Introduction

In DeepSeek Harness (DSH) development, each visible tool definition consumes input tokens on every request. Frequent changes to the tool list alter the request prefix, reducing context cache reuse. Progressive disclosure requires two properties at the same time: a small initial request, and a byte-stable tool and system prefix across subsequent requests.

dsh-progressive-tools plugin provides this contract through the default stable-proxy mode: send a small and fixed tool surface, keep the full catalog in process memory, and execute discovered tools through the normal Harness pipeline.

Plugin Introduction

Name: dsh-progressive-tools
Owner: everclear077
License: MIT
Purpose: Cache-stable progressive tool discovery for DeepSeek Harness.

Core Functionality

This plugin provides a complete tool management mechanism, mainly including:

  • Stable proxy mode: Sends a small and stable tool surface while keeping the full catalog in process memory.
  • Real pipeline execution: Executes discovered tools through the normal Harness pipeline.
  • Searchable catalog: Search updates the conversation history rather than the top-level tool list.
  • Frozen request surface: Ensures the request surface remains stable across multiple calls.
  • Lazy catalog: Keeps the full catalog in process memory.
  • Safety and control: Supports approval, guards, parameter validation, timeout wrappers, result policies, deferred context, and cancellation.

Installation and Dependencies

This plugin runs as a Cordis plugin and requires specific runtime environment.

Dependencies:

{
  "@deepseek-ai/cordis": "~4.0.4",
  "@deepseek-ai/dsh-agent": "0.2.0-rc.1",
  "@deepseek-ai/dsh-llm": "0.2.0-rc.1",
  "@deepseek-ai/dsh-system-prompt": "0.2.0-rc.1",
  "@deepseek-ai/dsh-tools": "0.2.0-rc.1"
}

Runtime environment:
* Node.js: ^22.19.0 or >=24.0.0
* Package manager: pnpm@11.19.0

Typical Usage

In actual configurations, you can enable the plugin’s core capabilities in the following ways:

  1. Use stable-proxy mode to send a small and stable tool surface.
  2. Keep the full catalog in process memory and execute discovered tools through the normal Harness pipeline.
  3. Search updates the conversation history rather than the top-level tool list.
  4. Set maxResults: 2 to reduce the number of visible tools.
  5. Enable approval, guards, parameter validation, timeout wrappers, result policies, deferred context, and cancellation.

Applicable Scenarios and Notes

This plugin is highly effective in regular file and coding work scenarios. Measurement data shows that, in a 100-round test:

  • Visible tools: Reduced from 71 to 9 (87% reduction).
  • Cache read tokens: Reduced from 3,133,428 to 663,687 (approximately 79% reduction).
  • Total latency: Reduced from 45.2 seconds to 20.9 seconds (approximately 54% reduction).

Notes:
* Since discovery results enter the conversation history, uncached input tokens may increase by about 20%.
* The S8 scorer treats tool_search as a failure, which may lead to inaccurate scoring.
* Billing is usually calculated at the lower cache read rate rather than the fresh input rate.

Conclusion

DSH Progressive Tools addresses the impact of tool definitions on context cache stability by introducing stable proxy mode. It reduces the number of visible tools, significantly lowers cache read tokens and latency, and is suitable for DeepSeek Harness pipelines that require stability and performance optimization.

Repository: https://github.com/everclear077/dsh-progressive-tools
Catalog: https://www.skillhub.cn/plugins/everclear077/dsh-progressive-tools