dsh-qwen-skill-gate

dsh-qwen-skill-gate

Keep Qwen skills out of non‑Qwen sessions. A DeepSeek Harness (dsh) plugin that exposes the Qwen (qianwen*) skills only while the session runs on the Qwen Token Plan CN provider — and hides and blocks them on every other provider.

English | 简体中文

---

The problem it solves

If you install the QianWen skill pack (qianwen-text, qianwen-vision, qianwenai-deploy, …), those skills are advertised to every session, no matter which model you are actually running. On a non‑Qwen model they are dead weight at best — and at worst the model loads one and injects a wall of Qwen‑specific instructions into your context, polluting the conversation and burning tokens.

The usual workaround is to paste a "only use this on Qwen" warning into each SKILL.md. That is an instruction‑level gate: it relies on the model reading and obeying the note. This plugin replaces it with a harness‑level gate — the skills are simply not offered, and cannot be loaded, unless the provider matches.

What it does

For each agent, the plugin watches the session's model provider. When the provider is not the allowed one (default qwen-token-plan-cn), it shadows every matching skill with a locked "tombstone" entry. Because dsh's skill consumer filters on the invocation flags, one mechanism gates all three surfaces at once:

Surface	Behavior on a non‑allowed provider
Session skill catalog	the skill is omitted from the list the model sees
skill loader tool	returns Error: skill "<name>" is not available for model invocation
/name user gesture	the injection is skipped

On the allowed provider the plugin reports nothing, so the real skills pass through untouched — live bodies, resource guidance, and file watching all preserved.

flowchart TD
    A[Agent step / skill lookup] --> B{Session provider<br/>== allowed provider?}
    B -- yes --> C[Report nothing<br/>real Qwen skills pass through]
    B -- no / unknown --> D[Report tombstones<br/>modelInvocable=false · userInvocable=false]
    D --> E[Catalog: skill omitted]
    D --> F[skill tool: load refused]
    D --> G[/name gesture: skipped/]

Install

From the plugin market (recommended). In DeepSeek Harness open Plugins → Discover and search for dsh-qwen-skill-gate, or add the community index if you haven't already:

https://cdn.jsdelivr.net/npm/dsh-plugin-catalog/index.json

From the CLI:

dsh plugin add dsh-qwen-skill-gate
# or target a specific profile:
dsh plugin --profile <your-profile> add dsh-qwen-skill-gate

From npm directly:

npm install dsh-qwen-skill-gate

The plugin is Host‑only: no dependencies, no client module, no build step. It activates automatically and applies to every session in the profile.

Configuration

Defaults are shipped in the bundle's cordis.patch.yml. Override them in your profile's cordis.patch.yml by targeting the row id dsh-qwen-skill-gate (a config override replaces the whole config, so restate the fields you need):

- id: dsh-qwen-skill-gate
  name: dsh-qwen-skill-gate
  config:
    provider: qwen-token-plan-cn   # the only provider allowed to see/load gated skills
    prefixes: [qianwen]            # skill-name prefixes to gate
    skills: []                     # extra exact skill names to gate
    failClosed: true               # hide gated skills when the provider can't be determined

Field	Default	Meaning
provider	qwen-token-plan-cn	The only provider allowed to see / load the gated skills.
prefixes	["qianwen"]	Skill‑name prefixes to gate. The default covers both qianwen-* and qianwenai-*.
skills	[]	Additional exact skill names to gate, on top of the prefixes.
failClosed	true	If the session's provider cannot be determined, hide the gated skills (fail safe).

Skills gated by default

The 14 skills of the QianWen pack: qianwen-text, qianwen-vision, qianwen-image-generation, qianwen-video-generation, qianwen-audio-tts, qianwen-model-selector, qianwen-ops-auth, qianwen-usage, qianwen-payment, qianwen-find-skills, qianwen-update-check, qianwenai-deploy, qianwenai-observe, qianwenai-operate.

Any future skill whose name starts with a configured prefix is gated automatically.

Behavior at a glance

Session provider	Catalog	skill load
qwen-token-plan-cn	all skills visible	allowed
any other provider	gated skills hidden	refused
provider unknown	gated skills hidden (failClosed)	refused

Switching the model mid‑session is handled: the gate re‑evaluates on the next step and the catalog is refreshed, in both directions.

How it works (under the hood)

The plugin registers one skill provider per agent, in that agent's scoped context. dsh composes the filesystem skills per agent and lets the nearest scope layer win a name, so the gate lives in that same nearest layer and shadows the real entries with a lower rank.

Key implementation constraints:

• Providers must be async. The registry races list()/get() against the lookup signal via promise.then; a synchronous return is rejected, the provider is skipped, and discovery is marked incomplete.
• Global registration is not enough. A global provider sits in a farther layer than the per‑agent filesystem skills and loses the name merge. The gate must be registered per agent.
• The provider name is unique per module generation (qwen-gate-<suffix>). Reinstalling a linked bundle loads a new module generation, but the previous generation's per‑agent registrations can leak. With a fixed name the new generation's registration fails ("already registered in this scope"), it loses its invalidate() handle, and a later model switch can no longer refresh the registry cache. A unique name lets each generation register alongside any leaked ones, so the active generation always holds a working invalidate handle and the gate self‑heals across reinstalls. Restarting the app clears leaked generations.
• The decision follows the viewer. list() reads the viewing agent's provider (options.scope), not the agent the provider was registered for, because the registry also consults an ancestor's provider when snapshotting a descendant's catalog.

The current provider is read from agent.model.current() (the same source the agent loop uses for a step's selection), falling back to the session's modelSelection projection and then the global default. The registry caches catalogs per (cwd, scope, revision); a session/event listener invalidates the cache only when a session's gate state actually flips.

Limitations

• It gates skills, not files. The Qwen skill scripts still exist on disk; a model could in principle run them via a shell tool. Nothing advertises that path, but this is a skill‑registry boundary, not a filesystem/credential control.
• Default roots only. Name discovery scans the default skill roots (project .dsh/skills and .agents/skills, $DSH_HOME/skills, ~/.agents/skills). Skills placed only in a custom root configured on dsh-skill-filesystem are not auto‑discovered — add them to skills explicitly.
• Transient "discovery incomplete" notice. Registering a provider invalidates the shared catalog cache; a concurrent read can briefly report incomplete discovery. It self‑corrects on the next request and loses nothing.

Compatibility

• DeepSeek Harness 0.2.x (uses the ctx.skills provider registry, per‑agent scope layers, and agent.model.current()).
• Host‑only; works in the Web GUI and headless. No browser UI.

Development

The plugin is a single plain‑JS module (index.js) plus a bundle patch, locale files, and an icon. To iterate locally, install the checkout as a bundle with the plugin manager, then reload. See the implementation notes above before changing the registration or provider shape.

Release notes

Version	Changes
1.0.0	Initial release. Provider‑gated Qwen skills across the catalog, skill tool, and /name gesture; per‑agent async provider with a generation‑unique name; configurable provider / prefixes / skills / failClosed; bilingual documentation.

License

MIT © PaysNatal

---

_{Not affiliated with the QianWen skill pack or Alibaba Cloud. "Qwen" and related names belong to their respective owners.}