DeepSeek Harness (DSH) does not hard-code skills to one local folder. Its skill subsystem exposes a provider contract that separates catalog discovery from loading the full selected definition. The documentation explicitly leaves room for local, embedded, and remote providers without changing the model-facing contract.
That seam is an unusually good fit for SkillHub, iFLYTEK's self-hostable registry for publishing, versioning, reviewing, and governing reusable skills.
A thin adapter with a large enterprise effect
The proposed package name is @iflytek/dsh-skillhub. This is a product proposal, not an already published integration. Its job would be deliberately narrow:
list(options)
→ return only the SkillHub catalog entries visible to this identity
get(candidate, options)
→ fetch the full body only after DSH selects a candidate
invalidate()
→ discard completed catalogs after a governed registry change
A configuration could bind the provider to a registry endpoint, namespace, scoped token, version channel, and cache policy. The token should stay in the provider process or secret store—never in model-visible skill text, prompts, or SKILL.md.
The moat is the control plane
An ordinary folder can store skill files. Enterprise deployment needs more: who may publish, who reviews a new version, which namespace a team can see, which release is stable, how a vulnerable skill is withdrawn, and who loaded what.
SkillHub already brings the primitives that matter here: self-hosting, namespaces, semantic versions and tags, review gates, RBAC, scoped API tokens, and audit trails. DSH remains the runtime and model-facing harness; SkillHub becomes the private distribution and governance plane.
Controlled refresh, not uncontrolled mutation
A remote provider should not download every full skill on every turn. It can fetch lightweight catalog metadata, lazily load the selected body, cache by version or ETag, and invalidate the completed catalog after a registry change. The next observation sees updated metadata while the model-facing contract stays the same.
That creates a practical enterprise hot-update path: publish a reviewed version, observe the registry change, invalidate cached discovery results, and let the next turn load the new body. Stable tags and version locks can keep production controlled. Polling, webhooks, and ETags are implementation recommendations; they are not claimed as already shipped DSH behavior.
Solve the cold start with iFly-Skills
A registry needs useful content. iFly-Skills already provides ready-made capabilities such as speech synthesis, OCR, image understanding, transcription, proofreading, translation, voice cloning, contract review, and animated diagrams. Teams could use those as seed content, then add private finance, legal, support, engineering, and operations skills.
The resulting stack is simple:
iFly-Skills and internal skills
→ SkillHub: publish, review, version, govern
→ @iflytek/dsh-skillhub: discover, lazily load, refresh
→ DeepSeek Harness: expose one stable contract to the model
The durable differentiation is not another folder of prompts. It is the organization's reviewed, permissioned, auditable capability layer.
Private, governed skills for DeepSeek Harness.
References: DeepSeek Harness · SkillHub · iFly-Skills
Top comments (0)