Claude Code 2.1.191

對話歷程的掌控度變得更高了。現在就算你已經用 /clear 清空了畫面，你還是可以用 /rewind（時光倒流）功能，回到清空之前的對話狀態，不用擔心手滑按錯。閱讀長篇對答時，以前只要 AI 還在吐字（streaming），畫面就會失控一直被強迫拉到最底，害你看不到前面的內容；或是當你在代理人名單捲動超過上限時，整個面板會詭異地跳動一行，現在這些惱人的畫面干擾都修好了。這就像看網路直播，以前只要有人一直洗留言，你的畫面就會被強制往下拉；現在你可以安心往上滑看舊留言，畫面穩如泰山。

從任務面板停掉背景代理後，它不會再自己復活。以前你從任務面板停掉一個背景代理人，它過不久又會像殭屍一樣復活，現在停掉就是永久終止了。如果你在代理人面板（claude agents）貼上圖片，它現在會乖乖顯示 [Image #N] 佔位符，不再把整串又臭又長的路徑印出來佔版面；如果你對著背景任務輸入 /usage 這類內建指令，它現在會提示你正確用法，而不是笨笨地把指令當成聊天文字傳出去。其他修復還包含：登入網址在 Windows 終端機跨行時不會再被截斷、解決了特定環境（Ghostty）下全螢幕無法點擊連結的問題、修正用逗號分隔的掛鉤（hooks）條件會默默失效的 Bug、開場動畫不再超出 Mac 終端機預設大小、組織禁用 /voice 語音時會給出明確原因，而且在「最近拒絕」的權限面板裡點擊同意，關閉後設定終於會確實保留了。這就像把辦公室裡各種不順手的工具都換新：原本打死不退的離職員工被徹底請走，超長的公文附件會幫你自動縮排整理，印表機出錯也會告訴你到底哪裡卡紙，一切都變得俐落直覺。

面對網路問題和外部權限，系統變得更有韌性。現在沙盒（sandbox）裡的網路權限只要你點過一次「允許」，整個對話期間系統都會記住，不會每次連線都跳出來煩你。對於 MCP（外部工具伺服器），如果在讀取功能清單或要求 OAuth 登入憑證時遇到網路閃斷，它現在懂得稍等一下自動重試，不會馬上宣告失敗；如果回報 404 找不到網頁，錯誤訊息也會明確列出網址並指引你檢查設定檔。此外，在無頭環境（headless，沒有螢幕或瀏覽器的環境下運行）中，系統會直接跳過叫不出瀏覽器的步驟，讓你直接貼上登入網址。在企業管控設定上，強制刷新設定（forceRemoteSettingsRefresh）現在會確實發送不快取的指令，保證抓到最新規則，不會被代理伺服器騙。這就像你帶人進公司大樓，以前過每個閘門警衛都要重新查一次證件，現在只要大門登記過就能暢通無阻；就算內部系統偶爾連線不穩，也會幫你自動重整幾次，不會直接把你踢出門外。

長時間使用的效能大幅改善。系統現在會把生成文字的畫面更新頻率整合到每 100 毫秒一次，這樣不僅人眼看起來一樣順暢，還能大幅降低約 37% 的 CPU 佔用率。同時也減少了終端機輸出快取在長時間對話下的記憶體膨脹問題。這就像把水龍頭的噴頭從散狀改成集中水流，雖然出的水量一樣，卻能省下一大半的加壓馬達耗電，即使水開一整天也不會讓水管壓力破表。

You now have tighter control over your conversation history. Even if you've wiped your screen with /clear, you can still use /rewind to resurrect the conversation from exactly where it was before the wipe, saving you from accidental clicks. When reading long answers, the app used to forcefully yank your scrollbar to the bottom while the AI was still streaming new words, making it impossible to read earlier text; similarly, scrolling past the overflow cap in the agent roster would cause the whole panel to jump weirdly. These annoying visual interruptions are now fixed. It's like watching a live video stream where the chat used to violently auto-scroll whenever someone posted a new message; now you can peacefully scroll up to read older comments while the chat stays perfectly anchored.

Stopping a background agent from the tasks panel will now prevent it from resurrecting. Previously, stopping a background agent from the tasks panel would see it resurrect like a zombie shortly after; now, stopping an agent permanently terminates it. If you paste an image into the claude agents panel, it will politely display a [Image #N] placeholder instead of littering the screen with the full, obnoxiously long file path; and if you type a builtin slash command like /usage to a background session, it now shows a helpful hint instead of blindly sending it as chat text. Other fixes include: login URLs no longer truncate when wrapping across lines in Windows Terminal, fullscreen Cmd-click links now work in Ghostty, comma-separated hook matchers no longer silently fail, the splash art won't overflow the default macOS Terminal window, disabled /voice modes now properly explain organizational restrictions, and approving a permission in the 'Recently-denied' tab finally saves properly upon closing. It's like replacing all the clunky tools in your office: the ex-employee who wouldn't leave is finally escorted out permanently, absurdly long email attachments auto-collapse, and the printer finally tells you exactly where the paper jam is. Everything is just crisp and intuitive.

The system is now much more resilient when facing network hiccups and external permissions. Network permissions inside the sandbox now remember your 'Yes' for the remainder of the session, sparing you from annoying popups on every single connection attempt. For MCP (external tool servers), if a transient network drop occurs while fetching capability lists or requesting OAuth tokens, it now smartly waits a beat and retries instead of instantly failing. If it hits a 404 Not Found error, the message explicitly names the failing URL and points you to your config. Additionally, in headless environments (running without a display/browser), it seamlessly skips the impossible browser popup and goes straight to prompting you to paste the URL. On the enterprise management side, forceRemoteSettingsRefresh now correctly sends a no-cache header, ensuring proxies don't feed you stale policies. It's like bringing a guest into a corporate building: instead of guards checking IDs at every single door, one sign-in at the front desk grants access for the whole visit; and if the internal intercom gets staticky, the system automatically dials again instead of throwing you out on the street.

Performance during prolonged use has been significantly optimized. The system now coalesces terminal text updates to once every 100 milliseconds while the AI is streaming—keeping the animation perfectly smooth to the human eye while slashing CPU usage by an impressive 37%. We've also addressed the terminal output cache, stopping memory from ballooning out of control during exceptionally long conversations. It's like replacing a garden hose nozzle from a frantic wide spray to a steady stream; you get the exact same amount of water exactly where you want it, but the pressure pump uses half the electricity, ensuring the pipes never burst even if left running all day.