dsh-gateway-provider

DeepSeek Harness ships one adapter per official provider. When your models live behind a gateway such as newapi, LiteLLM, Higress, or any OpenAI-compatible endpoint, the manual workaround is a hand-maintained static model list with guessed context windows and output caps. dsh-gateway-provider mounts the gateway itself: at startup it pulls the live model list via GET /v1/models, then enriches every entry with real parameters from models.dev, so adding a model on the gateway side immediately surfaces it in dsh with no re-deploy and no YAML edits.

Each model is routed over its own native wire protocol—OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, or Gemini—translated by the same pi-ai SDK the official dsh adapter uses, so tool calls and streaming behave exactly as the model maker intended. You can register multiple gateways simultaneously, each with its own route, credential, and cache. Per-model overrides let you pin the protocol, tune context windows, exclude non-chat models by regex, and add fully custom entries, all through the Settings → Gateway Models web page. API keys are stored in dsh's credential store, never in configuration files.

If you manage dozens or hundreds of models through an LLM gateway, want each model to speak its own protocol, and refuse to maintain static model lists by hand, this plugin is built for that workflow. A single gateway API key is all you need to get started.