cost-metering

📁 scientiacapital/skills 📅 4 days ago
1
总安装量
1
周安装量
#43001
全站排名
安装命令
npx skills add https://github.com/scientiacapital/skills --skill cost-metering

Agent 安装分布

mcpjam 1
claude-code 1
replit 1
junie 1
windsurf 1
zencoder 1

Skill 文档

<quick_start> Check current spend:

cat ~/.claude/daily-cost.json 2>/dev/null || echo "No tracking yet"

Initialize tracking:

mkdir -p ~/.claude
echo '{"date":"'$(date +%Y-%m-%d)'","spent":0,"budget_monthly":100,"budget_daily":5}' > ~/.claude/daily-cost.json

</quick_start>

<success_criteria>

  • Daily cost tracking initialized at ~/.claude/daily-cost.json
  • Budget alerts fire at 50% (info), 80% (warn), and 95% (block) thresholds
  • Model routing applied: Haiku for search/classify, Sonnet for code, DeepSeek for bulk
  • Cost-per-feature metric available via portfolio-artifact integration
  • Monthly spend stays within configured budget ($100/mo default) </success_criteria>

Model Rates (Current)

Model Input/1M tokens Output/1M tokens Typical Use
Claude Opus 4 $15.00 $75.00 Architecture, complex reasoning
Claude Sonnet 4.5 $3.00 $15.00 Code generation, standard tasks
Claude Haiku 4.5 $0.25 $1.25 Search, classification, simple
DeepSeek V3 $0.27 $1.10 Bulk processing
GROQ Llama 3.3 70B $0.59 $0.79 Fast inference
Voyage Embeddings $0.10 — Embeddings

Budget Configuration

~/.claude/daily-cost.json

{
  "date": "2026-02-07",
  "spent": 2.40,
  "budget_monthly": 100,
  "budget_daily": 5,
  "alerts": {
    "info": 0.5,
    "warn": 0.8,
    "block": 0.95
  }
}

Alert Thresholds

Threshold % Budget Action
Info 50% Display: “50% of monthly budget used”
Warn 80% Yellow alert: “⚠️ 80% budget — consider model downgrade”
Block 95% Red alert: “🛑 95% budget — require explicit override to continue”

Cost Optimization Strategies

1. Model Routing (biggest impact)

Task Expensive Optimized Savings
File search Sonnet ($3/1M) Haiku ($0.25/1M) 92%
Code review Sonnet ($3/1M) Haiku ($0.25/1M) 92%
Classification Sonnet ($3/1M) Haiku ($0.25/1M) 92%
Bulk processing Sonnet ($3/1M) DeepSeek ($0.27/1M) 91%

Rule: If the task doesn’t generate code, use Haiku. If it doesn’t need Claude, use DeepSeek.

2. Context Management

  • Keep SKILL.md files under 200 lines (progressive disclosure)
  • Load reference files only when needed
  • Use Explore agent with haiku model for codebase search
  • Avoid reading entire files — use Grep to find specific lines

3. Task Batching

  • Group related searches into one Explore agent call
  • Use parallel subagents (haiku) instead of serial sonnet calls
  • Combine file reads when possible

Tracking Commands

Daily Spend Check

cat ~/.claude/daily-cost.json | jq '{date, spent, remaining: (.budget_daily - .spent), pct: ((.spent / .budget_monthly) * 100 | floor)}'

Weekly Report

# Aggregate daily logs
cat ~/.claude/cost-log.jsonl | jq -s 'group_by(.phase) | map({phase: .[0].phase, total: (map(.est_cost) | add), count: length})'

Monthly Summary

cat ~/.claude/cost-log.jsonl | jq -s '{
  total: (map(.est_cost) | add),
  by_model: (group_by(.model) | map({model: .[0].model, cost: (map(.est_cost) | add)})),
  by_phase: (group_by(.phase) | map({phase: .[0].phase, cost: (map(.est_cost) | add)}))
}'

Integration Points

System How
workflow-orchestrator Cost gate checks budget before workflows
subagent-teams Model selection uses cost tiers
agent-capability-matrix Includes model recommendations
portfolio-artifact Reports cost-per-feature metrics
End Day protocol Logs daily costs, updates MTD
TaskCreate/TaskUpdate Zero API cost — local UI tool for progress tracking
TeamCreate/SendMessage Zero API cost — local coordination (but spawned agents incur model costs)

Storage

~/.claude/
├── daily-cost.json          # Current day's budget + spend
├── cost-log.jsonl           # Append-only operation log
└── portfolio/
    └── daily-metrics.jsonl  # Includes cost-per-feature

Deep dive: See reference/cost-tracking-guide.md, reference/budget-templates.md