Figure Spec

Core Properties

• Deterministic: identical FigureSpec JSON always produces identical SVG output (for a fixed renderer version + fonts)
• Editable: SVG output is plain-text, can be post-edited by hand or programmatically
• Validated: renderer enforces schema, rejects malformed specs with clear error messages
• Shape-aware: edge clipping works correctly for rect/rounded/circle/ellipse/diamond
• CJK support: multi-line labels with proper Chinese character width estimation
• No external API: runs fully local, no network, no API keys

Tool Location

tools/figure_renderer.py (from ARIS root). Invoke via:

python3 tools/figure_renderer.py render <spec.json> --output <out.svg>
python3 tools/figure_renderer.py validate <spec.json>
python3 tools/figure_renderer.py schema

Workflow

Step 1: Understand the Diagram Goal

From $ARGUMENTS (description or path to PAPER_PLAN.md / NARRATIVE_REPORT.md), identify:

• Purpose: architecture, workflow, pipeline, audit cascade, topology?
• Main entities: what are the boxes?
• Relationships: how do they connect? (uses, produces, calls, verifies, chains)
• Grouping: do entities cluster into named regions?
• Hierarchy vs network: stacked layers, left-to-right flow, or central hub?

Step 2: Draft the FigureSpec JSON

Canvas sizing guide:

• Single-column figure: ~500×350 px
• Two-column (full-width): ~900×500 px
• Tall topology: ~700×700 px

Start from a template based on the diagram type:

Architecture (stacked rows):

{
  "canvas": {"width": 900, "height": 520},
  "nodes": [
    {"id": "layer1_label", "label": "Layer 1", "x": 450, "y": 60, ...},
    {"id": "node_a", "label": "A", "x": 180, "y": 120, ...},
    {"id": "node_b", "label": "B", "x": 350, "y": 120, ...}
  ],
  "edges": [...],
  "groups": [
    {"label": "Layer 1", "node_ids": ["node_a", "node_b"], "fill": "#F0F9FF", "stroke": "#BAE6FD"}
  ]
}

Workflow (left-to-right chain):

{
  "canvas": {"width": 900, "height": 300},
  "nodes": [
    {"id": "step1", "label": "Step 1", "x": 100, "y": 150, "shape": "rounded"},
    {"id": "step2", "label": "Step 2", "x": 280, "y": 150, "shape": "rounded"}
  ],
  "edges": [
    {"from": "step1", "to": "step2", "label": "produces"}
  ]
}

Decision diamond:

{"id": "check", "label": "Passes?", "shape": "diamond", "x": 450, "y": 200}

Step 3: Render and Validate

# Validate first
python3 tools/figure_renderer.py validate /tmp/spec.json

# Render to SVG
python3 tools/figure_renderer.py render /tmp/spec.json --output figures/fig_arch.svg

# Convert to PDF for LaTeX inclusion
rsvg-convert -f pdf figures/fig_arch.svg -o figures/fig_arch.pdf

If validation fails, inspect the error (missing field, duplicate ID, overlap warning, invalid hex color) and fix the JSON.

Step 4: Visual Review

Open the SVG/PDF and check:

• No overlaps: nodes don't collide with each other or group boundaries
• Readability: font sizes are consistent, labels aren't clipped
• Edge clarity: arrows hit nodes at clean angles, labels near edges are legible
• Group alignment: background rectangles frame their members cleanly
• Color distinction: categories are visually distinct in both color and grayscale

If issues found, edit the JSON spec (never the generated SVG) and re-render.

Step 5: Iterate with Codex Review (Optional, for High-Stakes Figures)

For paper architecture figures, invoke cross-model review:

mcp__codex__codex:
  model: gpt-5.4
  config: {"model_reasoning_effort": "xhigh"}
  prompt: |
    Review this SVG figure for a technical paper (architecture / workflow diagram).

    Spec file: /path/to/spec.json
    Rendered: /path/to/fig.svg

    Evaluate:
    1. Clarity (C): can a reader understand the system from this figure alone?
    2. Readability (R): font sizes, label placement, visual hierarchy
    3. Semantic accuracy (S): do relationships match the described system?

    Score each axis 1-10 and list specific issues to fix.

Iterate until all three axes ≥ 7/10. The ARIS tech report figures went through 5 rounds of this loop to reach C:7/R:7/S:8.

Schema Quick Reference

Run python3 tools/figure_renderer.py schema for the authoritative schema.

Nodes

Edges

Groups

Rectangular background regions framing a set of nodes:

{"label": "Layer Name", "node_ids": ["a", "b", "c"], "fill": "#EFF6FF", "stroke": "#BFDBFE"}

Design Patterns

Pattern 1: Layered Architecture

Stack rows of related nodes, each row is a group, add inter-layer arrows with semantic labels (uses↓, produces↑, checks↓).

Pattern 2: Hub-and-Spoke

Central node (e.g., Executor), peripheral nodes (skills, tools), solid arrows for primary relations, dashed for feedback.

Pattern 3: Pipeline with Feedback

Left-to-right main flow, feedback arrows curve below with curve: true.

Pattern 4: Audit Cascade

Three-stage horizontal cascade with inputs feeding in from top, outputs exiting right, each stage in its own group.

Anti-Patterns

• Don't use groups as hierarchy: groups frame peer nodes, not containment
• Don't nest groups: renderer draws them as background rectangles; nested groups look like Russian dolls
• Don't cross-draw long diagonals: if an arrow crosses 3+ rows, rethink the layout
• Don't mix font sizes for same role: keep one size per node category

Output Contract

• SVG file in figures/ (vector, editable, hand-tweakable)
• Source FigureSpec JSON saved in figures/specs/ for reproducibility
• PDF version via rsvg-convert for LaTeX inclusion

Integration with Other Skills

• /paper-writing (Workflow 3): when illustration: figurespec (default for architecture figures), this skill handles Phase 2b
• /paper-figure: handles data plots; they complement each other (data + architecture = complete figure set)
• /paper-illustration: fallback for figures that need natural/qualitative style (method illustrations with photos, qualitative result grids)
• /mermaid-diagram: lighter alternative for simple flowcharts

Review Tracing

After each mcp__codex__codex or mcp__codex__codex-reply reviewer call, save the trace following shared-references/review-tracing.md. Use tools/save_trace.sh or write files directly to .aris/traces/<skill>/<date>_run<NN>/. Respect the --- trace: parameter (default: full).