Claude Code 2.1.197

你打開 Claude Code 現在預設用的就是 Sonnet 5，原生支援一百萬 token（AI 讀寫文字時切出來的最小單位，大約等於零點幾個字）的記憶視窗（context window，AI 一次能記住多少對話內容的容量），等於同一個對話能塞進更多的程式碼或文件還不會忘記前面講過什麼。即日起到 8 月 31 日，每百萬 token 只要 2 到 10 美元的促銷價。

以前 AI 在對話裡提到或附上一個檔案，你想看實體檔案在哪還得自己切到 Finder（macOS 的檔案總管）裡找。現在按住 Cmd（Mac）或 Ctrl（Windows/Linux）點一下附件，就會直接在檔案總管裡跳出來、選好給你看。像收到包裹不用自己去倉庫翻箱子，按一下貨單就直接帶你走到那一格架子前面。

/code-review（叫 AI 幫你審查程式碼找問題的指令）背後原本拆成五個各自獨立的「抓問題小幫手」分別跑一輪，現在合併成一個一起跑，審查品質不變，但消耗的 token（AI 讀寫文字計費的最小單位，你的用量額度就是用這個算的）少了大約四分之一。這就像原本派五個保全分頭巡邏五次，現在改成一個熟練的保全一次巡完全部範圍，一樣巡得到、但人事費省了四分之一。

在 Claude Code 裡按 ! 切換到 bash 模式（直接打終端機指令、不經過 AI 對話的模式）打字時，現在輸入檔案路徑會即時跳出自動完成（autocomplete，打到一半就幫你列出可能的選項讓你選）建議清單，不用自己記清楚資料夾名稱再一個字一個字打對。就像你在搜尋引擎打字，還沒打完它就先幫你列出你可能要找的東西。

/clear（清空目前對話、讓 AI 忘記前面講過什麼的指令）以前一旦按下去，前面的對話就真的回不去了。現在 /rewind（倒帶，把對話拉回之前某個時間點的指令）支援連 /clear 之前的內容都能拉回來，等於多了一次反悔的機會。像你以前按下碎紙機就真的沒救了，現在碎紙機旁多了一台還原機，可以把紙條重新拼回去。

AI 用串流（streaming，AI 一邊想一邊把文字逐字吐出來、不用等全部想完才顯示的方式）方式輸出回答時，系統以前是每收到一點新文字就立刻重新畫一次畫面，很耗電腦資源。現在改成把文字更新合併、每 100 毫秒（千分之一秒）統一刷新一次，CPU（處理器，電腦的大腦，負責執行運算）使用量降低了大約三分之一強。這就像以前服務生每端一口菜就跑一趟廚房，現在改成湊滿一輪菜再一起端出來，跑的趟數少了，人也不會那麼喘。

Claude Code now defaults to Sonnet 5, with a native 1-million-token context window (context window, how much of the conversation the AI can remember at once; a token is the small chunk of text the AI reads and writes in) — meaning one conversation can now hold more code or documents without the AI forgetting what came before. From now through August 31, pricing is a promotional $2-$10 per million tokens.

When the AI mentions or attaches a file in chat, you used to have to switch to Finder or Explorer yourself and hunt for it. Now Cmd-click (Mac) or Ctrl-click (Windows/Linux) on the attachment jumps straight to that file, already selected, in your file manager. It's like a delivery slip that walks you straight to the right shelf instead of making you dig through the warehouse yourself.

/code-review (the command that has the AI review your code for problems) used to run five separate 'issue-finder' passes behind the scenes. Now they're merged into one pass — same review quality, but token usage (token, the unit the AI's reading/writing is metered in, which is also what your usage quota counts) drops by roughly a quarter. It's like replacing five separate security guards each patrolling once with one experienced guard who covers the whole route in a single pass — same coverage, a quarter less payroll.

When you press ! in Claude Code to drop into bash mode (typing raw terminal commands directly, bypassing the AI), typing a file path now triggers live autocomplete (autocomplete, a suggestion list that pops up partway through typing) — no need to remember and spell out exact folder names yourself. It's like a search engine that starts suggesting what you're looking for before you've finished typing.

Once you ran /clear (the command that wipes the current conversation so the AI forgets everything before it), there used to be no going back. Now /rewind (the command that rolls the conversation back to an earlier point) can reach back to before a /clear, giving you a second chance. It's like a paper shredder that used to be final — now there's a reconstruction machine next to it that can piece the strips back together.

When the AI streams its answer (streaming, writing text out token by token instead of waiting until it's fully done), the system used to redraw the screen every single time a little new text arrived, which was hard on your machine. Now text updates are batched and redrawn once every 100 milliseconds (a thousandth of a second), cutting CPU usage (CPU, your computer's 'brain' that does the actual computing) by roughly a third. It's like a waiter who used to run to the kitchen for every single bite of food, now waiting to carry out a full round of dishes at once — fewer trips, less out of breath.