Introduction

One of the core design principles of DeepSeek Harness (DSH) is “Everything is a plugin.” When developing agents, models can easily hallucinate or answer solely from training memory rather than grounded evidence. When it is necessary to cite research results in machine learning, it is more reliable to have the model directly retrieve and verify paper records on arXiv.

What Is This

This is a DeepSeek Harness plugin maintained by user babkiny. Its core value is that agents can verify their claims through arXiv papers when answering engineering or theoretical questions, instead of relying only on internal memory.

Core Features

The plugin provides two tools and one skill file for paper retrieval and verification:

  1. arxiv_search tool: Supports parameters such as keywords, authors, categories, date ranges, sorting, and pagination. Search results return truncated abstracts to control cost.
  2. arxiv_get tool: Retrieves the full metadata and abstract for a specific paper ID. Supports segmentation (max_chars + segment) to handle long text.
  3. SKILL.md skill: Defines the workflow for how the agent uses the tools above. The skill instructs the model to restate claims, run 2-3 orthogonal queries, read short lists correctly, report a verdict (with IDs), and stay strictly within the returned text without asserting content that was not mentioned.
  4. Automatic loading: The plugin automatically loads this skill when handling general engineering issues, without the user needing to explicitly request papers.
  5. Dependencies: Depends only on @deepseek-ai/dsh-tools.

Installation and Activation

Installation is completed through DSH’s plugin command. After installation, the plugin must appear in the profile’s dsh.profile.bundles for the system to load it.

# 安装插件
dsh plugin --profile web add dsh-arxiv-ml-search

If you need to configure parameters in a specific profile (for example, limit results per page or timeout duration), you can set the following configuration items in the profile’s plugin entries:

{
  "dsh": {
    "profile": {
      "bundles": ["dsh-arxiv-ml-search"],
      "plugins": {
        "dsh-arxiv-ml-search": {
          "limit": 10,
          "timeoutMs": 20000,
          "abstractChars": 350
        }
      }
    }
  }
}

Typical Usage

After installation, the tools automatically appear in the toolset without additional parameters.

Command-line example:

dsh --profile headless "Do any papers show RLHF hurting calibration? Cite ids."

Direct invocation example:

node scripts/smoke.mjs "rlhf"

Skill trigger example:
When the agent encounters an engineering problem, even if “finding papers” is not involved, the skill is loaded. For example:

“My RL agent gets a set of state features but the reward never makes it rely on some of them. How do I increase a specific feature’s influence?”

At this point, the model loads the skill and retrieves potential methods from real papers (such as potential-based reward shaping, curiosity, auxiliary tasks, etc.), along with the retrieved arXiv IDs.

Use Cases and Notes

  • Use cases: Scenarios that require verifying machine learning theory or engineering claims; scenarios that require citing arXiv papers.
  • Peer Dependency warning: @deepseek-ai/dsh-tools must remain a peer dependency; otherwise, tool registration will break (Symbol conflict). dsh has autoInstallPeers: false configured in the profile, and the plugin must not declare it as a regular dependency.
  • Bundle configuration: The plugin must appear in dsh.profile.bundles to be loaded. dsh plugin add usually handles this automatically, but if build scripts are ignored (such as ERR_PNPM_IGNORED_BUILDS), the installation may succeed while the plugin does not appear.
  • Skill description limit: The Skill description is truncated to 500 characters or fewer, meaning the full whenToUse logic may not be directly readable from the description.
  • Windows development: When developing locally on Windows, paths must be absolute and use the file:// URL format (for example, file:///C:/absolute/path/...); otherwise, the ESM loader will throw an error.

Conclusion

This plugin addresses the problem of agents lacking externally verifiable knowledge sources by standardizing arXiv retrieval as tools and skills. Developers can integrate it into existing workflows with a simple installation command.

  • GitHub repository: https://github.com/babkiny/dsh-arxiv-ml-search
  • Plugin directory: https://www.skillhub.cn/plugins/babkiny/dsh-arxiv-ml-search