Usage 0.22.0

想回味過去這一年你到底多用力在用 AI，現在報表會幫你做成一份年度回顧。HTML 報表新增兩個段落，把你一整年的本地用量濃縮成一眼看完的摘要。第一個是仿 GitHub 風格的 52 週貢獻熱度圖（contribution heatmap，那種一格一格、顏色越深代表那天越活躍的年度方格圖），每天依你燒掉多少 token（AI 計量的單位，回答越長花越多）來著色，旁邊還標出你目前跟歷來最長的連續活躍紀錄（streak，連續好幾天沒中斷使用），以及你最忙的那一天。第二個是一張「Wrapped」總結卡，把這一年算總帳：總共燒了多少 token、花了多少錢、活躍了幾天、最長連續紀錄，外加你最常用的模型跟最常跑的專案；最後還會依你這一年到底比較常用哪個 AI，封你一個守護精靈——跑 Claude 多的是鳳凰（phoenix），跑 Codex 多的是龍（dragon）。全部只用你電腦裡的本地檔案算出來，不碰網路。就像 Spotify 年度回顧那樣，把你一整年的習慣做成一張精彩榜單，只不過這回統計的是你跟 AI 一起熬夜幹活的足跡，最後還會依你偏愛哪個 AI 給你一隻代表神獸。

就算你的紀錄檔已經累積得超巨大，報表開起來還是飛快。上述年度回顧背後那一整年的資料，現在會快取（cache，先把算好的結果暫存在磁碟裡、下次要用直接拿、不必重算）到磁碟上，你重複打開報表時幾乎是即時呈現，只有當資料過期了才會重新計算一次。所以即使你本地的 log 檔案龐大無比，報表依然能快速打開。就像一家餐廳把常客每天點的菜先整理成一本速查表，熟客一進門照表上菜立刻端出來，只有當你今天新點了別的菜，才需要重新謄寫一次。

Want to look back on how hard you leaned on AI over the past year? The report now turns it into a year-in-review. The HTML report gains two new sections that distill a full year of local usage into a single glance. The first is a GitHub-style 52-week contribution heatmap (that yearly grid of squares where darker means a busier day), shading each day by how many tokens (the unit AI usage is metered in; longer answers cost more) you burned, with your current and all-time longest active streaks (streak — consecutive days of unbroken usage) and your single busiest day called out beside it. The second is a "Wrapped" summary card that tallies up the year: total tokens burned, total cost, active days, longest streak, plus your most-used model and most-run project; finally it crowns you a guardian spirit based on which AI you actually ran more this year — a phoenix if you leaned on Claude, a dragon if you leaned on Codex. Everything is computed purely from local files on your machine, with no network. It's like a Spotify year-end recap turning a year of listening habits into a vivid rundown — except this time what's tallied is your trail of late nights grinding with AI, capped with a spirit beast chosen by which AI you favored.

Even if your logs have grown enormous, the report still opens in a flash. The full year of data behind the year-in-review above is now cached (cache — stashing the already-computed result on disk so it can be grabbed directly next time without re-computing) to disk, so when you reopen the report it appears almost instantly, recomputing only once the data has gone stale. So even with an enormous local log file, the report still opens fast. It's like a restaurant that pre-compiles its regulars' daily orders into a quick-reference sheet, so the moment a familiar face walks in the usual dishes come straight out — and only when you order something new today does it have to rewrite the page.