Struct PreCompactHandler 

pub struct PreCompactHandler {
    orchestrator: CaptureOrchestrator,
    llm: Option<Arc<dyn LlmProvider>>,
    use_llm_analysis: bool,
}

Handler for the PreCompact hook event.

Analyzes context being compacted and auto-captures valuable memories. Integrates with DeduplicationService to check for duplicates before capture.

Deduplication

When a deduplication service is configured, each candidate is checked against:

1. Exact match: SHA256 hash comparison
2. Semantic similarity: Embedding cosine similarity (if embeddings available)
3. Recent capture: LRU cache with TTL

Duplicates are skipped and reported in the hook output.

LLM Analysis Mode

When an LLM provider is configured and use_llm_analysis is enabled, content that doesn’t match keyword-based detection patterns will be analyzed by the LLM for classification. This provides more accurate namespace assignment at the cost of increased latency.

Configure via environment variable: SUBCOG_AUTO_CAPTURE_USE_LLM=true

Fields

orchestrator: CaptureOrchestrator

Capture orchestrator for coordinating captures with deduplication.

llm: Option<Arc<dyn LlmProvider>>

Optional LLM provider for content classification.

use_llm_analysis: bool

Whether to use LLM for classifying ambiguous content.

Implementations

impl PreCompactHandler

pub fn new() -> Self

Creates a new PreCompact handler.

The use_llm_analysis setting is loaded from the environment variable SUBCOG_AUTO_CAPTURE_USE_LLM (default: false).

pub fn with_capture(self, capture: CaptureService) -> Self

Sets the capture service.

pub fn with_deduplication(self, dedup: Arc<dyn Deduplicator>) -> Self

Sets the deduplication service.

When set, candidates are checked for duplicates before capture. Duplicates are skipped and reported in the hook output.

pub fn with_llm(self, llm: Arc<dyn LlmProvider>) -> Self

Sets the LLM provider for content classification.

When set (and SUBCOG_AUTO_CAPTURE_USE_LLM=true), content that doesn’t match keyword-based detection will be analyzed by the LLM.

pub const fn with_llm_analysis(self, enabled: bool) -> Self

Enables or disables LLM analysis mode.

This overrides the SUBCOG_AUTO_CAPTURE_USE_LLM environment variable.

fn analyze_content(&self, input: &PreCompactInput) -> Vec<CaptureCandidate>

Analyzes content and extracts capture candidates.

fn classify_with_llm(&self, section: &str) -> Option<CaptureCandidate>

Uses LLM to classify content that didn’t match keyword detection.

Returns Some(CaptureCandidate) if the LLM suggests capturing, None if LLM is unavailable or suggests not capturing.

fn extract_from_text(&self, text: &str) -> Vec<CaptureCandidate>

Extracts potential memories from text.

Uses keyword-based detection first, then optionally falls back to LLM classification if use_llm_analysis is enabled.

fn record_metrics( status: &str, duration_ms: f64, capture_count: usize, skip_count: usize, )

Records metrics for the hook execution.

Trait Implementations

impl Default for PreCompactHandler

fn default() -> Self

Returns the “default value” for a type.

impl HookHandler for PreCompactHandler

fn event_type(&self) -> &'static str

The hook event type this handler processes.

fn handle(&self, input: &str) -> Result<String>

Handles the hook event.