🎭 Deco

Multi-Agent AI Coding System | 多智能体 AI 编程系统

English | 中文

English

What is Deco?

Deco is a TODO-driven multi-agent AI coding system built as a plugin for OpenCode. It orchestrates 8 specialized AI agents to complete software development tasks with persistent tracking that survives session interruptions.

Key Philosophy: The Conductor plans and delegates, specialized agents execute. Work never stops until all TODOs are complete.

Features

🎭 8 Specialized Agents

| Agent | Model | Role | |-------|-------|------| | Conductor | Claude Opus 4.5 | Orchestrator - plans, delegates, never writes code | | Coder | GPT-5.2 Codex | Backend, API, database, CLI | | Artist | Gemini 3 Pro | Frontend, UI, styling, components | | Scout | Claude Haiku 4 | Fast codebase exploration | | Researcher | Claude Sonnet 4.5 | External docs, OSS examples | | Architect | Gemini 3 Pro | System design, architecture decisions | | Reviewer | GPT-5.2 Codex | Code reviews (batch and final) | | Scribe | Gemini 3 Flash | Documentation |

📋 TASK Persistence System

Tasks are stored in llmdoc/tasks/ with full state:

llmdoc/tasks/active/{task-id}/
├── README.md           # Overview and status
├── requirements.md     # Requirements
├── plan.md            # Technical plan
├── progress.json      # Machine-readable state
├── subtasks/          # Individual subtasks
└── reviews/           # Code reviews

Session interrupted? No problem. Resume exactly where you left off.

🔄 TODO-Driven Execution

The iron law: never stop until done.

1. Identify executable subtasks
2. Dispatch to agents in parallel
3. Monitor and update progress
4. Trigger reviews at thresholds
5. Archive on completion

🔌 3 Core Hooks

| Hook | Function | |------|----------| | todo-continuation | Forces completion of pending TODOs | | task-driver | Drives TASK system execution | | checkpoint | Saves progress periodically |

Installation

# Add to your OpenCode config (~/.config/opencode/opencode.json)
{
  "plugin": ["@disdjj/deco"]
}

Configuration

Create .opencode/deco.json (project) or ~/.config/opencode/deco.json (global):

{
  "$schema": "https://raw.githubusercontent.com/TokenRollAI/deco/main/assets/deco.schema.json",
  "agents": {
    "conductor": { "model": "anthropic/claude-sonnet-4-5" },
    "scout": { "model": "opencode/grok-code" }
  },
  "disabled_agents": [],
  "disabled_hooks": [],
  "task": {
    "batchReviewThreshold": 3,
    "checkpointIntervalMs": 300000
  }
}

Usage

Once installed, the Conductor becomes the primary agent. Simply describe your task:

Build a REST API for user authentication with JWT tokens

The Conductor will:

1. Create a TASK structure
2. Break down into subtasks
3. Dispatch to Coder, Scout, Reviewer as needed
4. Track progress until completion

Development

# Install dependencies
bun install

# Type check
bun run typecheck

# Build
bun run build

# Run tests
bun test

中文

什么是 Deco？

Deco 是一个基于 TODO 驱动的多智能体 AI 编程系统，作为 OpenCode 的插件构建。它协调 8 个专业 AI 智能体来完成软件开发任务，并具有持久化追踪功能，即使会话中断也能恢复。

核心理念：Conductor 负责规划和分配，专业智能体负责执行。工作永不停止，直到所有 TODO 完成。

特性

🎭 8 个专业智能体

| 智能体 | 模型 | 职责 | |--------|------|------| | Conductor | Claude Opus 4.5 | 协调者 - 规划、分配，从不直接写代码 | | Coder | GPT-5.2 Codex | 后端、API、数据库、CLI | | Artist | Gemini 3 Pro | 前端、UI、样式、组件 | | Scout | Claude Haiku 4 | 快速代码库探索 | | Researcher | Claude Sonnet 4.5 | 外部文档、开源示例 | | Architect | Gemini 3 Pro | 系统设计、架构决策 | | Reviewer | GPT-5.2 Codex | 代码审查（批量和最终） | | Scribe | Gemini 3 Flash | 文档编写 |

📋 TASK 持久化系统

任务存储在 llmdoc/tasks/ 中，包含完整状态：

llmdoc/tasks/active/{task-id}/
├── README.md           # 概述和状态
├── requirements.md     # 需求文档
├── plan.md            # 技术方案
├── progress.json      # 机器可读的状态
├── subtasks/          # 子任务详情
└── reviews/           # 代码审查记录

会话中断？没问题。 从上次离开的地方精确恢复。

🔄 TODO 驱动执行

铁律：永不停止，直到完成。

1. 识别可执行的子任务
2. 并行分配给智能体
3. 监控并更新进度
4. 在阈值触发审查
5. 完成后归档

🔌 3 个核心 Hooks

| Hook | 功能 | |------|------| | todo-continuation | 强制完成待处理的 TODO | | task-driver | 驱动 TASK 系统执行 | | checkpoint | 定期保存进度 |

安装

# 添加到你的 OpenCode 配置 (~/.config/opencode/opencode.json)
{
  "plugin": ["deco"]
}

配置

创建 .opencode/deco.json（项目级）或 ~/.config/opencode/deco.json（全局）：

{
  "$schema": "https://raw.githubusercontent.com/TokenRollAI/deco/main/assets/deco.schema.json",
  "agents": {
    "conductor": { "model": "anthropic/claude-sonnet-4-5" },
    "scout": { "model": "opencode/grok-code" }
  },
  "disabled_agents": [],
  "disabled_hooks": [],
  "task": {
    "batchReviewThreshold": 3,
    "checkpointIntervalMs": 300000
  }
}

使用

安装后，Conductor 成为主要智能体。只需描述你的任务：

构建一个带有 JWT 令牌的用户认证 REST API

Conductor 将会：

1. 创建 TASK 结构
2. 分解为子任务
3. 根据需要分配给 Coder、Scout、Reviewer
4. 追踪进度直到完成

开发

# 安装依赖
bun install

# 类型检查
bun run typecheck

# 构建
bun run build

# 运行测试
bun test

License

MIT © DJJ