The core of DeepSeek Harness is “everything is a plugin.” When building long-running agents, maintaining skill files (SKILL.md) often requires human intervention, which can lead to maintenance lag or disconnection from actual conversations. dsh-skill-curator aims to solve this repetitive labor problem by using a background agent to automatically review conversations and adjust the skill library.

This is a DeepSeek Harness (DSH) plugin maintained by runfali. It uses background subagents to periodically review conversations and automatically create, modify, or delete SKILL.md files. Its design philosophy draws on Nous Research’s Hermes Agent self-improvement loop, but integrates into DSH in a zero-intrusion way.

Core Features

  • Background automatic curation: The plugin listens for conversation turns. After reaching the configured interval (default 3 turns), a background subagent reads the session summary and edits the skill files.
  • Manual intervention: Provides the /skill-refine [focus] command, allowing a review to be manually triggered at any time.
  • Safe write mechanism: Before writing to a file, it must first call skill-library-read to obtain and echo the SHA256, preventing overwriting unread content; each write is automatically backed up to ~/.dsh/skill-curator/backups/.
  • Version rollback: Each review automatically commits changes to Git, preserving a rollback point.
  • Tool whitelist: The subagent can only call the skill-library-* family of tools, limiting its operational scope.

Installation and Enablement

  1. Enter the plugin directory.
    cd /path/to/dsh-skill-curator
  1. Install the plugin.
    dsh plugin --profile web add ./
  1. Restart dsh for the changes to take effect. After installation, find the “Skill Curator” card in Settings -> Plugins tab.

Typical Usage

  • Manually trigger a review: Enter /skill-refine [focus] in the conversation to immediately review the current session.
  • Adjust curation frequency: Modify skillNudgeInterval in the plugin settings to control the trigger interval for automatic curation.

Applicable Scenarios and Notes

  • Applicable scenarios: Suitable for scenarios where the agent needs to continuously optimize its skill library in long conversations and reduce manual maintenance costs.
  • Runtime environment: Depends on Node.js ^22.19.0 || >=24.0.0, as well as the @deepseek-ai/dsh-* series of dependencies. Ensure the plugin runs with the permissions of the current dsh process.
  • Configuration: Use excludedSkills to configure a list of protected skills, preventing the curation process from accidentally deleting critical skills.

This plugin enables DSH agents with continuous self-evolution through background automatic curation and strict safety mechanisms, without intruding on the source code.

  • Project homepage: https://github.com/runfali/dsh-skill-curator
  • Ecosystem directory: https://www.skillhub.cn/plugins/runfali/dsh-skill-curator