SuperAgent General: Integrating Feishu Chat into DSH via a Universal Plugin

Preface

Developers creating DSH plugins generally have two paths: either write their own message entry and task scheduling from scratch, or find a base that has already implemented session splitting, message determination, and recovery mechanisms to modify. SuperAgent General is an off-the-shelf option for the second path.

It did not rewrite the underlying logic of the message entry, but changed the use case: The official SuperAgent prototype defaults to revolving around the maintenance and collaboration of a single GitHub repository. This plugin brings it back to general-purpose tasks and replaces the Lark integration with the official SDK long connection by filling in APP_ID and APP_SECRET, instead of relying on a pre-configured bot CLI. Below is an introduction to what this plugin is, what changes were made relative to the prototype, and how to install and configure it.

What is it

SuperAgent General (dsh-superagent-general) is a DSH plugin designed for general-purpose tasks, maintained by comeonzhj. It is categorized as a “Client” in the community directory, has 4 GitHub stars, and is under the MIT license. After running, the plugin receives private and group chat messages via the Lark bot. A resident Router coordinates the tasks, executing them in independent DSH Worker Sessions, while the main panel uses the native DSH interface.

Applicable work includes research, analysis, writing, translation, information organization, file operations, and software development. Actual capabilities depend on the tools, models, and user authorizations provided by DSH.

It is a community adaptation project, developed based on the official DeepSeek Harness SuperAgent prototype (deepseek-ai/dsh-superagent-prototype), preserving the upstream copyright and license.

Adjustments relative to the official prototype

The plugin does not establish a separate task orchestration system but modifies the use case, access method, and initialization based on the prototype. The comparison is as follows, item by item, from the table in the README.

Item Official Prototype Current Plugin
Use Case Maintenance and collaboration of a single GitHub repository General-purpose tasks by default, GitHub as optional source
Lark Access Via configured bot lark-cli Fill in APP_ID, APP_SECRET, use official SDK to establish long connection and call API
Initialization Need to configure repository, CLI account, and run directory Automatically provide default data directory and task directory, model inherits DSH settings
Message Filtering Source group, mentioned or not, sender access and model evaluation Retain prototype access and evaluation, add configurable keyword attention/ignore
Default Requirements Centered on code modification, PR, and review General business requirements prioritize Chinese, GitHub-specific requirements moved to optional scenario skills
Scenario Customization Team prompts and skills Configure applicable conditions and delivery requirements for translation, project analysis, knowledge base maintenance, etc.
Management Interface DSH native panel and automatic polling Default overview, execution settings front-loaded, settings tab last, configuration instructions use ? hover tooltips
Lark Resource Operations Primarily via bot CLI Worker handles authorized messages, documents, knowledge bases, and tables via API using application identity

Keyword changes added to message filtering are included in the strategy version, meaning the strategy and keywords are not modified independently.

Retained Core Design

Retaining the prototype’s Engine, Router/Worker sessions, Trigger/Threads, direct subscriptions, on-demand context reading, SQLite, Effect/Outbox, and recovery mechanisms. src/core/ stays consistent with the upstream baseline, with adaptation changes concentrated in the periphery.

The main panel allows viewing messages, judgments, tasks, threads, and subscriptions with automatic updates. Task details can open the actual worker-… execution session to view reasoning, tool calls, and delivery processes; the Router unifies reply entry messages.

Upstream sources, baseline commits, and adaptation boundaries are written in the repository’s docs/UPSTREAM.md. Before finding “idle talk” for the plugin, you can first look at this boundary.

Installation

The installation method is installing the plugin package, not an npm package. The steps provided in the README are as follows.

  1. Download the pre-compiled .tgz plugin package from Releases.
https://github.com/comeonzhj/SuperAgent-General/releases
  1. Install this compressed package in the DSH plugin page, enable dsh-superagent-general, and open the SuperAgent in the sidebar.

The current version is developed based on the DSH 0.2.1-alpha.2 interface. See the repository’s package.json for the declared host compatibility range; Node.js requirements are ^22.19.0 or >=24. The repository version itself is 0.2.2, with runtime dependencies including @larksuiteoapi/node-sdk 1.76.0, zod 4.4.3, and proper-lockfile 4.1.2.

If you want to package it yourself from source, the flow provided in the README is to get a clean repository, then run pnpm install --frozen-lockfile, pnpm check, and pnpm pack --pack-destination dist in sequence. The artifact is dist/dsh-superagent-general-<version>.tgz. pnpm pack builds before packing. The repository also has other documentation like docs/ARCHITECTURE.md and docs/PROMPTS.md.

Configuring the Lark Application

After installing the plugin, you need to complete the connection in the SuperAgent panel in the sidebar.

  1. Connect the Lark application in the “Settings” tab at the end, fill in APP_ID and APP_SECRET, and select the groups to receive and people who can be assigned tasks.

  2. Add team skills in “More Settings” if needed, and fill in scenario requirements.

On the Lark application side, enable the bot, turn on the long connection to receive messages, apply for the corresponding message permissions, and publish. If you want to read and write documents and knowledge bases, the target resources also need authorization; message permissions alone are not enough. Use ? hover tooltips for configuration instructions, put the introduction content in “About”, and place the settings tab last.

Typical Workflow

After installation and Lark connection, a task generally goes through four steps.

  1. Send a message from the Lark side. Messages in private chats or allowed groups will be received by the bot.

  2. Router Determination. The Router is responsible for message filtering, access judgment, and model evaluation, deciding whether to convert to a task, and leaving a judgment record in the panel.

  3. Worker Session Execution. Accepted tasks run in an independent worker-… session, using DSH’s existing models, tools, workspace, and permissions to produce reasoning and tool call records.

  4. View and Reply in the Panel. The panel displays tasks, threads, and subscription status. Task details can open the actual execution session. The Router unifies reply entry messages to avoid multiple replies.

Reading and writing of Lark resources uses the API tools of the application identity, falling under the permission scope of the plugin’s application.

Verification and Current Boundaries

The verification scope recorded in the README is: passed automated tests, type checking, and builds, and verified isolation of installation, native panel, automatic updates, and main session entry in DSH. Real-time Lark message sending and receiving and resource reading and writing require joint debugging with your own application permissions; this step has not been pre-verified.

Upgrading from early versions will migrate configurations like access and models; the old task library is retained, and historical tasks are temporarily not converted to the current prototype’s domain model. See docs/USER-GUIDE.md for specific changes.

Applicable Scenarios and Notes

An individual or small team who wants to assign work directly in Lark while hoping the execution process remains visible and auditable in a DSH Worker session is suitable for using this plugin. Those who already have a bot CLI in use, or whose task boundaries are tightly bound to a specific GitHub repository, are better suited for staying with the official prototype.

Points to clarify first:

  1. The plugin runs alongside the DSH process and uses the permissions of the current DSH process, not an isolated sandbox. Tools and directories authorized to the plugin are actually authorized to the DSH process running it. Before installation, check the source code to confirm which permissions will be used.

  2. On the Lark side, prepare a published application credential, and grant message, document, and knowledge base permissions as needed. The plugin will not apply for them on your behalf.

  3. The current host interface baseline is DSH 0.2.1-alpha.2. package.json declares ^0.2.1-alpha.2 || ^0.2.2-alpha.3. Check the compatibility range before upgrading DSH.

Source and License

The plugin adopts the MIT License and is a community adaptation project, preserving the upstream copyright and license of the DeepSeek Harness Team. See the repository’s package.json for host compatibility, dependency versions, and package names. See docs/USER-GUIDE.md for complete installation steps and upgrade instructions, and docs/PROMPTS.md for task assignment rules.

Directory page: https://www.skillhub.cn/plugins/comeonzhj/SuperAgent-General

Repository: https://github.com/comeonzhj/SuperAgent-General

羽毛球分组比赛记分
小程序二维码

欢迎使用《羽毛球分组比赛记分》微信小程序

Xiaoye