dsh-plugin-modality-fallback

Running multimodal conversations in dsh has a small but recurring friction point: the entire session is locked to a single model, and if that model does not declare image-input support, the moment an image appears in the conversation history the built-in read_image tool refuses outright while ApiProxy blocks the outbound send. The usual workaround is switching the entire session to an image-capable model, which means surrendering the specific model chosen for code generation, long-form reasoning, or any other non-visual task in the same conversation.

dsh-plugin-modality-fallback takes a narrower approach: leave the session alone, redirect only the single request that actually needs the missing modality. The plugin hooks into the dsh request waterfall and, before a message is dispatched, inspects the session-derived history for content that exceeds the resolved model declared input modalities (currently images). When a mismatch is found and a fallback route is configured, it swaps provider and model for that one request only. Once the request completes, the session original model selection is untouched; subsequent requests resolve normally as soon as the image scrolls out of context or the user changes models. No core deepseek-harness code is modified; this is a standard Cordis plugin that loads alongside the rest of the composition.

If the dsh workflow mixes visual-understanding tasks with work that benefits from a specific model reasoning or coding strengths, and bouncing between two models every time an image shows up feels like unnecessary overhead, this plugin removes that friction. Configuration is a single modality-to-model mapping in cordis.patch.yml; the routing hook handles the rest automatically.