Introduction

In the DeepSeek Harness (DSH) workflow, there is often a need to handle citations pasted directly from documents or web pages. These citations are often poorly formatted and incomplete in fields. Manual cleanup is time-consuming and error-prone. The dsh-research-refs plugin automates this process, using code to parse, validate, deduplicate, and format data, enabling the agent to quickly generate standardized reference lists.

What It Is

dsh-research-refs is a DSH plugin maintained by Parker-xia and belongs to the workflow category. It is built on the open-source research-refs skill, but refactored into a deterministic toolchain. The core logic is code-driven, and the agent is only responsible for workflow orchestration.

Core Features

The plugin contains four tools and one orchestration skill, processing citation data sequentially:

  1. refs_parse: Parses messy citation text. Supports strategies such as numbering, blank lines, and type markers, and makes a best effort to recognize fields. Entries that cannot be parsed or have missing fields are marked with flags.
  2. refs_verify: Validates and completes information via Crossref.
    • If a DOI is available, query works/{DOI} to retrieve volume, issue, pages, journal, and year.
    • If no DOI is available, search by title + author. Similarity ≥ 0.9 is applied automatically; otherwise, the top 5 candidates are returned for selection.
    • Chinese citations are marked as “not verifiable online”.
  3. refs_dedup: Automatic deduplication. The criteria are identical DOI, or title similarity ≥ 0.85 and the same first-author surname. The most complete or verified entry is retained, and removed items are reported.
  4. refs_format: Outputs standardized formats.
    • Supported formats: GB/T 7714-2015 / APA / Vancouver.
    • English authors are normalized to “surname + initials” (e.g., ZHANG S); Chinese authors keep their full names.
    • Pending items are listed at the end. Optional writing to .md files.
  5. research-refs skill: Responsible for orchestrating the workflow (parse -> confirm -> verify -> deduplicate -> request style -> output) and enforcing the rule of “never fabricating fields”.

Installation and Enablement

Install dependencies:

npm i github:Parker-xia/dsh-research-refs

Add the plugin to the DSH composition configuration. Append the following content in cordis.yml or the agent preset’s agent.cordis.yml:

- id: research-refs
  name: 'dsh-research-refs'

Or use the bundle patch mechanism (dsh bundle patch).

Dependency Notes:
* The host must provide the tools service. fs and skills are optional; if missing, it degrades and marks the condition.
* Node version requirements must be met: Node >= 22.19.0 or >= 24.0.0.
* Tested successfully on @deepseek-ai/dsh@0.1.0-rc.6.

Typical Usage

In the conversation, paste a messy citation snippet and issue a cleanup command.

User Input Example:

clean up these references:
1. Zhang S et al. 2021. Deep learning for CT reconstruction...
2. Wang S (2024) Image reconstruction algorithms...

Agent Response Workflow:
The agent loads the research-refs skill and reports processing status step by step:

[refs_parse] 2 entries split, all fields recognized...
[refs_verify] #1 verified (DOI), #2 candidates awaiting choice...
[refs_dedup] 2 kept, none removed
[refs_format] GB/T 7714-2015 output...

Technical Details

  • Network Verification: Online verification uses Node native fetch (DSH Node ≥ 22). The plugin does not register an external web fetch provider, so it will not conflict with the host’s network configuration. If native fetch is unavailable (e.g., in a sandbox environment), it automatically degrades to the ctx.web service.
  • License: MIT.

Ecosystem Context

DSH follows the “everything is a plugin” philosophy. This plugin is listed on the independent site catalog_url, and has no affiliation with the official DeepSeek or Huanfang (High-Flyer) organizations.

References