Module compress

Expand description

Context compression for conversation history.

This module implements intelligent compression of conversation history to reduce token usage while preserving important information.

§Module Structure

AiCompressor: AI-based compressor using a Provider.
CircuitBreakerState: State for circuit breaker to prevent infinite retry loops.
CompressionBias: Compression bias - controls what to prioritize during compression.
CompressionConfig: Configuration for context compression.
CompressionHistoryEntry: Compression history entry for session metadata.
CompressionOutcome: Result of compression with metadata.
CompressionPipeline: Compression pipeline that orchestrates all modules.
CompressionResult: Result of a compression operation.
CompressionThresholds: Thresholds for content compression.
DependencyBuilder: Builder for dependency graphs.
DependencyGraph: Graph of message dependencies for preserving conversation coherence.
MessageDependency: Dependency relationship between ToolUse and ToolResult.
PhaseDetector: Detector for conversation phase.
PhaseWeights: Weight configuration for scoring based on conversation phase.
ScoredMessage: Message with its preservation score.
Scorer: Scorer for message preservation decisions.
SummarizedSegment: A segment of conversation history that has been summarized.
Summarizer: Summarizer for large content.
ToolCompressor: Compressor for tool results.

AiCompressionMode: Mode for AI-assisted compression.
CompressionStrategy: Strategy for compressing conversation history.
ConversationPhase: Conversation phase detected from message history. Determines which weights to apply during scoring.
ThresholdLevel: Threshold level for compression warnings.
ValidationError: Validation errors for compression.

AUTOCOMPACT_BUFFER_TOKENS: Token buffers for threshold levels (from Claude Code).
DEFAULT_COMPRESSION_THRESHOLD: Compression trigger threshold (percentage of context window). Lowered to 0.5 to compress earlier for long conversations (128K context -> 64K threshold)
DEFAULT_COMPRESSOR_MODEL: Default model for summarization.
DEFAULT_TARGET_RATIO: Target ratio after compression (keep this fraction of tokens).
ERROR_THRESHOLD_BUFFER_TOKENS
MANUAL_COMPACT_BUFFER_TOKENS
MAX_CONSECUTIVE_FAILURES: Maximum consecutive compression failures before stopping retries. Claude Code: “1,279 sessions had 50+ consecutive failures, wasting ~250K API calls/day”
MIN_MESSAGES_TO_KEEP: Minimum messages to keep after compression. Increased to preserve more recent context for continuity
TIME_BASED_MC_CLEARED_MESSAGE: Message to replace cleared tool result content (from Claude Code).
TIME_BASED_MC_GAP_THRESHOLD_MINUTES: Time-based microcompact threshold (minutes since last assistant message). When gap exceeds this, server cache has expired - clear old tool results.
WARNING_THRESHOLD_BUFFER_TOKENS

build_summary_prompt: Build a prompt for summarization.
compress_messages: Compress messages synchronously.
compress_messages_with_ai: Compress messages with AI assistance (async version).
compress_messages_with_full_ai: Compress messages with full AI support (async version).
compress_with_bias: Compress with bias-based scoring.
compress_with_pipeline: Legacy compression function (backward compatible).
estimate_tokens: Estimate token count for a message.
estimate_total_tokens: Estimate total tokens for a message list.
format_tokens: Format token count for display.
score_by_rules: Rule-based scoring for a message (public for pipeline use).
score_messages_only: Score messages without compressing (analysis only).
should_compress: Check if compression should be triggered.