VisorQA

VisorQA is an enterprise-grade autonomous visual QA and self-healing frontend agent that recovers 25%+ of weekly engineering sprint capacity by eliminating manual UI regression reviews. Traditional CI/CD test runners pass green builds as long as DOM nodes exist—leaving teams vulnerable to invisible z-index collisions, sticky banner occlusions, and responsive layout collapses. VisorQA closes this loop entirely through a deterministic 5-Node Finite State Machine (FSM): 1. SENSE: Renders pull requests across dual viewports (Mobile 375×812 and Desktop 1440×900) inside isolated sandboxes (Modal / E2B) and prompts Meta SAM 3.1 to generate pixel-accurate binary visibility masks. 2. TRIAGE: Extracts the Lowest Common Ancestor (LCA) DOM subtree and prompts Meta Muse Spark to isolate the 1–2 responsible layout files, reducing prompt tokens by 88.5% and shielding shared UI primitives (src/components/ui/*). 3. DIAGNOSE: Meta Muse Wonder performs multimodal reasoning over the red collision-mask overlay and float64 discrepancy metrics (Visibility Ratio Vr, Collision IoU, Horizontal Bleed Δx) to produce a declarative RepairSpec JSON without touching code. 4. PATCH: Meta Muse Code synthesizes a surgical, atomic Unified Diff enforcing strict anti-hack CSS policies (no !important, no arbitrary z-index escalation). 5. VERIFY: Injects the patch into the sandbox via Hot Module Replacement (HMR) and performs dual-viewport re-sensing. If Composite Health H ≥ 0.98, VisorQA auto-commits the signed fix directly to the GitHub PR. If unresolved after a 3-strike circuit, it safely posts the highest-scoring patch as an advisory PR review.