dsh-gateway-provider

DeepSeek Harness provides one adapter per official provider out of the box. The moment your models are served through a gateway like newapi, LiteLLM, Higress, or any other OpenAI-compatible endpoint, you are left to hand-maintain a static model list with guessed context windows and output caps. dsh-gateway-provider replaces that fragile setup by attaching directly to the gateway: it fetches the live catalog at startup, cross-references every model against models.dev for real parameters, and reflects new gateway models instantly—no re-deploy, no YAML wrangling.

Under the hood each request is translated into the native wire protocol of the model you picked—OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, or Gemini—using the same pi-ai SDK that powers the official dsh adapter, so tool calls and streaming behave exactly the way the model's maker intended. You can register multiple gateways side by side, each with its own route, API key, and 30-minute catalog cache. Fine-grained per-model controls (protocol pinning, context window, output cap, reasoning levels, include/exclude) are all available through the Settings → Gateway Models web page, and API keys are written into dsh's credential store rather than living in a config file.

This plugin is aimed at teams and power users who route large model portfolios through an LLM gateway, need each model to speak its own protocol for reliable tool calling and streaming, and want to retire the manual model-list maintenance that static adapters force upon them. One gateway API key is all it takes to make every catalog model appear in the dsh picker with correct parameters.