In the use of DeepSeek Harness, when an Agent loads a Skill, regular users often only see the result and do not know what it loaded or how it worked step by step. The plugin keeps two things separate: what the Skill declares (only from the SKILL.md body) and what was actually loaded in this session (only from DSH runtime facts).

The plugin uses Skill Trace (local loading evidence) to turn the runtime process into a reviewable, learnable local receipt.

Plugin Introduction

The plugin is named polinnizhong/dsh-skill-trace and belongs to the Memory category.

  • Core Value: See the Skills the Agent actually loads, and turn the runtime process into a reviewable, learnable local receipt.
  • Owner: PolinniZhong
  • License: MIT

Core Capabilities

The plugin provides the following core features:

  1. Skill Insights: Provides DeepSeek Harness plugin capabilities.
  2. Skill Trace (local loading evidence): Records the process by which the Agent actually loads Skills.
  3. Workflow: Supports a skill distillation workflow of understanding → reading → translating → replicating → evolving.
  4. List Display:
    • This Session’s Skills: Opens to the list of Skills that were actually loaded in this session.
    • Installed Skills: The list of Skills that can be discovered and invoked in the current environment.
  5. Detail View:
    • Skill Detail: Includes the original SKILL.md and a Chinese Reading Version.
    • Replicate Skill: Provides the replication function in the left-side object area.
  6. Evidence Scope: Observes the invocation/result of skill(name), distinguishing requests, successes, failures, unknowns, and manual judgments.

Installation and Enablement

Run the following command in the CLI to install it:

dsh plugin --profile web add dsh-skill-trace@0.9.2

After installation is complete, restart DeepSeek Harness Desktop and open Skill Insights in the session.

Typical Usage

  1. Check Loading Status: Open the plugin. By default, it enters the “This Session’s Skills” list, which shows the Skills that were actually loaded in this session. Only Skills with observed real loading evidence are shown.
  2. View Details: Click any card to open the Skill Detail page. The right side shows the original SKILL.md, and the left-side table of contents is extracted from the body headings.
  3. Switch Reading Language: When you need to read in Chinese, switch to “Chinese Reading Version”. This passes the body text to the host model for translation. The translation is stored locally, and the same version can still be displayed when reopened after exiting DSH.
  4. Replicate Skill: When you find a Skill worth learning, find “Replicate Skill” in the left-side object area and choose to replicate the entire package or only SKILL.md.
  5. Let the Agent Load Naturally: In the session, let the Agent naturally load a Skill and observe the loading evidence recorded by the plugin.

Notes and Boundaries

The plugin runs with the permissions of the current dsh process. Be aware of the following boundaries when using it:

  • Evidence Boundary: A successful skill(name) call only proves that the Agent requested and successfully loaded the Skill; it does not prove that the Agent fully followed the Skill’s instructions, nor does it prove that the Skill led to correct results.
  • No Execution: It does not discover/install/synchronize/route Skills; it does not automatically run scripts; and it does not automatically rewrite, commit, or publish Skills.
  • Privacy Protection: Local-first. It does not store full Prompts, full Skill bodies, Tokens, Cookies, absolute paths, or project content.
  • No Inference: It does not store full Skill instructions, access Tokens, Cookies, credentials, absolute paths, project files, or generated content.

References

  • Plugin directory: https://www.skillhub.cn/plugins/PolinniZhong/dsh-skill-trace
  • Source repository: https://github.com/PolinniZhong/dsh-skill-trace