發布日期：2026-07-14
現在你能用滑鼠直接點選全螢幕選單，也能讓公司統一控管啟動流程，還多了給輔助工具與 Vim 愛好者的新設定。如果你習慣用 Vim 編輯器，現在可以設定把連續按兩下 jj 等按鍵組合對應成退出插入模式的 Escape 鍵。若你需要使用無障礙工具，這次新增了螢幕閱讀模式（screen reader mode，把終端機畫面轉成純文字，讓報讀軟體能順利讀出畫面內容），可以透過指令參數、設定檔或環境變數來開啟。對於企業用戶，系統現在支援使用行程包裝器（process wrapper，一種能在軟體真正啟動前先套用公司內部規則的攔截程式），確保每一次背景啟動都會先經過公司的安全控管。就像為一輛車換上了更豐富的控制面板，不管你是習慣用方向盤、搖桿，還是需要專屬的輔助踏板，現在都能找到適合自己的駕駛方式。
現在你在切換不同 AI 模型時，再也不用手動把快速模式打開了。快速模式（fast mode，一種犧牲些微精準度來換取極速生成答案的運作狀態）是一個可以讓你省下等待時間的設定。以前如果你在設定檔裡開著這個模式，但中途切換去使用一個不支援此模式的模型，系統就會把它關掉；最煩人的是，當你切換回有支援的模型時，系統竟然會忘記你本來是開著的，導致你必須手動再開一次。這次更新修好了這個遺忘症，只要你的設定裡有開啟，換回支援的模型時它就會自動歸位。就像車子上的運動檔設定，以前你一旦開去不平的碎石路（不支援的模型）系統就會幫你解除，但回到高速公路時你還得自己重新按鈕；現在只要你開回平坦的道路，系統就會自動幫你把運動檔加回來，讓你一路順暢。
現在你交代給背景代理的任務與回覆，再也不會因為連線不穩而石沉大海了。背景代理（background agent，在背後默默幫你跑長時間任務的 AI 助手）有時候會因為網路波動導致訊息傳送失敗。以前遇到這種狀況，你辛辛苦苦打好的回覆就會直接人間蒸發；現在系統會妥善把你打的字存下來，等你重新啟動工作階段時再次幫你送達。此外，過去如果系統在你跑任務的途中進行了自動更新、替換掉了底層的執行檔，負責留守的常駐服務（daemon，一個安靜待在背景維持系統運作的隱藏程式）就會因為找不到原本的檔案而永久拒絕你重新連線，噴出「無法啟動」的錯誤；現在這種短暫的脫節不會再變成永久的死當。就像你把紙條塞給在另一個房間加班的同事，以前門卡住遞不過去紙條就飛了，現在會幫你壓在門縫下等他開門；而公司即使趁夜裡換了新打卡機，原本在裡面加班的同事也不會因此被鎖死在辦公室裡出不來。
當你帶著超長對話紀錄熬過一次系統更新後，不會再被錯誤的「容量已滿」警告嚇到了。模型在一次對話裡能記住的資訊總量叫做上下文視窗（context window，就像模型腦海裡的短期記憶暫存區）。以前當命令列介面在背景完成自動更新時，這個記憶體上限會短暫地重置退回 200k 的舊標準。如果你剛好在進行超長篇的對話，重新連線時系統就會因為這個舊標準，錯誤地亮出紅燈，告訴你「上下文已 100% 耗盡」，甚至連帶讓自動壓縮（auto-compact，一個在容量快滿時，會自動把舊對話濃縮成摘要來騰出空間的機制）的指示燈也跟著亂亮；現在這些儀表板都已經修好，能穩穩保持正確的讀數。就像一輛加掛了超大油箱的長途貨車，以前只要進廠快速保養一下，儀表板就會短暫故障，硬用小客車的油箱刻度來算，嚇得你以為油底殼見底了；現在不論怎麼更新，油表都會真實反映你那巨大的油箱容量，不再謊報軍情。
當網路伺服器單方面決定要關閉連線時，你的長時間工作任務再也不會被無辜牽連到直接崩潰了。在網路傳輸中，伺服器有時需要進行維護或清理連線，這時它會透過 HTTP/2（一種讓資料傳輸更有效率的新一代網路通訊協定）發送一個叫做 GOAWAY（名稱字面意思就是「走開」，用來宣告這條連線即將收線）的訊號。以前，如果你有在背景獨立運作的工作階段，或是正在你眼皮下執行的任務，剛好正在向伺服器發送請求時撞上這個訊號，整個任務就會因為承受不住這個意外的切斷而當場死機。現在系統變得更有韌性，它能聽懂這個訊號只代表當下這條連線結束，並把它當成一個可處理、可重試的正常斷線事件。就像你在跟客服通電話，以前只要客服說一句「不好意思本分機要保養了」然後掛斷，你手上正在填寫的表格就會跟著整張著火燒掉；現在你只會默默把電話放下，換播另一支號碼，手上的工作完全不受影響。
當你把 Claude 的輸出接到其他程式裡自動處理時，再也不會收到缺件的資料，系統也不會再把極大的數字看錯了。如果你要求系統輸出 stream-json（串流 JSON，一種把資料切成一小段一小段流出來，方便其他程式邊收邊吃的格式），以前遇到超長的回覆時，管線常常會在中途截斷，導致你不但拿不到完整的結果訊息，資料還殘缺不全。另一個隱藏的陷阱發生在你設定環境變數來限制系統輸出長度時：如果你為了方便，用了科學記號來寫數字（例如用 1e6 來代表一百萬），系統以前會眼瞎，只讀取 mantissa（尾數，也就是科學記號 e 前面的那個有效數字 1），結果你的「一百萬」字數上限就這樣莫名其妙變成了「1」個字。現在這些管線都已經疏通，不但能把超大包的資料原封不動地送達，也終於學會正確看懂科學記號了。就像請貨運公司送一列長長的火車模型，以前太長的車廂會被丟在半路，而且你在託運單上寫「10 的 6 次方」克，過磅員還會只看第一個數字算成「1」克；現在貨運公司不僅保證整列火車完整送達包裹不漏件，過磅員也終於看得懂數學符號了。
現在你看見 Claude 產出超大表格時，電腦不再因為記憶體耗盡而卡死了。以前如果 Markdown（一種用純文字來排版文件的格式寫法）的表格超過 200 列，系統會試圖一次畫出整張表，導致渲染進度停滯甚至拖垮記憶體；現在它只會優先顯示前 200 列，並在底下貼心地加上「還有 N 列」的提示。這就像翻開一本厚重的電話簿，它幫你先打開第一頁供你查閱，而不是硬生生把整本幾千頁的書攤開塞滿你的小書桌。
現在這些檔案處理工具不再因為遇到小狀況就胡亂報錯，讓你一頭霧水了。以前如果一個檔案在你讀取後被修改過，Edit 工具就會直接拒絕編輯，即便你要改的那段目標文字在檔案裡還是唯一相符的；現在只要內容對得上，它就能安全改寫。同時，其他的工具也變誠實了：Read 不會再把空檔案誤報為「比位移量（offset，用來指定從檔案哪裡開始讀取的數字）還短」，Grep 不會在你打錯正則表達式（regex patterns，一種用特殊符號來搜尋特定文字規則的語法）時默默騙你「找不到檔案」，而且在分頁（paginated，把大量結果拆成多頁顯示）模式下也不會少算總數。另外，Glob（一種用萬用字元比對路徑的工具）在遇到路徑或工作目錄裡混進空字元（null byte，程式用來標示字串結尾的特殊隱藏符號）時，也不會再不明不白地崩潰。就像個終於學會好好說話的圖書館員，他不會因為書架本來就是空的就嫌你「找得太深」，也不會在你給錯書名規則時兩手一攤說「沒這本書」，而是精確告訴你問題出在哪。
現在當你登入連線出狀況時，系統會直接告訴你真正的病因，不再讓你像個無頭蒼蠅了。以前當 apiKeyHelper（一個替你拿取 API 金鑰的輔助腳本）執行失敗時，系統會在背後默默重試大約十次，最後只丟給你一個毫無幫助的 401（HTTP 401 錯誤，代表未授權登入的通用網路代碼）；現在它只要失敗三次內，就會把腳本自己真正的錯誤訊息原封不動端出來。另外，如果你使用 AWS Bedrock 的串流請求（streaming requests，也就是讓 AI 一邊想一邊逐字吐出回覆的連線方式）時，資料不小心被公司的閘道器（gateway，夾在服務之間幫忙轉送資料的網路節點）從中改寫了格式，系統也不會再給你一句誤導人的「收到被截斷的訊息」，而是會精確點出是哪個內容類型（content-type，用來標示網路資料格式的標籤）出錯，並把矛頭正確指向代理伺服器（proxy，代替你轉送網路請求的中間站）。就像你去醫院看病，醫生不會再敷衍你「身體不舒服」，而是直接拿出 X 光片告訴你「是第三節脊椎有點問題」，讓你馬上知道該怎麼對症下藥。
現在系統裡的自動化與背景任務變得更強壯，不再因為微小的異常卡死或吵你了。首先，如果你的電腦瀏覽器壞了打不開，輸入 /upgrade（用來升級系統的指令）時會直接把更新網址（URL）印出來給你，而不是莫名其妙把你踢回登入畫面。對於在 Windows 環境寫程式的人，系統現在處理 stream-json（一種讓程式一小段一小段讀取 JSON 資料的格式）時，就算遇到空白的 CRLF（回車換行，Windows 特有用來表示換行的兩個隱藏控制字元）也不會導致整個工作階段崩潰。此外，如果無頭模式（headless，沒有顯示介面、純在背景默默運作的模式）的工作階段收到夾帶了格式錯誤的 set_model（用來切換 AI 模型的內部指令）時，現在會乖乖回報錯誤，而不是永遠當機懸在半空中。最後，當你重啟時，那些孤立的背景任務（orphaned background tasks，因為中斷而沒人管的殘留任務）也不會再洗頻狂刷「找不到完成紀錄」，而是會整齊地收攏成一行摘要。這就像是火車的自動駕駛系統升級了，遇到鐵軌上有一片落葉或是不小心收到錯誤的無線電訊號時，它懂得亮個警示燈繼續開，不會動不動就讓整台列車急煞甚至停駛。
現在遠端連線與 AI 團隊合作變得更即時、也更清晰了。以前當你用遠端控制客戶端（Remote Control clients，你用來從外部連線回主機的手機或網頁介面）接上本機終端的工作階段（terminal-hosted session，直接在你電腦黑色視窗裡跑的任務）時，你必須傻傻等到某個任務開始或結束，畫面才會跳出背景代理和工作流程的進度；現在只要一連上就會立刻同步顯示最新狀態。此外，當你使用代理工具（Agent tool，用來喚醒其他小幫手 AI 的功能）時，如果因為某些原因，子代理（subagent，被主 AI 呼叫出來幫忙的下屬 AI）的工具清單（tools list，一份告訴它能拿什麼工具的列表）裡全是一些無法辨識的廢字串，系統不會再笨笨地發給它一個「空空如也的工具箱」並硬著頭皮上工，而是會直接回報清楚的錯誤，並指名道姓告訴你是清單裡的哪些項目不認得。就像公司裡的主管出差，現在一打電話回辦公室就能立刻聽到各部門的即時進度報告，而不會只聽到一片死寂；且當主管派給新人任務清單時，如果上面寫的工具倉庫裡根本沒有，新人也會立刻反映「我找不到這些扳手」，而不是兩手空空就跑去工地發呆。
現在這幾個介面小毛病被修好了，不會再拿過期的舊資訊來忽悠你。首先，當你輸入 /usage（查詢 API 用量的指令）時，系統不會再用過期快取的長條圖（stale cached bars，系統為了省事沒更新的舊圖表）去蓋過最新的數據；當你在 /mcp（管理連線伺服器的指令）編輯完設定後，系統也會馬上重新判定那些佔位伺服器（placeholder servers，設定檔裡先挖個坑保留位置的伺服器名稱），不會再把它們當成死路。針對 SDK 主機（SDK hosts，像是 Claude Desktop 這種包著 AI 跑的外殼程式），即使背後還有背景任務在跑，只要你當下是閒置狀態（idle sessions，你沒在下指令發問的空檔），你就能隨意切換資料夾，不會再被系統用「正在處理這回合（A turn is in progress）」為由死板拒絕。當你要存檔時，工作流程儲存對話框（workflow save dialog）也會正確指向你的 CLAUDE_CONFIG_DIR（你指定的個人設定主目錄），而不是死硬綁在預設路徑。最後，如果你用 /release-notes（看更新日誌的指令）點開了「顯示全部」，它不會再把幾萬字的整本日誌硬塞進模型的上下文（model's context，AI 的短期記憶區）裡，害你之後的每一句話都帶不動。這就像是汽車導航終於會即時更新路況了，不會再拿昨天的舊地圖指路，而且你在等紅燈的空檔也能放心切換廣播頻道，系統更不會把整本車主手冊塞進儀表板裡卡死你的視線。
現在當你讓 Claude 進行長時間的對話或處理大量檔案時，它不會再越來越慢或吃光你的記憶體了。以前在許多地方會有記憶體漏洞，例如在 agent view（顯示 AI 工作狀態的介面）中貼上的圖片會一直卡在記憶體裡；MCP（Model Context Protocol，讓模型與外部工具溝通的協定）伺服器的 stderr（標準錯誤輸出）最高會累積到 64 MB；LSP（Language Server Protocol，提供自動完成等功能的協定）的文件會無限制地開著（現在最多只保留 50 份）；headless（沒有視覺介面的運作模式）或 SDK（Software Development Kit，軟體開發套件）也會因為工具回傳的巨大結果而無限膨脹。另外，背景 hook（特定事件發生時觸發的程式碼）的輸出也會殘留，如果在客戶端接上前提早重新載入 plugin（擴充功能的附加元件），SDK 工作階段甚至會遺失剛初始化的代理。現在這些問題都修復了，送出圖片後就會釋放記憶體，各種紀錄也會嚴格設定上限。如果你設定了很多權限拒絕或詢問的規則，系統現在會把它們預先編譯並快取起來，不再讓每個回合慢上好幾秒。讀取含有極長單行的檔案時，如果使用 offset/limit（指定讀取起迄位置的參數），現在會直接報錯，而不會把整行硬塞進記憶體導致當機。這就像一間辦公室，以前員工會把所有看過的文件和收到的傳真無限期堆在桌上，最後連轉身都有困難；現在他們學會了看完就歸檔或丟掉，讓辦公空間永遠保持寬敞。
現在當你裝滿各種工具或進行大範圍的檔案編輯時，操作介面會更加流暢，同時大幅降低對電腦效能的消耗。以前代理在更新任務清單時，會笨笨地把整個 UI（使用者介面）重新繪製一次，導致你打字時感覺卡頓；現在它只會更新變動的部分。如果你使用大量的 MCP（Model Context Protocol，讓模型與外部工具溝通的協定）工具，系統現在會快取 tool-pool（可供模型選擇的工具集合）的組裝過程，讓工具處理速度最高提升七倍。在記憶體管理方面，檔案編輯的讀取快取現在被嚴格限制在 16 MB 以內，不再像以前那樣死死釘住高達一千份完整檔案不放。系統也會主動修剪已經被覆蓋的檔案歷史備份，讓 checkpoint（對話狀態的存檔點）與 transcript（完整對話紀錄）大瘦身，在頻繁編輯的情況下檔案大小最多可縮減 79 倍。此外，從包含巨大對話紀錄的背景代理或 fork（從既有對話分岔出來的獨立進程）恢復工作階段時，也會佔用更少的記憶體。這就像一個專業修車廠，以前師傅每拿一個扳手就要把整個工具箱清空重排，而且還把每次換下來的廢料全堆在車庫裡；現在他懂得把常用工具掛在手邊，並且定時把沒用的廢鐵清走，讓修車廠永遠保持高效運作。
現在你能確實追蹤做完的背景任務，也能更安全地管理多份工作副本。以前 background agent（默默在背後幫你做事的 AI 助手）一完工就會瞬間從清單中消失，讓你無從確認結果；現在它們會乖乖留在 /tasks 列表中，直到你主動清理。當你重新接上一個已經停止的任務時，也不會再看著「工作階段啟動中」的空白畫面發呆，系統會立刻顯示它之前的 transcript（完整對話紀錄）。針對 daemon（躲在系統背景維持運作的服務），現在舊版的程式不會再自作主張，把新版產生的 worker（實際執行工作的行程）強行用舊版 binary（執行檔）重啟。在 Agent view（顯示代理工作狀態的介面）中，當你用 Ctrl+X 刪除 worktree（Git 讓同一個專案平行開啟多份副本的功能）時，它會正確處理被改過名字的分支，並且絕對不會摧毀你還沒推送的 commit（程式碼變更的存檔紀錄）；如果你選擇保留工作樹，工作階段的那一列也會被保留下來，重複使用的名字也會以當前基底重置。這就像你請助理去跑腿，以前他買完東西就人間蒸發，讓你不知道事情辦妥了沒，現在他會把收據和物品留在桌上等你核對；而你在不同房間同時進行的多個勞作專案，也不會因為清理走廊就被意外當成垃圾全部丟掉。
現在即使你開啟了略過權限或自動模式，系統也能精準攔截藏得很深的毀滅性指令，保護你的電腦不被意外清空。以前如果你在 auto mode（全自動模式）或啟用了 skip-permissions（讓系統不問你就直接跑指令的危險設定）時，如果像 rm -rf ~ 這種會把使用者資料夾全部刪除的災難性指令，是被包裝在 $()、反引號或 <() 這類指令替換語法（讓指令結果變成另一個指令輸入的寫法）裡面，系統的安全檢查可能會漏看而直接讓它過關。現在不管這些危險指令藏得有多深，系統都會把它揪出來，並且跟直接輸入危險指令一樣，強制停下來要求你手動確認。這並不是為了阻擋你正常的自動化工作，而是修補了安全漏洞。這就像是機場的安全檢查，以前如果危險物品被層層包裝在俄羅斯套娃裡，X 光機可能就只看最外層的盒子而讓它通關；現在安檢系統升級了，無論你包了多少層紙箱，裡面的違禁品都會被立刻掃描出來並扣留。
現在你在處理系統設定與查看用量時，不會再遇到莫名其妙的報錯或卡關。首先，系統不會再允許你在 background session（在背後默默運作的對話）裡彈出 /install-github-app 或 /mcp 的設定選單，避免你在看不見畫面的地方觸發需要互動的設定。當你的 MCP（Model Context Protocol，讓模型與外部工具溝通的協定）伺服器設定檔裡 URL（網址）留空時，系統現在會客氣地在清單顯示「未設定」，而不是直接丟出一個嚇人的設定錯誤。當你查看使用量時，如果 endpoint（提供數據的伺服器終端位址）遇到 rate-limited（因為太多人同時要求而被伺服器暫時擋下），系統現在會顯示你上一次成功抓取到的長條圖，並標註「截至某某時間」，而不是直接賞你一個錯誤畫面。另外，如果你使用 AWS SSO 登入 Bedrock，且 sso_region（單一登入的區域）與 Bedrock 服務區域不同，系統也不會再誤判並阻擋你，修復了之前總是報錯「找不到工作階段權杖」的退步問題。這就像去圖書館借書，以前如果你想查借閱紀錄時剛好系統繁忙，櫃台會直接拉下鐵門不理你；現在櫃台人員會拿出早上的備份清單告訴你當時的紀錄，而且也不會因為你的身分證戶籍地跟居住地不同就把你趕出大門。
You can now use your mouse to click through fullscreen menus, comply with corporate startup rules effortlessly, and enjoy new settings tailored for assistive tools and Vim users. If you are accustomed to the Vim editor, you can now map two-key combinations like jj to act as the Escape key for exiting insert mode. For users relying on accessibility tools, the new screen reader mode (a setting that renders the terminal interface as plain text so screen reading software can parse it) can be enabled via a command flag, an environment variable, or your settings file. For enterprise environments, the system now honors a process wrapper (an interceptor program that applies corporate rules before the actual application starts), ensuring every background task launch runs through your company's required launcher. It's like upgrading a vehicle's dashboard with more versatile controls—whether you prefer a standard steering wheel, a joystick, or specialized accessibility pedals, you can now drive exactly the way you want.
You no longer have to manually toggle fast mode back on every time you switch between different AI models. Fast mode (an operational state that trades a tiny bit of precision for blazing-fast response generation) is a great setting for saving wait time. Previously, if you had this mode enabled in your settings but temporarily switched to a model that didn't support it, the system would understandably turn it off. However, the frustrating part was that when you switched back to a supported model, the system would forget your preference and leave it disabled, forcing you to manually turn it back on. This amnesia has been fixed: as long as it is enabled in your settings, fast mode will now automatically restore itself the moment you select a compatible model. It is exactly like a car's sport mode—it might turn off automatically when you detour onto a rugged dirt road, but now, the moment you steer back onto the smooth highway, the car remembers your preference and kicks sport mode back on without you lifting a finger.
Your typed replies and ongoing tasks handed off to background agents will no longer vanish into thin air just because of a connection hiccup. A background agent (an AI assistant running long tasks silently out of sight) can sometimes fail to receive a message if the network drops. Previously, any carefully crafted reply you typed during such a failure would simply evaporate; now, the system securely saves your text and automatically delivers it as soon as the session restarts. Additionally, if the system auto-updated and replaced its underlying binary files while you had a session running, the background daemon (a hidden, always-running service that keeps things operating behind the scenes) used to panic, throwing a "Couldn't start" error and permanently locking you out of your session. This temporary disconnect is no longer a fatal error. It is like sliding a note under a coworker's door while they work in the next room: if the door was stuck, the note used to blow away, but now it waits safely on the mat. And even if the building management replaces the front door locks overnight, the coworker inside won't be permanently trapped in their office anymore.
You will no longer suffer a heart attack from a false "capacity full" warning when resuming a massive conversation right after a system update. The total amount of information an AI model can hold in its active memory during a single chat is called the context window (think of it as the model's short-term working memory). Previously, whenever the command-line interface finished a background auto-update, this capacity ceiling would briefly glitch and reset to an older, smaller 200k limit. If you were in the middle of a massive, long-context session, reconnecting would trigger a false red alert screaming "100% context used". This bug even caused the auto-compact indicator (a feature that condenses older chat history into a summary to free up space when you near the limit) to light up erratically. Both of these dashboard gauges have now been fixed and will hold their correct readings across updates. It is like driving a long-haul truck equipped with an extended fuel tank; previously, a quick pit stop for a software patch would make the dashboard glitch and measure your fuel against a tiny sedan's tank, terrifying you into thinking you were running on empty. Now, the fuel gauge reads your massive tank accurately at all times, no matter how many updates run in the background.
Your long-running tasks will no longer spectacularly crash just because a web server unexpectedly decided to close its connection. During routine network communications, servers sometimes need to perform maintenance or clean up connections, doing so by sending a GOAWAY signal (literally a farewell message declaring that the current line is shutting down) via HTTP/2 (a modern network protocol designed for more efficient data transfer). Previously, if your supervised tasks or background sessions happened to be mid-request when this sudden farewell arrived, the entire session would panic and crash on the spot, unable to handle the abrupt cutoff. The system has now been reinforced to understand this signal properly; it treats the closure as a manageable, recoverable disconnect rather than a fatal error. It is like talking to a customer service agent on the phone—previously, if the agent suddenly said, "Sorry, this line is undergoing maintenance," and hung up, the physical paperwork you were filling out would spontaneously combust in your hands. Now, you simply put the phone down, redial a different line, and continue filling out your forms completely undisturbed.
When you pipe Claude's output into other automated programs, you will no longer receive corrupted data fragments, nor will the system grossly misread your large configuration numbers. If you asked the system to output in stream-json (a format that streams data out piece by piece so other programs can consume it on the fly), massive responses used to get arbitrarily chopped off mid-pipe, leaving you with missing result messages and mangled data. Another hidden trap occurred when you set environment variables to limit output length: if you used scientific notation out of convenience (like typing 1e6 to mean one million), the system used to be notoriously bad at math. It would blindly read only the mantissa (the significant digit before the exponent e, which is just 1), silently shrinking your "one million" token allowance down to a microscopic "1". The data pipes have now been reinforced to deliver gigantic payloads fully intact, and the system finally understands scientific notation correctly. It's like hiring a courier to deliver a long model train: previously, they would drop half the carriages on the highway, and if you wrote "10 to the power of 6" grams on the manifest, the clerk would only read the "1" and charge you for a feather. Now, the entire train arrives flawlessly, and the shipping clerk finally knows how to read math.
Your computer will no longer freeze from memory exhaustion when Claude generates massive tables. Previously, if a Markdown (a lightweight formatting syntax for plain text) table exceeded 200 rows, the system would attempt to render the entire thing at once, stalling the display and devouring memory. Now, it smartly shows only the first 200 rows and appends a helpful "… N more rows" notice at the bottom. It is like opening a massive phone directory to the first page for you to read, rather than trying to physically unfold thousands of pages all at once to cram onto your small desk.
Your file processing tools now explain their failures accurately instead of throwing confusing or misleading errors. Previously, the Edit tool would flat-out refuse to modify a file if it had been altered after reading, even if your target text was still a unique match; now it safely edits it as long as the match holds. The other tools have also stopped bluffing: Read no longer misreports completely empty files as "shorter than offset" (a number specifying where to start reading), Grep won't silently lie that "No files found" when you feed it invalid regex patterns (a syntax using special characters to search for text patterns), and it correctly tallies the total in paginated (split across multiple pages) mode instead of under-reporting. Furthermore, Glob (a tool for matching file paths) will no longer crash with a cryptic error if the pattern, path, or working directory happens to contain a null byte (a special hidden zero-value character used by programs to mark the end of a string). It is like a librarian who finally learned to communicate properly: instead of claiming "you looked too far" when a shelf is simply empty, or shrugging "no such book" when you mispronounced the title rules, they now point out the exact issue.
You will finally see the real root causes of connection issues instead of chasing misleading ghosts. Previously, if the apiKeyHelper (a helper script that fetches your API key) failed, the system would silently retry about ten times before shrugging with a generic 401 (the standard HTTP code for unauthorized access); it now surfaces the script's actual error within three attempts. Furthermore, if your AWS Bedrock streaming requests (the connection method where the AI streams its answer word-by-word) were altered mid-flight by a corporate gateway (a network node that forwards traffic between services), it used to spit out a completely misleading "Truncated event message received" error. Now, it accurately identifies the content-type (a label indicating the format of network data) that went wrong and correctly points the finger at the proxy (the intermediary server handling your requests). It is like visiting the doctor and finally seeing the actual X-ray showing the exact fractured bone, rather than the receptionist just turning you away with a vague "you are not feeling well" note.
Your background automations and internal processes are now highly resilient and will no longer halt or spam you over minor hiccups. First, if your browser fails to open, typing /upgrade (the command to upgrade the system) now simply prints the upgrade URL instead of bizarrely throwing you back to a login screen. For Windows developers, processing stream-json (a format for programs to consume JSON data in small chunks) is now immune to completely crashing the session just because it encountered a blank line with CRLF (Carriage Return Line Feed, the two hidden characters Windows uses for line breaks). Furthermore, headless (running purely in the background without a visual interface) sessions will no longer hang permanently if a control request contains a malformed set_model (an internal directive used to switch the AI model) payload; the CLI now cleanly answers with an error. Finally, when you resume your work, any orphaned background tasks (leftover tasks abandoned after an interruption) will quietly collapse into a single clean summary rather than endlessly spamming "No completion record was found." It is like a train's autopilot receiving a massive upgrade: it no longer slams the emergency brakes just because a single leaf fell on the tracks or it received a slightly garbled radio signal, keeping your journey smooth and uninterrupted.
Remote connections and AI team collaborations are now instantly synchronized and communicative. Previously, when your Remote Control clients (the mobile or web interfaces you use to connect back to your host) attached to a terminal-hosted session (the core process running directly in your computer's black terminal window), you were kept completely blind to background agents and workflow progress until a task either started or stopped to trigger an update; it now synchronizes the exact state the moment you connect. Additionally, when using the Agent tool (the function used to summon helper AIs), if a subagent's (a subordinate AI called upon to handle specific tasks) tools list (a catalog telling the AI what commands it can use) resolved to entirely unrecognized garbage, the system would bizarrely launch it anyway with an entirely empty toolbox. Now, it smartly halts and returns a clear error explicitly naming the entries it couldn't recognize. It is like a boss calling the office while traveling and instantly getting a full status board, instead of awkward silence until someone finishes a job; and when that boss hands a new worker a list of requested tools that don't exist, the worker immediately speaks up saying "I can't find these wrenches," instead of showing up to the construction site empty-handed and standing around doing nothing.
Several annoying interface quirks have been squashed so the system no longer feeds you outdated maps. First, typing /usage (the command to check your token usage) will no longer paste stale cached bars (old graph data the system kept around to save effort) over your fresh API data; and if you edit settings in /mcp (the command to manage server connections), the system will properly reclassify placeholder servers (names keeping a reserved spot in the config file) instead of treating them as dead ends forever. If you are using SDK hosts (outer shell programs like Claude Desktop that run the AI), you can now successfully use the "Change directory" option during idle sessions (when you aren't actively prompting the AI), even if a background task is running, instead of getting a stubborn "A turn is in progress" rejection. The workflow save dialog will now correctly point to your CLAUDE_CONFIG_DIR (your specified primary configuration directory) instead of hardcoding a default path. Lastly, if you expand "Show all" in /release-notes (the command to read updates), the system will no longer aggressively inject the entire massive changelog (the update history) into the model's context (the AI's short-term memory limit), which previously bloated every single subsequent request you made. It is like a car navigation system that finally uses live traffic instead of yesterday's maps, allows you to change the radio station while waiting at a red light, and stops forcefully taping the entire vehicle owner's manual to your windshield.
When you keep Claude running for extended conversations or massive file processing, it will no longer gradually slow down or devour your system memory. Previously, several memory leaks plagued long sessions: pasted images in the agent view (the interface showing the AI's working state) would remain in memory for the screen's lifetime; MCP (Model Context Protocol, the standard for connecting models to external tools) servers would accumulate up to 64 MB of stderr (standard error output) per server; LSP (Language Server Protocol, the standard providing features like autocomplete) documents would stay open indefinitely (now capped at 50); and headless (operating without a visual interface) or SDK (Software Development Kit) sessions would grow unboundedly from large tool-result payloads. Additionally, async hook (code triggered by specific events) outputs were retained after backgrounding, and SDK sessions could lose initialized agents if a plugin (an add-on extending capabilities) refreshed before the client attached. Now, these issues are fixed: images are released after sending, and strict boundaries are set for various records. Furthermore, if you have many permission deny or ask rules, the system now pre-compiles and caches them, preventing multi-second delays per turn. When reading files with extremely long single lines using offset/limit (parameters specifying the start and end of a read), it now returns a clean error instead of blowing up memory by loading the entire line. It is like an office where workers used to indefinitely pile every read document and received fax on their desks until they could barely turn around; now they file or shred things immediately, keeping the workspace perpetually clear.
When you load up on tools or perform extensive file edits, the interface will now feel much snappier while consuming significantly fewer computer resources. Previously, when the agent updated the task list, it blindly re-rendered the entire UI (user interface), making typing feel sluggish; now it only redraws what actually changed. If you use many MCP (Model Context Protocol, the standard for external tool connections) tools, the system now caches the assembly of the tool-pool (the collection of tools available for the model to choose), speeding up tool rounds by up to seven times. For memory management, the file edit read cache is now strictly capped at 16 MB, rather than pinning up to 1,000 full files in memory. The system also actively prunes superseded file-history backups, significantly shrinking checkpoint (saved conversation state) disk usage and reducing transcript (complete conversation record) sizes by up to 79x in edit-heavy sessions. Additionally, resuming sessions with background agents or forks (independent processes branching off from an existing conversation) spawned from massive conversations now uses far less memory. It is like an auto mechanic who used to unpack and repack his entire toolbox every time he needed a wrench, while hoarding every scrap of replaced metal in the garage; now he keeps essential tools at hand and routinely throws out the scrap, keeping the shop running at peak efficiency.
You can now reliably trace completed background tasks and safely manage parallel working copies without fearing accidental data loss. Previously, background agents (AI assistants working silently behind the scenes) would vanish from the list the moment they finished, leaving you wondering if the job was done; they now stay put in /tasks until you explicitly clean them up. When you attach to a stopped background task, you will no longer stare at a blank "Session is starting" screen while it warms up; it instantly displays the previous conversation transcript (complete conversation record). Regarding daemons (services running quietly in the background), an older daemon will no longer stubbornly force workers (child processes executing tasks) spawned by a newer version to restart on the old binary (executable file). In the Agent view (the interface showing the AI's working state), when using Ctrl+X to delete a worktree (a Git feature that lets you open multiple parallel copies of the same project), it now correctly handles renamed branches and strictly protects your unpushed commits (saved records of code changes) from destruction. If a worktree is kept, its session row remains, and reused worktree names properly reset to the current base. It is like sending an assistant on an errand—they used to vanish upon returning so you never knew if they bought the items, but now they leave the receipt on your desk; and your multiple craft projects spread across different rooms will no longer be thrown in the trash just because someone was sweeping the hallway.
Even when running in auto or bypass modes, the system now accurately intercepts deeply hidden catastrophic commands, shielding your computer from accidental erasure. Previously, if you were in auto mode (fully automatic mode) or running with skip-permissions (a dangerous setting allowing the system to execute commands without asking), devastating commands like rm -rf ~ (which deletes your entire user folder) could slip right through the safety checks if they were wrapped inside command substitution syntax (methods like $(), backticks, or <() that feed one command's output into another). Now, no matter how deeply these dangerous commands are nested, the system extracts them and forces a manual confirmation prompt, matching the behavior of plainly written threats. This does not hinder ordinary automation; it simply closes a critical security loophole. It is like airport security: previously, if a dangerous item was packed inside a set of Russian nesting dolls, the X-ray machine might only check the outer shell and wave it through; with the upgraded scanners, no matter how many boxes you wrap it in, the contraband inside is instantly detected and seized.
You will no longer run into baffling errors or dead ends when managing system settings or checking your usage. First, the system no longer permits /install-github-app or /mcp settings menus to open within a background session (a conversation running invisibly behind the scenes), preventing interactive prompts from triggering where you cannot see them. When your MCP (Model Context Protocol, the standard for connecting models to external tools) server configuration has an empty URL (web address), the /mcp list now politely displays it as "not configured" rather than throwing an alarming config error. When checking your usage, if the endpoint (the server address providing the data) hits a rate-limit (being temporarily blocked by the server due to too many simultaneous requests), /usage now gracefully displays your last successfully fetched bar charts with an "as of" timestamp, instead of slapping you with an error screen. Additionally, if you use AWS SSO to log into Bedrock and your sso_region (the region for single sign-on) differs from your Bedrock service region, the system will no longer falsely block you with a "Session token not found" error, fixing a recent regression. It is like visiting a library: previously, if the computer system was busy when you asked for your checkout history, the clerk would just slam the window shut; now, they pull out the morning's backup ledger to show your status as of then, and they no longer kick you out just because the address on your ID does not match your current zip code.
Added screen reader mode: opt-in plain-text rendering for screen reader users. Run `claude --ax-screen-reader`, set CLAUDE_AX_SCREEN_READER=1, or add "axScreenReader": true to settings.
Added `vimInsertModeRemaps` setting: map two-key insert-mode sequences like `jj` to Escape in vim mode
Added `CLAUDE_CODE_PROCESS_WRAPPER`: agent view and the background service now honor a corporate launcher by running every Claude Code self-spawn through a required wrapper executable
Added mouse-click support for multi-select menus and "Other" input rows in fullscreen modeFixed fast mode staying off after switching back to a model that supports it — it now restores automatically when enabled in settingsFixed replies typed to a background agent being lost when delivery fails — the text is now saved and delivered when the session restarts
Fixed background-session attach failing permanently ("Couldn't start the background daemon") after an update replaced the binary a running `claude agents` process was launched fromFixed the context window (and auto-compact indicator) briefly resetting to 200k after the CLI auto-updates, causing a false "100% context used" when resuming long-context sessionsFixed supervised and background sessions crashing when a server closed an HTTP/2 connection with a GOAWAY while requests were in flightFixed truncated stream-json/JSON output and missing result message when piping large responses from `claude -p`
Fixed `CLAUDE_CODE_MAX_OUTPUT_TOKENS` and similar env vars silently using the mantissa of scientific-notation values (`1e6` became `1`)Fixed very large markdown tables stalling rendering or using excessive memory; tables over 200 rows show the first 200 with a "… N more rows" noticeFixed the Edit tool failing on files modified after reading when the target text still matches uniquely
Fixed Read reporting empty files as "shorter than offset", Grep silently returning "No files found" for invalid regex patterns, Grep count mode under-reporting totals when paginated, and Glob crashing with an unclear error when the pattern, path, or working directory contained a null byteFixed `apiKeyHelper` script failures being hidden behind a generic 401 after ~10 silent retries; the script's own error is now shown within 3 attempts
Fixed Bedrock streaming requests failing with a misleading "Truncated event message received" when a gateway transforms the response — the error now names the content-type and points at the proxyFixed `/upgrade` showing a login flow instead of the upgrade URL when the browser fails to open
Fixed stream-json input killing the session on blank CRLF or whitespace-only lines from Windows-style SDK hosts
Fixed headless stream-json sessions hanging permanently when a `control_request` carried a non-string `set_model` payload; the CLI now answers with an error response
Fixed repeated "No completion record was found" notices on session resume — orphaned background tasks now collapse into a single summaryFixed Remote Control clients attaching to a terminal-hosted session not seeing background agents and workflow progress until a task started or stopped
Fixed the Agent tool launching with no tools when a subagent's `tools` list resolves to nothing — it now returns a clear error naming the unrecognized entriesFixed `/usage` showing stale cached bars over fresher data, and `/mcp` not reclassifying placeholder servers after config edits
Fixed "Change directory" in SDK hosts (e.g. Claude Desktop) failing with "A turn is in progress" on idle sessions that have a running background task
Fixed the workflow save dialog showing `~/.claude/workflows/` instead of the `CLAUDE_CONFIG_DIR` location for user-scope saves
Fixed `/release-notes` adding the viewed notes to the model's context — "Show all" previously injected the entire changelog into every subsequent requestFixed a memory leak in the agent view where pasted images were retained for the screen's lifetime after sending peek replies
Fixed SDK sessions losing agents defined via the initialize request when a plugin refresh ran before the client attached
Fixed several memory leaks in long sessions: MCP stdio server stderr accumulating up to 64 MB per server, LSP documents staying open indefinitely (now LRU with 50-doc cap), async hook output retained after backgrounding, and unbounded growth in headless/SDK sessions from large tool-result payloads
Fixed a memory blowup when reading files with extremely long single lines using offset/limit — the read now returns a clean error instead of loading the whole line
Fixed multi-second per-turn slowdowns in sessions with many permission deny/ask rules — rule matchers are now compiled once and cachedImproved input responsiveness while agent task lists update — task updates no longer re-render the entire UI
Reduced per-tool-call CPU overhead in print/SDK sessions with many MCP tools by caching tool-pool assembly (up to 7x faster tool rounds at high tool counts)
Reduced memory usage by bounding the file edit read cache to 16 MB instead of pinning up to 1,000 full files
Reduced session transcript size (up to 79x in edit-heavy sessions) and bounded checkpoint disk usage by pruning superseded file-history backups
Reduced memory usage when resuming sessions with background agents or forks spawned from large conversationsCompleted background agents now stay listed in `/tasks` until cleanup instead of vanishing the moment they finish
Attaching to a stopped background agent now shows its transcript immediately while the session warms up, instead of a blank "Session is starting" screen
Background sessions: an older daemon no longer silently restarts workers spawned by a newer version onto the older binary
Agent view: Ctrl+X now deletes renamed-branch worktrees, never destroys unpushed commits, keeps the session row when a worktree is kept, and reused worktree names reset to the current baseCatastrophic removals (e.g. `rm -rf ~`) in commands containing `$(…)`/backticks/`<(…)` now prompt in `--dangerously-skip-permissions` and auto mode, matching the plain form`/install-github-app` and the `/mcp` settings menu no longer open in background sessions
MCP servers configured with an empty URL now show as "not configured" in `/mcp` instead of a config error
`/usage` now shows your last-known usage bars with an "as of" note when the usage endpoint is rate-limited, instead of an error screen
Fixed Bedrock auth failing with "Session token not found or invalid" for AWS SSO profiles whose sso_region differs from the Bedrock region (2.1.207 regression)