新功能
權限提示標明待審總數，指令與工具授權畫面統一
當 Claude 連續執行多項操作而跳出一連串授權詢問時，你現在能一眼掌握還剩幾項要確認。
權限提示在多筆請求堆疊時會顯示「2 of 5」等計數，讓你知道整體審核進度。網路擷取、技能、檔案讀取、沙盒網路、Claude in Chrome、工作流程腳本與筆記本編輯的授權面板，外觀都對齊檔案編輯提示。Bash、PowerShell 與 Monitor 指令在詢問時會以虛線框起，指令範圍更清楚。
繁體中文
新功能
全螢幕清單支援滑鼠點選，排版對齊且捲軸增設箭頭
在全螢幕下瀏覽落落長的項目時，清單操作與閱讀變得更加直覺。
「↑ N more」與「↓ N more」等折疊列現在支援滑鼠懸停、按下狀態與點擊直達列表端點。點擊「Thought for 4s」等收合列的文字間隙時，會正常展開內容而非僅反白。/artifacts、/mcp、/skills 與 /hooks（掛鉤，特定事件發生時被觸發的小程式）等清單中的詳細資訊皆統一對齊於同一欄位；沒有詳細資訊的列，例如名稱很長的操作列，不再把其他列的詳細資訊推向右側。捲軸新增可點擊或按住持續捲動的上下箭頭；在大多數清單中，省略列出現或消失時，捲軸不再跟著位移。
問題修復
憑證過期不再重複跳轉瀏覽器，精確標示登入狀態
同時開啟多個視窗或編輯器擴充功能時，帳號登入與憑證狀態不再打架。
當 gcpAuthRefresh 或 awsAuthRefresh 憑證過期時，多個行程不會再各自彈出瀏覽器登入頁面。macOS 工作階段在另一個視窗以 /login 登入成功後，即使殘留舊版 ~/.claude/.credentials.json 檔案，也不會誤判為未登入。claude auth status 針對主控台儲存的金鑰會如實回報為 api_key，VS Code 擴充功能也會正確識別為金鑰工作階段。/status 在同時存在 API 金鑰時，會將 Anthropic 設定檔明確標註為未使用。
問題修復
工作階段恢復不再遺失對話，修復非文字回傳引發的錯誤
遇到終端機當機或非預期中斷時，重新接關不再面臨對話進度遺失的窘境。
使用 claude --resume 或 --continue 恢復先前崩潰的工作階段（session）時，不會再遺失平行工具呼叫後的對話紀錄。工具或掛鉤回傳物件、數字或布林值等非純文字內容時，不再引發 API 400 錯誤；恢復的工作階段也適用。歷史紀錄極為龐大的雲端工作階段，容器不會在載入對話紀錄時提前停止而無法喚醒。附加檔名超長的檔案到雲端工作階段時，也能正常送達。
新功能
模型不可用時自動切換同級備援，重試次數改採單一上限
當 API 拒絕預設模型或模型別名對應的模型時，你不必每輪都手動換模型。
Claude Code 只重試一次，改用同等級的前一版模型備援（fallback）。備援模型與 --fallback-model 重試若無法高速運行，會以標準速度執行，並只在互動式工作階段顯示一次通知。若備援使上下文視窗從 1M 縮減至 200K token（AI 處理與計費的文字單位），備援通知與自動壓縮反覆觸發時的錯誤訊息都會標明容量變化。失敗的 API 請求改以單次模型呼叫為單位計算重試上限，預設設定下最多發送 14 次請求。僅持續一輪的備援也不會重複發送 Git 提交註記提醒。
問題修復
遠端操控配合組織政策自動斷線，排隊訊息於重啟後接續處理
透過遠端操控（Remote Control）跨裝置連線時，連線狀態與訊息傳遞更加可靠。
當組織政策停用遠端操控時，包含 claude remote-control 在內的工作階段會主動中斷連線並顯示通知。透過遠端連線傳送的附加檔案若未成功送達，Claude 會獲知，且檔案最後一次下載嘗試有時只剩 10 秒的問題已修好。在程式即將退出時抵達的訊息不會被誤標為已送達，而是保留在佇列中供下次啟動時回覆。
問題修復
隱藏錯誤日誌中的敏感金鑰，修正回報檔案的格式
在匯出對話紀錄或查看錯誤訊息時，機密資訊遮蔽變得更加滴水不漏。
MCP（讓 AI 連接外部工具的協定）錯誤訊息在 Bearer 或 Basic 位於鍵名前方時，不會再暴露憑證內容。百分比編碼的 Bearer token（授權令牌）不再只有部分被遮蔽。鍵名含有零寬空格等不可見字元的機密設定，在日誌與對話紀錄中會確實隱藏。URL 密碼若包含引號、括號、& 或第二個 @ 等標點，在日誌與對話紀錄中不會再殘留部分字元。若密碼在 ssh 網址中跨過 /，延伸至 [::1] 這類方括號主機，也會完整遮蔽。/feedback 儲存的壓縮檔在脫敏後，工作階段對話紀錄不再出現格式錯誤的 JSON 行。
新功能
MCP 伺服器重新驗證不再重複提醒，正常記錄工具用量
串接 MCP（讓 AI 連接外部工具的協定）伺服器時，連線握手與登入授權流程更為穩定。
無頭模式（沒有終端機視窗的執行方式）的工作階段重新驗證成功後，不再重複顯示提醒；這項修復適用於已啟用探索快取的情況。快取是把查過的結果暫存起來直接用。當伺服器停止舊版握手協定時，MCP 連接器不會再發生長達一天找不到工具的情況。重複發送的登入請求不會覆蓋尚未完成的驗證連結。MCP 伺服器仍在連線中或剛連上時，工具呼叫都會正確計入 /usage。主機重複啟用已連線伺服器時，SDK（供程式串接的開發工具包）與 -p 模式反應更快。
新功能
與子代理溝通更即時，操作背景對話紀錄前會先確認目標
派遣子代理在背景執行繁重任務時，即時發送訊息與檢視進度的體驗更加順暢。
在子代理視窗按下立即發送快捷鍵（Ctrl+Enter）時，會將子代理目前執行的指令移至背景，以便立刻讀取你的輸入。輸入給子代理的訊息在被讀取後，對話紀錄中不會再重複出現兩次。背景代理人回覆時，開頭不再多餘複述你剛剛說過的話。未註冊名稱的子代理交回工作時，會顯示代理人代稱而非原始任務識別碼。在檢視背景代理人或隊友的對話紀錄時輸入 /compact、/clear 或 /rewind，系統會跳出對話框標明目標並詢問確認，防止誤改主對話。
問題修復
修復子代理工具遺失與重複載入，精確反映審批狀態
當你啟動子代理處理專案時，背景環境的工具配置與生命週期更加健全。
在已啟用任務追蹤的工作階段中，前景子代理不會再偶爾遺失 TaskCreate、TaskGet 或 TodoWrite 等工具。透過工作樹（worktree，Git 讓同一專案平行開多份副本的功能）隔離建立的子代理，在首次讀取檔案時不會從副本重複載入專案的 CLAUDE.md 及其匯入設定。工作流程工具子代理在回應中途遇到幾分鐘的連線停頓時，不會再從初始提示詞被重頭重啟。等待使用者核准的背景工作，在狀態畫面上不再被錯誤標記為已完成。
問題修復
外掛安裝限制套件來源，清楚說明市集載入錯誤
管理與安裝外掛擴充時，套件相依性安全與錯誤排查資訊都更加清楚。
外掛安裝現在會拒絕來自 Git 儲存庫或本地資料夾的 npm 來源，且僅允許從套件註冊表（registry package）安裝相依套件。當系統拒絕載入特定外掛市集時，錯誤訊息會明確說明原因與修復步驟，不再僅顯示「not found」。在 /plugin 的 Discover 頁面中，市集名稱的引號顯示格式保持一致。伺服器發生短暫異常時，在 claude.ai 上啟用的外掛不會再於當前工作階段中突然消失。
新功能
外部編輯器精準定位游標行，斜線指令與設定選單更流暢
在終端機中撰寫長提示詞或調整外觀設定時，各項互動細節更加得心應手。
使用外部編輯器（Ctrl+G）時，支援行號參數的編輯器會直接開啟在游標所在的提示詞行數。安裝大量技能或外掛時，斜線指令的即時自動補齊更靈敏，說明文字改採單字前綴匹配。輸出風格選擇器會直接定位在目前使用的樣式，並在名稱下方顯示詳細說明；數字鍵不再選取輸出風格。主題選擇器改為符合終端機高度的可捲動清單，預覽畫面不再被推出螢幕外；數字鍵也不再選取主題。/hooks（掛鉤，特定事件發生時被觸發的小程式）現在依事件分組顯示於單一清單，只要按一次 Enter 即可檢視細節。詳細畫面的結尾也改說「this hook」，不再說「it」。
新功能
極簡模式限制連線與背景任務，提交前提示驗證技能
在乾淨環境執行任務或把關品質時，指令列與背景流程配合得更緊密。
專案或個人技能若有 verify 技能，除純改文件或測試外，Claude 提交前會收到先執行它的指引。啟用 --bare 旗標時僅連線指定的 MCP 伺服器、不發系統提醒且不開背景工作；逾時 Shell 指令直接停止而非轉入背景。閒置時發送提示詞，文字立即顯示為標準顏色而非灰色。按下立即發送鍵時，技能自帶的 Shell 指令會轉入背景而非被終止。/exit 移除工作樹時會先停止 Claude Code 在該工作樹啟動的伺服器與 Shell，避免 Windows 因檔案佔用無法刪除資料夾。/ultrareview 與 claude ultrareview 輸出不再附瀏覽器連結。
新功能
統一成品讀取與網路授權流程，精準計算閘道快取費用
讀取 claude.ai 成品或抓取網路資料時，權限詢問與計費更為準確。
WebFetch 讀取 claude.ai 成品連結時比照成品工具：工作階段開啟網路時不詢問，關閉時各問一次；自動模式不能代答僅你能確認的問題。網域安全檢查若遇速率限制，錯誤訊息會提醒 Claude 不要反覆重試。Claude apps 閘道原本把 1 小時提示快取（把算過的結果暫存起來直接用）誤算為 5 分鐘費率，現已修正。伺服器端工具串流也計入後續呼叫的輸入 token（計量文字的單位）。閘道 /protocol 頁面已對齊最新規格，並註明不應拒絕未知輸入。claude-api 技能的 Managed Agents 範例改用受限網路環境。
問題修復
Windows 解決資料夾信任大小寫不一致的判定問題
在 Windows 系統上使用背景代理人時，不再因為資料夾路徑大小寫不同而遭拒絕存取。
以前在 Windows 環境下，若資料夾信任紀錄儲存時使用了不同的大小寫形式，執行 claude --bg 或開啟代理人檢視畫面時，會被誤判為未受信任的資料夾而遭拒絕；現在已修復路徑判定邏輯，能正確識別既有的信任狀態並順利運作。
新功能
VS Code 新增回覆書籤與問題預覽，整合終端機與選取程式碼
在 VS Code 編輯器中使用 Claude 時，對話紀錄的整理與上下文檢視變得更加方便。
書籤功能讓你能將 Claude 的精彩回覆保存至側邊的 Bookmarks 面板，隨時固定檢視。在聊天面板回答問題卡片時，目前反白選項的畫面示意圖或程式碼片段，會顯示在選項旁或下方。回答完問題卡片後，對話中會新增 Questions 欄位完整保留問題與你的選擇紀錄。訊息下方新增專屬列，可直接點擊展開隨訊息一同傳送的終端機輸出、瀏覽器分頁、瀏覽器操作指令及選取的程式碼。
新功能
VS Code 狀態列常駐所有視窗，停止鍵僅結束當前輪次
在 VS Code 進行多工開發時，擴充功能的控制按鈕與分頁管理更符合操作直覺。
「✻ Claude Code」狀態列（status bar）圖示現在會在所有視窗顯示，即使未開啟任何檔案也能直接開啟。按下 Stop 按鈕或 Escape 鍵改為僅結束當前輪次的生成，背景代理人會繼續運行，且可於代理人地圖中單獨停止。當對話已在側邊欄開啟時，點選不會再於編輯區重複開啟同一份對話分頁，而是直接切換過去。設定對話框在輸出超過 1 MB 時不再未經確認就誤報儲存失敗。擴充功能無回應時卡在「Teleporting session…」的旋轉圖示問題已獲得修復。外掛管理視窗會說明停用的外掛為何仍因其他設定而啟用，也會說明外掛資料夾衝突。
問題修復
雲端工作階段精準反映排程與執行狀態，修復音訊預覽與授權卡片
在雲端環境中運行自動化例行排程或長時間任務時，狀態看板與操作介面更加精確。
當例行排程逾期尚未啟動時，頁面會顯示「Due」與預定時間，不再顯示過期的下次執行時間；若工作階段根本未啟動，狀態會如實標示為 Failed。工作階段（session）在 Claude 送出訊息轉為閒置後，使用者回覆問題卡片或核准工具呼叫能正常觸發後續回應。管理員設定中清除組織環境的建置指令稿後，新建立的雲端工作階段不會再沿用舊指令稿。點擊 Outputs 卡片中的音訊或影片檔案時，會直接進行播放而非跳出空白檔案搜尋。自託管環境管理頁面上的 Runner 選單不再於開啟數秒後自動關閉。
問題修復
Claude Tag 支援個別頻道花費上限，修復 Slack 記憶讀取與對話標題
在 Slack 標記 Claude 協作時，管理設定與對話延續性更加健全。
管理員可在花費上限頁點選「Add channel」，以頻道 ID 或 Slack 連結為包含私人頻道的任何頻道設定額度（quota）。無法使用預設 Sonnet 模型的組織，現在也能正常讀取長期記憶。企業網格管理員將頻道移到其他工作區後，即使首則貼文沒提到 Claude，頻道設定也不會遺失。在討論串回覆 Claude 剛問的問題時，不再偶爾重開新機器而遺失未推送進度。大型組織的花費上限頁會正常顯示公開頻道名稱而非原始代碼；Slack 發起的工作階段在 claude.ai 上的標題也改為純文字，去除跳脫字元與使用者識別碼。
English
FEATURE
Permission prompts show total pending count and unify command and tool authorization screens
When Claude performs multiple operations in a row and triggers a series of authorization requests, you can now see at a glance how many items remain to be confirmed.
Permission prompts now display counts such as "2 of 5" when multiple requests stack up, keeping you informed of the overall review progress. Authorization panels for web fetching, skills, file reading, sandbox network, Claude in Chrome, workflow scripts, and notebook editing all align their appearance with file edit prompts. Bash, PowerShell, and Monitor commands are framed with dashed lines during prompts, making the command boundaries clearer.
FEATURE
Fullscreen lists support mouse clicks with aligned layouts and scrollbar arrows
Browsing lengthy items in fullscreen mode is now more intuitive to navigate and read.
Collapsed rows like "↑ N more" and "↓ N more" now support mouse hover, pressed states, and clicks that navigate directly to the list ends. Clicking the text gaps of collapsed rows like "Thought for 4s" now properly expands content instead of merely highlighting text. Detailed information across lists such as /artifacts, /mcp, /skills, and /hooks is aligned in a single column. Rows without details, such as operation rows with long names, no longer push the details of other rows to the right. Scrollbars gain up and down arrows that can be clicked or held for continuous scrolling. In most lists, the scrollbar no longer shifts when ellipsis rows appear or disappear.
BUGFIX
Expired credentials avoid duplicate browser launches and accurately reflect login status
When running multiple windows or editor extensions simultaneously, account logins and credential states no longer conflict.
When gcpAuthRefresh or awsAuthRefresh credentials expire, multiple processes will no longer each open separate browser login pages. After a macOS session logs in successfully via /login in another window, it is no longer misidentified as logged out even if a legacy ~/.claude/.credentials.json file remains. claude auth status accurately reports console-stored keys as api_key, and the VS Code extension correctly identifies them as key sessions. When an API key is present, /status explicitly marks Anthropic profiles as unused.
BUGFIX
Session resumption preserves conversations and fixes errors from non-text returns
When encountering terminal crashes or unexpected interruptions, resuming work no longer risks losing conversation progress.
Resuming a previously crashed session with claude --resume or --continue will no longer lose conversation history following parallel tool calls. Non-text returns from tools or hooks—such as objects, numbers, or booleans—no longer trigger API 400 errors, which also applies to resumed sessions. For cloud sessions with massive histories, containers no longer terminate prematurely during transcript loading and fail to wake. Files with extremely long filenames attached to cloud sessions can now also be delivered properly.
FEATURE
Model unavailability triggers same-tier fallback with a unified retry limit
When an API rejects the default model or the model mapped by an alias, you no longer need to switch models manually every turn.
Claude Code retries only once, falling back to an earlier model of the same tier. If the fallback model and --fallback-model retries cannot run at high speed, they execute at standard speed and display a notification only once in interactive sessions. If a fallback reduces the context window from 1M to 200K tokens, both the fallback notification and error messages from repeated auto-compaction will state the capacity change. Failed API requests now calculate their retry limit per individual model call, sending at most 14 requests under default settings. Single-turn fallbacks will also not send duplicate Git commit note reminders.
BUGFIX
Remote Control disconnects on organization policy and queued messages resume on restart
Connecting across devices via Remote Control is now more dependable for connection states and message delivery.
When an organization policy disables Remote Control, sessions including claude remote-control proactively disconnect and display a notification. Claude is informed if attached files sent over remote connections fail to deliver, and an issue where the final download attempt sometimes had only 10 seconds remaining is resolved. Messages arriving right as the program exits are no longer mislabeled as delivered, remaining in the queue to be answered on the next run.
BUGFIX
Sensitive keys masked in error logs and report file formats fixed
When exporting transcripts or inspecting error messages, secret redaction is now far more airtight.
MCP error messages no longer expose credentials when Bearer or Basic precedes key names. Percent-encoded Bearer tokens are no longer only partially redacted. Secret settings with key names containing invisible characters like zero-width spaces are properly hidden in logs and transcripts. When URL passwords contain punctuation such as quotes, parentheses, &, or a second @, no partial characters linger in logs and transcripts. Passwords extending across / to bracketed hosts like [::1] in ssh URLs are also fully redacted. Responding to /feedback, redacted archives no longer produce malformed JSON lines in session transcripts.
FEATURE
MCP server re-authentication avoids duplicate prompts and logs tool usage properly
Connecting to MCP servers now offers more stable connection handshakes and login authorization workflows.
Headless sessions no longer display redundant prompts after successful re-authentication, a fix that applies when discovery caching is enabled. When servers deprecate legacy handshake protocols, MCP connectors will no longer lose track of tools for up to a day. Duplicate login requests will not overwrite pending verification links. Tool calls are accurately counted in /usage while MCP servers are connecting or just connected. When a host re-enables an already connected server, both the SDK and -p mode respond faster.
FEATURE
Subagent communication is more responsive and background transcript actions require confirmation
Dispatching subagents to handle heavy background tasks now provides a smoother experience when sending immediate messages and tracking progress.
Pressing the immediate send shortcut (Ctrl+Enter) in a subagent window moves the subagent's running command to the background to read your input right away. Messages typed to subagents no longer appear twice in the transcript after being read. When background agents reply, they no longer needlessly repeat what you just said. Subagents without registered names display an agent alias instead of raw task IDs when yielding work. Typing /compact, /clear, or /rewind while viewing a background agent or teammate transcript opens a dialog naming the target to confirm, preventing accidental modifications to the main conversation.
BUGFIX
Subagent tool loss and duplicate loading fixed while approval statuses reflect accurately
Launching subagents to work on projects now offers a sturdier tool configuration and lifecycle in background environments.
In sessions with task tracking enabled, foreground subagents will no longer occasionally lose tools like TaskCreate, TaskGet, or TodoWrite. Subagents created with worktree isolation will not redundantly reload the project's CLAUDE.md and imported settings from the copy on the first file read. Workflow tool subagents experiencing a multi-minute stall mid-response will no longer restart from scratch using the initial prompt. Background tasks awaiting user approval are no longer falsely marked as completed on status screens.
BUGFIX
Plugin Installation Restricts Package Sources and Clarifies Marketplace Loading Errors
When managing and installing plugin extensions, package dependency security and troubleshooting information are much clearer.
Plugin installation now rejects npm sources from Git repositories or local folders, allowing dependencies to be installed only from registry packages. When the system rejects loading a specific plugin marketplace, error messages explicitly explain the cause and remediation steps instead of merely showing "not found". On the /plugin Discover page, quotation marks around marketplace names maintain consistent formatting. Plugins enabled on claude.ai no longer abruptly disappear during the active session when brief server hiccups occur.
FEATURE
External Editor Precision Line Navigation and Smoother Slash Commands and Settings Menus
Writing long prompts or tweaking appearance settings in the terminal now feels significantly more responsive.
When using an external editor (Ctrl+G), editors supporting line number arguments open directly at the prompt line where your cursor was positioned. When many skills or plugins are installed, slash command autocomplete is more responsive, matching description text by word prefix. The output style picker now focuses directly on the active style with detailed descriptions shown beneath each name; numeric keys no longer select output styles. The theme picker is now a scrollable list sized to terminal height so previews are not pushed off-screen, and numeric keys no longer select themes. /hooks is now grouped by event in a single list, allowing you to inspect details with a single Enter keystroke. The detailed view ending now says "this hook" instead of "it".
FEATURE
Bare Mode Limits Connections and Background Tasks while Guiding Skill Verification Before Commit
Running tasks or enforcing quality in clean environments now integrates command-line tools and background workflows much tighter.
If a project or personal skill includes a verify skill, Claude is prompted to run it before committing changes, except for pure documentation or test edits. Enabling the --bare flag connects only to specified MCP servers, sends the model no system reminders, and skips launching background tasks; timed-out shell commands stop immediately instead of moving to the background. Submitting prompts while idle instantly displays text in standard coloring instead of gray. Pressing the instant-send key sends skill-invoked shell commands to the background rather than terminating them. Removing a worktree via /exit now terminates servers and shells started by Claude Code inside that worktree first, preventing Windows folder deletion failures caused by locked files. Output from /ultrareview and claude ultrareview no longer includes browser links.
FEATURE
Unified Artifact Reading and Web Permissions with Precise Gateway Cache Billing
Reading claude.ai artifacts or fetching web data now features more accurate permission prompts and billing calculations.
When WebFetch reads claude.ai artifact links, it follows artifact tool rules: no prompts appear when session web access is enabled, and one prompt per artifact is shown when disabled; auto-mode cannot answer prompts that require your direct confirmation. When domain safety checks hit rate limits, error messages advise Claude against repeated retries. The Claude apps gateway previously miscalculated 1-hour prompt caching at the 5-minute rate, which is now fixed. Server-side tool streaming now counts toward input tokens for subsequent calls. The gateway /protocol page now aligns with the latest specification and notes that unknown inputs should not be rejected. The Managed Agents example in the claude-api skill now uses a restricted network environment.
BUGFIX
Windows Resolves Case Sensitivity Mismatches in Folder Trust Checks
Running background agents on Windows no longer encounters access denials caused by folder path casing differences.
Previously on Windows, if stored folder trust records used inconsistent casing, running claude --bg or opening agent views would misclassify the directory as untrusted and deny access; path evaluation logic is now fixed to correctly recognize existing trust states.
FEATURE
VS Code Adds Response Bookmarks, Question Previews, and Integrated Context Drawers
Organizing chat history and inspecting context in the VS Code editor is now significantly more convenient.
The bookmark feature lets you save standout Claude responses to the sidebar Bookmarks panel for pinned viewing at any time. When answering question cards in the chat panel, screenshots or code snippets for the currently highlighted option appear beside or beneath that choice. Once answered, a Questions section is appended to the conversation to preserve the prompt and your selection history. A dedicated bar beneath messages lets you click to expand attached terminal outputs, browser tabs, browser actions, and selected code.
FEATURE
VS Code Status Bar Appears Across All Windows and Stop Action Halts Only Active Turn
Multitasking in VS Code now offers more intuitive extension controls and tab management.
The "✻ Claude Code" status bar icon now displays across all windows, allowing you to open the interface even without any active files. Pressing the Stop button or Escape key now cancels only the active generation turn; background agents continue running and can be stopped individually from the agent map. When a session is already open in the sidebar, clicking it switches to the conversation in the sidebar instead of opening a duplicate tab in the editor area. Settings dialogs no longer falsely report save failures without confirmation when output exceeds 1 MB. An issue where unresponsive extensions remained stuck on a spinning "Teleporting session…" prompt is resolved. The plugin management view now clarifies why disabled plugins remain active due to other settings and explains plugin folder conflicts.
BUGFIX
Cloud Sessions Accurately Reflect Schedule and Run States with Audio Preview and Card Fixes
Running automated schedules or long-running tasks in cloud environments now features more accurate dashboards and controls.
When a scheduled routine is overdue to run, the page displays "Due" with its scheduled time rather than an outdated next run time; if a session fails to start altogether, its status is accurately marked as Failed. After Claude sends a message and goes idle, replying to a question card or approving a tool call reliably triggers subsequent responses. Clearing the organization environment build script in admin settings ensures newly created cloud sessions no longer inherit obsolete scripts. Clicking audio or video files inside the Outputs card now plays media directly rather than opening an empty file search. The runner menu on the self-hosted management page no longer closes automatically seconds after opening.
BUGFIX
Claude Tag Supports Per-Channel Spending Limits with Fixes for Slack Memory and Session Titles
Collaborating with Claude via tags in Slack features sounder administrative controls and conversation continuity.
Admins can click "Add channel" on the spend limits page to assign quotas by channel ID or Slack link for any channel, including private channels. Organizations without access to the default Sonnet model can now retrieve long-term memory normally. When Enterprise Grid admins move channels between workspaces, channel settings remain intact even if the initial post does not mention Claude. Responding to Claude's questions in a thread no longer occasionally spawns a new machine and loses unpushed progress. Spend limit pages for large organizations properly display public channel names instead of raw IDs; session titles initiated from Slack on claude.ai are converted to plain text, stripping escape characters and user IDs.
原始 CHANGELOG
Original changelog權限提示標明待審總數，指令與工具授權畫面統一
Added a count such as "2 of 5" to the permission prompt when several permission requests stack up
Improved fetch, skill, file read, sandbox network, Claude in Chrome, workflow script and notebook edit permission prompts to match the look of file edit prompts
Improved Bash, PowerShell and Monitor permission prompts to show the command between dashed lines, matching file edit promptsOriginal changelog全螢幕清單支援滑鼠點選，排版對齊且捲軸增設箭頭
Added mouse support for the "N more" rows of lists in fullscreen mode: click one to jump to that end of the list, with hover and pressed states
Fixed a click on the space between words of a collapsed row (such as "Thought for 4s") highlighting the row without expanding it in fullscreen mode
Fixed a row with no details, such as an action row with a long name, pushing every other row's details to the right in list screens
Improved list scrollbars in fullscreen mode: in most lists the bar no longer shifts as the "N more" rows come and go, and it now has ↑/↓ arrows you can click or hold to scroll
Changed list screens (`/artifacts`, `/mcp`, `/skills`, `/hooks` and others) to always line up each row's details in one column after the names
Changed the overflow rows of lists to read "↑ N more" / "↓ N more" instead of "N more above" / "N more below"Original changelog憑證過期不再重複跳轉瀏覽器，精確標示登入狀態
Fixed several Claude Code processes and IDE extensions each opening a login browser when gcpAuthRefresh or awsAuthRefresh credentials expire
Fixed macOS sessions still showing "Not logged in" or "Login expired" after `/login` succeeds in another Claude Code window when a leftover `~/.claude/.credentials.json` exists
Fixed `claude auth status` reporting a Console sign-in's stored API key as `claude.ai`; it now reports `api_key`, and the VS Code extension treats that session as an API key session
Fixed `/status` listing an Anthropic profile beside an API key as if both were in effect; the profile is now marked as not in useOriginal changelog工作階段恢復不再遺失對話，修復非文字回傳引發的錯誤
Fixed `claude --resume` and `--continue` sometimes losing every turn after a batch of parallel tool calls when the earlier session crashed or was killed
Fixed API 400 errors after a tool or hook returned an object, number or boolean instead of text, including in resumed sessions
Fixed cloud sessions with very large histories never waking up because the container was stopped while the transcript was still loading
Fixed files with very long names not reaching a cloud session when attached to itOriginal changelog模型不可用時自動切換同級備援，重試次數改採單一上限
Fixed every turn failing when the Anthropic API refuses the model your default or a model alias resolves to: Claude Code now retries once on the previous model of the same tier
Fixed refusal and `--fallback-model` retries failing when the fallback model can't run fast; they now run at standard speed, with a one-time notice in interactive sessions
Fixed the commit attribution reminder being re-sent inside tool output when a model fallback lasts only one turn
Improved the model fallback notice and the autocompact-thrashing error to say when a fallback dropped the context window from 1M to 200K tokens
Changed how failed API requests are retried: one limit now covers a whole model call, so with the default retry settings a failing call sends at most 14 requestsOriginal changelog遠端操控配合組織政策自動斷線，排隊訊息於重啟後接續處理
Fixed Remote Control sessions (including `claude remote-control`) staying connected after your organization's policy turns Remote Control off; they now disconnect with a notice
Fixed Claude not being told when a file attached to a message sent over Remote Control did not arrive, and a file sometimes getting only 10 seconds for its last download try
Fixed a Remote Control message that arrived while Claude Code was exiting being marked delivered and then never answered; it now stays queued for the session's next runOriginal changelog隱藏錯誤日誌中的敏感金鑰，修正回報檔案的格式
Fixed MCP error messages showing a credential's value when "Bearer" or "Basic" came before its key name
Fixed percent-encoded Bearer tokens being only partly masked in error messages
Fixed redacted logs and transcripts showing a secret whose key name has an invisible character inside, such as a zero-width space
Fixed logs and transcripts showing part of a URL password that contains punctuation such as `)`, quotes, `]`, `&` or a second `@`, or that runs past a `/` to a bracketed host such as `[::1]` in an ssh URL
Fixed the session transcript in the zip that `/feedback` saves to disk containing invalid JSON lines after secret redactionOriginal changelogMCP 伺服器重新驗證不再重複提醒，正常記錄工具用量
Fixed headless sessions repeating the "MCP servers require authentication" reminder after a successful re-authentication when the MCP discovery cache is enabled
Fixed MCP connectors listing no tools for up to a day after their server dropped the older MCP handshake
Fixed a repeat MCP sign-in request from Claude replacing the pending sign-in link, which could stop that link from working
Fixed `/usage` not crediting an MCP server for a tool call made while that server was still connecting or had only just connected, such as right after a restart
Improved responsiveness of SDK and `-p` sessions when a host re-sends an MCP server enable for a server that is already connectedOriginal changelog與子代理溝通更即時，操作背景對話紀錄前會先確認目標
Fixed a message typed into a running subagent showing twice in its transcript after the subagent read it
Fixed subagent hand-back messages showing a raw task id instead of the agent's name when the subagent had no registered name
Fixed `/compact`, `/clear`, and `/rewind` typed while viewing a background agent's or teammate's transcript silently acting on the main conversation: a dialog now names the target and asks first
Improved send now (ctrl+enter) in a subagent's view: it now moves the subagent's running command to the background so your message is read right away
Improved replies from background agents to your messages so they no longer open with a separate recap of what you saidOriginal changelog修復子代理工具遺失與重複載入，精確反映審批狀態
Fixed foreground subagents sometimes missing the task-tracking tools (TaskCreate/Get/Update/List, TodoWrite) in sessions that have them enabled
Fixed subagents spawned with worktree isolation loading the project CLAUDE.md and its imports a second time from the worktree copy on their first file read
Fixed Workflow tool subagents being restarted from their original prompt when a connection stalled for a few minutes mid-response
Fixed background jobs showing done while waiting for your approvalOriginal changelog外掛安裝限制套件來源，清楚說明市集載入錯誤
Fixed plugins enabled on claude.ai occasionally going missing from Claude Code for a session after a transient server error
Fixed plugin errors for a marketplace Claude Code refuses to load: they now say why and how to fix it instead of "not found"
Fixed `/plugin`'s Discover tab showing a marketplace name unquoted in its "Checking … for new plugins" line when its rows already show that name in quotes
Changed plugin installs to refuse npm sources that are git repositories or folders, and to install plugin dependencies only from registry packagesOriginal changelog外部編輯器精準定位游標行，斜線指令與設定選單更流暢
Improved the external editor (Ctrl+G): editors that take a line number now open on the line your cursor is on in the prompt
Improved slash command suggestion responsiveness while typing when many skills or plugin commands are installed; command descriptions now match by word prefix
Improved the output style picker: it now opens on your current style instead of Default, with each style's description on the line under its name; number keys no longer pick a style
Improved `/hooks`: the closing line of a hook's detail screen now says "this hook" instead of "it"
Changed `/hooks` to open on one list of your configured hooks grouped by event, so viewing a hook takes one Enter instead of three
Changed the theme picker to a scrolling list that fits your terminal instead of pushing the preview off screen; number keys no longer pick a themeOriginal changelog極簡模式限制連線與背景任務，提交前提示驗證技能
Improved commit guidance: when your project or user skills include one named `verify`, Claude is now told to run it right before committing, except for docs-only and tests-only commits
Changed prompts sent while nothing is running or queued to show in the normal text color right away instead of gray
Changed `--bare` to connect only the MCP servers named on the command line, send the model no system reminders, and start no background tasks; under `--bare`, a shell command that reaches its timeout now stops instead of moving to the background
Changed the send-now key (ctrl+enter) to move a skill's own shell command to the background instead of ending it
Changed `/exit`'s Remove worktree to run after Claude Code stops the servers and shells it started there, which on Windows could keep the folder from being deleted
Removed the browser link from `/ultrareview` and `claude ultrareview` outputOriginal changelog統一成品讀取與網路授權流程，精準計算閘道快取費用
Fixed the Claude apps gateway's spend meter pricing 1-hour prompt cache writes at the cheaper 5-minute rate, and counting only the first model call's input tokens on streamed turns that run a server-side tool such as web search
Improved claude.ai artifact link reads: WebFetch now asks the same questions as the Artifact tool's read (no artifact prompt while the session's network access is on, one per artifact while it is off), and an auto-mode yes no longer counts where only you can answer
Improved the protocol page a Claude apps gateway serves at `/protocol`: it now says not to reject unknown input and matches what Claude Code sends today
Changed the WebFetch error for a rate-limited domain safety check to tell Claude not to retry it in a loop
Changed the `claude-api` skill's Managed Agents examples to create environments with limited networkingOriginal changelogWindows 解決資料夾信任大小寫不一致的判定問題
Windows: Fixed `claude --bg` and the agents view refusing a folder that `claude` already trusts when its trust record was saved with different letter caseOriginal changelogVS Code 新增回覆書籤與問題預覽，整合終端機與選取程式碼
[VSCode] Added bookmarks: save Claude's responses and keep them in view in a Bookmarks side panel
[VSCode] Added the questions Claude asks and your answers to the conversation: after you answer a question card, a Questions row shows each question with your picks
[VSCode] Added option previews to question cards in the chat panel: the highlighted choice's mockup or snippet shows beside or under the options
[VSCode] Added rows under a message that open to the terminal output, browser tab, browser instructions and selected code sent with itOriginal changelogVS Code 狀態列常駐所有視窗，停止鍵僅結束當前輪次
[VSCode] Fixed a second copy of a conversation opening in a tab when it was already open in the side bar; the side bar now switches to it
[VSCode] Fixed settings dialogs reporting a failed save, without re-checking, when Claude Code printed more than 1 MB of output
[VSCode] Fixed an endless "Teleporting session…" spinner when the extension stops responding
[VSCode] Improved the Manage plugins dialog: it says when a turned-off plugin is still on because of other settings, and explains a plugin folder clash
[VSCode] Changed Stop and Escape to end only the current turn; background agents keep running and can be stopped one by one from the agent map
[VSCode] Changed the "✻ Claude Code" status bar item to show in every window, so you can open Claude when no file is openOriginal changelog雲端工作階段精準反映排程與執行狀態，修復音訊預覽與授權卡片
[Cloud sessions] Fixed an answered question card or approved tool call getting no reply when the session had gone idle after Claude sent a message
[Cloud sessions] Fixed clearing an organization environment's setup script in admin settings leaving new cloud sessions still running the old script
[Cloud sessions] Fixed the Runner actions menu on the self-hosted environments admin page closing on its own a few seconds after it opened
[Cloud sessions] Fixed routine runs whose cloud session never started showing as Succeeded in the Runs pane, the routine's page and the sidebar; they now show as Failed
[Cloud sessions] Fixed clicking an audio or video file in a cloud session's Outputs card opening an empty file search instead of playing the file
[Cloud sessions] Changed a routine's page to read "Due" with the scheduled time, instead of a next run time in the past, when a scheduled run is late and hasn't startedOriginal changelogClaude Tag 支援個別頻道花費上限，修復 Slack 記憶讀取與對話標題
[Claude Tag] Added an Add channel button to Claude Tag's spend limits page in admin settings, so a limit can be set on any channel, including a private one, from its channel ID or Slack link
[Claude Tag] Fixed memory recall finding nothing in organizations that cannot use the default Sonnet model
[Claude Tag] Fixed a Slack channel losing its Claude settings when an Enterprise Grid admin moves it to another workspace and the first post afterward doesn't mention Claude
[Claude Tag] Fixed Claude in a Slack thread occasionally starting over on a fresh machine, losing work it hadn't pushed, when your reply answered a question it had just asked
[Claude Tag] Fixed public channel names on Claude Tag's spend limits page in admin settings showing as raw Slack IDs in larger organizations
[Claude Tag] Improved the titles of sessions started from Slack as shown on claude.ai: they now read as the words you typed, without Slack user IDs or escape codes