Skip to content

token-optimization.md

Latest commit

 

History

History
88 lines (59 loc) · 2.37 KB

token-optimization.md

File metadata and controls

88 lines (59 loc) · 2.37 KB
Raw

Token Optimization Guide

Overview

Managing token consumption in Gemini CLI is important for both cost and performance.

Model Selection

Recommended Models

Task Model Reason
Most coding tasks gemini-2.0-flash Fast, cost-effective
Complex architecture gemini-2.5-pro Deep reasoning
Simple queries gemini-2.0-flash-lite Minimal cost

Configuration

Set your default model in ~/.gemini/settings.json:

{
  "model": "gemini-2.0-flash"
}

Or per-session via environment variable:

export GEMINI_MODEL=gemini-2.5-pro

Context Window Management

MCP Server Optimization

Each active MCP server consumes context tokens. Best practices:

  • Enable fewer than 10 MCP servers per project
  • Disable unused servers in project settings
  • Use project-specific MCP configurations in mcp-configs/

Session Management

  • Start fresh sessions for unrelated tasks
  • Use /egc-checkpoint to save state before context-intensive operations
  • Use /egc-sessions to review and resume past sessions

Skill Loading Optimization

Skills are loaded on-demand when referenced. To optimize:

  1. Be specific: Reference only the skills you need
  2. Use core skills: coding-standards, backend-patterns, tdd-workflow cover most cases
  3. Defer niche skills: Load language-specific skills only when working in that language

Hook Runtime Controls

Control which hooks run to reduce overhead:

# Minimal hooks (fastest)
export ECC_HOOK_PROFILE=minimal

# Standard hooks (recommended)
export ECC_HOOK_PROFILE=standard

# Disable specific hooks
export ECC_DISABLED_HOOKS="hook-id-1,hook-id-2"

Practical Tips

Situation Action
Large codebase review Use targeted agents (@go-reviewer not full review)
Repeated build fixes Use language-specific resolver (/egc-go-build vs generic)
Documentation updates Use /egc-update-docs instead of manual prompting
Architecture decisions Use @architect with clear scope boundaries

Continuous Learning for Token Efficiency

The /egc-learn command extracts patterns from your sessions into reusable skills, reducing the need to re-explain context in future sessions:

/egc-learn          # Extract patterns from current session
/egc-evolve         # Cluster instincts into efficient skills