Module terse 

Terse Compression Engine

Unified, 4-layer token compression pipeline that replaces the legacy compress_terse/compress_ultra functions and the TerseAgent prompt system.

Architecture

Tool Output → [Layer 1: Deterministic] → [Layer 2: Residual] → Compressed Output
Model Prompt → [Layer 3: Agent Shaping] → Optimized Prompt
MCP Tools List → [Layer 4: Description Terse] → Compact Descriptions

Layers

• Layer 1 (engine.rs): Deterministic output compression — surprisal scoring, content/function word filtering, domain dictionaries, quality gate.
• Layer 2 (residual.rs): Pattern-aware post-terse — applies after pattern compression, avoids double-compression, tracks attribution.
• Layer 3 (agent_prompts.rs): Agent output shaping — scale-aware brevity prompts, Telegraph-English-inspired format, adaptive levels.
• Layer 4 (mcp_compress.rs): MCP description compression — shrinks tool descriptions, lazy-load stubs, on-demand expansion.

Modules

agent_prompts

Layer 3: Agent output shaping prompts.

counter

Token counting utilities for terse compression.

dictionaries

Domain-specific abbreviation dictionaries for terse compression.

engine

Layer 1: Deterministic output compression.

mcp_compress

Layer 4: MCP tool description compression.

pipeline

Central compression pipeline — the single entry point for ALL integration modes.

quality

Quality gate for terse compression.

residual

Layer 2: Pattern-aware residual compression.

rules_inject

Injects compression prompts into agent rules files for all integration modes.

scoring

Surprisal-based line scoring for deterministic compression.

Structs

TerseResult

Result of a compression pipeline run with full attribution.