Deterministic agent-token-cost intelligence from real agent usage data. Turns DeepSeek/OpenRouter usage CSVs into a JSON data product: per-request economics, cache-hit split, and the actionable cost tail.
Run: python3 analyze_usage.py → emits a deterministic JSON report (no LLM,
reproducible byte-for-byte).
- 962 req | 278,113,375 input tokens | 289,099 tok/req
- 99.22% cache-hit (native) | 2,171,677 cache-miss input tokens/day
- Cache-miss input ≈ $0.96/day ≈ $29/mo on one model — the slice a client-side dedupe layer (mcp-token-saver) reclaims.
This is the raw-material discipline that produces honest, verifiable numbers — the same contract as the rest of the sudo-ai-git family.