Introduction

When using DeepSeek Harness for multi-agent collaboration, a common problem is that multiple models can only answer individually, lacking a long-term session space where they can discuss continuously, assign tasks, compare solutions, and validate results.

dsh-agent-arena is an interactive multi-AI collaboration plugin for DSH. It places human users, multiple AI users, and administrators into the same long-term session, ensuring that discussions, task execution, solution selection, and result validation happen within a unified interface.

Below, we introduce its positioning, installation methods, core capabilities, and usage boundaries.

What is it

dsh-agent-arena is a DeepSeek Harness plugin maintained by Tikzen, licensed under the MIT License. Its core scenario is an interactive multi-agent discussion and prompt collaboration space, suitable for scenarios requiring multiple AI roles to participate in tasks, compare solutions, and produce results together.

It is not simply sending the same sentence to multiple models; instead, it preserves persistent identities, model configurations, and session capabilities for each AI user, and supports task, decision, and result management within the long-term session.

Environment Requirements

Before using this plugin, confirm that the environment meets the following conditions:

  • Node.js 22 or higher
  • DeepSeek Harness Web profile
  • At least one configured model provider in DSH settings

Installation

Below is the installation command using the GitHub pinned version to install v0.4.0:

<DSH 安装目录>\manage-dsh.bat plugin --profile web add "github:Tikzen/dsh-agent-arena#v0.4.0"

This command mounts the pinned version of Tikzen/dsh-agent-arena to the DSH Web profile.

Core Capabilities

Below, we introduce the verified plugin capabilities organized by use case.

AI Users and Long-term Sessions

dsh-agent-arena supports a persistent AI user library, which can maintain the following information:

  • Nickname
  • Avatar
  • Personality
  • Provider
  • Model
  • Predefined quick conversations

AI users can participate in one-on-one private chats or join group chats composed of 2–12 AI members. The group chat automatically includes the administrator.

Within the same long-term session, human users, multiple AI users, and administrators can participate continuously without having to re-enter the provider or model every time.

Session Interaction

The session supports the following interaction methods:

  • Free speech
  • @mention AI to assign tasks or request a response
  • Silence a role using natural language
  • Resume a role’s speech using natural language

This means humans can temporarily control who continues to speak and who pauses during the discussion without needing to switch to complex meeting workflows.

Task Board

The Task Board is used to break down work in the chat into trackable tasks.

Supports:

  • Creating tasks
  • Setting assignees
  • Managing task status

When specific execution items are discussed in the group chat, these items can be added to the Task Board, reducing the problem of discussions staying only in message bubbles without status tracking.

Decision Board

When multiple solutions appear during discussion, the Decision Board can be used for structured comparison.

The Decision Board displays:

  • Multiple candidate solutions
  • Each role’s reasoning
  • Risk
  • Confidence

Ultimately, the human selects the solution. This design separates the “basis for comparison” from the “final decision,” avoiding the scenario where multi-AI discussions only leave emotional support or opposition.

Result Repository

The Result Repository is used to register verifiable results produced by AI or humans.

Supports registering:

  • Files
  • Links
  • Conclusions
  • Phase summaries

Humans can perform the following actions on results:

  • Validation
  • Rejection
  • Request additional evidence

Rejected or results requiring additional evidence are returned to the collaboration workflow rather than staying only in the chat history.

Collaboration Console

The Collaboration Console on the right displays runtime status during collaboration, including:

  • Role dynamics
  • Tools
  • Sub-agents
  • File locks
  • Recent actions

This panel is used to observe what each AI is currently doing, what capabilities it has called, and whether it is currently occupying file resources.

Collaboration Board and File Protection

The plugin provides a shared arena_coordination collaboration board.

Before writing to a file, an atomic lock is performed, and file writing is protected. This reduces the occurrence of overwrites or conflicts when multiple AI roles edit the same file simultaneously.

Channel and Auto-respond Control

The plugin supports collaboration control at the channel level, including:

  • Channel rate limiting and cooldown
  • Same-channel request queue
  • Auto-respond switch

These configurations are used to control the request rhythm and auto-respond behavior when multiple AI share provider channels.

Permission Control

Each conversation can have its permission mode set independently:

  • Read Only
  • Workspace Write
  • Full access

Permissions are isolated per conversation, and settings for a role in one conversation will not carry over to another. When a restricted role performs a high-risk action, approval is required, decided by a human allowing, denying, or adding execution requirements.

Typical Usage

Below is a reproducible usage workflow:

  1. First, configure model providers in DSH settings.
  2. Create members in the AI User Center, setting names, avatars, models, and optional personalities.
  3. Create private chats, group chats, or select meeting formats such as Roundtable, AI Court, Code Review, or Roast Session.
  4. Free speech in group chats or @AI to assign work; complex work is split into the Task Board.
  5. When multiple solutions appear, place them in the Decision Board so roles can provide reasoning, risk, and confidence.
  6. Register files, links, or conclusions to the Result Repository for final human validation.

Data and Records

Meeting, private chat, and group chat records are saved in:

DSH_HOME/agent-arena/meetings.json

These data are written to the local DSH Home by default. Records are not automatically cleaned up and can only be manually deleted by human users in History Management.

Permissions and Runtime Boundaries

Once enabled, the plugin runs alongside the DSH process, and the tools AI members can call are still constrained by DSH’s current workspace access mode and tool approval rules.

Please note the following boundaries:

  • The plugin does not read or save API keys in DSH.
  • Meeting topics, public statements, and interjections are sent to the selected model provider.
  • AI members can use tools provided by DSH, but are constrained by DSH’s workspace access mode and tool approval rules.
  • Administrators can only perform meeting actions on the server whitelist and cannot modify keys, model permissions, or arbitrary files via chat commands.
  • When a restricted role performs a high-risk action, approval is required, decided by a human allowing, denying, or adding execution requirements.
  • By default, only one meeting runs at a time; this can be adjusted via maxConcurrentMeetings in cordis.patch.yml.

It is recommended to check the source code, dependencies, and license before installation. AI brand preset avatars come from Lobe Icons, used under its MIT License; brand logos and trademarks belong to their respective rights holders.

Use Cases

dsh-agent-arena is suitable for scenarios such as:

  • Multiple AI roles need to discuss around the same issue continuously
  • Humans need to temporarily mention, silence, or resume a role during the conversation
  • Discussion results need to be broken down into tasks and tracked for status
  • Multiple solutions need to be compared based on reasoning, risk, and confidence
  • Files, links, or conclusions require manual validation

It is better suited for DSH multi-agent workflows where the collaboration process is visible and the final decision-making power is retained with humans.

Links

GitHub repository:

https://github.com/Tikzen/dsh-agent-arena

The directory page address was not confirmed in the verified materials for this session.