Introduction

In the DeepSeek Harness (DSH) ecosystem, all features exist as plugins. One of the most common questions developers ask when writing a new plugin is: Has the community already implemented similar functionality? Manually searching GitHub topics, Awesome lists, or npm packages is inefficient. dsh-dejaview aims to solve this problem by front-loading the “is there already a similar plugin” check, helping developers avoid reinventing the wheel.

Plugin Overview

dsh-dejaview is a DeepSeek Harness plugin maintained by jiang4wqy. It registers a model-facing tool called check_plugin_novelty. Before building a new plugin, this tool searches the awesome-dsh-plugin registry, GitHub dsh-plugin topic, and npm packages to answer whether a similar plugin already exists. The plugin is responsible for retrieval and scoring, while the Verdict is made by the Harness model based on the evidence.

Core Features

  1. Register a check tool: Registers the check_plugin_novelty tool in the Harness environment for the Agent to call before building a plugin.
  2. Multi-source retrieval: Searches three public sources: the awesome-dsh-plugin registry, the GitHub dsh-plugin topic, and npm packages.
  3. Similarity scoring: Returns ranked candidates with similarity scores and evidence. Scoring is based on IDF (inverse document frequency) weighted semantic similarity.
  4. Verdict guidance: Provides six-dimensional verdict guidance and disclaimers to assist the model in making a decision.
  5. Environment dependency: Requires DeepSeek Harness (engines.dsh >= 0.1.0-rc.6).

Installation and Enablement

Run the following command in the project directory to install it:

dsh plugin --profile dejaview add .

After installation, the plugin automatically registers the corresponding tools in the Harness context.

Typical Usage

In a DSH prompt or configuration, the Agent can call the check_plugin_novelty tool. The tool accepts the following parameters:

  • idea: A one-line description of the plugin you want to build.
  • name: The proposed plugin name.
  • keywords: Keywords for key capabilities.

For example, the Agent can input: “replace rotating phrases with a turn-status label”.

The tool returns a JSON-formatted result, including a list of candidates, similarity scores, sources, and verdict prompts.

Notes

  • Evidence limitations: The evidence only covers the sources queried at the time of the request and does not guarantee absolute novelty. Not found does not mean it does not exist.
  • Network restrictions: Network access is limited to GitHub’s public REST API and raw file hosts. The plugin does not send credentials or read local files.
  • Retrieval degradation: If retrieval degrades (for example, due to rate limiting), the tool returns partial results and lowers confidence instead of failing outright.

Closing

dsh-dejaview brings the evidence-first philosophy of DejaView into the DSH ecosystem. By standardizing retrieval and scoring, it enables the model to make verdicts based on objective data. For DSH plugin developers, this is an effective way to perform duplicate checks before building and ensure development efficiency.