In the plugin ecosystem of DeepSeek Harness (DSH), building multi-Agent collaboration systems typically faces problems such as task loss caused by Agent crashes, heavy Token consumption from long outputs, and latency caused by frequent Agent cold starts. The dsh-swarm plugin v0.4.0 addresses these issues by introducing resident processes and a task ledger mechanism.

Core Features

  1. Resident Agents: Each Agent is a standalone Node.js process that continuously reads its inbox, executes tasks, and writes to its outbox, eliminating the need for a cold start every time.
  2. Independent API configuration: Each Agent independently configures its model, Provider, apiBaseUrl, and apiKey, which are written to its own state.json.
  3. Artifact-reference messaging: Long outputs are automatically persisted to a shared artifact store; messages between Agents only carry an art://<id> reference plus a 200-character summary, and the full text is expanded by reference when needed, saving Tokens significantly.
  4. Task ledger + watchdog: The task state machine is pending → claimed → done/failed. Failed tasks are automatically retried (up to 3 times); if a task remains in the claimed state for more than 5 minutes without progress, it is automatically reassigned, preventing task loss due to Agent crashes.
  5. Visualized UI: The DSH interface includes a built-in team panel that supports visually creating/deleting blank Agents and viewing task ledger progress.
  6. Built-in web access: Agents include web_search / web_fetch (jina.ai) without requiring additional plugins.

Installation and Activation

# 复制到本地插件目录
cp -r dsh-swarm ~/.dsh/local-plugins/dsh-swarm

# 同步到 desktop profile 副本（Windows 下 link: 不生效，是物理拷贝）
robocopy "local-plugins\dsh-swarm" "profiles\desktop\node_modules\dsh-swarm" /MIR

# 重启 DSH（引擎启动时自动加载）

Typical Usage

  1. Add an Agent: You can add one directly in the conversation, for example: “Add an Agent with ID researcher1, using deepseek-chat, and API https://api.deepseek.com/v1”. You can also click “Add” in the team panel.
  2. Work as a team: When encountering a complex task, the leader automatically splits it and assigns subtasks to Agents. For example: “Help me research the pros and cons of framework X and framework Y, and write a comparison report.”
  3. Check progress: The built-in team panel in the DSH interface displays Agent statuses in real time (⏳pending 🔄claimed ✅done 💀dead) and task ledger progress.

Differences from agent-teams

Dimension agent-teams dsh-swarm
Agent form Same-process subagent Independent resident process + independent API
Result passing Full text relayed through the main session Artifact reference + summary, expanded on demand
Failure handling Task loss Ledger retry + watchdog automatic reassignment
Suitable scenarios Simple delegation Complex multi-role collaboration / long-running tasks

Tool List

User tools

  • swarm_add_agent: Add a blank Agent (specifying agent_id / model / provider / api_base_url / api_key).
  • swarm_remove_agent: Remove an Agent and stop its process.
  • swarm_list_agents: View all Agent statuses and role assignments.

Captain tools

  • swarm_create_team: Form a team: automatically assigns roles to blank Agents (researcher/coder/analyst/reviewer) and starts their processes.
  • swarm_set_agent_prompt: Set a role prompt for an Agent (built-in 4 templates or custom full text).
  • swarm_send_task: Dispatch a task: write to inbox → poll outbox → return result (long results return an art:// reference).
  • swarm_read_results: Non-blocking read of results already produced by an Agent.
  • swarm_read_artifact: Expand the full artifact text by an art://<id> reference (supports max_chars chunking).
  • swarm_status: Team status + task ledger (triggers one watchdog scan).
  • swarm_disband: Disband the team and stop all Agent processes (does not delete the Agents).

Applicable Scenarios and Notes

This plugin is suitable for complex task scenarios that require multi-role division of labor and long-process outputs. Because this plugin runs with the permissions of the current DSH process, inspect the source code and license before installing.

Summary

dsh-swarm builds a robust multi-Agent collaboration environment through resident processes, independent APIs, artifact-reference passing, and a task ledger watchdog mechanism. It solves problems of Token waste and low task reliability in traditional multi-Agent systems.

  • Catalog page: https://www.skillhub.cn/plugins/wanghj040530/dsh-swarm
  • GitHub: https://github.com/wanghj040530/dsh-swarm