dsh-vision-sidecar

Text-only models in DeepSeek Harness inherently lack visual perception, and traditional solutions either require replacing the entire reasoning engine or deploying local vision models with GPU resources and API keys. The dsh-vision-sidecar elegantly sidesteps these limitations by acting as a lightweight plugin that injects visual capabilities into existing text models. Without altering the core reasoning engine, it forwards images to a free LLM7.io endpoint or a custom vision API, generates detailed descriptions, and persists them as durable session messages for the reasoning model to use. This allows text models to indirectly 'see' image content, making the process transparent to users with minimal resource overhead.

The plugin's core strengths lie in its 'no-key' default setup and durable design. Out of the box, it leverages LLM7.io's anonymous vision service, eliminating the need for API keys or local environments, which significantly lowers the barrier to entry. The generated visual descriptions are saved as formal session messages in DSH's persistent layer, ensuring that previous image analyses can be accurately replayed even after process restarts, avoiding repeated calls and quota waste. Additionally, the plugin supports flexible configuration, allowing users to easily swap the underlying vision provider or reasoning target model to accommodate different privacy needs or performance preferences.

This plugin is especially suited for developers, researchers, and everyday users looking to quickly expand DeepSeek Harness's multimodal capabilities without complex model migrations or infrastructure changes. If you're building applications that process mixed text and image inputs but are constrained by the text-only interface of reasoning models, dsh-vision-sidecar offers a clean, low-cost upgrade path. It makes visual understanding as straightforward as calling a regular API while maintaining system stability and maintainability.