dsh-autovision

In DeepSeek Harness, only models that explicitly declare image input can receive pictures. Pure-text models like DeepSeek and GLM either silently fail or throw an error the moment a session carries an image. dsh-autovision registers a transparent twin adapter for every text-only model at runtime, routing image-bearing requests to a multimodal model you configure yourself, transcribing them to text, and feeding the result back into the original text model. No model switching, no relay, no hardcoded credentials.

The workflow is nearly frictionless: paste an image into any session and send, and the text model receives a faithful transcription while the original picture stays visible as a thumbnail. In agent mode the model can proactively call the autovision_read_image tool to read a local image file with its own per-task instruction, such as transcribing every word or describing a UI state. The model picker stays clean because the twin returns an empty list, so no extra entries clutter the selector.

It is built for users who rely on text-only models in DeepSeek Harness but occasionally want to drop a screenshot and let the model recognize its content. Pick any multimodal model you already have as the transcription engine in the plugin settings, and everything else happens automatically. The plugin is purely additive to the dsh core, so upgrades do not break your setup.