dsh-sight

Running several models side by side in DeepSeek Harness exposes a handful of friction points: multimodal models cannot receive a pasted image natively and fall back to a lossy image-to-text pipeline; once a single image enters the conversation DSH blocks the switch back to a text-only model; freshly added third-party channels leave the model picker with no reasoning-effort slits; and the official Figma MCP gates design-to-code and canvas editing behind paid seats and per-tier rate caps, making a free account nearly unusable.

dsh-sight folds all of this into a single plugin. The image-paste feature sends a native image content block straight to the model, with a built-in dictionary covering Qwen-VL, GPT-4o/5, Gemini, Claude, Kimi, GLM, DeepSeek-VL, LLaVA and others—one click batch-declares them and the badge appears in the input frame. The session-image clear button removes images from the model-visible history without touching the rendered transcript, so you can drop back to a text-only endpoint instantly. Auto reasoning-effort injection walks the model-family table (GPT-5.x, o3/o4, Grok 4.x, DeepSeek V4, GLM-5.x, Kimi K3, Qwen 3, MiniMax), writes each family's official reasoningEfforts into the llm-pi-ai config, and leaves already-declared or natively-capable models untouched. The Figma MCP ships two independent blocks: one needs only a Personal Access Token and lets the model read a design via the REST API and emit React or CSS; the other adds a desktop-bridge plugin that unlocks figma_write, figma_read and figma_status so the model can draw, edit and screenshot on the canvas. Both paths run on the community-open figma-developer-mcp and figma-ui-mcp, work with a free Figma account, and carry no seat or tier prerequisite.

Built for developers who orchestrate multiple model channels in DSH and rely on multimodal conversations daily, as well as engineers and designers who want AI to participate directly on the Figma canvas. No build step, no OAuth ceremony, no paid tier—install, restart, and go.