dsh-llm-openai-compatible

Connecting DeepSeek Harness to an OpenAI-compatible inference endpoint—local vLLM, LM Studio, llama.cpp, Ollama's /v1 compatibility layer, or remote gateways like OpenRouter, Together, and DeepSeek's official API—used to mean patching the core or writing glue code every time you switched providers. dsh-llm-openai-compatible is the universal plug for exactly this: install it, point the base URL at your endpoint, and chat works immediately with zero changes to the Harness core.

Core capabilities include automatic /v1 normalization of the base URL, optional API-key auth (anonymous requests for local services, Bearer auth for remote gateways), a declarative models array recording each endpoint's real model IDs along with vision, thinking, and token-window metadata, a discoverModels() helper that fetches GET /v1/models and imports real IDs in one click, and a live Settings panel where you change the endpoint, key, model catalog, and retry policy with instant effect—no restart required.

It is built for teams already running vLLM, LM Studio, or llama.cpp locally; individual developers who prefer Ollama's models but want to stay on the open protocol; and platform admins routing through aggregators like OpenRouter or Together for unified multi-model management. In short: if a service speaks the OpenAI Chat Completions protocol, plug it in and chat.