v2.2.2: Ollama Local AI Provider

What's New

Ollama Local AI Provider

• Run AI models locally on the Jetson using Ollama — no API key needed
• Preset models: Llama 3.2 3B and Qwen2.5 3B, optimized for 8GB RAM
• Search and download any Ollama model filtered for 8GB devices
• Delete installed models with trash icon button

Memory Optimizations for 8GB Jetson

• Cap Ollama context window to 16K via openclaw.json (models.mode: replace)
• Q8_0 KV cache quantization halves KV memory (via systemd override)
• Flash attention enabled by default
• Single model / single parallel to prevent OOM
• Disable snapd service to free ~22MB RAM

Other Improvements

• Factory reset now deletes all Ollama models
• Switching from Ollama to cloud provider resets models.mode to merge
• Installer applies all Ollama optimizations automatically

Full Changelog: v2.2.1...v2.2.2

v2.2.1

What's New

• WiFi status fix: Use wifi_configured from config to properly track WiFi setup status
• WiFi skip handling: Mark WiFi as configured when user skips in setup wizard
• Hotspot improvements: Disable hotspot configuration on WiFi connect, improved warning text
• Hotspot validation: Fix validation when hotspot is disabled
• Text escaping: Fix apostrophe escaping in UI strings
• MAXN power mode: Improved regex matching for power mode detection
• Disk fallback: Use constant for disk space fallback value
• OpenClaw updates: Always install latest version from npm

v2.2.0

What's New

• OAuth improvements: Auto-close browser tabs on successful connection, mobile fallback messaging
• OpenAI device code flow: Added to dashboard (parity with setup wizard)
• AbortController support: All AI provider fetches in dashboard are now cancellable
• UI polish: Spinner icons on all Save/Connecting buttons, auto-collapse sections after save
• WiFi dashboard: Replaced scan list with simple SSID/password inputs
• WiFi persistence: Fix WiFi not reconnecting after reboot (tries all saved profiles before AP mode)
• OpenClaw updates: Installer now fetches latest version from npm instead of hardcoded pin
• Google OAuth: Credentials loaded from .env, set automatically by installer
• Performance mode: Persistent MAXN power mode via systemd service
• Footer: Added Discord and ClawBox logo buttons
• Code quality: Shared OAuth utilities, isGoogleConfigured guard, polling error logging

v2.1.0

What's New

Google Gemini Subscription Support

• Connect your Google One AI Premium subscription via OAuth
• Full Cloud Code Assist API integration with automatic project onboarding

Factory Reset Improvements

• Saved WiFi networks are now properly cleared during factory reset — device returns to AP mode cleanly
• Fixed dashboard showing WiFi as "connected" when only the setup AP is active
• Proper error reporting when file cleanup fails

System Update UI

• Shows both ClawBox and OpenClaw versions with current → target indicators

Reliability & Security

• OAuth credentials centralized and secured (Google secrets via environment variables)
• Dynamic UID/GID detection instead of hardcoded UID 1000
• Gateway restart failures now surface to the user instead of being silently ignored
• Device code auth validates session before polling
• Google OAuth onboarding has a 2-minute timeout to prevent infinite loops
• Insecure control UI auth changed to explicit opt-in
• Subscription tab hidden for providers without backend OAuth configured
• Version cache properly cleared on errors

CI/CD

• Test workflow rebuilt from scratch — runs on every push and PR
• Properly captures test exit codes through pipe