Architecture

Layers

UI (screens / widgets)
        │
   ModelOrchestrator  ←── MemoryService (RAG), MCP, Tasks, Notes
        │
   flutter_gemma / LiteRT-LM
        │
   Platform channels → Android (tools, screenshot, assistant key)

Key Dart services

Service	Role
ModelOrchestrator	Model load/unload, chat stream, tool loop, idle release
ModelManager	Download, disk register, prefs sync
MemoryService	Local RAG / conversation memory
ToolExecutorService	MethodChannel + progress EventChannel
ScreenshotService	MediaProjection frames via dedicated channel
McpService / McpClient	External MCP (HTTP/SSE, stdio)
ParallelSessionManager	Multi-chat (capped on Android)
PlatformAdaptationService	Platform capability flags + RAM warnings

Inference pipeline

1. User sends message (optional image / attachments)
2. RAG context retrieved
3. Model selected (ModelSelector) or override from UI
4. Engine loaded with supportImage when the model has vision
5. Stream tokens via generateChatResponseAsync
6. Tool calls executed; results fed back (multi-round)
7. Idle timer / lifecycle may unload the model only after streaming ends

Native Android

Component	Purpose
MainActivity	Flutter host, channels, MediaProjection result
AssistantActivity	System assistant entry + screenshot handoff
ScreenCaptureHelper	MediaProjection, on-demand JPEG frames
MediaProjectionService	FGS for API 34+ projection
ToolExecutor	Alarms, apps, SMS, weather, etc.

Stability rules (critical)

• Never close() the LiteRT engine while _isStreaming is true (SIGABRT)
• Vision models must load with supportImage: true even for text-only first turns
• Screenshot bytes travel via ScreenshotService, not large MethodChannel maps
• Prefer SmolLM / Gemma 3 1B on ≤6 GB RAM devices for long debug sessions