 系统设计 → 代码实现 → 质量审计 → 交付验收。支持从零开发、断点恢复、中途需求变更。

✨ 核心能力

能力	说明
全流程自动化	需求→设计→执行→审计→交付，零人工干预
断点恢复	中断后自动校验状态，从断点继续，不丢进度
需求变更	开发中途改需求？`pivot` 命令无缝切换
夜间模式	崩溃自动恢复 + 模块级独立进程 + DAG 并行，睡一觉项目搞定
TDD 严格流程	先写测试再写代码，每任务完成跑全量测试
API 契约验证	前后端字段自动比对，消灭「接口对不上」的经典 bug
自学习闭环	犯过的错误自动记录，下次项目自动预防
配额管理	自动追踪 token 消耗，快超限时智能降级

🚀 用法

完整开发（从零开始）

/autodev "开发一个人事管理系统，包含员工管理、考勤、薪资计算"

其他命令

/autodev run            # 从已有设计文档直接执行
/autodev resume         # 从中断恢复（含自动状态校验）
/autodev pivot "加一个导出功能"  # 中途需求变更
/autodev status         # 查看当前进度
/autodev doctor         # 自检诊断
/autodev overnight      # 夜间模式：崩溃恢复 + DAG 并行

轻量入口（已有项目）

/autodev review         # 全面代码审查
/autodev audit          # 安全审计
/autodev optimize       # 性能/质量优化
/autodev fix "登录报500" # 根因定位 → 修复 → 回归测试
/autodev iterate "加个搜索" # 小功能迭代

定时任务

/autodev watch "每5分钟检查PR状态"
/autodev watch "持续修复失败的测试，直到全部通过"

🔄 执行流程（5 Phase）

Phase 0  需求澄清
  │  深度访谈 → 技术选型 → 架构设计 → API 契约 → 任务拆解
  ▼
Phase 1  自动执行
  │  TDD 编码 → 增量测试 → 冒烟验证 → 契约验证 → 自动 commit
  ▼
Phase 2  质量审计
  │  静态分析 → 安全扫描 → 死代码检测 → 覆盖率检查 → 对手审查
  ▼
Phase 3  交付验收
  │  文档生成 → Git 收尾 → 环境清理 → 交付报告
  ▼
Phase 3.5  复盘归档
     经验提取 → 自学习闭环 → 知识沉淀

🛡️ 九大执行纪律

#	纪律	为什么
1	增量验证 — 每任务完成后跑全量测试	攒批跑测试，错误堆积定位成本指数级增长
2	模块接口协议 — 执行前读取确认	隐式依赖是跨模块 bug 的头号来源
3	审查闭环 — 每模块回到 checklist 逐项验证	不显式验证，CRITICAL 问题必然遗漏
4	端到端冒烟 — 每模块完成后跑最小 E2E	模块间数据流断裂越晚发现越难修
5	自学习闭环 — 犯过的错不再犯	同类错误不应犯两次
6	连续性保障 — Stop Guard + Auto Mode 三重保障	慢就是快，每步踏实才下一步
7	影响面预检 — 改枚举/模型前先 grep 扫描引用	枚举改动引发测试雪崩占失败 ~40%
8	Mock 透明性 — mock 测试必须验证真实签名	270 个 mock 测试全过但生产崩溃
9	API 契约一致性 — 前后端字段自动比对	Dashboard 显示 0 人，根因是字段名不匹配

🔧 技术架构

错误处理（自动分类 + 熔断）

错误类型	动作
网络抖动 / SDK 崩溃	自动退避重试
校验失败	自动修复
语义错误	重新规划
配额不足	暂停等待

熔断机制：同一任务失败 3 次 → 升级到更强模型；5 次 → 标记阻塞，写入报告。

上下文管理

每个 Phase 按需加载，不预加载
任务执行走独立子 Agent，主上下文只保留摘要
每完成 3 个模块自动重启新会话，防止上下文膨胀

配额管理

使用率	级别	动作
50-69%	SLOW	简单任务用轻量模型
70-84%	THROTTLE	自动降级
85-94%	WARN	完成当前模块后暂停
≥95%	CRITICAL	立即暂停，保存状态

📊 适用场景

✅ 适合	❌ 不适合
中大型全栈项目开发	一次性脚本 / 小工具
需要严格质量保证的项目	不需要测试的快速原型
需要断点恢复的长周期项目	只需简单问答的场景
团队项目的自动化开发流水线	不信任 AI 编码的保守场景

📁 文件结构

autodev/
├── SKILL.md              ← 核心引擎（路由 + 规则 + 纪律）
├── phase0.md             ← Phase 0：需求 → 架构
├── phase1.md             ← Phase 1：自动执行
├── phase2.md             ← Phase 2：质量审计
├── phase3.md             ← Phase 3：交付验收
├── phase3.5.md           ← Phase 3.5：复盘归档
├── phase-lite.md         ← 轻量模式（review/audit/fix 等）
├── error-handling.md     ← 错误处理协议
├── hooks/                ← 15+ 自动化 Hook 脚本
├── templates/            ← 模板文件
└── learnings.json        ← 自学习经验库

🏆 版本历史

版本	亮点
v1.20	Auto Mode 适配，权限体系双轨制，DAG 并行提升
v1.19	Stop Guard 执行连续性保障
v1.18	API 契约验证、SPA 路由冒烟、Mock 透明性加固
v1.17	夜间模式、崩溃恢复、模块级独立进程
v1.16	模块边界分段重启，根治上下文膨胀
v1.13	错误自动分类、per-module 熔断、自动 commit
v1.12	DAG 并行执行、上下文预算、影响面预检

AutoDev — 让 AI 真正能交付大型项目的开发引擎。

AI Automation Product Ops Builder

Build a bilingual digital-product or productized-service system covering offer validation, a seven-day build plan, payment and delivery, client intake, content repurposing, launch copy, tracking, and measurable experiments.

US$49

Autowin : One shot any task

We built Autowin by putting models in a loop: discover upgrades → test them → judge them with an assembly of adversarial sub-agents — and we ran that very process on itself, iterating the kit until each piece held up under its own review. The result: from one loosely-defined need, you get back work that has been built, tested, and scored by a panel of judge sub-agents — with closure authority kept outside the model (deterministic checks + you), so an "it's done" is never just taken on faith.

US$29.99

A.I. Integration

AI Integration — Plans and builds LLM API connections (OpenAI, Anthropic, and more) into real apps. Covers model selection, cost estimation, key safety, error handling, and working code. Adapts to non-engineers and developers alike. Stops the #1 mistake: shipping your API key to the browser.

US$10