A multi-agent personal decision-making system inspired by the Tang Dynasty's "Three Departments and Six Ministries" (三省六部) structure.
唐朝三省六部制启发的多智能体个人决策系统。
Palace decomposes complex decisions into a structured four-phase pipeline — Draft → Review → Finalize → Execute — powered by 9 AI agents collaborating as the Three Departments and Six Ministries.
Palace 将复杂决策分解为「起草 → 审议 → 定稿 → 执行」四阶段流程，由九个 AI 智能体（三省六部）协同完成，帮助个人做出更全面、更理性的决策。
| Phase | Agent | Role |
|---|---|---|
| Draft | Zhongshu (中书省) | Analyze the issue, generate multiple options |
| Review | Menxia (门下省) | Audit the draft, identify risks and gaps |
| Finalize | Shangshu (尚书省) | Synthesize draft + review into a final decision |
| Ministry | Role |
|---|---|
| Libu (吏部) | Personnel & role assignments |
| Hubu (户部) | Budget & resource planning |
| Libu Ritual (礼部) | Process & communication |
| Bingbu (兵部) | Execution steps & milestones |
| Xingbu (刑部) | Risk mitigation & compliance |
| Gongbu (工部) | Tools, templates & deliverables |
1. As a Claude Code Skill (Recommended) | 作为 Claude Code 技能（推荐）
No API key needed. Claude itself acts as all 9 agents.
无需 API Key，Claude 自身扮演全部 9 个智能体角色。
/palace Should I quit my job to start a business?
/palace 帮我决定要不要辞职去创业
See Claude Code Plugin section below for installation.
2. As a Standalone CLI | 作为独立命令行工具
Requires a DashScope API key (Alibaba Cloud).
需要阿里云 DashScope API Key。
export DASHSCOPE_API_KEY="your-api-key"
# Batch mode | 批处理模式
python3 mvp.py --issue tests/fixtures/issue_tech_stack.json
# Interactive mode | 交互模式
python3 interactive.py --issue tests/fixtures/issue_tech_stack.jsonInteractive mode pauses at each checkpoint for user input:
交互模式支持在每个关键节点暂停，让用户参与决策：
- Draft selection | 中书省草案选择
- Review feedback | 门下省审议意见注入
- Final approval | 尚书省定稿确认
- Ministry output review | 六部执行方案审批
Option A: From GitHub | 从 GitHub 安装
/plugin marketplace add zx12671/palace
/plugin install palace@palace
Option B: Project-level (auto) | 项目级（自动生效）
Clone this repo and work inside it. The skill at .claude/skills/palace/ activates automatically.
克隆此仓库并在其中工作，.claude/skills/palace/ 下的技能自动生效。
合：想在自己的项目里直接用 /palace 命令。
# 1. 在你的项目根目录下创建 .claude/skills/ 目录
mkdir -p /path/to/your-project/.claude/skills/
# 2. 从 palace 仓库复制 skill
cp -r /path/to/palace/.claude/skills/palace /path/to/your-project/.claude/skills/palace
# 3. 进入你的项目，打开 Claude Code
cd /path/to/your-project
claude进入 Claude Code 后直接输入：
/palace 帮我决定要不要辞职去创业
就这样，不需要 API Key，不需要安装任何依赖。Claude 自身扮演全部 9 个智能体。
/palace I need to decide between React and Vue for my frontend
/palace 我应该选 React 还是 Vue 做前端
- Build Issue — Claude helps structure your decision as JSON, you confirm
- Three Departments — Draft (options) → Review (risks) → Finalize (recommendation)
- Emperor's Ruling — You approve, reject, or revise the final decision
- Six Ministries — Execute with roles, budget, process, steps, risks, tools
- Report — Full
decision.mdwith all analysis
- 构建议题 — Claude 帮助将决策结构化为 JSON，皇上确认
- 三省流程 — 起草（方案）→ 审议（风险）→ 定稿（推荐）
- 皇上裁决 — 批准、打回或修改定稿
- 六部执行 — 分工、资源、流程、执行、风险、工具
- 报告 — 生成完整的
decision.md决策报告
palace/
├── palace/ # Core package | 核心包
│ ├── llm.py # DashScope API client (stdlib only)
│ ├── agents.py # Agent definitions & runner
│ ├── session.py # Interactive session state machine
│ └── renderer.py # Markdown report generator
├── templates/
│ └── agent_prompts/ # 9 agent prompt templates
├── schemas/ # JSON schemas for input/output
├── tests/
│ └── fixtures/ # Sample issue JSON files
├── palace-plugin/ # Claude Code plugin (self-contained)
│ ├── .claude-plugin/
│ │ └── plugin.json
│ └── skills/palace/
│ ├── SKILL.md
│ └── references/
├── .claude/skills/palace/ # Project-level skill
├── mvp.py # Batch CLI entry point
├── interactive.py # Interactive CLI entry point
└── scripts/
└── package_skill.py # Skill packaging tool
Default: qwen3-max (Alibaba Cloud Qwen, 262K context window) via OpenAI-compatible DashScope API.
默认使用 qwen3-max（阿里云千问系列旗舰模型），上下文窗口 262K tokens。
When used as a Claude Code skill, no external model is needed — Claude handles everything.
作为 Claude Code 技能使用时，无需外部模型，Claude 自身处理全部流程。
# Offline tests (no API key needed)
python -m pytest tests/test_offline.py -v
# Lint
ruff check --select E,F,W .MIT