DSH Expert Mode

One agent preset that turns DSH into a "1 Coordinator + 11 Experts" multi-agent team.

中文版 | English

What it does

Install this preset and DSH automatically becomes a "Chief Coordinator" mode:

• Receives task → identifies domain → delegates to the best expert subagent
• Complex tasks can dispatch multiple experts in parallel
• Simple tasks handled directly by the Coordinator — no forced delegation
• Experts stay online after completion for follow-up modifications

No custom prompts to write. No multi-config to maintain. Just install and use.

Demo

Select the "Expert Mode" preset in DSH workspace to use

5 expert subagents working in parallel, with real-time token usage and timing

11 Experts

Core Mechanisms

Progressive Disclosure (v0.5.0)

The Coordinator holds a complete expert methodology index and injects on-demand — not dumping all expert personas into context at startup.

How it works:

• Coordinator persona contains a methodology index for all 11 experts (methodology + output format + iron rules)
• Expert subagent personas stripped to minimum (role + iron rule + completion criteria only)
• During delegation, Coordinator extracts the relevant methodology from the index and injects it via Near-distance Guidance
• Subagent starts with complete "who I am, what to do, how to deliver" context

Anchored Non-degradation (v0.5.0)

Solves the "trajectory flip" problem caused by system prompt mutations, ensuring reasoning style stability:

Five Anchor Constraints

The Coordinator self-checks five anchor points every turn to prevent drift and loops:

Near-distance Guidance

During delegation, the Coordinator extracts the relevant methodology from the index and fills in the structured guidance template:

── Near-distance Guidance ──
Identity: You are [expert role]
Task: {specific task description}
Input: {input data}
Output format: [extracted from methodology index]
Completion criteria: {clear delivery standard}

Subagents start with clear "who I am, what to do, how to deliver."

Expert Communication Protocol

When cross-expert collaboration is needed, the Coordinator routes structured messages:

[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

High-risk tasks (architecture selection, contract review, financial analysis) trigger multi-expert independent review. The Coordinator synthesizes conclusions.

Experience Pool

After completing important tasks, experts extract lessons to .expert-mode/experts/{name}/lessons.md, automatically injected for similar future tasks.

Expert Persistence

Experts stay online after task completion. The Coordinator can wake them up for follow-up modifications with full context preserved.

Installation

Method 1: dsh plugin add (recommended)

dsh plugin --profile web add github:Asher-2000/dsh-expert-mode

Method 2: git clone

mkdir -p ~/.dsh/.agent-presets
git clone https://github.com/Asher-2000/dsh-expert-mode.git ~/.dsh/.agent-presets/expert-mode

Want English? The repo includes an expert-mode-en/ preset:

cp -r ~/.dsh/.agent-presets/expert-mode/expert-mode-en ~/.dsh/.agent-presets/expert-mode-en

Method 3: Manual download

Download from Releases and place preset.yml + agent.cordis.yml into ~/.dsh/.agent-presets/expert-mode/.

Usage

After installation, select the Expert Mode preset when creating a new session in DSH Web GUI.

FAQ

Q: Does Expert Mode consume extra model quota? A: Delegating to expert subagents triggers subagent model calls (DSH subagent mechanism), same as the official subagent feature. Simple tasks are handled directly by the Coordinator with no extra calls.

Q: Can I add my own experts? A: Yes. Copy any expert entry in agent.cordis.yml, modify the tool name and persona.

Q: How does it relate to the official standard preset? A: Built on top of the official standard preset combination, preserving the full toolset and only adding the expert delegation layer.

File structure

.
├── preset.yml          # Preset metadata (name + description)
├── agent.cordis.yml    # Coordinator persona + methodology index + 11 expert subagent tool definitions
├── expert-mode-en/     # English preset
├── README.md
├── README.zh.md
└── LICENSE

Changelog

• v0.5.0 — Progressive disclosure + anchored non-degradation (prompt tokens -28%, reasoning stability 9/10)
• v0.4.0 — Five anchors + near-distance guidance + cross review + experience pool
• v0.3.0 — Expert persistence + communication protocol

Community

• 📢 Featured in Awesome DSH Plugin: https://github.com/awesome-dsh-plugin/awesome-dsh-plugin
• 🏷️ GitHub dsh-plugin topic: https://github.com/topics/dsh-plugin

Tags

dsh deepseek-harness agent-preset expert-mode multi-agent subagent ai-agent dsh-plugin

License

MIT License — see LICENSE.