<div align="center">

# Draw.io Reconstruction Skill

[English](README.md) | [中文](README.zh-CN.md)

**A Codex skill for reconstructing diagram images into editable Draw.io files, with examples strongly recommended to be reproduced using Codex + GPT-5.5 xhigh.**

[](https://arxiv.org/abs/2605.15677) [](https://huggingface.co/datasets/sxy1620348809/VCG-Bench)

</div>

This repository contains a Codex skill and helper scripts for converting reference diagram images into editable `.drawio` files. It is the practical reconstruction workflow used in the VCG-Bench release examples: an agent inspects a reference image, creates a visible-element inventory, rebuilds text and structure with Draw.io primitives, uses crops or SVG where appropriate, exports a PNG preview, and verifies the result. The bundled examples are packaged so others can reproduce them from the original PNG inputs. For faithful reproduction of the displayed examples, we strongly recommend using Codex + GPT-5.5 xhigh mode; weaker models or lower reasoning modes may not match the same visual fidelity.

The workflow uses two lightweight, independently reviewed repair loops. One clean Icon Producer prepares the icon set and a separate Icon Reviewer accepts or rejects it. Each FIX starts a fresh repair Producer on the current artifacts, followed by a fresh Reviewer. The complete reconstruction follows the same pattern. Review evidence is split into small `icons-review*.png` and `placement-review*.png` shards (at most eight icons each), with literal 1:1/2x views and surrounding source/final context. Reviewers must return one verdict per icon; Producers never accept their own work.

The companion benchmark repository is released at https://github.com/sxy1499894281/VCG-Bench.

## Recommended Reproduction Configuration

The example reconstructions in this repository are best reproduced