dsh-expert-mode
Agent 与工作流Asher-2000/dsh-expert-mode
DeepSeek Harness 的专家模式代理预设:首席协调官加 16 位领域专家子代理,组成全栈多智能体团队,自动将任务委派给最佳专家。
- agent-preset
- ai-agent
- deepseek-harness
- dsh-plugin
- expert-mode
- multi-agent
- subagent
README
🧠 DSH Expert Mode
✨ What it does
Install this preset and DSH automatically becomes a "Chief Coordinator" mode:
| Scenario | Behavior |
|---|---|
| Receives task | Identifies domain → delegates to the best expert |
| Complex tasks | Dispatches multiple experts in parallel |
| Simple tasks | Coordinator handles directly — no forced delegation |
| Task complete | Experts stay online for follow-up modifications |
No custom prompts to write. No multi-config to maintain. Just install and use.
🖼️ Demo
🧩 16 Experts
🎯 Full-Stack Core (6)
| Expert | Tool | Domain |
|---|---|---|
| 🖥️ Frontend Dev | expert_frontend_dev | Web frontend, React/Vue, CSS/UI |
| 🖥️ Backend Dev | expert_backend_dev | API, server logic, authentication |
| 🗄️ Database | expert_database | Schema design, SQL, optimization |
| 🏗️ Architect | expert_architect | System design, tech selection |
| 🛠️ DevOps | expert_devops | CI/CD, Docker, K8s, deployment |
| 🧪 QA Engineer | expert_qa_engineer | Testing strategy, automation |
🔒 Security & Data (3)
| Expert | Tool | Domain |
|---|---|---|
| 🔒 Security | expert_security | Code audit, vulnerabilities, hardening |
| 📊 Data Analyst | expert_data_analyst | Statistics, visualization, insights |
| 🎨 UI/UX Design | expert_uiux_design | Interface design, design systems |
💼 Business (7)
| Expert | Tool | Domain |
|---|---|---|
| 📋 Product Manager | expert_product_manager | PRD, requirements, competitor research |
| ✍️ Copywriter | expert_copywriter | Marketing copy, content creation |
| ⚖️ Legal Review | expert_legal_review | Contract review, legal risk |
| 📱 Social Media | expert_social_media | Multi-platform distribution |
| 🚀 Growth Hacker | expert_growth | Growth strategy, A/B testing |
| 💹 Quant Finance | expert_quant_finance | Quantitative models, risk |
| 💰 Finance | expert_finance | Financial analysis, budget |
🛡️ Features
| Feature | Description |
|---|---|
| 🎯 Smart Delegation | Auto-identifies task domain and routes to the best expert |
| 🚀 Fast Track | Simple tasks handled directly — no forced delegation |
| 🔄 Five-Anchor Constraint | Prevents topic drift with per-turn self-check |
| 🤝 Cross Review | High-risk tasks get multi-expert independent review |
| 💾 Experience Pool | Lessons learned are saved and injected next time |
| 💬 Inter-Expert Bus | File-based message bus (bus.py): experts send/read directly, zero coordinator relay, P2P capable |
| ⚡ Fault Recovery | Auto-retry on timeout, strategy switch on failure |
| 📉 Progressive Disclosure | Methodology injected on-demand, 28% token savings |
| 🌐 Bilingual | Complete EN/ZH documentation |
📦 Installation
Option A: dsh plugin add (won't crash, but ❌ does NOT activate expert mode)
# In DSH workspace
dsh plugin add github:Asher-2000/dsh-expert-mode
⚠️ Important: this is an agent-preset plugin, not a Cordis service plugin.
dsh plugin addonly guarantees the bundle loads without crashing — it does not mount expert mode into your sessions. The preset only works via Option B (~/.dsh/.agent-presets/). Use Option B. This entry is a compatibility placeholder only.
Option B: Manual install from GitHub (current)
Clone the repository, then copy the preset into DSH's agent-presets directory:
# 1. Clone anywhere
git clone https://github.com/Asher-2000/dsh-expert-mode.git
cd dsh-expert-mode
# 2. Copy the preset into DSH's agent-presets directory
mkdir -p ~/.dsh/.agent-presets/expert-mode
cp -r agent.cordis.yml preset.yml cordis.patch.yml ~/.dsh/.agent-presets/expert-mode/
# If you want the full methodology docs (methods/, experts/, comm/ bus), copy the whole tree:
# cp -r .expert-mode ~/.dsh/.agent-presets/expert-mode/
# 3. Restart DSH web, then select "专家模式" in the workspace preset selector
dsh web
Note:
~/.dsh/.agent-presets/is DSH's preset discovery directory. Each subdirectory = one preset. The preset name comes frompreset.yml'snamefield.
Then select "专家模式" in the workspace preset selector.
Optional: Cross-session memory (recommended)
The expert-mode preset itself does not register the cross-session memory service — it is a HOST-PLANE plugin, and registering it inside a preset conflicts with the host composition (causing preset mount failure). To enable cross-session memory, install dsh-memory-connect separately into the host composition:
# 1. Clone the memory plugin
git clone https://github.com/Asher-2000/dsh-memory-connect.git
cd dsh-memory-connect
npm install github:Asher-2000/dsh-memory-connect#v0.4.0 # or place it into the dsh dependency tree manually
# 2. Register it in the host composition (e.g. append to ~/.dsh/profiles/web/cordis.patch.yml):
# - id: cross-session-memory
# name: '@deepseek-ai/dsh-memory-connect'
# config:
# path: ~/.dsh/memory.db
# openAt: startup
# 3. Restart DSH web
dsh web
⚠️ Important: Do NOT add
@deepseek-ai/dsh-memory-connectinto this preset'sagent.cordis.yml. It is a HOST-PLANE plugin (injectssessions+systemPrompt); registering it inside the preset throwsservice has been registered at <cross-session-memory>, which makes the expert-mode preset fail to mount and the UI fall back to the default preset. This preset ships with an explanatory comment about it.
🚀 Quick Start
- Install the plugin
- Select "专家模式" preset
- Ask any question — the coordinator auto-delegates to the right expert
Example
User: 帮我设计一个用户认证系统
Coordinator:
→ 识别领域: 后端开发 + 安全
→ 委派 Backend Dev: API 设计、JWT 实现
→ 委派 Security: 安全审计、漏洞防护
→ 汇总输出完整方案
🏗️ Architecture
┌─────────────────────────────────────────────────────────────────┐
│ Expert Mode Architecture │
├─────────────────────────────────────────────────────────────────┤
│ │
│ ┌──────────────────────────────────────────────────────────┐ │
│ │ Chief Coordinator (协调官) │ │
│ │ • Task analysis • Domain identification │ │
│ │ • Expert routing • Result aggregation │ │
│ └──────────────────────────────────────────────────────────┘ │
│ │ │
│ ┌───────────────┼───────────────┐ │
│ ▼ ▼ ▼ │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ Frontend │ │ Backend │ │ DevOps │ │
│ │ Database │ │ Security │ │ QA │ │
│ │ Architect │ │ ... │ │ ... │ │
│ └─────────────┘ └─────────────┘ └─────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────┘
💬 Inter-Expert Communication Bus (v0.8.0)
┌──────────────────────────────────────────────────────────────┐
│ File Message Bus (comm/bus.py) │
│ .expert-mode/comm/mailboxes/<expert>/*.msg │
├──────────────────────────────────────────────────────────────┤
│ │
│ data-analyst ──send──▶ frontend-dev (direct, async) │
│ copywriter ──send──▶ social-media (direct, async) │
│ coordinator ──broadcast──▶ all experts (global sync) │
│ expert A ──P2P subagent──▶ expert B (synchronous) │
│ │
│ • Zero relay: content flows between experts, NOT through │
│ coordinator context │
│ • Durable: every message persisted as .msg file │
│ • Auditable: full log at comm/logs/bus.log │
│ • Commands: send / read / ack / broadcast / stats │
└──────────────────────────────────────────────────────────────┘
Communication Modes:
| Mode | How | Use case |
|---|---|---|
| A. Relay | Expert A sends result → Expert B reads | Sequential collaboration |
| B. Parallel | Experts send results to coordinator → read --all | Independent collection |
| C. Broadcast | One message → all mailboxes | Global state changes |
| D. Review | Experts send "agree/partial/disagree + reason" | Cross review |
| E. P2P | Expert spawns subagent for direct Q&A | Synchronous clarification |
📚 Documentation
| Document | Description |
|---|---|
| Communication Protocol | Inter-expert message bus protocol v1 |
| Expert Methods | 16 expert methodology docs |
| Experience Pool | Lessons learned per expert |
| README.zh.md | 中文文档 |
🤝 Contributing
- Fork the repository
- Create a feature branch
- Commit your changes
- Push to the branch
- Open a Pull Request
📄 License
MIT License - see LICENSE for details.
🙏 Acknowledgments
- DeepSeek Harness - The core framework
- Cordis - Plugin system
- Awesome DSH Plugin - Community listing