Introduction

DeepSeek Harness (DSH) adopts an “everything is a plugin” architecture. When developing agents or running experiments, developers often need to have the agent continuously modify code, measure results, and iterate based on feedback. Manually managing such iterative sessions is inefficient and makes it difficult to maintain context consistency over long-running sessions.

dsh-autoresearch is designed to solve this problem. It provides DSH with a persistent automated research experiment loop, allowing experiments to be created, run, and monitored directly in the official Web GUI.

What It Is

This is a workflow plugin maintained by aa2246740 and licensed under the MIT License. It adds a closed-loop capability to DSH: “create a goal → the agent continuously modifies, measures, and keeps or rolls back changes → results are displayed in the top panel.” With this plugin, developers can define research goals in a session and let the agent automatically run multiple rounds of experiments.

Core Features

  • Persistent loops: Supports creating, running, and monitoring experiment loops, with persistent session state.
  • Web GUI support: Supports operating and managing the plugin through the official Web GUI.
  • Flexible experiment control: Supports modifying, measuring, keeping, or rolling back intermediate experiment states.
  • Result visualization: Displays experiment results in real time in the top panel of the session.

Installation and Enablement

Before installing, make sure the environment meets the requirements: the DeepSeek Harness version must be >=0.2.0-rc.1 <0.2.1, and the Node.js version of the host peer must be >=22.19.0.

Desktop / DSH Studio

  1. Open DSH Studio.
  2. Go to Settings → Plugins.
  3. Click Add Plugin and enter the following:
github:aa2246740/dsh-autoresearch#v1.0.10
  1. Follow the plugin manager prompts to complete the installation.

Web

Run the following command in a terminal (make sure dsh, Node.js, and pnpm are available in your PATH):

dsh plugin --profile web add github:aa2246740/dsh-autoresearch#v1.0.10

After installation, restart the Web Host and open the page to use the plugin.

Typical Usage

After installation, you can start using it in any project session.

1. Create and Run Research

  1. Open a project session.
  2. Type /autoresearch in the input box.
  3. Select New Research.
  4. Fill in the research goal and number of rounds, then confirm.

2. Execution Process

After running, the plugin will modify project files and execute local commands. The entire process is driven by the model in the current session, and continuous execution will consume model quota. The plugin does not upload or push project code to remote repositories.

3. Common Commands

The plugin provides the following commands for controlling the experiment loop:

  • /autoresearch: create, continue, check status, stop, or clear a research run.
  • /autoresearch resume: resume a paused research round.
  • /autoresearch status: view the current persistent state.
  • /autoresearch off: stop automatic continuation while keeping existing results.
  • /autoresearch clear: clear the Autoresearch ledger for the current project.

4. Data Storage

The experimental ledger data is stored under the .auto/ directory in the project directory, including prompt.md, measure.sh, optional checks.sh, log.jsonl, ideas.md, and config.json. When issues such as discard, crash, or checks_failed occur, the system attempts to restore protected files from the Git baseline or plugin-private snapshots.

Use Cases and Cautions

  • Use cases: Suitable for experimental development workflows that require long-running, multi-round iteration.
  • Security considerations: The plugin runs with the permissions of the current process, so review the source code and license before installation.
  • Prerequisites: Git is required to save baselines, or you can use plugin-private snapshots.
  • Network and quota: No code upload is involved, but model quota will be consumed.

Summary

By persisting the experiment loop and integrating it into the Web GUI, dsh-autoresearch lowers the barrier to iterative testing in agent development. It allows developers to focus on experiment design instead of cumbersome session management.

For more information, visit the DeepSeek Harness plugin catalog or the GitHub repository.