Let DeepSeek Be the Host, Claude, Qwen, Kimi Sit Around a Roundtable — I Built a DSH Roundtable Plugin

In one sentence: Upgrade a DeepSeek Harness session from "you chatting one-on-one with an AI" to "you + a host + a circle of expert AIs from different vendors holding a roundtable."

---

Why I Built This

When you talk to a single model, you get "one model's one answer." But if you let DeepSeek act as the host, with Claude, Qwen, Kimi, and MiniMax sitting at the same table, each assigned a role and holding a stance—

The disagreement itself becomes information. And the final decision still rests with you.

I wrote a plugin based on DeepSeek Harness (DSH): dsh-plugin-roundtable. In a chat, just say "hold a roundtable to discuss XX proposal," and DeepSeek will automatically assemble a circle of experts, organize a debate, and finally converge on a conclusion.

What It Looks Like

Ring topology diagram: The host (DeepSeek) sits at the center, with a ring of expert nodes connected by directional arrows. When an expert is working, its node shows a breathing light; when speaking, message pulses flow along the connections.

Drag-to-connect: Drag from the port on the edge of an expert node to another node to create a debate channel (supports one-way/two-way; right-click to change).

Right sidebar: Expert status / role assignments / knowledge base / speaking timeline, all visible at a glance.

Core Gameplay

1. One Sentence to Start a Meeting

Hold a roundtable to discuss: our mini-program cold-start strategy, bring one expert each for product, architecture, and operations.

DeepSeek will automatically roundtable_create the meeting, roundtable_add_node to invite people by role, and wire up the topology.

2. Add/Remove People Mid-Meeting, Just Tell the Host

Add one more security compliance expert, using GLM.

The host will pull the new expert into the meeting in real time.

3. Experts Can't Agree? You Make the Call

When disagreement becomes irreconcilable, the meeting pauses and a human decision card pops up: the host lays out the options, you choose, and the meeting continues.

4. Full Visualization + Persistence

Topology, speeches, and summaries are all on the interface; meeting state is persisted to disk in the workspace, not lost even if you restart DSH, and you can switch back to view historical meetings at any time.

What's New in v0.2.0

• Expert management UI: Click the「＋」in the right sidebar to add/remove experts, pick a vendor from the model dropdown (DeepSeek / GLM / Qwen / Kimi / MiniMax). Your changes are first recorded in a behavior document, and the host automatically executes them in the next round — the UI and the host's awareness no longer drift apart.
• Real vendor logo avatars: Real logos for DeepSeek blue whale, GLM, Gemini, Claude, Kimi, MiniMax, replacing the old abbreviation color blocks.
• Knowledge base (reading edition): Fill in a folder path, and it automatically lists files and formats; when an expert needs material, the host reads and forwards it on demand, no bulk transfer of the entire library, saving tokens.
• Response limits: You can set a "per-round output token cap" and "max opinions per round" for each expert, so multi-model meetings don't burn money; exceeding the budget automatically mutes them.
• Meeting deletion + cross-talk toggle: One-click delete a meeting from the right sidebar (confirmation dialog + cascading cleanup of expert sub-agents); in settings, you can show only meetings opened by the current conversation.

[Screenshot placeholder: Settings page "Response Limits" section — expert per-round output cap + max opinion count]

Tech Internals (Short Version)

• Plugin system: DSH's Cordis plugin (Host + Client dual-end), TypeScript + React + CSS Modules, packaged with tsdown.
• Host side: Registered 13 roundtable_* tools (create meeting / add person / connect / speak / summarize / decide…), plus a global system prompt protocol telling the host "when to use a roundtable and how to orchestrate it."
• Experts = real sub-agents: Each expert is an independent continuable sub-agent, can run different vendor models, and has an independent persona ("Global Collaboration Charter" + role rules).
• State persisted to disk: <workspace>/.roundtable/<meetingId>/ (meeting.json + transcript.jsonl), atomic writes + in-process lock.
• Three token-saving tactics: Budget circuit breaker (mute), expert per-round max_tokens cap, host on-demand relay for knowledge base.

How to Start

1. In your DSH environment, mount the plugin to the web profile (link: pointing to the repo directory).
2. Refresh the page → say "Hold a roundtable to discuss…" in chat.
3. In the right-side "Roundtable" tab, view the topology, add experts, and follow progress.

Open-sourced (MIT):

https://github.com/9931666/dsh-plugin-roundtable

Closing

A single model is an "answer generator"; a group of models is a "perspective generator." Making disagreement visible and keeping the final decision in human hands — that was my original intent for building this plugin.

If you're also playing with DeepSeek Harness / multi-agent orchestration, feel free to come chat in the repo, or just start a roundtable and experience it firsthand.

---

If you find it useful, a like and bookmark is the greatest support for me; if you have any thoughts, see you in the comments.