Struct TransformContext 

pub struct TransformContext<'a> {
    pub signal: &'a CancellationToken,
    pub model_id: &'a str,
    pub iteration: usize,
    pub last_provider_usage: Option<&'a Usage>,
    pub estimator: &'a dyn TokenEstimator,
}

Read-only context handed to a ContextTransform hook.

Carries the cancellation signal plus a few cheap observables that transforms key on (model identity, iteration index, last-turn token usage, the loop’s configured token estimator). Gathering these on the hook context — rather than widening the trait one parameter at a time — keeps the trait stable as later compaction layers (per-tool-result cap, cache-aware microcompact, auto-compact) come online.

New fields are additive: transforms that don’t care can ignore them.

Fields

signal: &'a CancellationToken

Cancellation signal for the current run.

model_id: &'a str

Model identifier the run is targeting (e.g. provider/model). May be empty when the host runtime doesn’t surface one — tests, fixture-replay transports, etc. Plugins that key per-model behavior should treat empty as “unknown”.

iteration: usize

Zero-indexed iteration within the current run. Same semantics as ToolGateContext::iteration: the very first LLM call of the run is 0.

last_provider_usage: Option<&'a Usage>

Token usage reported by the provider on the most recent assistant turn that surfaced a Usage block. None on the very first turn or when the provider didn’t surface usage. Useful for cache-aware decisions (read cache_read_input_tokens to see if the prompt prefix actually hit cache last turn).

estimator: &'a dyn TokenEstimator

Estimator the loop is configured with. Plugins use this to count tokens for budgeting and compaction without duplicating the loop’s tokenizer choice.

Implementations

impl<'a> TransformContext<'a>

pub fn for_test(signal: &'a CancellationToken) -> Self

Convenience constructor for tests and ad-hoc callers that don’t have a model id, iteration counter, or usage data. Picks the default char-heuristic estimator.