// Standard usage
let builder = PromptBuilder::new(config);
let prompt = builder.build_system_prompt(&tools);

// With soul identity
let prompt = builder.build_system_prompt_with_soul(&tools, &soul, profile);

// Sub-agent usage (replaces old prompt_sections::resolve())
let prompt = builder.build_for_agent(&agent_def, &tools, &soul);

// Mode + budget control
let result = builder.build_with_budget(&tools, &soul, profile, PromptMode::Compact, &budget);

pipeline.invalidate("soul");    // Invalidate one layer by name
pipeline.invalidate_all();      // Clear all cached sections
pipeline.cache_stats();         // CacheStats { hits, misses, entries }

pub enum ThinkingLevel {
    Off,        // No extended thinking
    Minimal,    // budget_tokens: 1024
    Low,        // budget_tokens: 2048
    Medium,     // budget_tokens: 4096 (default)
    High,       // budget_tokens: 8192
    XHigh,      // budget_tokens: 16384
}

User requests: thinking = High
    │
    ├─▶ Claude Opus → ✓ Native extended thinking
    │
    ├─▶ GPT-4o → ✗ No support → Fallback to o1
    │
    └─▶ Gemini → ✗ No support → Use thinkingPreface prompt

LLM Response Stream
    │
    ▼
┌─────────────────────────────────────────┐
│ BlockStateManager                        │
│   • Track current block type             │
│   • Detect block boundaries              │
└─────────────────────────────────────────┘
    │
    ▼
┌─────────────────────────────────────────┐
│ BlockReplyChunker                        │
│   • Split into semantic chunks           │
│   • Handle code blocks, lists, etc.      │
└─────────────────────────────────────────┘
    │
    ▼
┌─────────────────────────────────────────┐
│ BlockCoalescer                           │
│   • Merge small chunks                   │
│   • Emit complete blocks                 │
└─────────────────────────────────────────┘
    │
    ▼
Event: StreamChunk { content, block_type }