發布日期：2026-07-13
當 OpenAI 伺服器突然把 Codex 的限流（rate-limit，限制你多久能發送幾次請求的機制）從「5 小時與每週」改成「只剩每週」時，你的選單列卡片不會再留下一行醜醜的空白了。以前系統只會傻傻照著位置填數字，現在它會認明限流的時長，伺服器沒給的就不顯示，整張卡片會自動縮排；而且選單列上的百分比也會聰明地退回顯示每週的額度。另外，系統現在把零碎的檔案掃描紀錄集中起來，最多每 5 分鐘才寫入硬碟（disk I/O，程式讀寫硬碟的動作）一次，不再頻繁吵醒你的硬碟。就像是原本收納盒裡有兩個專屬格子，如果廠商突然不送其中一款贈品，以前格子會空在那裡長灰塵，現在收納盒會自動縮小不佔位；而且原本你每丟一張發票記帳軟體就要存檔一次，現在它會先記在腦袋裡，每五分鐘才寫進本子裡一次。
查看 Antigravity 的額度倒數時，數字現在會即時跟著時間跳動，不會像以前那樣卡在上次更新的時間點。這背後是因為系統現在改為直接讀取你電腦上已有的登入憑證（唯讀不修改），去向官方的額度介面（API）要資料，不再透過背景偷跑指令來檢查。這讓資料更新頻率從 15 分鐘縮短到 5 分鐘，倒數時間無比精準，也免去了多開額外處理程序的負擔；而且這完全只讀取額度資訊，不會扣到你的模型扣打（quota）。此外，HTML 報表裡的 Antigravity 使用紀錄也不會再被誤標成「unknown」了。就像搭火車，以前月台看板每 15 分鐘才人工翻牌一次，有時還會寫著「未知車種」；現在直接連線行控中心，每分每秒都精準倒數，而且車次名稱清清楚楚。
When OpenAI's servers abruptly switch the Codex rate-limit (the mechanism restricting how many requests you can make) payload from showing both a 5-hour and a weekly window to weekly-only, your menu bar card won't leave an ugly blank row anymore. Previously, the system mapped numbers purely by their position, but now it classifies limits by their duration—if the server stops reporting one, that row disappears completely, the card shrinks, and the menu bar percentage smartly falls back to the weekly value. Additionally, the system now coalesces file scan records and flushes them to your disk (disk I/O) at most once every 5 minutes, stopping the constant disk waking. It's like a bento box with dedicated slots: if a side dish stops being served, the box used to just leave an empty, dusty compartment, but now it automatically resizes to save space; and instead of writing to the ledger every single time you hand over a receipt, the accountant now batches them up and writes everything down once every five minutes.
When checking your Antigravity quotas, the reset countdowns now tick down live second by second, rather than freezing at whatever time it was last fetched. Behind the scenes, the system now reads the sign-in token already on your machine (strictly read-only) to query the official quota endpoint directly, instead of running background commands. This shortens the refresh cycle from 15 to 5 minutes, makes countdowns perfectly exact, and eliminates the overhead of spawning helper processes—and since it only reads metadata, it never consumes your model quota. Additionally, Antigravity usage rows in the HTML report are now properly labeled instead of defaulting to "unknown". It's like taking a train: the platform display used to only manually flip every 15 minutes and sometimes said "Unknown Train"; now it's hooked directly to dispatch, counting down perfectly to the second with the exact train name displayed.
**Codex card no longer shows an untitled "--" row**: mid-day on 2026-07-13, OpenAI's server-side rate-limit payload switched to weekly-only (the 5-hour window vanished with the CLI version unchanged), and the loader — which mapped fields by position — shoved the weekly number into the session slot, leaving a blank row behind. Rate-limit windows are now classified by their duration, a window the server no longer reports is hidden entirely (the card shrinks by one row), and the menu-bar Codex percentage falls back to the weekly value while the 5-hour limit is absent. If OpenAI brings the 5-hour window back, both rows return automatically.
**Less idle disk I/O**: the JSONL scan caches are flushed to disk at most once per 5 minutes (with a final flush on quit), rapid file-event bursts coalesce into a single refresh, and the four per-project ranking windows are aggregated in one pass over history.**Antigravity reset countdowns tick down live** instead of freezing at the last-fetched value until the next refresh.
**Antigravity quota now comes straight from the official quota API**: usage reads the sign-in token the Antigravity CLI already stores on your machine (strictly read-only — the file is never modified) and queries the official quota endpoint directly, replacing the background `/quota` CLI probe. Numbers refresh every ~5 minutes instead of 15, countdowns are exact, no helper process is spawned, and the call reads quota metadata only — it never consumes your model quota. Any fetch failure still falls back to the cached snapshot with the usual stale badge. The data-source description in all five README languages is updated to match.
Antigravity token rows in the HTML report are labeled "Antigravity" instead of "unknown".