The core design philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” For PDF processing needs in scientific research scenarios, f1star/dsh-research is a DSH bundle plugin. It is not a Codex plugin, but an independent bundle running within an existing DSH profile. Its core value is to provide evidence-first paper reading, persistent paper records, and cross-paper synthesis capability.

Core Features

The plugin mainly provides the following capabilities:

  • Paper Reading: provides tools such as paper_import, paper_reading_pack, paper_outline, paper_search, and paper_read. It supports importing local PDFs, performing a bounded first scan of identified key sections, navigating the heading structure, performing lexical matching searches, and recovering exact context blocks through physical page anchors, parser revision anchors, and citation hash anchors.
  • Paper Library: provides paper_library_register, paper_library_list, paper_library_get, and paper_library_alias. It is used to preserve paper identity, citation sources, exact source versions, parser observations, and reversible aliases in profile storage.
  • Research Integration: through operations such as research-question, evidence, note, claim, entity, observation, comparison-protocol, synthesis, matrix, audit, and research_review_render, it establishes traceable records across papers. Synthesis reasoning can retain explicit comparison protocols, and the renderer presents that comparison basis.
  • Scientific Extraction: provides the paper_structure tool for table, formula text, and figure/table extraction via precise parser revision anchors.
  • Research Workspace: provides a “Research Workspace” in the Web sidebar, allowing registered researchers to browse archived PDFs, preserve reading positions, write notes, inspect matrices, and review claims and numerical observations.
  • Reports and Tasks: supports exporting Markdown, LaTeX, BibTeX, CSL-JSON, and source files. research_task_list, research_task_get, and research_task_write preserve explicit workflow checkpoints, pause/resume states, and stale source warnings.

Installation and Enablement

Environment Requirements

Use Node.js 24 LTS (including npm and npx). Supported Node versions are in the range ^22.19.0 || >=24.0.0.

Installation Steps

  1. Select a standalone storage directory (a new directory is recommended for version 0.4, because 0.3 profile data cannot be opened directly in 0.4):
    export DSH_HOME="$HOME/.dsh-research-v040"

This directory is independent from the existing DSH installation and is used to store configuration, credentials, and default storage. Do not delete old data; just use the new `DSH_HOME` path.
  1. Install the released plugin:
    npx @deepseek-ai/dsh@0.1.1-rc.2 --version
    npx @deepseek-ai/dsh@0.1.1-rc.2 plugin --profile web add https://github.com/F1star/dsh-research/releases/download/v0.4.0/dsh-research.tgz
    npx @deepseek-ai/dsh@0.1.1-rc.2 --profile web --dump-config

Note: DSH uses pnpm to install profile plugins, even if you use npx to start DSH. `--dump-config` is used to confirm that the plugin has been correctly added to the configuration.
  1. Start the research workspace:
    npx @deepseek-ai/dsh@0.1.1-rc.2 web

After running the above command in the terminal, open the local URL in a browser as prompted. In the browser, configure the model API key, select the workspace, and open the "Research Workspace" in the sidebar. If the paper has not been registered yet, you can use `paper_import` to import a PDF.

Typical Usage

In the CLI, the recommended workflow is as follows:

  1. Import a PDF and perform a bounded scan:
    export DSH_HOME="$HOME/.dsh-research-v040"
    npx @deepseek-ai/dsh@0.1.1-rc.2 web

In the conversation, invoke:
*   `paper_import`: import a local PDF.
*   `paper_reading_pack`: perform a bounded first scan of identified key sections.
*   `paper_outline`: inspect the heading structure.
*   `paper_search`: find additional text blocks.
*   `paper_read`: call this command before dependent paragraphs.

  1. Register and manage the paper library:

    Use paper_library_register to register persisted documents, and retain the returned paper ID, source version, document, blocks, parser version, physical page, and citation hash identifiers.

  2. Perform research operations:

    Use tools such as research-question, evidence, and claim to build cross-paper records.

Notes

  • Version compatibility: 0.3 profile data cannot be opened directly by 0.4. If you need to use 0.4, back up existing storage and use a completely new, independent profile storage directory.
  • Python dependencies: The default PDF reader does not require Python; only the optional OCR functionality requires Python.
  • Package manager: DSH uses pnpm to install profile plugins, even if you launch the DSH main program through npx.
  • Permission scope: Installing the plugin effectively grants each agent started in that profile permission to use the research tool modes and their stable prompt guidance.

Summary

f1star/dsh-research provides DSH users with a complete scientific research toolchain, from PDF import and evidence-anchored reading to cross-paper synthesis. By combining the Web sidebar with CLI tools, it helps researchers establish rigorous, traceable research records with model assistance.