dsh-custom-provider

Pointing a client at a model endpoint is rarely as simple as entering a base URL and an API key. The wire protocol might be OpenAI Chat Completions, OpenAI Responses, or Anthropic Messages; the model catalog is often known in advance rather than discovered at runtime; context windows and max output sizes need explicit declaration; and request fields for reasoning effort or message replay can differ from one provider to the next. dsh-custom-provider turns all of that into provider routes that DeepSeek Harness picks up directly. Configured models appear in DSH's native model picker and flow through the standard streaming, tool-call, usage, finish-reason, cancellation, and reasoning-effort pipelines.

The plugin ships a bilingual Web settings page alongside a declarative settings.yaml section. Each route independently selects one of the three protocols, and the model catalog is static with no dependency on a remote /models endpoint. API keys are resolved per request through DSH's credentials service and are never persisted into configuration files. Chat Completions routes accept provider-level and per-model compatibility overrides covering the store parameter, developer role, thinking-content format, maxTokens field name, and related details. A full configuration validation gate runs before any route replacement takes effect, so a bad edit cannot leave routes half-activated.

If you run a compatible gateway, a self-hosted LLM service, or an aggregator endpoint and want DSH to treat it as a first-class model source, this plugin is built for that. It introduces no telemetry, requests no workspace, shell, or browser access, and manages only the llm-custom configuration namespace and the DSH credentials it references.