發布日期：2026-07-24
你靠自動功能守著 5 小時工作階段時，第一個提醒過後不會再整天沒動靜。先前它改用「跨過時段交界才提醒」的判斷，卻因為背景執行的 claude -p 不會觸發 statusLine（顯示在終端機最底下的狀態列）掛鉤，usage-status.json 一直留著已處理過的交界時間，於是自動功能永遠不再啟動。現在除了保留交界判斷來即時抓住真正換時段，也加上 5 小時 5 分鐘的冷卻計時；資料太舊時便重新準備下一次提醒。像鬧鐘既會在整點響，也有備用計時器；其中一個卡住，另一個仍會叫你起床。原本避免已開啟工作階段又亂提醒的關卡都還在。
程式在寫入資料時若突然當掉或被強制關閉，~/.usage/ 裡留下的暫存檔現在會被收走，不會越堆越多。正常情況下，原子寫入（先寫好新檔再一次換上去，避免寫到一半壞掉）結束時會刪除 mkstemp 產生的檔案；但遇到 SIGKILL（作業系統立刻強制終止程式的訊號）或崩潰，清理步驟來不及跑。一台電腦曾在 12 天累積 40 MB，主要是寫得較慢的 JSONL 快取（把算過／查過的結果暫存起來直接用）。現在啟動時會掃描 ~/.usage/ 和它直接底下的 *.d 資料夾，只刪符合名稱、超過 24 小時未碰過的一般檔案；符號連結和資料夾一律不碰。像倉庫每天開門先清昨夜遺下的包材，仍在裝箱的貨不會被誤丟。
If you rely on automatic starts to watch a 5-hour session, it will no longer fall silent for the rest of the day after its first alert. The recent switch to detecting a session boundary got stuck because headless claude -p does not trigger the statusLine hook, leaving usage-status.json stuck on a boundary it had already handled. The boundary check remains, so a real rollover is still caught immediately, but a separate 5-hour-and-5-minute cooldown now rearms the alert once that record is stale. Think of an alarm clock with both a chime and a backup timer: if one gets stuck, the other still wakes you. All the existing checks that prevent an alert during a live session remain in place.
If the app crashes or is forcibly stopped while saving data, its leftover temporary files in ~/.usage/ will now be cleared instead of piling up. An atomic write normally deletes the file made by mkstemp after safely putting the completed new file in place, but SIGKILL or a crash prevents that cleanup from running. One installation gathered 40 MB in 12 days, mostly from slower JSONL caches. On startup, Usage now checks ~/.usage/ and its immediate *.d children, deleting only ordinary files with the matching temporary-file name that have not been touched for 24 hours. Symlinks and directories are always left alone. It is like a warehouse clearing yesterday's packing scraps each morning without throwing away a box still being packed.
**Auto-start 5-hour Session stopped firing at all after its first ping.** 0.28.17 swapped the keeper's elapsed-time throttle for reset-boundary deduplication, which only holds up while the boundary keeps moving — and it doesn't. The ping runs `claude -p` headless, which never triggers Claude Code's statusLine hook, so `usage-status.json` goes on reporting the boundary that was already handled and the keeper wedges permanently. One machine went 23 hours with no auto-opened window while the cooldown-based Antigravity keeper kept working normally. The boundary check stays, since it is what catches a real rollover the moment it happens; a 5h5m cooldown is added alongside it as a second, independent way to re-arm once the payload has gone stale. Every existing gate is untouched, so a live window still never draws a spurious ping.**Orphaned temp files piled up in `~/.usage/`.** Every atomic write unlinks its `mkstemp` file in a `finally` block, and that block never runs when the app is SIGKILLed or crashes mid-write. One install had collected 40 MB of them over twelve days, dominated by the large JSONL caches whose slow writes are the likeliest to be interrupted. Startup now sweeps `~/.usage` and its direct `*.d` children, deleting only regular files that match the `mkstemp` name shape and have gone untouched for 24 hours — long enough that a second usage process cannot lose a temp file it is still writing. Symlinks and directories are skipped no matter how well their name matches.