visual-review

In AI chat interfaces, images are powerful carriers of information, yet standard text models often fail to directly perceive visual content. This limitation hinders seamless interaction in environments like the DeepSeek Harness (DSH) Web interface, where users may upload images expecting insightful analysis. The visual-review plugin addresses this gap by enhancing the DSH Web interface with dual capabilities. It renders user-pasted or uploaded images directly within chat bubbles, ensuring visual clarity. Moreover, through intelligent interception and annotation conversion, the plugin enables any text model to "see" images and invoke the built-in visual_review tool for interpretation, returning descriptive text in Chinese that covers elements like text, objects, and scenes. Featuring a dual-engine approach, it prioritizes cloud-based OpenAI-compatible APIs for zero local dependencies, while automatically falling back to a local Qwen3-VL-8B model when unconfigured, safeguarding data privacy. This plugin is ideal for DSH Web users, including AI developers, researchers, and enthusiasts who regularly handle image content. It integrates image capabilities seamlessly without requiring a model switch, making it suitable for tasks ranging from quick local file analysis to customized queries (e.g., extracting chart trends). Users prioritizing data security benefit from the local engine, as all processing occurs on-machine with no external network requests.