發布日期：2026-07-03
你現在會發現歷史讀取更穩更快，而且壞掉時也會坦白告訴你了。背後整理歷史紀錄的機制做了大掃除，現在只要掃描一次目錄，就能同時算好數字並找出有變動的地方，不用重複多跑好幾趟；而且如果你把 Codex 對話封存（archive）了，它也能即時跟上變化，不會再顯示舊數字。此外，萬一系統讀不到你的對話紀錄，以前它會悄悄拿上次的舊資料充數、讓你以為沒人在講話，現在它會在卡片旁邊掛一個小徽章，老實告訴你「我讀不到資料」；如果背後負責更新狀態的程式（hook）壞了，它也不會再像無頭蒼蠅一樣瘋狂重掃整個資料夾。就像以前計分板壞了會一直顯示 0 分讓你以為沒人得分，現在它會掛個牌子說「機器故障」，而且修理工人不會再為了找問題把整個球場的地板掀開。
You'll notice that reading your history is now faster, more stable, and more honest when things break. The engine that crunches your history in the background has been completely overhauled. It now scans your directories just once to both calculate the numbers and spot any changes, rather than making multiple redundant trips; also, if you archive your Codex chats, the system now tracks those edits instantly instead of showing stale data. Additionally, if the system fails to read your chat logs, it used to quietly fall back to the last known data, making you think you simply hadn't typed anything; now, it hangs a small warning badge on the card, honestly confessing that it can't read the data. And if the background updater program (the hook) breaks, it will no longer frantically rescan your entire folder tree over and over. It's like a broken stadium scoreboard that used to freeze at zero and make you think no one was scoring, but now it hangs a "Broken" sign — and the repair crew no longer rips up the entire field trying to find the problem.
**Archived Codex sessions could show stale data**: usage counted `~/.codex/archived_sessions/` when computing totals but didn't watch it for changes, so edits there might not refresh the menu bar until an unrelated refresh happened; the archived folder is now included in the same change-detection that drives live updates.
**A broken usage-data hook no longer causes constant background rescanning**: when Claude Code's status hook goes stale, usage now caches the "is there recent activity" check briefly instead of rescanning your entire `~/.claude/projects` tree on every poll.
**A small warning badge appears if local usage history can't be read**: previously, a failed read silently fell back to the last known data with no visible sign anything was wrong; the Project Usage card now shows a brief note (hover for detail) so you can tell "no new data" apart from "something broke."
**Faster per-refresh history scans**: each refresh previously walked the Claude/Codex session directories multiple times (once to detect changes, again per data source); it's now scanned once and reused.