mirror of
https://github.com/DayuanJiang/next-ai-draw-io.git
synced 2026-10-11 03:59:58 +08:00
dayuan.jiang
1c51a9f31f
The model now writes a shape as one self-closing mxCell with x, y, w and h and an edge as one with source and target; compact-cells.ts fills in vertex, edge, parent="1" and the mxGeometry element after validation. The long form stays accepted, so waypoints, edge labels and anything else keep their mxGeometry. The diagram shown to the model (the system context, the edit error texts and the MCP get_diagram result) is folded into the same notation, so the model reads what it is asked to write; the canvas and the saved file stay standard draw.io XML. Folding then expanding the 652 valid production diagrams gives the same cells back for every one of them. The auto-fix gains one step for the slip this notation invites: a missing closing quote before the self-closing slash (h="112/>). Prompts and the shared examples use the compact form; the single waypoint example keeps the long form on purpose. Measured on the five start-screen examples, same prompt, against the named-styles build: gpt-6-luna 19% fewer output tokens, Opus 5.5 26% fewer (35% and 41% against main). Every shape was written compactly on the first try; edits used it too and all applied.