dsh-plugin-skill-manager

English | 中文

A DeepSeek Harness plugin that manages your skills — the Claude Code-style SKILL.md instruction files that DSH loads from your project and user directories.

It adds two surfaces to any DSH profile built on @deepseek-ai/dsh-base:

• skill_manager tool — the model can list, inspect, create, update, delete, validate, enable, and disable skills directly in a conversation.
• /skills command — you can manage the same library from the chat input without spending a single token.

Every file this plugin writes follows the exact discovery contract of @deepseek-ai/dsh-skill-filesystem (same roots, same precedence, same frontmatter fields), so mutations are picked up by the live skill registry within its usual watcher stability window — the session's available-skills catalog updates without a restart.

Install

dsh plugin --profile web add dsh-plugin-skill-manager

Then restart the profile (dsh web). Any other profile works the same way. To install a specific version or a git checkout:

dsh plugin --profile web add dsh-plugin-skill-manager@0.1.0
dsh plugin --profile web add github:user/dsh-plugin-skill-manager

To remove:

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

What it manages

The same roots, in the same precedence order, as the shipped filesystem provider:

Rank	Source	Path
100	project-dsh	<projectRoot>/.dsh/skills
200	project-agents	<projectRoot>/.agents/skills
300	custom	configured customSkillDirs
400	user-dsh	~/.dsh/skills
500	user-agents	~/.agents/skills

Skills are <name>/SKILL.md bundles or flat <name>.md files with YAML frontmatter:

---
name: my-skill
description: What this skill does, shown in the catalog.
whenToUse: Optional extra routing hint.
disable-model-invocation: false   # true hides the skill from the model
user-invocable: true              # false hides it from human commands
---
Body instructions the model reads verbatim when the skill is invoked.

name + description are required; whenToUse, metadata, and the two invocation keys are optional (the plugin only writes the invocation keys when restricting, so default files stay Claude Code-clean).

The skill_manager tool

Action	Key arguments	What it does
list	—	Every skill file across all roots with path, format, source, invocation state, plus shadowed-duplicate and live-registry flags.
inspect	name	Full frontmatter and body preview for every entry declaring that name.
create	name, description, content?, scope? (user | project), when_to_use?, model_invocable?, user_invocable?	Scaffolds a valid skill (refuses to clobber an existing name). Default scope comes from config.
update	name + any of description?, when_to_use?, content?, model_invocable?, user_invocable?	Merges frontmatter in place, preserves unknown keys, replaces body only when content is given.
delete	name, confirm: true	Removes the bundle directory or flat file. The explicit confirm flag is mandatory.
validate	name?	Reports broken skill files the loader silently skips (bad YAML, missing fields, bad names) plus ignored directories and unreadable roots.
roots	—	The resolved root list with ranks.

Example — ask the model in a DSH session:

用 skill_manager 建一个 skill：名字 pdf-report，描述"生成 PDF 周报"，正文按我们刚讨论的模板写。

The /skills command

Runs entirely in the UI command plane (output never enters model history):

/skills                              # list everything with status
/skills show pdf-report              # frontmatter + body preview
/skills create pdf-report 生成 PDF 周报  # scaffold, then ask the model to fill the body
/skills disable pdf-report           # hide from the model (disable-model-invocation: true)
/skills enable pdf-report
/skills delete pdf-report --yes      # permanent; --yes is required
/skills validate                     # check every skill file
/skills roots                        # show scanned directories

Configuration

Patch the row in your profile's cordis.patch.yml (dsh plugin installs it with defaults; no configuration is required):

- id: skill-manager
  config:
    defaultScope: user        # where /skills create and bare create put skills
    bodyPreviewChars: 600      # preview length in inspect/show output
    customSkillDirs: []        # extra roots, scanned at rank 300
    # dshHome / agentsHome override $DSH_HOME / $DSH_AGENTS_HOME

Design notes

• Reads cross-check the live registry. list annotates each entry with whether the ctx.skills registry currently serves it, so a shadowed duplicate (e.g. a project skill overriding your user skill) is visible at a glance.
• Deletes are guarded. delete only accepts names discovered by the plugin's own scan — never caller-supplied paths — re-reads the target before removing, and refuses when the file's declared name changed since the scan.
• Diagnostics over silence. The shipped provider hides malformed entries with a log line; validate surfaces them with exact reasons.
• No build step. Plain ESM JavaScript; npm test runs the unit and integration suites (the integration suite boots a real cordis context with the actual dsh services).

Compatibility

• DeepSeek Harness @deepseek-ai/dsh 0.1.0-rc.x profiles built on dsh-base (web, headless, …).
• Node.js ≥ 20.

License

MIT