Claude Code 2.1.207

現在你看 Claude 吐出落落長的表格或程式碼時，打字和捲動都會滑順無比。以前只要遇到超長的清單或段落串流（streaming，答案一邊生成一邊逐字顯示），終端機畫面就會卡住，而且等它講完還會莫名其妙往上跳回開頭；現在這些顯示問題都修好了。就像看新聞台底下的跑馬燈，以前字太多會卡住甚至整條重播，現在不管塞多少字都順順跑過。

用 Bedrock、Vertex 或 AWS 上 Claude 平台的人，系統現在預設會幫你掛上最新的 Opus 4.8 模型，而且也不會像以前那樣，每講一句話就跑去向 IAM Identity Center 要一次新憑證（credentials，證明你是合法使用者的金鑰），速度快多了。對 Windows 使用者來說，如果 AWS 憑證卡住（例如背後處理程序當掉），系統現在最多等你 60 秒就會自己中斷，不會整個程式死當在那裡。就像進遊樂園玩，以前每玩一項設施都要重新查驗一次身分證，而且驗票機當機你只能站著枯等；現在買一次票就能玩到底，驗票機卡住一分鐘就會有人來放行。

讓多隻 Claude 互相合作的代理團隊現在更強韌了。以前只要團隊信箱（mailbox，代理們互相傳話的檔案）裡混進一則格式錯誤的壞訊息，整個團隊就會陷入無限崩潰的迴圈，每秒報錯直到你手動刪檔；現在它們懂得出錯時直接忽略並繼續工作。而針對遠端遙控，網路斷線或是憑證（證明你是合法使用者的金鑰）更新重連後不再漏接任務進度，你在手機跟網頁版上也能正常看見桌機版 App 跑的背景任務與工作流進度了。就像一間辦公室，以前收到一封亂碼傳真全公司就停擺，現在會把它扔進垃圾桶繼續辦公；出差在外的主管，也能順暢看到辦公室裡的進度白板。

這次收緊了外掛（plugin，擴充 Claude 功能的附加元件）的安全鎖。以前外掛設定如果放在專案資料夾裡，心懷不軌的專案就能利用環境變數的寫法，偷偷把惡意指令混進你的終端機裡（這叫 shell injection 攻擊）；現在直接禁掉這種寫法，並且不再讀取專案層級的外掛設定，只讀取個人設定、--settings 參數以及 managed settings。此外，在非互動模式（non-interactive run，像 SDK 那樣不跳提示直接跑的模式）下，也不會再跳過安全授權對話框，把你偷偷標記成「已同意」。就像公司的門禁系統，以前會不小心讀到訪客自己印的假識別證，或者讓走後門的人免簽名就混進來；現在只認總公司發的卡，而且所有人一律得過警衛這關。

你混用其他 Git 工具時，不會再被 Claude Code 留下的殘留設定卡住了。以前當你把工作樹（worktree，讓同一個專案多開幾份平行工作副本的 Git 功能）全砍光時，Git 設定檔裡會殘留一行設定，導致其他第三方 Git 工具讀不懂而卡死；現在它懂事後會順手清乾淨。另外，在讀取忽略檔（告訴系統別管哪些檔案的名單）或路徑規則時遇到括號打錯，或者只是要把多餘的輸出丟進虛無的 /dev/null，系統以前會嚇得停止運作或是大動作問你權限，現在它變聰明了，能正確分辨這些無害的動作。就像搬家公司以前撤走後會在牆上留下殘膠害下一個房客裝潢失敗，或者看到你把垃圾丟進垃圾桶還要跳出來問「確定要丟嗎」，現在他們搬得一塵不染，也不會管你怎麼丟垃圾。

這次修復了幾個常讓你嚇一跳的系統行為。儲值額度時如果手滑貼上時間戳（timestamp，像 2026-07-13 這種日期時間字串），系統以前會不聲不響剝掉非數字的部分、把剩下的數字整串黏成金額——貼個日期可能瞬間變成兩千多萬的天價；現在會直接報錯擋下，而且一次買超過一千美金會要求你多打字確認，避免荷包大失血。如果你為了方便自己改過啟動 Claude 的捷徑檔，自動更新也不會再粗暴地把它蓋掉。同時，系統內部正常的更新對話，不會再被防火牆誤判為你在進行 prompt injection（提示詞注入攻擊，把惡意指令藏進對話騙 AI 亂做事的手法）了。就像你去提款機，亂按密碼不會再隨便吐錢出來，而且銀行升級軟體時不再把你自訂的快速鍵洗掉，防盜警鈴也不會因為銀行員自己補鈔票而亂響。

When Claude spits out massive tables or code blocks, typing and scrolling will now stay buttery smooth. Previously, streaming (responses appearing word by word as they generate) very long lists or paragraphs would freeze the terminal, and when the answer finished, the screen would inexplicably jump back to the top; these display glitches are now fixed. Like a news ticker at the bottom of a TV screen — it used to stutter or reset when the headline was too long, but now it scrolls steadily no matter how much text you feed it.

For people using Bedrock, Vertex, or the Claude Platform on AWS, the system now defaults to the latest Opus 4.8 model, and it no longer frantically begs IAM Identity Center for fresh credentials (the keys proving you're a valid user) on every single API request, speeding things up significantly. For Windows users, if AWS credential fetching stalls (like a stuck background process), the app now bails out after 60 seconds instead of hanging indefinitely. It's like visiting a theme park: you used to have to show your ID at every single ride, and if the scanner broke, you just stood there forever; now one ticket covers the whole park, and if a scanner hangs for a minute, a worker waves you through.

Agent teams that collaborate on tasks are much more robust. Previously, a single malformed message ending up in a teammate's mailbox (the file agents use to pass notes to each other) would throw the entire team into a catastrophic crash loop, tossing an error every second until you manually deleted the file; they now know how to shrug it off and keep working. For Remote Control users, task status updates no longer drop into the void when your network hiccups or your credentials (the keys proving you're a valid user) refresh, and mobile/web clients now correctly mirror the background agent and workflow progress hosted on your desktop app. It's like a bustling office that used to freeze completely if it received a single unreadable fax — now they just toss it in the bin and keep working, while the boss traveling abroad can still perfectly see the progress whiteboard back at headquarters.

We've battened down the security hatches for plugins (add-ons that extend Claude's capabilities). Previously, if plugin configs lived in a project folder, a malicious repository could use variable syntax to sneak nasty commands straight into your terminal (a shell injection attack); this syntax is now completely rejected, and plugin settings are no longer read from project-level files at all, only user, --settings, and managed settings are honored. Furthermore, non-interactive runs (like SDKs that run without prompting the user) will no longer silently record you as "consented" to security policies without actually showing you the dialog. Like a corporate security gate that used to occasionally accept a visitor's home-printed badge or let back-door VIPs bypass the sign-in sheet; now it strictly honors only HQ-issued cards, and everyone signs the book.

Your other Git tools and everyday commands no longer trip over a setting Claude Code left behind. When you completely removed your last worktree (a Git feature that gives one project several parallel working copies), a stray line of configuration used to be left behind in your Git config, breaking other third-party Git tools; it now sweeps up after itself. Also, when parsing malformed brackets in ignore files (lists telling the system which files to skip) or path rules, or dealing with commands that just harmlessly piped output into the /dev/null void, the system used to panic, break file reads, or stop everything to ask for permission. Now it accurately assesses these safe scenarios. Like a moving crew that used to leave glue residue on the walls ruining the next tenant's paint job, or a security guard who jumped out to ask for ID just because you tossed a crumpled napkin into the trash can — they now leave the place spotless and let you throw away garbage in peace.

A few startling system behaviors have been smoothed out. If you accidentally pasted a timestamp while topping up your credits, the system used to blindly strip it down to digits and charge some random amount; it now halts with an error, and any amount over $1,000 requires you to explicitly type a confirmation to save your wallet from a catastrophic slip. If you customized the launcher script that opens Claude, the auto-updater will stop rudely overwriting it on every release. Also, totally normal internal system chat updates will no longer falsely trigger the "prompt injection attack" alarms. It's like going to an ATM: mashing the keypad won't randomly withdraw a weird amount of cash, the bank's software update won't erase the custom PIN shortcuts you set up, and the burglar alarm stops screaming every time a bank teller legitimately refills the money tray.