Skip to content

aetox-skills/token-calc

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

24 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Token Auditor

รู้ก่อนเสีย — วัด token consumption จริงของ AI coding tools ก่อน deploy

License: MIT Platform Python PowerShell


📏 Measuring system prompt...
  Platforms detected: OpenCode, ZCode
  TOTAL system prompt                            12,012 tok
  
⚠️  FIRST CALL SHOCK
  Opening this session costs you 14,012 tok before you type.
  
📊 TOKEN BREAKDOWN (per call)
  Input Sent                              12,012
    ↳ Cache Hit (reused)                   3,604
    ↳ Cache Miss (fresh)                   8,408
  Output Received                          2,000

📈 CACHE EFFICIENCY
  Cache Hit Rate                    30.0%
  
🔄 100 calls → 1,044,404 tok processed (not 1,401,200 sent)
💡 Top optimization: History (5,000 tok), MCP tools (2,700 tok)

ตรวจจับ system prompt ของคุณ จริง จากไฟล์ config จริง
แล้วบอกว่าแต่ละ session เผาผลาญ tokens เท่าไหร่ — แยกตาม cache hit/miss, projection ถึง 100 calls, และแนะนำว่าควร optimize ตรงไหน


Features

🔍 Auto-detect อ่าน system prompt จาก OpenCode, ZCode, Claude Code โดยตรง
💰 First Call Shock รู้ว่าเปิด session แต่ละครั้งเสีย tokens เท่าไหร่ก่อนพิมพ์
📊 Token Breakdown แยก input = cache hit + cache miss + output
📈 Cache Efficiency Hit rate, miss rate, tokens ที่ประหยัดได้จริง
🔮 Processing Projector 1 → 10 → 20 → 50 → 100 calls จะเสียเท่าไหร่
💡 Recommendations แนะนำสิ่งที่ควร optimize แยกตาม platform
🚨 Threshold Guard ตั้ง alert ถ้า system prompt เกิน limit (exit 2)
📉 Baseline Tracking --save + --diff เปรียบเทียบเมื่อเปลี่ยน config
🔄 Multi-model วัด AI coding tool ได้ทีเดียว — เทียบ prompt size

Quick Start

Python (cross-platform)

# Auto-detect + projection 100 calls
python token-calc.py --measure --calls 100 --output-tokens 2000

# Custom input
python token-calc.py --input-tokens 60000 --cached-input-tokens 52000 --calls 200

# Threshold guard (exit 2 if exceeded)
python token-calc.py --measure --threshold 50000

PowerShell (Windows)

.\token-calc.ps1 -Measure -Calls 100 -OutputTokens 2000
.\token-calc.ps1 -InputTokens 60000 -CachedInputTokens 52000 -OutputTokens 4000 -Calls 100

Use Cases

🎯 รู้ต้นทุนจริงต่อ session

python token-calc.py --measure --calls 50

→ "50 calls = 524K processed tokens, cache hit 30% → ลดได้อีก"

🚧 CI Guard

python token-calc.py --measure --threshold 50000
if ($LASTEXITCODE -eq 2) { throw "System prompt too large!" }

→ Deploy ล้มเหลวถ้า system prompt > 50K tok

📉 ติดตาม trend

python token-calc.py --measure --save baseline.json
# ... แก้ config ...
python token-calc.py --measure --diff baseline.json

→ "Cache hit rate ลดลงจาก 45% → 30% — MCP ตัวใหม่อาจเป็นปัญหา"


All Parameters

Flag Default Description
--measure off Auto-detect system prompt ของ AI coding tools
--platform auto opencode / zcode / claude
--input-tokens 0 Input tokens ต่อ call
--cached-input-tokens 0 ส่วนที่เป็น cache hit
--output-tokens 0 Output tokens ต่อ call
--calls 1 จำนวน calls ที่ projection
--milestones 1,10,20,50,100 ขั้น projection custom
--threshold 0 Alert threshold (exit code 1/2)
--save บันทึก baseline เป็น JSON
--diff เทียบกับ baseline JSON
--context-window 200000 Context window ของ model
--no-details แสดงแค่ summary

Philosophy

Token-only. ไม่มีราคา, ไม่มี model pricing, ไม่มีสกุลเงิน — วัดแค่ tokens
เพราะ token consumption คือ metric ที่ stable ที่สุดสำหรับ AI agent ops

คุณจัดการ cost เอง — Token Auditor ทำให้คุณรู้ก่อนว่าเท่าไหร่


Examples

  • OpenCode Steward → system prompt ~22K, cache ~77%, $0.07/20 prompts (DeepSeek V4 Flash)
  • ZCode backend-py → system prompt ~18K, cache ~65%
  • Claude Code → system prompt ~40K, cache ~50%

ดูรายละเอียดใน SKILL.md


License

MIT

About

Token savings calculator: measure system prompt, project cost savings with compounding multipliers. Platform-agnostic skill for OpenCode, ZCode, Codex, Claude, any AI agent.

Topics

Resources

License

Stars

1 star

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors