Introduction

In monitoring and evaluating model inference, averages often obscure the distribution details of the data. To more accurately assess performance bottlenecks or extreme cases of token consumption, calculating percentiles is a necessary approach. DeepSeek Harness (DSH), as a plugin-based architecture, allows community developers to extend its core capabilities. dsh-percentile focuses on calculating percentiles locally, without additional network requests.

What It Is

dsh-percentile is a plugin for DeepSeek Harness that calculates percentile values during the model inference process. The plugin is maintained by the developer uckkk. Its core purpose is to provide lightweight statistical computation capability, without relying on external network services, ensuring stable operation in offline or intranet environments.

Installation and Enablement

Installing this plugin requires a DeepSeek Harness runtime environment, and the environment must satisfy its peer dependencies.

  1. In the terminal, run the following command to install the plugin:
dsh plugin add github:uckkk/dsh-percentile
  1. Ensure that the library versions required by the current DSH process meet the following requirements:
    * The @deepseek-ai/cordis version must be ^4.0.1 or higher
    * The @deepseek-ai/dsh-tools version must be 0.1.0-rc.6 or later

Core Features

  • Pure Node.js implementation: The underlying logic is written entirely in native Node.js, resulting in efficient execution.
  • No network dependency: All computation is performed within the local process, without sending requests to external APIs, reducing latency and avoiding the impact of network failures.

Use Cases and Notes

  • Intended users: Developers or operations personnel who need to deeply analyze the distribution of key metrics in an inference chain, such as latency and token count.
  • Runtime permissions: The plugin runs with the permissions of the current DSH process. Before installation, it is recommended to review the source code and LICENSE (MIT) in the GitHub repository to confirm that its behavior meets your security and compliance requirements.

Conclusion

dsh-percentile provides a lightweight, local percentile calculation solution for the DSH ecosystem, making it suitable for monitoring scenarios that are sensitive to data distribution. For more details and source code, please refer to the GitHub repository or the DSH plugin directory.