dsh-auto-vision: Visual Bridging for Models Without Image Input Support
The DeepSeek Harness (DSH) plugin `dsh-auto-vision` provides visual capabilities to models that do not support image input (such as `deepseek-v4-pro`). It intercepts images, uses a configured vision model to recognize them and generate text descriptions, which are then written into the history as collapsed context to enable persistent memory. The plugin features automatic completion of input declarations, request-level stripping, and recursive recognition of tool results. It can be enabled through configuration alone, without modifying source code, allowing non-vision models to handle screenshots.
Read Moredsh-auto-vision: Equip DeepSeek Harness's text-only main model with vision tools
## Introduction When developing agents with DeepSeek Harness (DSH), you are likely to encounter this issue: the built-in `read_image` tool directly injects image chunks into the current model's context, and it is only usable if the current main model supports images. If your main model is a pure text model such as `deepseek v4 flash`, calls to `read_image` will be rejected—the main model cannot perceive the image, effectively breaking the image reading path. dsh-auto-vision adopts a different approach: it forwards images internally to a multimodal model for recognition, ensuring the main model only processes text throughout the process. This article explains its underlying principles, installation, and configuration. ## What is dsh-auto-vision
Read More