Preface

DeepSeek Harness (DSH) allows its functionality to be extended through plugins. Traditional long-running agents run silently until final output, citations are based on model memory rather than retrieval results, and corrections require restarting the work. wxxb789’s dsh-raven-research plugin aims to address these issues. It introduces a progressive, evidence-aware task abstraction in DSH for deep research, general writing, academic writing, and learning.

Plugin Introduction

The plugin is named dsh-raven-research, and the current version is v1 developer preview. It is a Cordis-based plugin, divided into agent role mode (raven_task, prompts, and task context) and optional host-role settings cards. Its core goal is to provide adjustable checkpoints, in-run steering, and citation validation that strictly matches the retrieved bytes.

Core Capabilities

The plugin provides the following core capabilities:

  • Progressive task abstraction: Provides a single, evidence-aware task abstraction covering deep research, general writing, academic writing, and learning.
  • Adjustable checkpoints: Provides early useful checkpoints during task execution, allowing in-run steering without restarting the entire task.
  • Citation validation: Citations are validated against retrieved bytes, rather than based on model memory.
  • Synthesis stage: Performs synthesis, exposes summary debt, and preserves inspectable insight candidates with claim sources.
  • Structure Studio: Used to build structure, presenting different argument architecture choices for long-form writing.
  • Source checking: Allows inspecting sources.
  • Conditional Markdown writing: Supports conditional Markdown writing.

Installation and Enabling

Because the plugin has not been released to npm yet, it must be installed from a source checkout.

  1. Run the following commands in the project root to build and pack:
    pnpm build && pnpm pack
  1. Add the generated tarball to the DSH deployment environment.
  2. Run the installation preset script and select Raven as the session mode:
    npx dsh-raven-install-preset

Typical Usage

In a DSH session, state requests directly, without special startup phrases or a separate task UI. The plugin’s default context guidance is auto, and it can also be set to off.

For example, when conducting deep research, you can make the following request:

Research the evidence for and against centralizing our recovery records. Show me an early useful evidence map, then turn the result into a decision memo for skeptical engineering directors. Preserve contradictions and keep the conclusion reusable.

During generation, you can reply naturally to steer the direction, for example:

reply naturally: Combine the recovery-risk opening with the uncertainty boundary, and focus on implementation risk.

Use Cases and Considerations

This plugin is suitable for scenarios that require rigorous evidence support and structured argumentation, such as deep research, academic writing, and complex decision analysis.

Notes:

  • This is currently the v1 developer preview version.
  • The plugin has been pinned and tested against DeepSeek Harness 0.1.2-alpha.1.
  • Context guidance defaults to auto and can be set to off in the configuration.

Summary

By introducing progressive tasks and strict source validation within the DSH main process, dsh-raven-research transforms the workflow for long-form research and writing. It allows adjusting direction during execution, checking citation sources, and selecting argument structure. It is suitable for complex writing tasks that require high confidence and reusable conclusions.

View plugin details
GitHub repository