Crate oharness_budget 

Budget middleware and shipped BudgetHandle implementations (plan §10).

The BudgetHandle trait itself lives in oharness-core; this crate provides the concrete implementations (TokenBudget, StepBudget, CostBudget, TimeBudget, CompositeBudget), a PricingTable for cost calculations, and BudgetMiddleware — an Llm wrapper that pre-checks, post-consumes on complete(), and observes Chunk::Usage events on stream().

BudgetMiddleware implements oharness_llm::Llm directly (the escape hatch per plan §5.6.2) rather than via a helper-trait wrapper, because it needs to thread the same budget counter through all three hook sites.

Feature flags per plan §16.1: token and step are on by default; cost and wall-clock opt-in.

Re-exports

pub use amount::amount_from_response;

pub use amount::amount_from_usage;

pub use amount::budget_request_from;

pub use composite::CompositeBudget;

pub use error::BudgetExceeded;

pub use middleware::BudgetMiddleware;

pub use pricing::ModelPricing;

pub use pricing::PricingTable;

pub use step::StepBudget;

pub use token::TokenBudget;

Modules

amount

Helpers to project CompletionRequest / CompletionResponse / per-chunk Usage values onto the BudgetAmount / BudgetRequest shapes that the oharness_core::BudgetHandle API expects.

composite

CompositeBudget — fans check / consume / snapshot across any number of inner BudgetHandles. Any single child returning Deny denies the whole composite (plan §10.2).

error

The sentinel error returned when a budget is exhausted.

middleware

BudgetMiddleware — an Llm wrapper that threads a shared BudgetHandle counter through three hook sites:

pricing

Per-model pricing table used by CostBudget and the cost component of BudgetAmount (plan §10.4).

step

StepBudget — hard cap on the number of completed LLM calls (plan §10.2).

token

TokenBudget — hard cap on the sum of input + output tokens consumed over the lifetime of the handle (plan §10.2).