dsh-model-modality

The DSH runtime enforces strict image-input gates: message images, the read_image tool, attachment uploads, and sub-agent image prompts all depend on the inputModalities field in model metadata. When using a third-party relay or a custom provider, that field is often missing or mis-declared, so a genuinely multimodal model gets blocked with Model does not support image input, image tools stay locked, and the underlying stream call silently projects images into text.

This plugin offers two conversational tools: declare_model_multimodal writes the image-support flag directly into the correct provider setting (inputModalities for llm-deepseek, input or modelOverrides for llm-pi-ai) and validates immediately via resolveModelInfo; list_model_multimodal gives an at-a-glance view of declared input modalities across providers. The declaration persists in the settings of the owning provider, independent of the plugin lifecycle, so uninstalling the plugin does not erase it.

Built for DSH users who work with relay stations, custom API endpoints, or third-party model catalogs. If the model itself supports image input but DSH has not unlocked the gate, a single tool call is enough to open it, with no source edits or manual JSON fiddling required.