Skip to content

Agent skill that turns natural language, code, Terraform/K8s, SQL, OpenAPI, AsyncAPI, Protobuf and GraphQL sources into editable, tested draw.io architecture diagrams: incremental sync, multi-view projection, drift diff, CI architecture tests, whiteboard derasterize, interactive HTML/PPTX/Mermaid exports.

feat: add image extraction algorithm to style-extraction reference

Agents365-ai30c846d157ae
diff --git a/references/style-extraction.md b/references/style-extraction.md
index cbd536a..7ff6d53 100644
--- a/references/style-extraction.md
+++ b/references/style-extraction.md
@@ -165,3 +165,86 @@ Input: a `.drawio` file path. Output: candidate preset JSON. Deterministic, no L
| Non-standard `shape=` keywords (e.g., `shape=mxgraph.aws4.*`) | These do not match the Step 4 precedence ladder, so the vertex falls through to `rounded=0` for shape-class purposes. Iconography is lost; color, label, and edge style are still captured. Role inference still runs via the label-regex rules. Summary notes: *"Non-standard shape library detected — iconography not preserved in preset (color and label captured)."* |
| Non-English labels | The English-keyword regexes in step 4 will mostly miss; most vertices collapse to `service`. Palette/shapes/font/edges still captured correctly (they don't depend on label text). `confidence` stays `"high"`. Summary notes: *"Role labels not in English — `service`/`database`/`decision`/`container`/`external` inferred from shape class; other roles not mapped."* |
| File has no `<mxCell vertex="1">` at all | Stop. Refuse to save. Message: *"Nothing to learn from — source file has no shapes."* |
+
+## Image extraction path
+
+Input: path to a PNG/JPG (or any vision-readable image format). Output: candidate preset JSON. Inference-based; `confidence: "medium"` at best.
+
+**Prerequisite:** the agent's vision capability must be available (same mechanism the main workflow's self-check uses). If vision is not available, stop and tell the user:
+*"Image-based learning needs a vision-enabled model (Claude Sonnet or Opus). Re-run on such a model, or provide the `.drawio` source file instead."*
+
+### Steps
+
+1. **Read the image.** Use the agent's vision input — the same path the main workflow's step 5 uses to read exported PNGs during self-check.
+
+2. **Extract palette by visual inspection.** Identify distinct fill-color regions on shape bodies.
+
+ For each distinct fill:
+ - `fillColor` — quantize: round each RGB channel to the nearest multiple of 16, harmonize lightness to ~85% (drawio-standard pastel look). Emit as `#RRGGBB`.
+ - `strokeColor` — read the matching border. If unreadable, derive from fill by darkening ~25% (match HSL, drop L by 0.25).
+
+ Map each `(fillColor, strokeColor)` pair to a named slot by hue of the fill:
+ - blue family → `primary`
+ - green family → `success`
+ - yellow family → `warning`
+ - orange family → `accent`
+ - red/pink family → `danger`
+ - grey family → `neutral`
+ - purple family → `secondary`
+
+ Two distinct colors in the same hue family: keep the more frequently-used one in its canonical slot; the other spills to the nearest empty slot (adjacent hue).
+
+3. **Extract shape vocabulary.** Classify every visible shape by silhouette:
+ - rounded rectangle → `rounded=1`
+ - sharp rectangle → `rounded=0`
+ - circle / oval → `ellipse`
+ - diamond → `rhombus`
+ - cylinder (rectangle with curved top/bottom) → `shape=cylinder3`
+ - titled container (header bar + nested children inside) → `swimlane;startSize=30`
+ - dashed-bordered rectangle → `rounded=1;dashed=1`
+
+ Role assignment uses the **same label-text + shape rules as the XML path step 4**. Visible labels are read via vision.
+
+4. **Extract fonts.** Best-effort. Distinguishable categories:
+ - clearly serif → `fontFamily: "Georgia"`
+ - clearly monospaced → `fontFamily: "Courier New"`
+ - otherwise → `fontFamily: "Helvetica"`
+
+ Size by relative appearance:
+ - small → `fontSize: 11`
+ - medium → `fontSize: 12`
+ - large → `fontSize: 14`
+
+ If titles/container headers are distinctly larger or bolder → set `titleFontSize` accordingly and `titleBold: true`.
+
+5. **Extract edge defaults.**
+ - Right-angle orthogonal arrows → `edges.style = "edgeStyle=orthogonalEdgeStyle;rounded=1;orthogonalLoop=1;jettySize=auto;html=1"`.
+ - Curved arrows → append `;curved=1` to `edges.style`.
+ - Filled triangle arrowheads → `edges.arrow = "endArrow=classic;endFill=1"`.
+ - Open V-shaped arrowheads → `edges.arrow = "endArrow=open;endFill=0"`.
+ - Any dashed arrows near labels like "optional", "async", "fallback", "secondary" → add those label tokens to `edges.dashedFor`.
+
+6. **Extract extras.**
+ - Visibly hand-drawn / rough / sketch look (wavy strokes, uneven fills) → `extras.sketch = true`.
+ - Heavy strokes (clearly >1.5× normal) → `extras.globalStrokeWidth = 2`.
+ - Otherwise default: `extras = { "sketch": false, "globalStrokeWidth": 1 }`.
+
+7. **Set provenance and confidence.**
+ ```json
+ {
+ "source": { "type": "image", "path": "<input absolute path>", "extracted_at": "YYYY-MM-DD" },
+ "confidence": "medium"
+ }
+ ```
+ Adjustments:
+ - <3 distinct shapes identifiable → `confidence: "low"`.
+ - Very clean image, ≥5 roles confidently mapped, all seven palette slots filled → may upgrade to `"high"`.
+
+### Image edge cases
+
+| Situation | Behavior |
+|---|---|
+| Vision unavailable | Stop as described above — do not fall back to guessing. |
+| Image has <3 identifiable shapes | Continue; mark `confidence: "low"`; summary explicitly warns the user that the preset is a loose approximation. |
+| Image has no visible labels | Role assignment collapses to shape-class only: cylinders → `database`, diamonds → `decision`, swimlanes → `container`, everything else → `service`. Palette/font/edges still captured. Summary notes: *"No labels readable — semantic roles beyond shape-class not inferred."* |
+| Two palette slots would land in the same hue family | Keep the more frequent one in its canonical slot; spill the other to the adjacent empty slot (rule in step 2). |