dsh-provider-qoder

DeepSeek Harness is a flexible model-inference orchestration framework, but wiring it to the Qoder CN gateway means dealing with COSY request signing, WAF-encoded streaming, and the PAT-to-short-lived-job-token exchange. dsh-provider-qoder packages that entire chain into a plug-and-play adapter so Qoder CN inference services slot into DeepSeek Harness as naturally as any other backend.

Under the hood the plugin registers an llm-qoder settings namespace and resolves connection facts per request. Public-cloud users only need to paste a PAT; enterprise VPC deployments can pin the vpcInstance in settings.yaml. Model discovery follows a live-first strategy: every time DSH enumerates the provider, the adapter fetches Qoder signed catalog, so newly released wire keys (e.g. gmodel) appear without a plugin update while known models keep stable selector ids (gm51model → glm-5.2). A static models override is available for deployments that deliberately pin the selector list, replacing live discovery rather than extending it.

The design is a good fit for teams that already run multiple inference backends under DeepSeek Harness and want a single scheduling entry point, particularly when traffic must go through Qoder CN public-cloud or enterprise-VPC gateway. Auth, transport, malformed-JSON, and empty-catalog failures are surfaced as discovery errors rather than silently falling back to a stale built-in list, keeping the troubleshooting path transparent and predictable.