bungkus-eat-token

🔥 Bungkus Eat Token

Pantau dan analisis token consumption AI. Inspired by CodeBurn.

Kenapa track token meskipun gratis?

• Quota harian free tier ada limit
• Context window efficiency — token besar = slow response
• Cost awareness — kalau switch ke model berbayar, tau berapa biayanya
• Waste detection — token terbuang = resource terbuang

Token Metrics

Metrics yang WAJIB di-track:

Metric	Cara Hitung	Target
Total tokens	input + output + cache_read + cache_write + reasoning	-
Input tokens	Token yang dikirim ke model	Minimalisir dengan prompt caching
Output tokens	Token yang di-generate model	<500 per response ideal
Cache read	Token dari cache (hemat!)	>80% dari total input
Cache write	Token untuk buat cache baru	Minimal setelah warm-up
Reasoning tokens	Token internal thinking (o3, Claude extended)	Monitor, bisa besar
Token per session	total_tokens / jumlah_session	Track tren naik/turun
Token per message	total_tokens / jumlah_message	<2K ideal
Cache hit rate	cache_read / (input + cache_read + cache_write)	>80%

Breakdown format:

🔤 Tokens:
   Input:        1.5M    ← yang dikirim ke API
   Output:       146K    ← yang di-generate
   Cache Read:   39.4M   ← dari cache (hemat!)
   Cache Write:  0       ← sudah warm, tidak perlu tulis ulang
   Reasoning:    0       ← model tidak pakai thinking
   ─────────────────────
   Total:        41M     ← total token yang lewat
   Cache hit:    96.4%   ← sangat bagus (>80%)

Waste Detection (CodeBurn style)

Pattern	Indikasi	Fix
Output > Input	Model terlalu verbose	Tambah "be concise" di prompt
Cache hit < 80%	Prompt/context tidak stabil	Cek system prompt, jangan ubah tiap turn
Reasoning tinggi	Model overthinking	Turunkan reasoning effort
Token/session naik	Context bengkak	Compress context, hapus history lama
Cache write terus	Cache tidak persist	Cek provider caching config

Cost Estimation

Reference pricing (per 1M tokens USD):

Model	Input	Output	Notes
MiMo / GLM / Qwen	$0	$0	Free tier
DeepSeek	$0.14	$0.28	Murah
Claude Haiku	$0.25	$1.25	Budget Claude
GPT-4o	$2.50	$10.00	Mid-range
GPT-5/5.4	$5.00	$15.00	Premium
Claude Sonnet	$3.00	$15.00	Balanced
Claude Opus	$15.00	$75.00	Top-tier

Theoretical Cost

Bahkan model gratis tetap dihitung "berapa jika bayar":

🤖 Models:
   xiaomi/mimo-v2-pro: $0.0000 (240x, 0%)
   gpt-5.4: $0.0000 actual · ~$11.69 if paid (6x)
   💡 Theoretical cost if paid: ~$11.69

Activity Classifier

13 kategori dari tool usage pattern:

Kategori	Trigger
Coding	edit, write_file, patch
Debugging	fix, debug, error, bug
Feature Dev	create, implement, add, build
Refactoring	refactor, rename, simplify
Testing	test, pytest, vitest
Exploration	read_file, search_files, browser
Planning	plan, writing-plans, todo
Delegation	delegate_task
Git Ops	git, commit, push, merge
Build/Deploy	deploy, docker, build
Communication	send_message, clarify
Data/Analysis	execute_code, jupyter
General	lainnya

Hermes Commands

# Summary
python3 ~/.hermes/scripts/cost-logger.py           # All-time
python3 ~/.hermes/scripts/cost-logger.py today     # Hari ini
python3 ~/.hermes/scripts/cost-logger.py week      # 7 hari
python3 ~/.hermes/scripts/cost-logger.py month     # 30 hari

# Status (one-liner)
python3 ~/.hermes/scripts/cost-logger.py status

# Optimize scan
python3 ~/.hermes/scripts/cost-logger.py optimize

# Budget
python3 ~/.hermes/scripts/cost-logger.py budget 50     # Set $50/hari
python3 ~/.hermes/scripts/cost-logger.py budget        # Cek budget

# Export
python3 ~/.hermes/scripts/cost-logger.py export csv    # CSV
python3 ~/.hermes/scripts/cost-logger.py report --json # JSON

Script: cost-logger.py

Script companion untuk Hermes. Taruh di ~/.hermes/scripts/cost-logger.py.

Fitur:

• Activity classifier (13 kategori ala CodeBurn)
• Cache hit rate tracking
• Theoretical cost estimate (jika model tidak gratis)
• Token breakdown per model/source/session
• Budget alerts
• JSON/CSV export

Code Burn Analysis (repo check)

Untuk cek code burn sebelum implementasi:

# Structure scan
find . -name "*.py" | xargs wc -l | tail -1
find . -name "*.js" | xargs wc -l | tail -1
tree -L 3 --dirsfirst -I 'node_modules|.git|__pycache__|venv'

# Dependencies
cat requirements.txt
cat package.json | python3 -c "import sys,json; d=json.load(sys.stdin); print(json.dumps(d.get('dependencies',{}), indent=2))"

# Security scan
grep -rn "password\|secret\|api_key" --include="*.py" --include="*.js"
grep -rn "eval\|exec\|os.system" --include="*.py" --include="*.js"

Output Format

## Token Analysis: [project]

### Overview
- Sessions: [X]
- Messages: [X]
- Tools: [X]

### Token Consumption
- Input: [X]M
- Output: [X]K
- Cache hit: [X]%
- Theoretical cost: ~$[X]

### Activities
- Coding: [X]%
- Exploration: [X]%
- ...

### Risks
- 🔴 [blocker]
- 🟡 [medium]
- 🟢 [low]

Pitfalls

• Token consumption ≠ cost (free models = $0 tapi token tetap terpakai)
• Cache hit rate target >80% (CodeBurn standard)
• Jangan cuma cek cost, cek token breakdown
• Reasoning tokens bisa besar tanpa disadari
• Session panjang = context bengkak = token boros