🎯 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

// 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%

---

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
Professional Depth	✅ Persona + Iron Rules	✅ Maintained
Near-distance Guidance	✅ Template system	✅ Maintained
Backward Compatibility	—	✅ 100%

---

🚀 Quick Start

Installation

# 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
dsh web

Usage

1. After starting DSH Web service, click the 🎯 floating button in the bottom right corner
2. 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

【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

[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

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

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

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](LICENSE) for details

---

## 🙏 Acknowledgments

- [DeepSeek Harness](https://github.com/deepseek-ai/dsh) - Core framework
- [dsh-anchored-flash](https://github.com/deepseek-ai/dsh-anchored-flash) - Anchored no-degradation mechanism
- [dsh-ai-solution-council](https://github.com/deepseek-ai/dsh-ai-solution-council) - Cross review pattern
- [dsh-memory-evolve](https://github.com/deepseek-ai/dsh-memory-evolve) - Experience pool mechanism

---

**Version**: 2.0.0  
**Last Updated**: 2026-08-18  
**Author**: Asher-2000