Module routing 

Stage: Model Routing Intelligence

Responsibility

Automatically route prompts to the optimal worker backend based on complexity scoring. Simple prompts go to the free local model (llama.cpp / Mistral); complex prompts go directly to the paid cloud model (Claude API). Prompts in the middle zone try local first with cloud fallback.

Guarantees

• Deterministic: the same prompt text always produces the same complexity score and initial routing decision.
• Thread-safe: all state (CostTracker, adaptive threshold) uses atomics or interior locking — safe under concurrent pipeline access.
• Non-blocking: route() is a pure O(n) scan over the prompt with no I/O, allocation-free hot path.
• Bounded: adaptive threshold adjustment is clamped between local_threshold and 1.0, so it can never oscillate unboundedly.

NOT Responsible For

• Actually calling the workers (that belongs to stages / worker)
• Cross-node routing (future distributed module)
• Semantic understanding of prompt quality (heuristic-only)

Re-exports

pub use arbitrage::ArbitrageEngine;

pub use arbitrage::ProviderProfile;

pub use arbitrage::ProviderSnapshot;

pub use cascade::CascadeEntry;

pub use cascade::CascadeError;

pub use cascade::CascadeResult;

pub use cascade::CircuitBreakerConfig;

pub use cascade::ProviderCascade;

pub use cascade::CascadeFailover;

pub use cascade::FailoverExhausted;

pub use cascade::FailoverResult;

pub use cascade::FailoverTier;

pub use config::RoutingConfig;

pub use cost_tracker::CostSnapshot;

pub use cost_tracker::CostTracker;

pub use cost_tracker::SessionBudgetTracker;

pub use health::ProviderHealthBuilder;

pub use health::ProviderHealthSnapshot;

pub use health::ProviderMetrics;

pub use pool_sizer::PoolSizer;

pub use pool_sizer::PoolSizerConfig;

pub use pool_sizer::ScaleAction;

pub use pool_sizer::ScaleRecommendation;

pub use router::ModelRouter;

pub use router::RoutingDecision;

pub use scorer::ComplexityScorer;

pub use scorer::ScoreBreakdown;

pub use semantic::ClassificationResult;

pub use semantic::RoutingCategory;

pub use semantic::SemanticRouter;

pub use semantic::SemanticRouterConfig;

Modules

arbitrage

Provider Arbitrage Engine

cascade

Multi-provider cascade fallback.

config

Routing configuration types.

cost_tracker

Cost tracking and savings computation.

health

Provider health dashboard snapshot.

pool_sizer

Adaptive Worker Pool Sizer

router

Model routing logic.

scorer

Prompt complexity scoring.

semantic

Semantic prompt router using keyword-based heuristics.