Usage 0.28.21

你打開 usage 想看 Claude Code 的用量，卻懷疑回應變慢是自己額度用完，還是服務真的故障？現在有頁尾的面板，會在 Claude Code、Claude API 或 Codex API 發生中斷或效能降低時，底部亮起橘紅警示條。它讀的是各家公開的 Statuspage.io 狀態頁，不會呼叫 LLM 用量 API；所以這盞燈只告訴你服務是否出事，不會替你查 AI 用了多少。每個工具用各自面板的警示色，藏起工具卡時也不會冒出它的通知。五種介面語言都有文字說明。這像車上各自獨立的故障燈：引擎、煞車哪一處亮了就看哪一處，沒裝感測器的 Antigravity 則不會硬亮一顆假的燈。

掃過八種主題面板時，現在更容易一眼同時看到用量速度和服務狀態。原本上下堆疊的兩項資訊，已改成和預設面板一致的左右並排；畫面像把兩張便條貼在同一排，而不是一張壓著一張。萬一服務警示條出現，面板也會自己長高，替它留出位置，不再把原有內容擠扁。就像電梯裡加了一則公告，電梯不會因此把樓層按鈕擠到看不清楚，而是多開一塊版面讓內容看清。

明明只是接著同一段對話問問題，usage 卻一直把你標成 Heavy，現在這種誤判會少很多。用量速度會以每分鐘 token（AI 處理與計費的文字單位）分類成 Idle、Normal、Active、Heavy；先前計算 Heavy 時，把 cache_read 也算進使用量。cache_read 是提示快取（把已讀過內容暫存，重用時不用重新處理）的重讀，幾乎免費，主要反映對話有多長，不代表你用得多猛。現在它已被排除，門檻也從每分鐘 50／250／1000 調成 500／2500／6000 tokens。像健身房把你反覆看同一張課表當成運動量，當然會虛報；現在只計真正的訓練，標籤比較接近實際強度。

When usage feels slow, it can be hard to tell whether you hit a limit or the service itself is struggling. Panels with a footer now show an orange-red banner when Claude Code, the Claude API, or the Codex API has an outage or degraded performance. The banner reads each service's public Statuspage.io page; it never calls an LLM usage API, so it reports service health rather than your AI consumption. Each tool uses its own panel colors, and a hidden tool card produces no banner. The wording is available in all five interface languages. Think of separate dashboard lights in a car: the light tells you which fitted component has trouble. Antigravity has no usable public status page, so it does not pretend to have one.

Scanning the eight themed panels now makes it easier to see usage rate and service status at the same time. Those footer details used to be stacked vertically; they now sit side by side, matching the default panel, like two notes placed on one row instead of one covering the other. If a service-status banner appears, the panel also grows tall enough to give it space rather than squeezing the existing content. It is like adding a notice to an elevator without cramming the floor buttons out of sight: the cabin simply makes room.

If usage kept labeling you Heavy while you were only continuing one conversation, that false impression should now be much rarer. The live rate sorts activity into Idle, Normal, Active, and Heavy by tokens per minute—tokens are the units of text an AI processes and bills for. Previously, the Heavy calculation included cache_read, a nearly free reread of prompt-cache content that mainly reflects conversation length rather than intensity. That value is now excluded, and the thresholds have moved from 50 / 250 / 1000 to 500 / 2500 / 6000 tokens per minute. It is like a gym counting rereading the same workout sheet as exercise; now it counts the training itself, so the label better matches the effort.