Skip to content
Vox: The AI-Native Programming Language
Unverified: This page has not yet been anchored to automated code tests.

GUI Visual AI Adversarial Review

GUI Visual AI Adversarial Review

Section titled “GUI Visual AI Adversarial Review”

The rule

Section titled “The rule”

Every vox-gui surface is screenshotted on the post-merge / full-ci Playwright sweep. A surface is AI-reviewed against the design principles ONLY when its screenshot content hash changes (or it is new). This is advisory — it never gates CI; it warns and collects reports.

How it works

Section titled “How it works”
  • Discover: SURFACE_REGISTRY (SSOT) → every viewKey with tier != 'none'.
  • Capture: e2e/visual-review.spec.ts shoots hi-DPI PNGs (2x), emits manifest.json with per-surface sha256 + timing.
  • Cache: contracts/reports/gui-visual-review/cache.v1.json maps viewKey → screenshot_sha256 + verdict. Same hash → reuse verdict, no AI call. Changed hash → warn + re-review.
  • Review: gui-visual-review (binary in vox-orchestrator-mcp) sends the PNG to an OpenRouter vision model and parses a JSON verdict against the design-principles rubric.
  • Report: versioned contracts/reports/gui-visual-review/<date>.json + ledger.jsonl (trend/spike). Spike = run review-time > spike_factor × trailing median → warning only.

Model (pluggable, not pinned)

Section titled “Model (pluggable, not pinned)”

contracts/orchestration/visual-review.config.v1.json lists vision models in priority order (default google/gemini-3-flash-preview; anthropic/claude-opus-4.8 for escalation). The resolver picks the first config entry the model registry marks supports_vision; per-surface model is recorded for future A/B learning.

What it never does

Section titled “What it never does”
  • Never fails CI (binary exits 0; CI step continue-on-error).
  • Never gates on appearance change — it warns and re-reviews; it does not pixel-diff-gate.