creating-dsh-plugins

An agent skill for building tools and plugins for DeepSeek Harness (dsh) — the official plugin API, project layout, and workflow, distilled from the real @deepseek-ai/dsh-tools type definitions. Written in the standard SKILL.md format so every mainstream AI coding agent can load it.

If your agent has ever refused to write a dsh plugin (unfamiliar API) or confidently emitted a wrong one (run instead of execute, missing output, no inject), this skill fixes both: it gives the agent authoritative signatures it can copy verbatim.

Install

The skill is a single SKILL.md folder. Install it in the skills directory your agent reads. The common locations:

Agent	Path
Claude Code	~/.claude/skills/creating-dsh-plugins/
Codex	~/.codex/skills/creating-dsh-plugins/ or ~/.agents/skills/creating-dsh-plugins/
Gemini CLI	~/.agents/skills/creating-dsh-plugins/
opencode	~/.config/opencode/skills/creating-dsh-plugins/ (or a plugin)
Cursor	~/.cursor/skills/creating-dsh-plugins/
superpowers (plugin ecosystem)	~/.config/opencode/node_modules/superpowers/skills/creating-dsh-plugins/

One-line install:

mkdir -p ~/.claude/skills && \
git clone --depth 1 https://github.com/dsh-io/dsh-plugin-skill.git ~/.claude/skills/creating-dsh-plugins
# swap the destination path for your agent (table above); for Codex/Gemini use ~/.agents/skills

For a single project, you can also drop the folder into the project's skills directory (.claude/skills/, .cursor/skills/, ...) so only that repo's agents load it.

Verify it works

Start a fresh session in your agent and ask:

Write a dsh plugin that registers a get_weather tool using the creating-dsh-plugins skill.

A correctly loaded skill produces src/index.ts with inject = ['tools'], ctx.tools.register(defineTool({ ... output: { schema, render } ... })) and an execute(args) method — never run, never a required array.

Contents

• SKILL.md — the skill itself: tool API reference, hard schema rules, project layout, workflow, common mistakes, quick reference

Authoritative facts baked in

• Tools register via defineTool from @deepseek-ai/dsh-tools; output (schema + render) is mandatory
• required is per-property (required: true); there is no required array, no required: false
• additionalProperties must be explicit on object schemas
• Container is @deepseek-ai/cordis (a cordis fork), not koishi
• pnpm users need a pnpm-workspace.yaml override for the unpublished upstream peer @deepseek-ai/dsh-type-meta
• Distribution: npm publish + GitHub dsh-plugin topic

Related

• @dsh-io/dsh-dev — the official dsh plugin toolchain CLI (scaffold, dev, check, publish)
• DeepSeek Harness — the agent framework this skill targets

License

MIT © 2026 dsh-io