"""
Prefetch Engine for dm20-protocol.
An intelligent prefetch system that anticipates likely DM responses by
monitoring game state context. In combat, pre-generates 2-3 narrative
variants (hit/miss/critical) using the campaign's main model, then uses
Haiku to select and refine the correct variant when the actual result
is known. This reduces response latency by 50%+.
Components:
- ContextObserver: Monitors game state and classifies context
- PrefetchCache: TTL-based cache for pre-generated variants
- PrefetchEngine: Pre-generation + refinement pipeline
- TokenUsage: Tracks token cost of prefetch operations
Usage:
from dm20_protocol.prefetch import PrefetchEngine, PrefetchCache
engine = PrefetchEngine(
main_model=campaign_llm,
refinement_model=haiku_llm,
intensity="conservative",
)
# Feed game state updates
engine.on_state_change(game_state)
# Resolve with actual result (fast path via cache)
narrative = await engine.resolve_with_actual(turn_id, actual_result)
# Get token usage summary
print(engine.get_token_summary())
"""
from .cache import PrefetchCache, PrefetchCacheStats, CacheEntry
from .observer import ContextObserver, GameContext, PlayerTurn
from .engine import PrefetchEngine, TokenUsage, LLMClient
__all__ = [
# Engine
"PrefetchEngine",
"TokenUsage",
"LLMClient",
# Cache
"PrefetchCache",
"PrefetchCacheStats",
"CacheEntry",
# Observer
"ContextObserver",
"GameContext",
"PlayerTurn",
]
