dsh-skill-manager
A Skill Manager plugin for DeepSeek Harness (Web/Desktop):
- Pick skills BEFORE the conversation starts — the new-chat screen shows a
"对话 Skill" chip above the composer: follow the workspace default, enable everything, or choose exactly which skills this conversation may load, then save before typing the first message.
- Global enable / disable — switch any skill off or on everywhere. Disabled
skills disappear from the model's session skill catalog, from the / input trigger, and cannot be loaded by the skill tool or the /name gesture. Their files are never touched, so re-enabling restores them instantly.
- Per-workspace defaults for new conversations — pick "all available" or
"only these skills" per workspace in Settings → Skill 管理. New conversations in that workspace start with exactly that selection.
- In-conversation management — the
Skillbutton in the conversation
header shows the current chat's effective skill set and lets you follow the workspace default, enable everything, or choose exactly which skills this chat may load. The session catalog republishes automatically on the next message ("对话中管理skill").
- Delete user skills — permanently remove a skill's files from
~/.dsh/skills / ~/.agents/skills (project, custom, and bundled skills are never deletable through this plugin).
How it works
The skill registry merges layers with the NEAREST scope winning a duplicate name, and on the Web surface the filesystem provider is mounted by the agent preset into the preset's STANDING scope layer. This plugin therefore registers a shadow provider named skill-manager with rank 10 (below every shipped provider rank) into each AGENT's own scope layer on agent/created — the nearest layer of all — so its synthetic non-invocable candidates for every effectively-disabled skill name always win that agent's catalog. The skill is hidden from the model and the user without touching its files, and the registration unwinds automatically when the agent is disposed.
The management surface never uses the plain registry read path (an unscoped ctx.skills.list() sees only the global layer — empty on the Web surface): listRealSkills() walks every layer (global + standing + agent) directly, so the settings page shows the real catalog with real descriptions, invocation flags, and paths. The "selected mode" complement uses that same cross-layer universe.
State lives in one small JSON document: <dshHome>/plugin-data/skill-manager/state.json.
Routes
| Method | Path | Body | Effect |
|---|---|---|---|
| GET | /plugins/dsh-skill-manager/state | — | global disabled set, workspace defaults, session overrides |
| GET | /plugins/dsh-skill-manager/skills?cwd=… | — | full real catalog with invocation flags, path, deletable |
| POST | /plugins/dsh-skill-manager/global | { names } | replace the global disabled set |
| POST | /plugins/dsh-skill-manager/workspace | { cwd, mode, selected } | set a workspace's new-conversation default |
| POST | /plugins/dsh-skill-manager/session | { sessionId, mode, selected } | override one conversation's skill set |
| POST | /plugins/dsh-skill-manager/delete | { name } | delete a user skill |
Mutating routes require the x-dsh-skill-manager: 1 header (CSRF hardening).
Build / install
Pack the artifact (the tarball root holds a package/ directory):
tar -czf artifact.tgz package/Install through the Harness plugin center, or manually: add the artifact as a dependency of the profile's package.json ("dsh-skill-manager": "file:…/artifact.tgz"), append dsh-skill-manager to dsh.profile.bundles, run pnpm install in the profile directory, then restart Harness.
Testing
The pure-logic suites (core, provider, state view, intro/detail) need nothing but Node 22+:
node --test test/core.test.js test/provider.test.js test/stateview.test.js test/intro.test.jsThe integration suite imports the harness's own cordis / dsh-skill / dsh-scope packages, so point DSH_HOST_NODE_MODULES at a DeepSeek Harness install's node_modules/@deepseek-ai directory first:
# Windows example
$env:DSH_HOST_NODE_MODULES = "$env:LOCALAPPDATA\Programs\@deepseek-aidsh-desktop\resources\host\node_modules\@deepseek-ai"
node --test test/integration.test.jstest/compose-check.mjs is a machine-specific offline check that composes the live profile (DSH_PROFILE, DSH_HOME and DSH_HOST_NODE_MODULES can point it at any profile) and asserts the skill-manager loader row is present.
Skill frontmatter metadata
The UI shows two Chinese introductions per skill, both read from the skill's own SKILL.md frontmatter (the metadata object passes through untouched and is never rendered into the model-facing catalog):
metadata:
intro: <一句话中文简介，列表展示>
detail: |-
<详细中文介绍，详情弹窗展示：用途/适用场景/工作流程/典型触发语/产出物>The bundled skill-intro-writer skill (see docs/ or your ~/.dsh/skills) automates writing both fields whenever a new SKILL.md is generated.
License
MIT