Struct ContextCompiler 

pub struct ContextCompiler { /* private fields */ }

Builder + entry point for context-window assembly.

Hosts construct one ContextCompiler per session (or per agent) and reuse it across turns. Optional dependencies (summarizer, sink) can be set once and shared via Arc.

Implementations

impl ContextCompiler

pub fn new(scorer: Arc<dyn RelevanceScorer>, budget: BudgetPolicy) -> Self

Construct a Tier 0 compiler with the supplied scorer and budget.

pub fn with_defaults() -> Self

Convenience: build with the default heuristic scorer and default budget.

pub fn with_summarizer(self, summarizer: Arc<dyn Summarizer>) -> Self

Inject a Tier 1 summarizer (M2). When absent, the orchestrator runs at Tier 0.

pub fn with_sink(self, sink: Arc<dyn ContextEmissionSink>) -> Self

Attach a structured-event sink (mirrors ainl_compression::with_telemetry_callback).

pub fn with_embedder(self, embedder: Arc<dyn Embedder>) -> Self

Inject a Tier 2 / M3 embedder for relevance reranking of non-pinned segments.

pub fn probe(&self) -> CapabilityProbe

Return the active capability probe based on injected dependencies.

pub fn compose( &self, latest_user_query: &str, segments: Vec<Segment>, existing_summary: Option<&AnchoredSummary>, vitals: Option<&CognitiveVitals>, ) -> ComposedPrompt

Compose a prompt window from segments, scored against latest_user_query, within self.budget. existing_summary carries over from the prior turn (Tier ≥ 1 only). vitals (when supplied) caps compression aggressiveness on low-trust turns per SELF_LEARNING_INTEGRATION_MAP §15.3.

Algorithm (per the plan):

1. Coarse selection — always-keep + heuristic scoring.
2. Budget allocation — apportion across kinds per BudgetPolicy.
3. Older-history compaction — Summarizer when present, else heuristic compression.
4. Tool-result fine-grained pruning.
5. Per-segment compression via ainl_compression.
6. Emit telemetry events.