Overview
dsh-vision-3090-fix
Local reverse proxy plugin for DeepSeek Harness that caps images per request to 1 and drops pinned max_tokens, for self-hosted vLLM backends that reject prompts with more than one image or with prompt_tokens + max_tokens beyond max_model_len
README / EN
Package documentation
Registry summary
Local reverse proxy plugin for DeepSeek Harness that caps images per request to 1 and drops pinned max_tokens, for self-hosted vLLM backends that reject prompts with more than one image or with prompt_tokens + max_tokens beyond max_model_len
dsh.pub verifies the pinned bundle contract, runtime facts, and distribution semantics. The complete README remains in the source repository.
Read the full README on GitHubLIMITATIONS
Known limitations
- Buffers the request body fully before forwarding (needed to parse and rewrite JSON); fine for chat/vision payloads, not meant for large file uploads. The response is streamed through without buffering. - No retry logic and no request queuing — it's a thin pass-through, not a load balancer. - Assumes the backend is plain HTTP/HTTPS `chat/completions`-shaped JSON; a provider using a different wire shape for images (not OpenAI's `image_url` content part) won't be recognized.