dsh-plugin-multimodal: Adding image support to text-only DSH sessions
dsh-plugin-multimodal is a visual sidecar plugin for DeepSeek Harness (DSH), designed to address the issue where text-only main models cannot process image inputs. This plugin receives image attachments via the GUI, invokes a vision-capable sidecar model to convert image content into text descriptions, and then sends the text to the main model, thereby bypassing the "current model does not support images" limitation. Core features include support for pasting images in the Web UI, a see_image tool for processing disk screenshots, and configuration of the vision model via environment variables or configuration files. Users need to install the plugin and restart the DSH Web process to ensure the sidecar points to a real multimodal model (such as GLM-4.5V) and avoids using a text-only model as the vision endpoint. This plugin is suitable for those who wish to...
Read Moredsh-read-image: Plug-and-play Image Reading for Text-only DeepSeek Harness
`dsh-read-image` is an MIT-licensed open-source plugin for DeepSeek Harness (DSH) designed to address the pain point where pure text models cannot directly read image pixels. This plugin enables text routing to accept image inputs, projects image chunks into `[Image #N]` text placeholders, and automatically registers the `read_image` tool. Through this tool, developers can invoke a configurable vision model to read specified images or file paths within a session, converting their content into text for model inference. If the routing itself supports native multimodality, the plugin intelligently falls back to the built-in tool to avoid conflicts. The plugin provides a "Read image" settings block, supporting the configuration of parameters such as the vision model and prompts. Installation requires DSH version 0.1.0-rc.6 or higher; simply add it via the command line and restart. The plugin features environment self-check and failure-safe mechanisms.
Read More