dsh-vision-tool: Adding Image Understanding Capabilities to a Text-Only Main Model
dsh-vision-tool is a plugin designed to address the issue that DeepSeek Harness (DSH) cannot handle images with text-only models. By relaxing admission controls, registering tools, and injecting guidance, it delegates image understanding to a user-configured multimodal model. The plugin includes an intelligent backoff mechanism, so models that natively support images are not affected. Installation requires running the command in the DSH environment and being aware of the security risks related to pnpm build authorization. The configuration includes the vision model provider and its parameters; during use, users paste an image and ask a question, while the main model automatically invokes the tool and obtains the vision model's analysis results.
Read Moredsh-vision-tool: A DeepSeek Harness Plugin That Adds Kimi Vision Capabilities to Text Models
The `dsh-vision-tool` plugin for DeepSeek Harness (DSH) enables visual capabilities for text-only models. It uses `vision-prompt` to intercept requests, automatically persists images as attachment references, and rewrites prompts. It also registers a global tool via `vision-tool` to invoke the Kimi vision model for image analysis. The plugin supports installation and configuration, allowing users to specify Kimi API parameters in `cordis.patch.yml` and supporting multiple image formats. After installation, the profile must be restarted and dependencies must be configured correctly. Operation logs are recorded in `$DSH_HOME/vision-trace.log`.
Read More