@nexus-aethra/dshell-buffer

Token usage

Settings → 用量 charts what the models have cost, from an index dshell keeps in its own data directory:

• 每日热力 — a fixed 26-week calendar of daily totals, filled top-to-bottom and left-to-right, so the newest day is always the bottom-right cell. It reads all time on purpose: the range chips move the charts, never the calendar.
• 合计 / 未缓存输入 / 输出 / 缓存读 / 调用 / 缓存命中率 — the window's headline numbers. The hit rate is the one to watch: cache reads over all input.
• 每日消耗（按模型） and 模型占比 — one line per model over the window, and each one's share.
• 按模型明细 — input, output, cache read, cache write, calls and total per model.

The index is incremental: dsh attaches provider-reported usage to each assistant/message, and a scan reads only the sessions whose logs have grown since last time. It runs when you open the page, and on 重新聚合. A session whose log dsh cannot decode is reported as skipped — the page shows that the totals are partial, and why, instead of quietly counting less.

---