DSH (DeepSeek Harness) uses a plugin-based architecture. Existing plugin marketplaces usually allow browsing by name or category. Users often know the task they want to solve, but cannot remember specific plugin names. dsh-need-finder uses local semantic matching to let the Agent directly understand requirements and find the most suitable plugins from a curated catalog.

Core Positioning

This plugin is positioned as a “requirement-driven plugin guide.” It is not a simple catalog index; instead, it uses a local scoring algorithm to match natural-language requirements against the plugin catalog, providing matching reasons and installation commands.

Installation

Install the plugin in the DSH environment:

dsh plugin --profile <profile> add dsh-need-finder

Feature Overview

  1. Local semantic scoring: Scores candidates within the curated catalog without internet access or LLM calls.
  2. Transparent scoring details: The response includes detailed matching weights (e.g., name hit +3, tag hit +2).
  3. Built-in community recipes: Provides 8 curated plugin combinations (dotfiles style), supporting one-click installation of a complete setup.
  4. Prompt injection: Automatically injects prompts into the Agent to guide it to invoke this tool when the user describes a requirement.

Usage

After installation, the DSH environment provides two tools: plugin_guide and recipe.

plugin_guide: Find by Requirement

The Agent describes the requirement in natural language, and the plugin returns the list of most relevant plugins.

plugin_guide need='notify me when a task is completed' limit=5

Parameter Description:
* need (required): Natural-language requirement; supports Chinese and English.
* limit (optional): Number of results to return; defaults to 5.
* category (optional): Exact category filter.
* tag (optional): Tag substring filter.
* diversify (optional): Limits the number of results from the same category.
* perCategory (optional): Maximum number of results per category in diversify mode.

The result includes a details field that shows the matched fields and weights, making it easier for the Agent to explain the recommendation rationale.

recipe: Install by Suite

For complex scenarios, you can use built-in community recipes.

# 列出所有配方
recipe action=list

# 按需求搜索配方
recipe action=search need='手机远程'

# 应用指定配方（输出安装命令）
recipe action=apply id=notify-suite

Built-in recipes include Notification Suite, Security Audit Suite, Mobile Remote Suite, and more.

Limitations and Dependencies

  • Schema compatibility: The tool schema must be compatible with the @deepseek-ai/dsh-tools value-schema DSL. Older versions may fail to start DSH due to schema issues; if this happens, the plugin should be removed.
  • Runtime environment: The plugin runs under the current DSH process permissions, with no network requests, no file writes, and no secrets operations.
  • Dependencies: Requires the @deepseek-ai/cordis and dsh-system-prompt environments.

Ecosystem and Maintenance

  • Owner: 863683348
  • Open-source license: MIT
  • Repository: https://github.com/863683348/dsh-need-finder
  • Catalog: https://www.skillhub.cn/plugins/863683348/dsh-need-finder

Summary

dsh-need-finder uses localized semantic matching and recipe management to solve the pain points of DSH users who “cannot remember plugin names” or “do not know how to combine plugins.” It transforms the Agent from “browsing a supermarket” to “ordering a dish,” directly obtaining matching solutions.