savingslisted

# QuotaCove — savings report Two sources. The retrospective is almost always the better one. ## Preferred: replay the real history ``` node "${CLAUDE_PLUGIN_ROOT}/scripts/what-if.mjs" --by-month ``` Also takes `--days 30` for a recent window and `--by-project` to split by working directory. It reads local Claude Code transcripts, scores the prompt behind each past turn, and reprices that turn's **actual** token counts at the tier the router would have chosen. No assumed token mix enters the arithmetic, and it never routes work upward — whatever model was in use is the ceiling. It reads only model names and token counts. Prompt text is scored in memory and is never printed, stored, or transmitted. ## Secondary: the live log If the hook has been running with logging on, `~/.claude/quotacove-savings.jsonl` holds one line per scored prompt: ```json {"ts":"2026-08-24T18:02:11.000Z","score":18,"tier":"haiku","capped":false,"savedVsPremium":0.0525} ``` It only covers prompts seen since the hook was installed, so it is thinner than the retrospective. Use it for the tier distribution of recent work, or when transcripts are unavailable. ## Presenting it Lead with the headline: total saved, and the percentage. Then whichever breakdown was asked for — by month, by project, or by tier — as one compact table. State the caveats once, at the end, without softening them: - Figures use list prices. On a subscription there is no invoice to shrink; the same effect shows up as usag