dsh-vision

Several models inside DeepSeek Harness (such as deepseek-v4-flash) have no native vision capability, and the stock DeepSeek adapter makes it worse by rejecting image uploads with a MODEL_DOES_NOT_SUPPORT_IMAGES error, so users never even get a chance. dsh-vision exists to plug that gap: it gives text-only models an extra set of eyes, letting them inspect a local or remote image mid-conversation and pull a Chinese-language description back into the context window.

The flow is simple. During inference the model automatically invokes a vision tool, handing a local absolute path or an http(s) URL to a vision LLM (default: Alibaba Cloud Bailian qwen3.7-flash over the OpenAI-compatible API; switching vendors is a one-line base-URL change) and receiving a text description to continue with. The plugin has zero extra runtime dependencies, relying solely on Node built-in fetch, and configuration follows a clean precedence chain: plugin config, then environment variables, then defaults.

It is built for two audiences: developers who are assembling agents on DeepSeek Harness and need to add vision to otherwise text-only models, and everyday users who simply want to paste a screenshot into the chat box and get an instant written description. Pair it with the small adapter patch described in the project PATCHES.md to lift the upload restriction, and you have a paste-image, get-description loop ready to go.