jolly-dsh-vision

A text-only LLM, no matter how capable, is blind by default. Hand it a screenshot, a wiring diagram, or a handwritten note and it can only guess—a near-useless answer in any serious workflow. jolly-dsh-vision closes that gap by pairing a text brain (DeepSeek-V4-Pro by default) with a vision eye (DeepSeek-V4-Flash-Vision-Exp) so the brain receives a structured evidence JSON spanning summary, OCR text, layout, semantics, visual description, and a confidence-uncertainty field, and answers from evidence instead of speculation. Two visual pathways are provided: a vision tool that accepts a local path or URL and returns that evidence block on demand, and a (ds vision) twin entry in the model picker that unlocks paste-and-drag image input directly in the chat box—the plugin silently converts the image to evidence text before the request leaves, so the brain never has to call the tool manually. Results are cached by content address so each image is converted only once per session, and a failed conversion degrades to a placeholder string rather than breaking the conversation. The plugin ships with zero runtime dependencies: no commander, no undici, no build step. Tool definitions use raw JSON-Schema registration and all message/stream/adapter interfaces are duck-typed, keeping the supply-chain attack surface at a minimum. On the security side, SSRF protection rejects private, loopback, link-local, and cloud-metadata addresses by default (with a configurable escape hatch), API credentials are resolved at call time by the harness credential seam so plugin code never reads process.env or builds an Authorization header, and OCR text extracted from images is framed as evidence-to-cite to mitigate prompt-injection risk. It is well suited for developers and heavy chat users who want their DeepSeek conversations to read images reliably without deploying a separate vision-inference service or managing multi-engine keys themselves, and is a natural fit for OCR forensics, layout analysis, chart-semantics workflows, or any stack where it needs to coexist cleanly with an existing modlens plugin.