dsh-plugin-dev

August 16, 2026 · View on GitHub

简体中文 | English

An Agent Skill for developing, maintaining, and publishing DeepSeek Harness (DSH) plugins. Tell DSH something like "build a plugin", "modify that plugin", or "publish the plugin to GitHub", and it will follow a complete develop → isolated test → publish pipeline.

What problem does it solve

DSH plugins run inside your DSH process. Installing an experimental plugin means injecting risk into your production environment: style pollution, memory leaks, conflicts with other plugins, even disrupting daily development. This Skill fully separates "experimental validation" from the "production environment":

  1. Local compile-only repo: experimental plugins are only compiled in a dedicated cloned repo, never installed into the production environment;
  2. Isolated-environment testing: deploy to an isolated test server and run independent snapshot tests in a docker container (no pollution of the main environment), passing the four test categories one by one;
  3. Main-environment install check: only after container tests pass, install into the server's main DSH environment and verify no conflicts with other plugins and correct rendering;
  4. Manual verification before installing back: bridge via SSH tunnel for final manual verification, and only after the user confirms the tests pass is the plugin installed into the local test environment.

What's inside

  • Full pipeline: compile → deploy → container snapshot test → main-environment install → bridge verification → install back → archive registration;
  • Four test categories:
    • Overflow tests (memory): abnormal memory growth after loading multiple sessions / memory overflow on page refresh / memory overflow on first click after page refresh;
    • Performance tests (lag): sending a command message / first click opening a long session / first click opening a short session / refreshing the page;
    • Interaction closed-loop tests: click / expand / collapse / input / switch full-chain closed loops;
    • Visual-model verification: whether UI rendering is correct, with misalignment / missing / pollution;
  • Development standards: choosing between dynamic Cordis plugins vs standalone bundle packages, the three package-structure essentials, client bundle contract (__ModuleLoader__, inline styles, slot priority), encoding constraints (no BOM);
  • High-frequency pitfalls: inject declaration, unbounded-loop deadlock, React Hooks order, outside-click close, data hygiene, host↔client communication, performance intuition;
  • Maintenance flow: when a plugin breaks, always follow "container isolated test → main-environment install test → bridge test → reinstall the production package", never patch the local production environment directly;
  • Publishing standards: official publish.md, GitHub essentials (LICENSE + README + package.json), fork / Discussion approach since the official repo does not accept external PRs.

Installation

Send this to DSH:

Please install the dsh-plugin-dev skill from https://github.com/GammaChineYov/dsh-plugin-dev

For manual install, copy the whole skills/dsh-plugin-dev/ directory to $DSH_HOME/skills/ (default ~/.dsh/skills/); for project-only use, copy to <project root>/.dsh/skills/. To share with other agents, you can also put it in <project root>/.agents/skills/. The directory watcher makes it effective immediately, no restart needed.

Skill / memory separation

This Skill only carries executable process (what to do, how to decide). All machine-specific facts (compile repo path, test server address, SSH keys, each plugin's repo and install state) should live in your own memory/archive area. The Skill's memory pointers only give the archive's Chinese display name, never a specific path — paths are maintained by your own memory index, and you can find the matching archive by name. Facts change only in the archive, the Skill never drifts, and the same Skill is reusable across machines.

Dependencies

  • DeepSeek Harness (DSH); dynamic plugin development needs the cordis-plugin-development skill, and composition authoring needs the editing-cordis-compositions skill (both ship with DSH);
  • An isolated test server capable of deploying DSH (docker recommended but optional, for clean snapshot tests).

License: MIT