Introduction¶

Developing with DSH for multi-agent applications often involves opening several sessions simultaneously, assigning tasks to each, and finally relying on manual log inspection to align who wrote what and who verified what. While it enables collaboration, the delivery evidence is scattered: verification commands are never re-run, implementers claim completion without proof, and task dependencies exist only within prompts.

The agent-bot plugin (npm package name @eternalnight/agent-bot, maintainer EternalNight996) is designed to address this gap. It transforms a single DSH session into a structured delivery team: a task DAG led by a Captain, independent reviews, and local execution of artifacts with consolidated evidence during delivery. The project is a derivative of AgentTeams and is licensed under MIT. Below, it is introduced in sections covering features, installation, usage, and gates.

What is it¶

One-sentence positioning: A DSH multi-agent team collaboration and delivery pipeline plugin, accompanied by a desktop pet component.

Project information (source from GitHub README and package.json):

• Maintainer: EternalNight996
• Current Version: v0.1.6 (npm latest tag)
• License: MIT
• Runtime Requirements: Node.js ^22.19.0 || >=24
• Category in Community Directory: “Workflow”
• Plugin core has no cloud dependencies; team state is written locally to the .agent-teams directory

The specific problem it solves is “delivery evidence.” Common AI UIs present a wall of logs; agent-bot places who is working, what is stuck, and what has been delivered, along with smoke test results, independent acceptance, and real rendering evidence, all in a single interface.

Core Features¶

Team Runtime¶

The underlying mechanism is not a fake state machine; it is a persistent runtime:

• A task DAG led by the Captain, with tasks dispatched only when dependencies are satisfied.
• Persistent member sessions, email, and attempt ownership, supporting repair rounds and cold recovery.
• Each role has an independent write domain; frontend / backend / integration do not overlap.
• Tasks carry quality contracts: goals / scope / acceptance / verification command / review conclusion.
• Implementers cannot approve their own implementations; the host will re-run the declared verification commands.

Team state directory structure:

.agent-teams/<teamId>/
  team.json
  inbox/
  delivery-report.md
# Local-first; no cloud account integration

Interface and Onboarding Guide¶

The plugin provides a “Cave” style interface containing six clickable facilities: Mountain Gate, Dao Tradition, Merit Stele, God List, Sky Mechanism Pavilion, and Sect History. The team formation process follows a fixed sequence: fill in goals and roster, AI interviews based on goals (runtime environment, visual theme, animation intensity, etc.), select and finalize, plan is reviewed, and creation only begins after approval; artifacts are run directly in the Driver Gateway during delivery. The real-device capture records in the README show that a “Snake Game” delivery was summarized as 2/2 pass, 0 blocked, 13m33s, 165.2K tokens, ¥1.66; this is a single run record, not a guarantee for any task.

Role Lineup¶

Teams are formed based on “what to do next,” with a maximum of 5 slots, and idle roles do not consume tokens. It includes 6 built-in task lineups, such as Feature Orthogonal (Product Manager, Senior Developer, Evidence Collector, Code Reviewer) and Bug Fix (Codebase Onboarding Engineer, Senior Developer, Evidence Collector). Definitions are located in src/client/lineup-store.ts, and role display names are taken from src/agency-role-seeds.ts.

Desktop Pet¶

The plugin comes with a separately installed desktop pet (image “Little Fat Fish”). Windows and DSH share the same role assets:

• 25 frame-by-frame actions: click response, sleep chain, 6 work states, drag feedback, random idle.
• Wanders a segment every 18–28 seconds when idle, stopping to change actions randomly.
• Sleep lasts 10 minutes; clicking wakes it up, otherwise it continues sleeping.
• Double-click opens DSH and expands the sidebar; DSH will automatically launch if not running.
• Single click triggers petting only, does not invoke or modify the DSH window; hover displays a status card (Level, Growth, Satiety, Mood, Energy).

The pet is a standalone installer, does not depend on a Python environment, installs to %LOCALAPPDATA%\Programs\DSH Pet by user and creates a desktop shortcut.

Installation and Activation¶

To install the plugin core in DSH, use the official command:

dsh plugin --profile desktop add @eternalnight/agent-bot
# Restart DSH → Form Team

The desktop pet is installed separately; see GitHub Releases for the latest pet package:

DSH-Pet-Setup-<version>.exe
# Install to %LOCALAPPDATA%\Programs\DSH Pet, create desktop shortcut

To install from source, the README provides cloning from GitHub followed by adding the local directory. Note that the plugin runs with the permissions of the current dsh process. You should check the source code and license before installing (the project is MIT, from a third-party maintainer, not an official DeepSeek / BigModel product).

The minimal closed-loop workflow after installation: Click “Form Team” → Fill goals and roster → Interview and finalize → Review plan then approve → Watch members run and view artifacts directly in the Driver Gateway during delivery.

Delivery Gate Configuration¶

The quality field controls how gates audit. The following fields and default values are all from the original README:

An example where the host stops at Pending Human Review after auditing and specifies a smoke script for a specific project:

quality:
  autoReview: true
  reviewMode: "human"   # Host stops at "Pending Human Review" after passing audit
  commandsByProject:
    "projects/贪吃蛇": ["node projects/贪吃蛇/smoke.mjs"]

The plugin’s own release gates use pnpm verify (full suite) and pnpm verify:web-routes (browser-side contract) for validation. The README also reminds that the applicability of these tools is parallel wall-clock time and independent verification, not “every task is cheaper”: small tasks are usually more cost-effective with a single agent.

Use Cases and Notes¶

Suitable scenarios:

• Tasks can be split into dependent parallel work, and deliverables can be independently verified by commands.
• Teams seeking a workflow where “implementers do not self-approve” and “the host re-runs verification commands.”
• Need to see run-time evidence of deliverables directly on-site during delivery (smoke tests, rendering, review conclusions).

Points to note:

1. The plugin runs with the permissions of the current dsh process. You should read through the source code before installing to confirm there are no unauthorized file system or network behaviors; the license is MIT, author info can be found in package.json (credited as “Dream Walker”, Company “Dream Walker Technology Co., Ltd.”).
2. For daily development, you should directly run its development version verification tool python desktop-pet/pet.py --check --workspace . to check the pet’s main flow. This check comes from the practical instructions in the README (exit code 0 counts as pass) and serves as a quick self-check method after installation.
3. quality defaults to autoReview: false, meaning review tasks default to manual confirmation, suitable for running through first and then gradually loosening.
4. Stars and versions are based on real-time data from GitHub/npm; the screenshot values in the README are single run records.

Conclusion¶

The value of agent-bot lies in bringing “multi-agent collaboration” down to auditable delivery gates and a clickable interface, rather than just停留在角色卡片和日志墙上. The desktop pet is an add-on. Project links can be found on the SkillHub directory page and GitHub repository: The SkillHub directory is an independent site and has no official affiliation with DeepSeek / BigModel.

• Directory Page: https://www.skillhub.cn/plugins/EternalNight996/agent-bot
• GitHub: https://github.com/EternalNight996/agent-bot
• npm: https://www.npmjs.com/package/@eternalnight/agent-bot