token-stack

name	token-stack
description	Unified router for the 14-Layer Master Token & Context Engine (Semantic Cache, Model Router, Skill Router, Code Topology, Data Lens, Ponytail, Caveman, RTK, Turn Folding, CoT Governor, Loop Breaker, Headroom, MemoraX, OpenViking).

Token Stack Router

Provides unified access to the 14-Layer Master Token & Context Engine:

• Layer -1: Zero-Token Semantic Cache (SQLite N-Gram Cosine similarity, <12ms, 0 tokens)
• Layer 0: Model Cascading Router (RouteLLM & Frugal Classifier, -85% routine cost)
• Layer 0.5: Dynamic Skill Router (SkillRouter - arXiv:2603.22455, arXiv:2605.24050, arXiv:2606.10388: Two-Stage Retrieve & Rerank, Anti-Skill-Shadowing, -98% prompt bloat)
• Layer 1: Code Topology (Graphify, GitNexus, CodeGraph - AST navigation)
• Layer 1.5: Data & Quant Topology (DataLens - DuckDB / ClickHouse Data Contracts & Tear-Sheets)
• Layer 2: Code Reduction (Ponytail - Anti-boilerplate, KISS, YAGNI)
• Layer 3: Word Reduction (Caveman - Concise responses & Git patch diffs)
• Layer 4: CLI Output Filter (RTK - 60-90% terminal log & test noise reduction)
• Layer 5: In-Flight Turn Folding (5-turn Epoch Freezing, preserves prompt cache)
• Layer 6: CoT Reasoning Governor (Dynamic task-aware thinking token throttler)
• Layer 7: Runaway Loop Breaker & Failover (SHA256 ring buffer + 500ms 429 failover)
• Layer 8: Context Proxy (Headroom - Lossless HTTP compression & prompt caching)
• Layer 9: Knowledge Memory Harvester (MemoraX Code - Post-task episodic slots)
• Layer 10: Context Database Platform (OpenViking, Obsidian Vault - L0/L1/L2 storage)

Sub-Skills

• token-stack:benchmark -> Run isolated and cumulative token savings benchmarks across 10 standardized scenarios (14 layers).
• token-stack:health -> Run live health probes across all 14 layers.
• token-stack:setup -> Configure and install components into profiles.
• token-stack:report -> Generate token savings analytics.