Crate rullama_call_policy 

Provider-layer resilience middleware for the rullama.

Wraps any rullama_core::Provider with composable decorators:

• RetryProvider — exponential backoff with jitter on transient failures.
• BudgetProvider — atomic token/USD/round caps with pre-flight rejection.
• CircuitBreakerProvider — half-open state machine, optional fallback.

Decorators wrap Arc<dyn Provider> and return Arc<dyn Provider>, so they compose freely. Typical stacking (outermost first):

CircuitBreaker → Retry → Budget → base Provider

Structs

BudgetConfig

Caps to enforce on a single BudgetGuard.

BudgetGuard

Shared atomic budget counters.

BudgetProvider

A Provider decorator that enforces a BudgetGuard around every call.

CacheKey

Key used to address a cached response.

CachedProvider

A Provider decorator that deduplicates identical chat() calls.

CachedResponse

Wire representation of a cached response, suitable for JSON storage.

CircuitBreakerConfig

Circuit-breaker configuration.

CircuitBreakerProvider

Circuit-breaker decorator around any Provider.

FailoverProvider

Decorator that walks through a list of providers, returning the first success.

HeuristicTokenizer

Cheap fallback: chars / 4. Matches the original approx_input_tokens behaviour; safe to use when no provider-specific tokenizer is wired up.

KeyedBudgetGuard

Per-caller budget tracking. Each unique K (e.g. user_id, tenant_id) gets its own private BudgetGuard with the same per-key caps.

MemoryCache

In-memory cache — the default backend. Cheap to Arc-clone.

RetryPolicy

Configuration for retry behavior.

RetryProvider

A Provider decorator that retries transient failures with exponential backoff and optional jitter.

Enums

CircuitState

Circuit-breaker state for a single provider/model key.

ErrorClass

Coarse error classification used by retry and circuit-breaker logic.

ResilienceError

Errors surfaced by resilience decorators.

Traits

CacheBackend

Pluggable storage backend behind CachedProvider.

Tokenizer

Counts tokens in a Message slice the way a particular provider would see them. Implementations are pure (no I/O) and cheap enough to call per pre-flight check.

Functions

cache_key_for

Compute a stable cache key from the inputs to a chat() call.

classify_error

Classify an anyhow::Error by string inspection.