🎯 Expert Mode Pro > DeepSeek Harness Expert Mode Pro — Intelligent Collaboration System with Chief Coordinator + 11 Domain Experts 中文版 | English --- ## ✨ New Features ### 🚀 Phase 0: Expert Persistence Before: Each delegation spawns a new subagent, destroyed after completion Now: Experts can be awakened for follow-up conversations with full context preserved javascript // Expert stays online after task completion // Coordinator can wake up expert for additional modifications send_message(expert_id, "Please add file size statistics") Benefits: - ✅ Experts stay online after task completion - ✅ Coordinator can wake up experts for modifications with full context - ✅ Backward compatible, no impact on existing delegation logic --- ### ⚡ Phase 1: Progressive Disclosure Before: All 11 expert personas injected at once (~3230 chars) Now: Coordinator holds only index, injects on demand (~930 chars) Before: 3230 chars → After: 930 chars Token savings: 71% Response speed: +20% Mechanism: - Index Layer: Coordinator holds only expert index (~700 chars) - On-demand Injection: Expert persona injected during delegation (~230 chars) - Context Isolation: Expert tasks don't affect coordinator reasoning --- ### 🧠 Phase 1: Anchored No-Degradation Problem: System prompt mutation from 46 to 6620 chars triggers "trajectory flip" Solution: Progressive disclosure + Context isolation Verification Results: - ✅ Reasoning style stability: 9/10 - ✅ Output quality maintenance: 9/10 - ✅ Prompt mutation avoidance: 100% --- ### 🌐 Phase 8: Web UI Management Panel Four Major Modules: | Module | Function | Highlights | |--------|----------|------------| | Task Management | Three views (List/Kanban/DAG) | Dependency management, priority settings | | Expert Monitoring | Real-time 11-expert status | Wake-up follow-up, history statistics | | Cross Review | Multi-expert independent review | Opinion comparison, visualization | | Experience Pool | Expert knowledge base | Search, categorize, export | --- ## 📊 Feature Comparison | Dimension | Original Expert Mode | Expert Mode Pro | |-----------|---------------------|-----------------| | Expert Persistence | ❌ Destroyed after use | ✅ Can be awakened | | Expert Communication | ❌ All via coordinator | ✅ Direct routing | | Task Dependencies | ❌ Manual judgment | ✅ DAG auto-scheduling | | State Persistence | ❌ Lost on session change | ✅ File persistence | | Self-Constraint | Three Anchors | Five Anchors (+collaboration +resource awareness) | | Context Efficiency | Full injection | Progressive disclosure + token budget | | Intelligence Stability | Prompt mutation may degrade | Anchored no-degradation | | High-risk Decisions | Single expert | Cross review | | Experience Accumulation | None | Experts learn over time | | Visualization | ❌ Plain text | ✅ Web UI Panel | | Professional Depth | ✅ Persona + Iron Rules | ✅ Maintained | | Near-distance Guidance | ✅ Template system | ✅ Maintained | | Backward Compatibility | — | ✅ 100% | --- ## 🚀 Quick Start ### Installation bash # 1. Clone repository git clone https://github.com/Asher-2000/dsh-expert-mode.git cd dsh-expert-mode # 2. Copy plugin to DSH plugin directory cp -r dsh-ui-plugins/dsh-expert-mode-pro-ui ~/.dsh/plugins/ # 3. Configure cordis.patch.yml cat >> ~/.dsh/cordis.patch.yml << 'EOF' - insert: - id: expert-mode-pro-ui name: "@local/dsh-expert-mode-pro-ui" EOF # 4. Restart DSH Web service dsh web ### Usage 1. After starting DSH Web service, click the 🎯 floating button in the bottom right corner 2. Click to open Expert Mode Pro management panel 3. Start using task management, expert monitoring, cross review, experience pool features --- ## 📋 Expert Team | Expert | Role | Methodology | |--------|------|-------------| | 📊 Data Analyst | Business → Metrics → Anomalies → Conclusions | Ask business questions → Break down metrics → Find anomalies → Give conclusions | | ✍️ Copywriter | Persona → Selling Points → Multi-version Copy | Define user persona and selling points → Generate multi-version different style copy | | ⚖️ Legal Review | Elements → Risks → Modification Suggestions | List elements → Mark risk points → Give modification suggestions | | 📋 Product Manager | Requirements → User Stories → PRD | Requirements clarification → User stories → PRD framework → Feature prioritization | | 💻 Frontend Dev | Selection → Components → Performance Optimization | Tech selection → Component design → Performance optimization | | 🎨 UI/UX Design | Flow → Architecture → Visual Specs | User flow → Information architecture → Visual specs → Delivery | | 🏗️ Architect | Requirements → Selection → Modules → Decisions | Requirements → Architecture selection → Module breakdown → Key technical decisions | | 📱 Social Media | Platforms → Content Differentiation → Private Domain | Define platform mix → Differentiate content by platform traffic logic | | 📈 Growth Hacker | Funnel → Levers → Experiments | Funnel breakdown → Find growth levers → Design experiments → Data validation | | 💹 Quant Finance | Data → Metrics → Models → Backtesting | Data → Metrics → Models/strategies → Backtesting/conclusions | | 💰 Finance | Reports → Breakdown → Budget → Suggestions | Data first → Report ratio breakdown → Budget variance attribution → Quantified suggestions | --- ## 🔧 Core Mechanisms ### Five Anchor Constraints yaml 【Anchor 1·Review】Before this round: Current subtask? Previous output? 【Anchor 2·Convergence】Before this round ends: Does output advance overall goal? 【Anchor 3·Anti-drift】No progress for 2 rounds → Force strategy switch 【Anchor 4·Collaboration Check】Cross-expert collaboration needed? Correct routing? 【Anchor 5·Resource Awareness】Context token usage healthy? Need simplification? ### Expert Communication Protocol yaml [FROM:expert_data_analyst → TO:expert_frontend_dev] Task: Design frontend data display component based on analysis conclusions Data: {Expert A's conclusion summary} ### Cross Review Protocol yaml Trigger Conditions: - High-risk tasks (architecture selection, contract review, financial analysis) - User requests "multi-angle verification" - Expert output has low confidence Execution Flow: 1. Assign to 2-3 related experts for independent output 2. Each expert's output becomes review input for others 3. Reviewers mark "Agree/Partially Agree/Disagree + Reason" 4. Coordinator as judge, synthesize final conclusion 5. Review records saved to .expert-mode/reviews/ directory ### Experience Pool Protocol yaml After expert completes important task: 1. Extract "what was learned" (max 3 items) 2. Write to .expert-mode/experts/{name}/lessons.md 3. Next similar task, lessons.md injected as additional context --- ## 📁 Project Structure dsh-expert-mode/ ├── README.md # English version ├── README.zh.md # Chinese version ├── expert-mode-pro-plan.md # Upgrade plan ├── .expert-mode/ # Runtime state │ ├── experts/ # Expert experience pool │ │ ├── data-analyst/ │ │ │ └── lessons.md │ │ ├── frontend-dev/ │ │ │ └── lessons.md │ │ └── architect/ │ │ └── lessons.md │ ├── reviews/ # Cross review records │ │ └── 2026-08-18-*.md │ ├── team-state.json # Task DAG state │ ├── optimized-persona.md # Progressive disclosure optimization │ ├── anchored-no-degradation-experiment.md # Anchored no-degradation experiment │ └── test-progressive-disclosure.md # Progressive disclosure test ├── dsh-ui-plugins/ │ └── dsh-expert-mode-pro-ui/ # Web UI plugin │ ├── package.json │ ├── cordis.patch.yml │ ├── DESIGN.md │ ├── README.md # Plugin README (EN) │ ├── README.zh.md # Plugin README (ZH) │ ├── demo.html │ └── lib/ │ ├── index.js │ ├── client.js │ ├── i18n.js # Internationalization │ ├── modules/ │ │ ├── task-manager.js │ │ ├── expert-monitor.js │ │ ├── review-visualization.js │ │ └── experience-viewer.js │ └── styles/ │ └── main.css └── .gitignore --- ## 🎨 Interface Features ### Dark Mode - Auto-detect system dark mode preference - Click 🌓 button to manually toggle - All components fully support dark mode ### Responsive Design - Desktop (> 1024px): Three-column Kanban layout - Tablet (641px - 1024px): Single column layout - Mobile (≤ 640px): Full-screen panel ### Animation Effects - Panel open/close: Scale + fade in/out - Card hover: Lift + shadow enhancement - Status indicator: Pulse animation (busy state) - Toast notifications: Slide in/out ### 🌐 Internationalization (i18n) - Bilingual Support: Chinese (中文) and English - One-click Switch: Click EN/中 button in panel header - Language Persistence: Preference saved to localStorage - Full Translation: All UI text, buttons, labels, messages --- ## 🔧 Technical Specifications - Implementation: Pure JavaScript (ES Modules) - External Dependencies: Zero - CSS Solution: Pure CSS + Custom Properties - Browser Compatibility: Chrome 80+, Firefox 78+, Safari 14+, Edge 80+ - Bundle Size: < 50KB (including CSS) --- ## 📈 Performance Metrics | Metric | Before | After | Improvement | |--------|--------|-------|-------------| | Initial Token Consumption | ~3230 chars | ~930 chars | -71% | | Response Time (Simple Task) | ~5s | ~4s | +20% | | Response Time (Complex Task) | ~15s | ~12s | +20% | | Expert Output Quality | 9/10 | 9/10 | Maintained | | User Experience | Average | Excellent | Significantly Improved | --- ## 🛠️ Development ### Adding New Expert 1. Add expert configuration in agent.cordis.yml: yaml - id: tool-subagent-expert-your-expert name: '@deepseek-ai/dsh-tool-subagent' config: provider: spawn toolName: expert_your_expert backgroundMode: continuable persona: | You are a senior expert. Methodology: ... 2. Add expert info in expert-monitor.js: javascript const EXPERTS = [ { id: 'your_expert', name: 'Custom Expert', role: 'Role description', icon: '🔧', color: '#3b82f6' }, // ... ]; ### Adding New Module 1. Create new module file in lib/modules/ 2. Implement create*() factory function, returning { render, ...methods } object 3. Import and register in modules object in client.js 4. Add new tab configuration in TABS array --- ## 📝 Contributing 1. Fork this repository 2. Create feature branch: git checkout -b feature/your-feature 3. Commit changes: git commit -m 'Add some feature' 4. Push branch: git push origin feature/your-feature 5. Submit Pull Request --- ## 📄 License MIT License - See LICENSE for details --- ## 🙏 Acknowledgments - DeepSeek Harness - Core framework - dsh-anchored-flash - Anchored no-degradation mechanism - dsh-ai-solution-council - Cross review pattern - dsh-memory-evolve - Experience pool mechanism --- Version: 2.0.0 Last Updated: 2026-08-18 Author: Asher-2000