Analyzes Claude Code and Codex usage logs to surface cost drivers such as low cache hit rates, unpaid cache writes, top-tier sub-agents, reasoning share, and bloated sessions, with evidence and what-if savings to help cut AI coding spend.
Savings are what-if repricings of the same tokens at another rate. They say nothing about quality; tell the user to spot-check before switching a model.
A finding with saving n/a has no defensible number (no measured delta exists). Do not invent one.
Do not change settings, models or CLAUDE.md on the user's behalf; present the action and let them choose.
Prefer the largest controllable lever first: cache discipline, then model tier for sub-agents, then reasoning effort, then context size.
Pair with cost-ledger for the totals and cost-session to drill into one expensive session.
