dsh-image-vision

When the main model in DeepSeek Harness runs through a text-only route, images pasted into chat or embedded in Feishu/Lark documents are blocked by the input-modality gate, and the model only ever sees plain text. dsh-image-vision solves exactly that: it replaces no adapter, disables no built-in plugin, and simply wraps llm.resolveModelInfo so images pass the admission check, then uses the official agent/pre-step waterfall to call a configurable vision model that transcribes each image into text. The text-only main model gains full image understanding with zero host code changes.

Core capabilities span three scenarios: pasting or dropping an image directly in chat, where the vision model produces a full description covering people, scenes, tables, charts, and verbatim text rather than simple OCR, with a user-overridable prompt; Feishu/Lark document images produced by lark_read_doc and similar tools, automatically downloaded via lark-cli and injected back as described tool output; and the describe_image model tool for describing a local image file path. Vision models can be selected from the ModelScope free community API or SiliconFlow, with routes auto-discovered from settings.yaml or explicitly pinned in the plugin config. Per-message image caps, Feishu image caps, and the description prompt are all tunable; transient vision failures return a placeholder note instead of aborting the turn.

Built for anyone whose main model runs a text-only route — pi-ai forwarded DeepSeek, or any backend that does not declare an image input modality — yet still wants the model to read images in chat and agent workflows. The plugin is plain JS ESM with no build step and minimal peer dependencies, fitting the everything-is-a-plugin philosophy of DeepSeek Harness.