Xenon V2: The Next Generation of Autonomous Mobile Infrastructure
We're thrilled to announce Xenon V2 — a ground-up evolution of the autonomous mobile infrastructure platform. This release brings AI-powered diagnostics, self-healing locators, visual intelligence, and enterprise-grade operational features.
This post documents Xenon V2 as it shipped on 2026-02-17 and is preserved as a historical record. Some details have evolved since:
- Tier 4 of the self-healing engine no longer uses Florence-2; it now runs OCR (Tesseract.js) with the configured AI provider as the visual fallback. See Self-Healing and Omni-Vision.
- Hub↔node communication is HTTP REST + Socket.IO (not gRPC). See Architecture and Remote Execution.
- Hub-node authentication uses per-node
(accessKey, token)pair auth (not a shared secret). See Security.
AI Root-Cause Diagnosis
Xenon V2 introduces multimodal AI failure analysis. When a session fails, the AI engine reasons over screenshots, device logs, and command sequences to produce actionable root-cause reports — no manual triage required.
Supported LLM Providers:
| Provider | Default Model | Multimodal |
|---|---|---|
| Google Gemini | gemini-3-flash-preview | [Yes] |
| OpenAI | gpt-4o | [Yes] |
| Anthropic | claude-3-5-sonnet | [Yes] |
| Ollama | llama3 | [No] (text-only) |
Configure via environment variables (GEMINI_API_KEY, OPENAI_API_KEY, ANTHROPIC_API_KEY) or the Dashboard Settings UI.
🩺 5-Tier Self-Healing Engine
When a locator breaks, Xenon's HealingOrchestrator cascades through five increasingly powerful strategies:
- ResilioTree — Structural tree-diff matching (free, instant)
- FuzzyXML — Attribute similarity scoring (free, instant)
- OCR — Text-based visual matching (free, local)
- Visual AI — Florence-2 element detection (local GPU)
- LLM Deep Healing — Full reasoning with GPT/Claude (API call)
Successful healings are autonomously learned via the Etalon Service, preventing the same break from recurring.
Omni-Vision (Visual Intelligence)
Powered by Florence-2, Omni-Vision enables:
- Coordinate-free element detection — Tap elements by description, not XPath
- Visual assertions — Verify screen state without brittle selectors
- AI-guided interaction — Natural language commands for exploratory testing
Network Conditioning
Simulate real-world network conditions during test execution:
| Profile | Latency | Download | Upload |
|---|---|---|---|
| 4G | 20ms | 15 Mbps | 7.5 Mbps |
| 3G | 100ms | 2 Mbps | 1 Mbps |
| Edge | 400ms | 250 Kbps | 150 Kbps |
| Offline | — | 0 | 0 |
Platform-specific: Android (ADB svc), iOS Simulator (simctl network), plus proxy-level latency injection.
Webhook Notifications
Configure real-time alerts via Slack or generic HTTP webhooks:
- Events:
device_offline,session_failed,device_new - Custom templates: Use
{{variable}}syntax for payload customization - Managed entirely through the Dashboard Settings UI
Deployment Flexibility
Three supported topologies:
- Standalone — Single node with SQLite
- Hub-Node — Distributed grid with PostgreSQL + gRPC
- Cloud — BrowserStack, SauceLabs, LambdaTest, HeadSpin, pCloudy
Dashboard Settings
The V2 dashboard includes a full settings panel for:
- AI Configuration — Provider selection, model, API key testing
- Maintenance — Retention policies, cleanup schedules
- Infrastructure — Device types, session limits, timeouts
Additional Improvements
- Session Heartbeat Service — Automatic idle session detection and release
- Request Logging — Full HTTP request/response audit trail
- Metrics & Profiling — CPU, memory, and thermal monitoring
- Resource Isolation — Per-session resource boundaries
- Video Pipeline — MJPEG proxy with reconnection handling
- OpenTelemetry — Distributed tracing via OTLP export
Upgrading
appium plugin update @xenon-device-management/xenon
See the Setup Guide and Configuration Reference for details.