dsh-tool-accurate-vision

When a text-only agent needs to reason about an image, the bottleneck is clear: it can only rely on a vague verbal description from a vision model to infer spatial relationships—how far apart two objects are, which label sits where, or what the true pixel-level distance looks like. dsh-tool-accurate-vision closes that gap by calling any OpenAI-compatible vision model, extracting bounding boxes normalised to the 0–1000 range, and formatting them into a structured text block the next model turn can consume. Object positions, layout, and OCR text become computable, verifiable numbers rather than impressions to trust blindly.

Every call also emits a self-contained SVG that overlays all bounding boxes and labels on the original image, so a developer can eyeball the annotations instead of accepting them on faith. The package ships with a ready-made edge-to-edge distance helper that pairs facing edges per axis and returns the true shortest gap, keeping downstream agents from accidentally measuring wrong edges.

It is built for developers constructing text-only agents that need precise spatial reasoning from images—chart and diagram interpretation, layout analysis, object-distance queries, or any workflow where geometric accuracy must be provable, not approximate.