lib/
├── index.ts          — Re-exports (Juce, isPlugin, cn, etc.)
├── utils.ts          — Shared utilities: cn() for class merging, isPlugin detection
├── native/           — Native bridge layer (nativeInvoke, SliderState, etc.)
├── ai/               — AI copilot (multi-provider, intent-driven)
└── theme/            — Dark/light theme CSS and system detection

ai/
├── types.ts           — Shared interfaces (LLMProvider, Intent, SpeedProfile, ToolCall)
├── index.ts           — Barrel exports for AI module
├── agents/            — Specialized AI agents
│   ├── base.ts        — Base agent class with shared behavior
│   ├── composer.ts    — Composition agent: writes .bird patterns, arrangements, and melodies
│   ├── mix-engineer.ts— Mixing agent: adjusts levels, panning, EQ, compression
│   ├── sound-designer.ts — Sound design agent: tweaks synth parameters, creates presets
│   └── audio-creator.ts — Audio creation agent: generates audio via Lyria
├── core/              — Core AI infrastructure
│   ├── copilot.ts     — Main orchestrator: shortcuts → intent → profile → provider
│   ├── router.ts      — Intent classifier (regex/keyword → Intent → SpeedProfile)
│   ├── prompts.ts     — Modular system prompt builder (conditional sections)
│   └── compactor.ts   — Conversation compaction for long threads
├── tools/             — Tool declarations and execution
├── bird/              — .bird file parsing and validation utilities
├── cache/             — Plugin parameter cache (pre-bake to skip round-trips)
├── docs/              — AI documentation and brainstorm notes
└── providers/
    ├── index.ts       — Provider registry & factory
    └── gemini.ts      — Gemini provider (LLMProvider implementation)

User message
  → shortcuts.ts     (instant: /bpm 120 → done, no API call)
  → router.ts        (classify intent: mixer/param/bird_edit/chat)
  → SpeedProfile     (model tier + thinking + tool subset)
  → prompts.ts       (build tailored prompt — 2KB for simple, 8KB for composition)
  → tools.ts         (select 1-3 tools instead of 9)
  → provider.stream  (Gemini/OpenAI/local with resolved config)
  → tool-executor.ts (handle function calls → native bridge)
  → result

interface LLMProvider {
  name: string;
  models: Record<ModelTier, string>;
  streamMessage(tier: ModelTier, request: StreamRequest): Promise<LLMResponse>;
  createCache?(prompt: string, tools?: ToolDeclaration[]): Promise<string | null>;
}