Introduction

DeepSeek Harness (DSH) is DeepSeek’s open-source local Agent work system, whose core design philosophy is “everything is a plugin.” dsh-research-kit is a DSH plugin that runs on the browser side. It maintains a catalog of research resources, assembles workflows and user parameters into editable prompts, and the actual sending and execution are handled by DSH’s current session.

Core Features

Workflows and Skill Library

The plugin includes human-reviewed scientific research workflows and skill entries:
* 351 research workflows: covering 28 process families such as papers and manuscripts, literature research, data analysis, genomics, and clinical research. Each workflow is a parameterized, editable prompt template with built-in anti-fabrication guardrails (no fabricated citations, data, page numbers, or author intentions).
* 109 skill entries: including 46 guidance modules (e.g., scientific writing, statistical review, citation verification) and 63 capability catalog entries (e.g., bulk RNA-seq, Nextflow, literature API search). Capability catalog entries clearly state their dependent toolchains and data egress boundaries; they do not inject prompts and do not execute on the user’s behalf.

Data Sources and Evidence Management

  • 129 scientific data sources: covering literature, clinical, genetics, omics, protein/chemistry, and more. Of these, 11 support direct querying by the plugin and return candidate records with links and stable identifiers; when not connected, the plugin explicitly falls back to DSH Agent/MCP.
  • Research evidence graph: connects the current session’s resources, workflows, query sources, assets, saved evidence, and automatically accumulated knowledge into traceable relationships. Supports zoom/pan, minimap, focus/upstream-downstream/two-point path modes, conclusion traceability panel, and de-identified snapshot export.
  • Claim review panel: manages the status of each claim (extracted, inferred, ambiguous, verified, rejected). Supports extracting candidates from paragraphs and registering them, or manual registration. Uses Agent review to generate independent suggestions and reasons, after which researchers confirm, flag as questionable, or reject items one by one and link relevant evidence.

Prompt Assembly and Editing

The plugin turns repeatedly executed research tasks into parameterized, editable, and visible-before-sending prompts:
* Parameter substitution generates a prompt preview in real time.
* After manual editing, it provides “Restore Auto-Generation.”
* It does not send when required parameters are missing.

Cross-Host MCP Tools

The plugin provides a standalone stdio MCP Server, compatible with hosts such as Codex, Claude Code, and Zed, and provides tools for literature search, evidence management, research runs, and review.

Typical Usage

1. Workflow Assembly and Execution

The browser-side operation flow in DSH is as follows:
1. Open the “Resources / Workflows” shortcut.
2. Select a workflow from the research capability catalog.
3. Fill in parameters and use DSH’s native @文件 to reference materials.
4. Preview and edit the final generated prompt.
5. Write it into the input box or send it to the current session, where DSH’s current session Agent executes it.

2. Draft Enhancement

Use the “Draft Enhancer” on the right side of the input box:
* Supports a lightweight mode (zero-token structured organization) and a semantic mode (reuses the current session model, streams output to the screen, and provides five-dimensional diagnostics).
* After enhancement, it can be undone and compared with the original draft.

Environment and Notes

  • DSH Status: DSH is currently in Developer Preview. The conversation.view slot and inputActions contract that it depends on evolve with DSH versions, so revalidation is required after upgrades.
  • Plugin Coexistence: Multiple views can be registered in the same slot. This plugin’s order in conversation.view is 91, so it can coexist with similar plugins, but their data states are not shared.
  • Permissions and Security:
    • It does not hold model keys. On the DSH side, it reuses the current session’s model routing; on the MCP side, it does not host models.
    • It does not recreate file upload. Material references use DSH’s native @文件.
    • All research judgments still require verification by the researcher; it does not execute any tools on the user’s behalf.

Summary

dsh-research-kit is a DSH plugin for research scenarios. Through standardized workflows, parameterized prompt assembly, and evidence graph management, it helps researchers complete a closed-loop workflow in a local environment, from literature search to conclusion verification. All research judgments still require verification by the researcher; it does not execute any tools on the user’s behalf.

Plugin catalog page
GitHub repository