outing terminal output through local proxies and semantic sandboxes drastically reduces API consumption without sacrificing debugging fidelity.

Approach	Metric 1	Metric 2	Metric 3
git status	3,000	600	80%
pytest / vitest	25,000	2,500	90%
ls -R / tree	2,000	400	80%
Full 30-min Session	111,000	23,200	~80%

Key Findings: The sweet spot for token optimization lies in intercepting stdout before it reaches the LLM, sandboxing verbose outputs locally, and replacing raw file reads with semantic dependency graphs. This pipeline consistently yields 80-90% token reduction while maintaining or improving agent task completion rates.

Core Solution

Implement a three-layer context optimization stack: terminal filtering, local context sandboxing, and semantic indexing.

1. RTK (Rust Token Killer): Terminal Proxy ⚡

RTK is a Rust binary that intercepts console output before it reaches the AI. It filters blank lines, strips comments, and groups repetitive logs to minimize token payload.

# Instalación vía Homebrew
brew install rtk-ai/tap/rtk

# Configurar el hook automático para Claude Code
rtk init --claude

2. Context Mode: Local Sandbox 🪨

Instead of dumping 50KB of failed test logs into the chat, Context Mode stores them in a local SQLite database. The AI receives only a summary and key terms, querying the database on-demand if deeper inspection is required.

# Ver cuánto has ahorrado en la sesión actual
/context-mode:ctx-stats

3. Tokensave: Semantic Knowledge Graphs 🔍

Tokensave builds a local map of your codebase (functions, classes, dependencies) so the agent understands architecture without reading every file.

# Indexar tu proyecto localmente
tokensave init
tokensave sync

4. Caveman Mode & CLAUDE.md Optimization

Force ultra-concise AI responses by stripping articles, pleasantries, and filler. Focus system instructions on the "why", not the "how". Replace verbose procedural rules with executable hooks. Optimize your CLAUDE.md using a 35-line template to boost cache hit rates up to 95%.

Pitfall Guide

1. Raw Log Dumping: Pasting full terminal outputs or stack traces directly into the chat window. Best Practice: Route all stdout through RTK or Context Mode to filter noise before API injection.
2. Ignoring the "Lost in the Middle" Effect: Assuming the LLM will retain instructions placed in the middle of a massive prompt. Best Practice: Anchor critical system directives at the beginning and end of the context window.
3. Over-Engineering CLAUDE.md: Writing 500-line procedural rule files that dilute cache efficiency and increase token overhead. Best Practice: Use a concise 35-line template focused on architectural constraints and executable hooks.
4. Skipping Semantic Indexing: Relying on the agent to grep through entire repositories for dependency mapping. Best Practice: Run tokensave init and tokensave sync to build a local knowledge graph, eliminating redundant file reads.
5. Verbose AI Conversations: Allowing the model to generate conversational filler ("Sure, I'd be happy to help..."). Best Practice: Enforce "Caveman Mode" in system prompts to strip pleasantries and force direct, token-efficient responses.
6. Context Window Blindness: Not monitoring token consumption during long debugging sessions. Best Practice: Use /context-mode:ctx-stats or npx codeburn to audit token leaks in real-time and adjust routing rules dynamically.

Deliverables

• Blueprint: Context Optimization Architecture Diagram (Terminal Proxy → Local SQLite Sandbox → Semantic Graph → LLM API)
• Checklist: Pre-Session Token Audit Checklist (RTK installed? Context Mode active? Tokensave synced? Caveman Mode enabled? CLAUDE.md under 35 lines?)
• Configuration Templates: CLAUDE.md 35-line optimization template, RTK hook configuration, Context Mode SQLite schema reference, and semantic indexing ruleset.