dsh-glm-vision

DeepSeek and other text-primary models excel at reasoning and tool orchestration but cannot directly interpret images. dsh-glm-vision plugs ZhiPu GLM vision models into DeepSeek Harness as an independent provider route, so the text model keeps its reasoning strengths while GLM simply serves as the eyes.

The plugin registers a glm_vision tool that the main model can invoke mid-conversation to describe a local image in words. It also provides a GLM text-and-vision supplementary route that supports multi-turn tool calling, letting even text-only GLM models process image blocks indirectly. The plugin can take over the official DeepSeek route seamlessly: when an image is uploaded, the prompt gate passes it through and the model is guided to call the vision tool automatically. A settings card lets you adjust thinking depth, timeouts, and the default model on the fly.

Designed for developers already running a text-primary model inside DeepSeek Harness who need image understanding without switching stacks. A single ZhiPu API key is all you need, and a built-in coexistence guard auto-sleeps the plugin when another vision plugin is detected to prevent route conflicts.