Struct LlmCondenser 

pub struct LlmCondenser { /* private fields */ }

Default Condenser implementation: summarizes a session’s tail via an LLM, reusing zeph_context::summarization::summarize_structured. Distinct from live in-memory compaction (owned by zeph-context) — this operates on the durable event log and is recorded as a crate::event::SessionEvent::Condensation event so replay can fold the same summary deterministically (spec §8, AC-6).

Implementations

impl LlmCondenser

pub fn new(deps: SummarizationDeps, threshold: f64, keep_recent: usize) -> Self

Construct a new condenser with the given LLM dependencies, trigger threshold (fraction of context budget used, 0.0..=1.0), and number of trailing messages to always keep.

Trait Implementations

impl Condenser for LlmCondenser

async fn should_condense( &self, state: &ReconstructedState, budget_used_fraction: f64, ) -> bool

Returns true if the reconstructed context has grown enough (relative to budget_used_fraction, the fraction of the context budget currently consumed) to warrant a condensation pass.

async fn condense( &self, events: &[SessionEventEnvelope], last_condensed_seq: u64, ) -> Result<CondensationResult, SessionError>

Condense events (typically the tail since last_condensed_seq), producing a CondensationResult whose replaced_range starts strictly after last_condensed_seq.