dsh-skill-manager

English | 中文

A DeepSeek Harness plugin that lets you manage your own skills — the ~/.dsh/skills directory the harness already discovers — straight from the chat, through slash commands and a model-facing tool.

DeepSeek Harness ships the full skill capability (the dsh-skill registry, the dsh-skill-filesystem provider, the skill loader tool, and the /name composer source). This plugin is a Consumer of that capability: it only reads and writes the user skill directory, so anything you create or change here shows up in the / menu and the model catalog on the next discovery.

Everything is a plugin: this package is a bundle (dsh.bundle → a cordis.patch.yml that inserts one host-plane row). It is plain ESM with no build step.

Install

dsh plugin --profile web add github:bitterSmilezzz/dsh-skill-manager
dsh web

Pin a commit for reproducible installs:

dsh plugin --profile web add github:bitterSmilezzz/dsh-skill-manager#<sha>

To remove it:

dsh plugin --profile web remove dsh-skill-manager

Usage

Slash commands (human, in the composer)

Command	What it does
/skills	List your user skills and their model/user invocation flags.
/skill-remove <name>	Delete a user skill by name.

The skill_manage tool (the model)

Ask the agent to manage a skill and it calls skill_manage:

• list — list user skills.
• get (name) — read one skill's instructions.
• save (name, description, content, optional whenToUse, modelInvocable, userInvocable, originalName for rename) — create or update a skill.
• remove (name) — delete a skill.

For example: "Create a skill named code-review that reviews a diff for correctness and style."

Skill file format

Skills are the exact format dsh-skill-filesystem discovers — a <name>/SKILL.md directory bundle with YAML frontmatter:

---
name: code-review
description: Review a code change
whenToUse: For pull-request review tasks
---

The instructions the model follows when this skill is loaded.

disable-model-invocation: true and user-invocable: false are written only when the corresponding surface is turned off, matching the provider's default-permit semantics.

How it works

dsh --profile web
  └─ skill-manager (host plane) injects commands + tools
       ├─ /skills, /skill-remove        → UI command plane (no model tokens)
       └─ skill_manage tool             → model calls it to read/write SKILL.md
            └─ dsh-skill-filesystem     → discovers the change on next catalog refresh

Configuration

Row	Field	Default	Meaning
skill-manager	dshHome	$DSH_HOME, then ~/.dsh	Harness home whose skills directory is managed.

Model Experience

• The tool result is model-visible only when the model calls skill_manage; it returns a compact string summary. The tool schema adds a fixed per-request cost when the tool is visible.
• Slash commands add no model tokens: command discovery, execution, and output stay in the UI command plane.
• Changed skills reach the model through the existing skill catalog and /name gesture, which dsh-tool-skill already republishes after the filesystem provider invalidates.

Known Limitations

• User-root skills only — this plugin manages ~/.dsh/skills; bundled, project, and preset skills are read-only elsewhere (they appear in the per-session / menu).
• No settings UI — management is command- and tool-driven, matching filesystem-backed agent conventions (Claude Code, Codex) rather than a graphical page.
• Text-only skill bodies — the instruction body is plain Markdown; there is no resource enumeration or attachment upload.
• Watcher latency — a change is visible to the model on the next catalog refresh (the filesystem provider watches the directory); the tool result itself does not force an immediate catalog republish.

Publishing

This repository is publish-ready as an installable bundle:

• No build step — index.js is plain ESM, so a git install needs no prepare script and no allowBuilds step.
• Add the dsh-plugin topic (already set on this repo).
• dependencies (yaml, @deepseek-ai/dsh-tools) are published, so dsh plugin add installs them with pnpm.

License

MIT