@wulongxi/dsh-skills-hub

Skills Hub is a DeepSeek Harness (DSH) plugin that adds a native, first-class panel to the web GUI for discovering, installing, and managing Agent Skills.

It ships three capabilities in one panel:

• Market — search online skill registries (default: skillhub.cn, plus skills.sh and any public GitHub repository), preview a skill's full documentation, and read its file manifest before deciding to install.
• My Skills — browse every skill the runtime has loaded, grouped by source; enable or disable skills, edit their SKILL.md, or remove them (deletes go to .trash, never straight to disk loss).
• Create — write a new skill into the user-level or project-level skill root, with frontmatter handling handled for you.

The panel reads the official skill roots (<project>/.dsh/skills, <project>/.agents/skills, ~/.dsh/skills, ~/.agents/skills, plus any configured customSkillDirs) and merges the host's skill registry on top, so what the panel displays is exactly what the model receives.

Status	Market search & preview: available. Skill installation: planned (see Roadmap).
Requires	DSH >= 0.2.0-rc.2
License	BSD-3-Clause
Languages	English, 简体中文 (follows the GUI language setting)

---

Installation

From a local checkout (link install)

dsh plugin --profile <profile> add link:<path-to-this-repo>

Restart the DSH service. The sidebar then shows a Skills Hub entry; selecting it opens the panel in the center column.

From npm (once published)

dsh plugin --profile <profile> add @wulongxi/dsh-skills-hub

Usage

Market tab

Type a query and press Enter. The source picker to the left of the search field selects which registry the search asks; the choice applies to the next search, so results always match the query that produced them.

• skillhub.cn (default) — a China-hosted skill store with 190,000+ skills, reachable directly with no network accelerator. The site explicitly invites agent integration: it publishes an agent-facing integration document (install/skillhub.md), allows crawlers in robots.txt, and its read path is unauthenticated. Rows carry Chinese descriptions, download counts, and icons.
• skills.sh — the Agent Skills registry's search endpoint. This source serves one page; the panel says so honestly instead of offering a load-more control that could only return empty.
• GitHub repositories — a repository can also be browsed directly: any owner/repo (a full URL or a .git suffix also works) is enumerated through the GitHub API, and every skill directory inside it becomes a row.

Long result lists load the next page automatically as you scroll, with an explicit "load more" button as the keyboard/click target. Selecting a skill's name opens a detail dialog rendered by the shell's own markdown renderer — the skill reads exactly as it does in a conversation — alongside its file list and total size.

Installing from the market

The detail dialog carries the install action. A skill is installed after its body has been read, so the licence line and the file list sit on the same screen as the button. Installing asks twice: the target root (user ~/.dsh/skills, shared by every project, or the current project's .dsh/skills), then a confirmation that states the licence situation at the moment of decision.

The install itself is engineered to fail cleanly:

• Staged then renamed. Files land in a staging directory beside the target and the directory is renamed into place only when every file arrived. A failed or cancelled install leaves the previous state of the disk untouched — no half-written skill the model could later load.
• Collision is an error. An existing skill with the same name is never silently replaced (the panel answers 409; updating an installed skill is M4's job).
• Bounded. Per-file 2 MB, 16 MB total, 200 files, and every manifest path is validated against a strict relative-path grammar before anything is written.
• Provenance travels with the skill. A .skills-hub.json record inside the installed directory names the market id, the source, the licence at install time and the timestamp — the anchor M4's update checks read back.

Licences are stated, never laundered. Every result row declares its licence: declared ones show it (MIT, GPL-3.0, …); undeclared ones show No licence declared. In a sampled catalogue more than half of the entries declare none, and that is a fact a user needs before deciding to copy a skill to disk.

My Skills tab

Skills are grouped by provenance (project, user, plugin providers, runtime registration). A skill backed by a file can be enabled or disabled — the toggle rewrites disable-model-invocation in its frontmatter — edited, or deleted. Skills registered by plugins at runtime have no file on disk; the panel says so once, per group, instead of decorating every row, and their enable state is stored in the panel's own configuration.

Preview

Every skill can be previewed, including bundled and runtime-registered ones that have no SKILL.md on disk — the host falls back to the runtime skill registry for those, and the dialog labels the source honestly ("from file" / "from skill registry"). Frontmatter fields are shown as a definition list above the body and are never mixed into the markdown, because a leading --- fence is a thematic break in GFM and would mangle the YAML into headings and stray prose.

Configuration

Plugin config (profile patch):

Field	Type	Default	Description
enabled	boolean	true	Master switch; false mounts no routes.
githubProxies	string[]	built-in chain	Accelerator prefixes for GitHub raw content, tried in order; include "" to try direct first.
customSkillDirs	string[]	[]	Additional skill root directories to scan.
dshHome	string	$DSH_HOME or ~/.dsh	Override for the user config root.

Security model

• Every route sits behind the shared loopback fence (socket address + Host header + browser same-origin markers). An unpaired LAN client receives 403 forbidden: loopback-only before any skill file is touched. When dsh-remote-web-ui is loaded, a live paired-device cookie is an additional allow path; unpaired and revoked devices still receive 403.
• A client-submitted path is only an identity claim. Before touching any file, the host re-scans and requires the same skill name to resolve to that exact path. A project skill that disappeared can therefore never be redirected onto a same-named user skill, and the routes cannot be used to read or write an arbitrary path.
• The browser never sends a URL. Market requests carry only the opaque skill id the host issued; repositories, refs, directories and file paths are re-derived host-side, and the id grammar is validated segment by segment, so a path traversal is rejected before it can reach any fetch.
• Remote fetching happens only in the host half, behind the trust fence. The unauthenticated GitHub API allows 60 requests/hour, so only discovery (tree listing) touches it, cached for 30 minutes; file bodies come from raw.githubusercontent.com through an accelerator chain, which does not count against that budget. skillhub.cn is reached directly.
• Registry icons are proxied, not linked. Icon URLs are fetched host-side under an explicit policy (https only, allow-listed host, image content-type, 256 KB cap) and re-served as data URLs, so the browser never loads a third-party origin: no referrer leak, no tracking pixel per render.
• Installs re-derive everything from the id. The install route receives only the opaque skill id and the workspace the panel is showing; the file list comes from the skill's own detail manifest (never client-supplied), the target directory is derived host-side, and every manifest path is validated before it is joined onto the staging root. An install can therefore only write a skill directory under a skill root.
• Writes are bounded (64 KB bodies), symlinks refuse delete/update (both would escape their skill root), and deletes move files into .trash rather than removing them.

Architecture

The package is a dual-half DSH plugin:

• Host half (lib/index.js, runs in the DSH server process) — serves the /api/skills-hub/* route family, owns all network access, file I/O, and the trust fence.
• Client half (lib/client.js, runs in the web GUI) — registers the sidebar entry and the center-column page through the shell's official slot seats (sidebar.panellist and the keyed main slot), so the shell owns the row box, active highlight and panel switching exactly as it does for its own Plugins and Schedule pages.

The client is built only on the host's frozen browser module table (react, @deepseek-ai/cordis, @deepseek-ai/dsh-client-ui-*); every control is the official shared primitive (Button, Switch, Input, Menu, Modal, MarkdownText, …) — nothing is hand-drawn. Registration failures are logged and degrade the panel; they never fail the GUI's boot.

Development

pnpm install
pnpm typecheck        # tsc --noEmit
pnpm build            # tsc -b tsconfig.build.json && tsdown

Test suites (run with node tests/<file>.mjs):

Suite	Scope
m0-smoke.mjs	Loads both bundles the way the host does; asserts the loader contract, injected-require usage and route registration.
m1-host.mjs	Drives the route family against a real temporary skill tree: scan grading, write-route identity checks, symlink protection, fence denials.
market.mjs	Market layer with fake transports, zero network: aggregation, dedup, id grammar (incl. path-escape attempts), source contracts, installer staging/path-safety/provenance.
ui-structure.mjs	UI guards: official-primitives-only, card anatomy, scroll height chain — each negative-controlled.
theme-tokens.mjs	Verifies every referenced CSS custom property exists in the shipped theme, because an undefined token makes var() silently drop the whole declaration.
market-live*.mjs, icon-live.mjs, install-live.mjs, tools/scroll-selftest.mjs	Manual, live-network / real-browser smoke tests; not part of the automated gate.

Project layout

src/
  index.ts                host entry: plugin contract, config, mount-once guard
  routes.ts               /api/skills-hub/* handlers
  access.ts, loopback.ts,
  pair-access.ts          trust fence (loopback + paired-device cookie)
  http.ts                 bounded body reader, JSON writer
  core/                   multi-root skill scan, frontmatter parse/rewrite,
                          provider toggle store
  market/                 source registry: fetch limits, TTL caches, accelerator
                          chain, skill-id codec, source adapters
  client/                 browser half: dictionaries, slot registrations, panel
tests/                    automated suites (see Development)
docs/                     design documents (Chinese)

Roadmap

Milestone	Scope	Status
M0	Plugin skeleton, sidebar entry, build pipeline	✅ Shipped
M1	Local management: browse / enable / edit / create / delete / preview	✅ Shipped
M2	Market: multi-source search, detail preview, paging, icons	✅ Shipped
M3	Installation engine: validation, size caps, atomic writes, provenance	✅ Shipped
M4	Lifecycle: provenance store, update checks, local-modification detection	Planned
M5	Custom registry manifests, offline import, accelerator settings UI	Planned

License

BSD-3-Clause © WLX2007