dsh-llm-openai-completions

DeepSeek Harness ships with pi-ai, whose compatibility layer assumes every non-standard baseURL supports the developer role. When a self-hosted proxy such as vLLM or LM Studio actually rejects that role, the request dies with a 400 before any inference happens. Compounding the problem, Qwen3-style models render thinking tokens directly into the content body on vLLM, making it impossible for the frontend to separate reasoning from the answer. dsh-llm-openai-completions acts as a fourth adapter class that takes over the wire behaviour of those providers with a zero-guessing policy: system is always sent as role "system", the thinking field is forwarded strictly according to each model's compat.thinkingFormat (enable_thinking, chat_template_kwargs, or reasoning_effort), and the receive path splits thinking text back out of content into a dedicated reasoning block.

Since v0.2.0 the adapter also closes the vision loop. When a model declares image input capability in its llm-pi-ai entry, user-uploaded single or multiple images are serialised as base64 data URIs inside an OpenAI-compatible multi-part content array, preserving source order alongside interleaved text parts. Models without vision capability are rejected loudly with UNSUPPORTED_CONTENT—nothing is silently dropped.

All configuration stays in the existing Settings → Models llm-pi-ai section. The plugin wraps the stream of the providers you list rather than re-registering them, so it coexists cleanly with the built-in llm-deepseek and llm-pi-ai adapters. It is built for developers and teams who run Qwen or other self-hosted models behind vLLM, LM Studio, or a custom OpenAI proxy and need correct thinking passthrough plus multi-image vision input without forking pi-ai.