In agent development and local knowledge management, reading papers is a core step for acquiring new knowledge. DeepSeek Harness (DSH) extends its capabilities through a plugin mechanism, and lessxi/paper-workbench is designed to address the gap between “skimming” and “in-depth study and reproduction.”

What Is It

This is a DSH plugin maintained by LessXi, aimed at turning agent tools such as ZCode, DeepSeek Harness, or Claude Code into a full-workflow workbench for paper reading, learning, and reproduction. It covers the complete pipeline from discovery, acquisition, parsing, and close reading to accumulation, reproduction, and synthesis.

Core Capabilities

The plugin mainly includes the following features:

  1. Full-Workflow Workbench: Covers the full lifecycle of paper study, from retrieval and storage to CPU-level reproduction.
  2. Learning Mode: Emphasizes “you read, you write.” The AI provides a guided reading map, Q&A, recitation and error correction, closed-book questions, and Code Review. Formulas and code must be worked through using the user’s own thinking.
  3. Multi-Source Search and Parsing: Supports multi-source search and storage. Parsing relies on MinerU, which has the strongest support for formulas, tables, and Chinese, and it supports CPU environments.
  4. Visualization and Tools: Provides citation graphs, BibTeX, method family trees, and comparison matrices. Built-in CLI tools are included (paper status / parse).
  5. Monitoring and Reproduction: Supports arXiv subscription digests for tracking new papers, as well as CPU-level reproduction based on the official repo.
  6. Diagnostics and Ledger: Provides the paper-lab command for ledger inventory and bottleneck diagnosis.

Installation and Enabling

Run the following command in the DeepSeek Harness source directory to install it:

dsh plugin --profile web add github:LessXi/paper-workbench

After installation, restart the session to use the /paper-* command series.

Typical Usage

The plugin provides the following commands to manage the paper workflow:

  • /paper-init [directory] [--lite]: Initializes the workbench. The --lite parameter enables second-level startup (plain-text extraction only), while the default full mode restores formulas and tables (the first run may take around ten minutes).
  • /paper-find <your topic>: Performs multi-source search. The system automatically tiers the results (must-read / optional-read / skip) and stores them.
  • /paper-read <paper>: AI read-on-behalf mode. It calls MinerU to parse the paper and performs three-pass reading with structured notes.
  • /paper-learn <paper>: Learning mode. The AI accompanies reading, poses questions, and corrects errors, making it suitable for deep mastery.
  • /paper-survey <topic>: Generates method family trees, comparison matrices, and evolution timelines for the specified topic.
  • /paper-watch <topic>: Subscribes to arXiv and tracks new-paper digests.
  • /paper-reproduce <paper>: Performs CPU-level reproduction, including scale reduction, implementation from scratch, and numerical comparison.
  • /paper-lab: Views ledger inventory, bottleneck diagnosis, and recommendations for the next steps.

Environment Dependencies and Notes

The plugin depends on the following environment:

  • uv: Used to automatically manage the Python 3.12 environment.
  • MCP Services: Depends on arxiv-mcp-server and paper-search-mcp. The installation script automatically writes the workspace configuration.
  • Parsing Engine: Depends on MinerU, which can run on CPU and provides the best support for formulas, tables, and Chinese.
  • Data Storage: Data is stored as an Obsidian Vault. The .claude-plugin and .zcode directories should be excluded.
  • Multi-Machine Environments: Be aware that antivirus software may intercept uv trampoline.

Directory Structure

After installation, the core content is located under:

  • skills/paper-lab/: Contains SKILL.md, command definitions, templates and scripts, and an environment troubleshooting archive.
  • .claude-plugin/: Contains the plugin manifest and is compatible with the ZCode/Claude Code marketplace specification.