Struct PressureMonitor 

pub struct PressureMonitor { /* private fields */ }

Monitors rho = used_tokens / max_tokens and recommends compression actions. All thresholds come from ContextConfig — no hardcoded constants.

Implementations

impl PressureMonitor

pub fn new(max_tokens: u32, config: ContextConfig) -> Self

pub fn max_tokens(&self) -> u32

pub fn pressure( &self, partitions: &ContextPartitions, engine: &ContextTokenEngine, observed_prompt_tokens: Option<u32>, ) -> f64

Current pressure rho ∈ [0, +∞). Uses provider-reported prompt tokens when available; otherwise estimates from partitions.

This is the raw rho (full partition weight). Making rho paging-aware — i.e. subtracting non-resident (Collapsed/SpooledOut/PagedOut) handle tokens so paging immediately relieves pressure — is not a drop-in here: [crate::context::manager::ContextManager::recompute_handle_residency] decides the Resident↔Collapsed projection from this very rho, so subtracting collapsed tokens would drop rho below collapse_threshold and immediately un-collapse (oscillation). That needs a deliberate split into raw rho (drives the collapse decision) vs effective rho (drives further compaction/renewal), tracked as remaining W1-1 design work — see ContextManager::rho.

pub fn recommend(&self, rho: f64) -> PressureAction