Claude Code 2.1.285：終端機指令直開桌面版，直接接續既有對話

繁體中文

新功能

終端機指令直開桌面版，直接接續既有對話

在黑底白字命令列工作到一半想換到視窗介面，現在只要一行指令就能切換。你在終端機使用 Claude Code 時，輸入 claude --desktop 就能直接在當前目錄啟動桌面版應用程式。若是想接續剛剛聊到一半的工作階段，只要加上 --continue 或指定代碼 --resume <id>，系統就會把既有對話完整搬進視窗中。

新功能

外掛安裝支援預先設定，MCP 伺服器即裝即用

安裝擴充外掛時如果每次都要事後手動翻找選單填寫參數，流程往往被打斷。現在透過 claude plugin configure <外掛名稱> 指令，你可以直接查看該外掛的所有設定項與未填欄位，也能搭配 --values-stdin 從標準輸入（stdin）一次灌入數值。安裝打包好 MCP（讓 AI 連接外部工具的協定）伺服器的外掛時，支援在 claude plugin install --config 後方直接加上 <伺服器>.<鍵>=<值> 帶入參數。過去如果外掛內建的伺服器缺少設定，系統常會默默跳過執行；現在不論在安裝提示或外掛清單中，都會主動提醒你前往設定。

新功能

外掛辨識更加嚴謹，SSH 與市集網址驗證更清楚

管理社群外掛時，名稱的細微差異可能讓安裝或開關動到別的外掛。以前透過 SSH 下載外掛或市集清單，會忽略環境變數 GIT_SSH 與 Git 設定 core.sshCommand 指定的程式，現在會套用。兩個外掛的識別名稱若只差 .、-、@，或在 macOS、Windows 上只差大小寫，安裝時可能誤用已安裝外掛的快取（把查過的結果暫存起來直接用）或資料資料夾；現在會拒絕安裝。以完整的 name@marketplace 名稱啟用或停用外掛，也不會改到另一種大小寫的設定。外掛、市集及目前版本庫遠端的 Git 網址都會接受更完整的檢查；市集網址被拒時，畫面會說明原因。

新功能

企業政策集中管控，專案設定不得覆寫安全沙盒

公司要限制 Claude Code 可走哪些 API 供應管道時，管理員可在集中管理設定用 allowedProviders 指定範圍。專案設定不能放寬或關閉管理員要求的沙盒（限制程式可存取範圍的環境），也不能替換受管理封鎖清單保護的代理伺服器、擴大嚴格允許清單，或重新開放被禁止讀取的檔案。透過 ANTHROPIC_AUTH_TOKEN 登入 Anthropic API 的工作階段，現在會載入組織政策。團隊、企業，以及無法判定登入方案的工作階段，若啟動時未能載入政策，會先暫停 WebFetch。作業系統拒讀管理設定檔時，Claude Code 會警告並在沒有該檔政策的情況下啟動；其他讀取錯誤或檔案無法解析，則會停止工作階段。

新功能

支援停用網頁擷取，網址安全檢查錯誤回報更精準

當你不希望 Claude Code 自行抓取網頁資料，可用環境變數 CLAUDE_CODE_DISABLE_WEB_FETCH 關閉 WebFetch（網頁擷取工具）。網址安全檢查的錯誤訊息也修正了：過去網域檢查遇到速率限制（短時間請求太多而被暫時限制），可能被誤報為網路問題或企業政策封鎖；現在會顯示請求頻率受限。

新功能

子代理完整繼承父層權限，背景執行回報不灌水

讓子代理（代理人分派的子工作單元）協助處理任務時，結果與權限現在更容易掌握。使用 CLAUDE_CODE_FORK_SUBAGENT=1 執行 claude -p，子代理再呼叫 Agent 時會在前台執行，因而拿得到下一層代理人的結果。使用 claude -p --permission-prompt-tool 時，背景子代理的權限請求會交給指定的審批工具，不再自動被拒。分支子代理會繼承工作階段中父層的權限模式，包括計畫模式與 dontAsk，也不能自行離開計畫模式。自動模式下，背景子代理回報後不再要求第二次回覆；子代理把報告交回呼叫者，也會直接結束。

新功能

自動模式支援第三方平台，指令執行預設更俐落

若你串接其他模型供應商，或關閉了遙測（傳回使用狀況的診斷資料），現在也會收到切換權限的一次性提示。這項提示會在使用者設定為其他模式時出現；自動模式（讓 Claude 依權限設定連續執行的模式）是可選的新預設。在這些環境下執行 claude -p 或 Python Agent SDK（供程式呼叫代理人的工具組）工作階段時，若未設定權限模式，會像互動式工作階段一樣直接以自動模式啟動，亦可用 --permission-mode 覆寫。

新功能

背景命令設定逾時上限，工作階段管理與切換更順手

把 Bash 或 PowerShell 命令放到背景執行時，現在預設 30 分鐘後停止，最長可設為 2 小時；命令因逾時停止，Claude 會收到通知。/tasks 會把 Claude Code 自己執行的背景工作收在「System tasks」一列，按 Enter 可展開。背景中的工作階段現在可用 /resume 或 claude --resume 開啟；使用 claude --resume <id> "prompt" 時，附帶文字才會送成下一輪對話。背景工作階段，以及由 Claude Code 工具啟動的工作階段，不能開啟自動記憶，但仍可將它關閉。從 claude agents 回覆等待權限的背景工作階段，也不會意外核准待審命令。

新功能

遠端控制訊息不遺漏，插話提問能掌握進行中進度

從其他裝置透過遠端控制送出訊息和附件時，檔案若因網路錯誤、逾時或伺服器錯誤下載失敗，現在最多會重試兩次。排隊訊息會等 Claude 開始處理才標為已讀；終端機若先關閉，未處理的訊息會在下次恢復工作階段時送達。在 Claude Desktop 等應用程式託管的工作階段，使用 /btw 問旁支問題，現在能看到正在進行的回合，不再只看到上一輪。遠端控制客戶端也不能執行 /claude-api。

新功能

Chrome 連線辨識主機名稱，設定指引更明確

讓 AI 操作瀏覽器進行網頁自動化時，跨電腦連線的管理介面變得更容易辨識。本機端程式現在會主動向瀏覽器回報電腦名稱，多台裝置連線時不再只顯示籠統的序號標籤。在終端機輸入 /config chrome=true 時，系統會引導你開啟設定面板完成配置，而輸入 chrome=false 則可直接關閉連線。同時，遠端控制的說明文件也修正了參數預設值說明：遠端衍生出的工作階段預設不會開啟瀏覽器控制，除非在啟動時明確附上 --chrome 參數。

問題修復

雲端工作階段全面修復，環境載入與工具快取更順暢

在雲端接續一段壓縮過的對話時，現在可以更新該工作階段先前讀取或發布的工件檔案。建立雲端工作階段或使用 /remote-env，也不再只看得到最新 20 個環境。若在雲端環境變數中設定 MCP_DISCOVERY_CACHE=1，重啟後會重用連接器的工具清單；其他 MCP（讓 AI 連接外部工具的協定）伺服器不再快取，並在啟動時連線。開場回覆的短暫延遲、還沒上傳檔案就誤報上傳資料夾遺失，也已修正。按下排程的「立即執行」時，若在啟動前被拒，畫面會顯示與該排程失敗通知相同的說明，不再露出內部錯誤文字。

問題修復

MCP 指令列輸出過濾跳脫字元，安全隱藏外掛設定值

在終端機管理 MCP（讓 AI 連接外部工具的協定）伺服器時，名稱或設定值裡的換行、終端機跳脫字元，現在不會弄亂 claude mcp list、claude mcp get 的輸出，也不會弄亂 remove、login、logout 找不到項目時的錯誤訊息。claude mcp list 會列出 WebSocket 伺服器的網址與連線狀態。標準輸入輸出伺服器的設定若省略 type，claude mcp get 仍會顯示 Type、Command、Args、Environment。外掛提供的這類伺服器則會隱藏命令、參數及環境變數值，只顯示變數名稱。

新功能

MCP 工具關閉即刻生效，自訂工具延遲載入更精準

在 SDK 或 -p 工作階段中，若中途新增 MCP（讓 AI 連接外部工具的協定）伺服器、之後再關閉，所屬工具現在會跟著移除，不再殘留給模型使用。工具若以 _meta['anthropic/alwaysLoad'] 設為 false，即使所屬的 --mcp-config、Agent SDK 或外掛伺服器設為 alwaysLoad，它仍會延遲載入。雲端工作階段與自架 Runner 也把 widgets 及 widgets_ 等相近名稱保留給系統；自訂伺服器若用了這些名稱，必須改名才能載入。

新功能

工件編輯發布一步到位，對話倒帶嚴防誤覆寫

請 AI 撰寫文件、網頁或產生展示工件（artifact）時，產出與發布的流程變得更快且安全。現在編輯或建立頁面時，系統會在同一步驟中主動建議發布，並精簡了後續提示文字以節省 Token（AI 處理與計費的文字單位）。為防止檔案內容被意外覆蓋，當先前讀取中斷、檔案外部變更、或使用者利用雙擊 Esc 倒帶對話時，系統會強制重新讀取最新原始檔後才允許發布。同時修復了伺服器短暫異常重試時，誤將成功發布判定為工作階段衝突的問題。

問題修復

工件清單關閉不中斷連結，目錄外檔案發布更受控

管理各項已產生的展示工件清單時，權限防護與版本關聯現在更加可靠。在 /artifacts 面板中按下叉號關閉單一項目，不會再解除原始檔案與工件的關聯，下次發布同一個檔案時依然會正確更新版本而非重複新建。即使事先勾選了免詢問許可，若工件檔案位於工作目錄之外，系統仍會先向你確認，除非事先透過 --add-dir 將該目錄加入允許清單。此外，在自動模式下讀取他人建立的工件或上傳資源時，系統會確實進行安全審查，不再因為你在其他模式下曾批准過就跳過檢查。

新功能

Ultrareview 支援部分複製快照，要求 Git 2.31 以上版本

在 macOS 或 Linux 用 /ultrareview 上傳本機版本庫時，現在要求 Git 2.31 以上。使用部分複製（只取回部分版本庫資料）的專案，在符合版本要求時會以工作樹快照上傳，不再退回舊方法。舊版 Git 若遇到工作樹缺少檔案的部分複製，會拒絕上傳並說明原因，不會先抓缺少的檔案；未使用 --filter 複製的版本庫仍可上傳。上傳時不納入符號引用（指向其他分支的別名）；目前分支若是符號引用，會拒絕上傳並說明原因；使用 --separate-git-dir 分離 Git 目錄的版本庫，也會拒絕上傳，不再用舊方法上傳。

問題修復

Ultrareview 跨平台檔案上傳修復，敏感憑證防護更周全

用 /ultrareview 上傳版本庫時，幾種特殊檔名與路徑不再出錯。WSL（在 Windows 中執行 Linux 的環境）的 Linux 磁碟區上，已變更檔案的名稱若含冒號，或以點號、空格結尾，不再讓上傳遭拒。macOS、Linux 上，某些特殊檔名造成的上傳變慢已修正；檔案或資料夾名稱帶有多個備份、編輯器標記時，憑證檔檢查也不再漏看。像 server:8443.key 這種副檔名前含冒號的憑證檔，其未提交變更不再混入上傳。專案資料夾暫時讀不到時誤報 core.worktree is set 的錯誤訊息、macOS 與 Linux 上 Git 工作樹（同一版本庫的另一份工作目錄）設定了 core.longpaths 而無法上傳，以及 Windows 上版本庫位於個人資料夾時誤傳連結工作樹，也已修正。

新功能

程式碼審查流程放寬，GitHub 組織選單滾動載入

設定自動修復 Pull Request 或排程審查時，GitHub 整合與規則套用提示更加明確。即使開啟 disableWorkflows，程式碼審查與 /ultrareview 仍會執行；但若執行審查的機器由管理員自行設定此鍵，則不在此限。在「新增存放庫」對話框中，組織清單現在支援向下捲動載入更多項目，不再只顯示前面幾筆。當專案自訂的 REVIEW.md 規則因為拉取請求規模過大或身為符號連結（指向另一個位置的捷徑檔）而未能套用時，檢查報告中會主動加註說明。此外，/autofix-pr 與 /schedule 不會再對尚未完成狀態檢查的存放庫誤報未安裝 GitHub App。

新功能

API 請求共享重試額度，非串流備援定時發送存活訊號

網路不穩、串流一再失敗時，會改走非串流備援（不逐段接收，改等完整回覆）；以前備援另有一套重試，同一個請求最多會重送 21 次，現在與原請求共用次數。可用環境變數 CLAUDE_CODE_NONSTREAMING_TIMEOUT_RETRIES 限制備援請求逾時後的重送次數。若 API 的輸出內容過濾器封鎖回覆，畫面會立即顯示該錯誤，不再反覆重送數分鐘。開啟部分訊息時，Anthropic API、Claude Platform on AWS 與閘道的非串流備援期間，每 30 秒會送出一次 ping 事件。

新功能

雲端模型支援同級自動降級，啟動狀態快取一整天

透過 Bedrock 或 Vertex AI 使用 Claude Code 時，若管理員移除預設模型的存取權，工作階段、對話標題與摘要會改用同級較舊的可用模型，不再直接失敗。兩者的啟動檢查也會記住帳號不能用的模型，最長一天，免得每次啟動都重查。Bedrock、Vertex AI、Mantle 的啟動檢查會帶上與一般請求一致的標頭，並表明請求來自 Claude Code。Bedrock、Mantle、Claude Platform on AWS 若使用非預設連接埠，簽名的 Host 標頭會包含連接埠。Bedrock 串流若遇到模型逾時或服務不可用，畫面會顯示錯誤訊息，不再是一整段原始 JSON。

問題修復

自訂伺服器啟用百萬視野，對話中途換模型限制即時生效

透過自訂 ANTHROPIC_BASE_URL 連線時，Opus 4.7 以上、Sonnet 5 以上與 Fable 等支援百萬上下文的模型，現在會使用 1M 上下文視窗（模型一次可處理的文字容量）。若閘道只支援 200K，可執行 /autocompact 200k 調整壓縮門檻。工作階段中途透過 set_model 請求切換模型，例如 Agent SDK 的 setModel，新模型的輸出 token（AI 處理文字的計量單位）上限和自動壓縮視窗會立即生效。若伺服器以客戶端未預期的另一個備援模型回覆拒絕訊息，/cost 與 SDK 的 modelUsage 會記到實際回覆所用的模型。

新功能

長對話畫面不再微幅卡頓，歷史紀錄展開與退出更滑順

在對話累積了大量檔案搜尋與程式碼改動時，終端機畫面的操作反應更為靈敏。在進行長對話時，非全螢幕模式下偶爾出現的一秒卡頓已獲修正，即使執行過 /clear 或壓縮紀錄也不再復發。配置大量權限拒絕規則與外部工具時，每輪對話的運算效能獲得改善。全螢幕模式下，按 Ctrl+O 開啟含數百次檔案讀取或搜尋的紀錄時，畫面不再短暫凍結；非全螢幕模式下，離開紀錄畫面也更順暢。開啟紀錄時仍在執行的工具，完成後會即時補上結果。此外，表格縮排中使用的 HTML 空白字元實體會正常渲染為空格，不再直接把原始碼字串印在畫面上。

問題修復

網址密碼遮蔽滴水不漏，修復歷史紀錄格式異常

檢視日誌或對話紀錄時，網址密碼若含 @，或把 @ 寫成 %40，現在都會完整遮蔽，不再露出部分或全部密碼。用 Vim 模式按 Ctrl+G 到外部編輯器修改後，回到 NORMAL 模式按 x 或 r，也不會弄壞提示文字末端的貼上預留標記。儲存的對話紀錄若含欄位缺失或格式錯誤的壓縮標記、喚醒項目，現在不會因此讓壓縮或恢復工作階段失敗、漏掉歷史，或直接當掉。

問題修復

掛鉤不再因背景子行程卡死，取消指令即刻攔截

透過掛鉤（特定事件發生時會執行的小程式）擴充流程時，若同步掛鉤啟動的背景行程一直開著輸出管道，Claude Code 現在會在掛鉤本身結束後不久繼續，不再一直等。命令或掛鉤若在準備啟動時被取消，也不會照樣跑到結束。離開計畫模式時，掛鉤與 SDK 權限回呼（程式收到權限請求時執行的函式）都能讀到同一次回覆剛寫好的計畫。工作流中的 agent()、parallel()、pipeline() 呼叫若失敗，即使稍後才等待結果，或根本沒有等待，也不再被當成未處理的 Promise 拒絕（非同步工作失敗卻無人接住），導致背景工作階段結束。

問題修復

沙盒放行規則更智慧，權限檢查與系統變數防護升級

執行 python3 -c、node -e 等行內腳本時，沙盒（限制程式可存取範圍的環境）不再只因內容含 = 就每次要求核准。版本庫外若設定了 Git trace2，.claude/settings.local.json 的放行規則也不再被擋住。PowerShell 命令解析器若啟動失敗，權限檢查不會跳過 deny 與 ask 規則，也不會快取（暫存結果供下次使用）該次失敗。從 IDE 差異檢視核准 Edit，或核准時更動了編輯內容，目標若是 /dev/null 等裝置，編輯現在能完成。Windows 的專案與本機設定不能再用 env 設定 ALLUSERSPROFILE、SystemDrive、CommonProgramFiles 相關變數；需要設定時可改用使用者或管理設定。

問題修復

登入驗證與指令列操作修復，支援多種圖片預覽轉檔

日常登入與接續工作階段時，瀏覽器顯示登入成功後，終端機不再一直等待；兩個工作階段同時接手當機後留下的登入鎖，也不再偶爾驗證失敗。CLAUDE_CODE_RESUME_INTERRUPTED_TURN 不會重跑因 --max-turns 而結束的回合。別名把選項放在 claude attach、logs、stop、respawn 或 rm 前面時，也不會誤把指令名稱當成新對話提示。工作樹與 /teleport 擷取遠端資料若遇到 SSH 金鑰密語或新主機確認提示，會迅速失敗，不再佔住終端機。Claude 送出 BMP、HEIC、HEIF、AVIF 或 TIFF 圖片時，Claude Code 若能轉換，Claude 應用程式現在可顯示預覽。

問題修復

評測工具防止符號連結穿透，截斷回覆獨立計分

用 /claude-api 評測模型時，評測執行架構與報表產生器已修好輸出檔被預置符號連結（指向另一位置的捷徑檔）或實體連結時，寫入會穿過連結的問題。達到 max_tokens 上限而被截斷的回覆，過去會混入平均分數；現在會標為截斷，另行計算。這樣檢視結果時，就能分清完整回覆與未寫完的回覆。

新功能

VS Code 支援主動診斷問題，外掛管理提供視覺化設定表單

在 VS Code 側邊欄處理專案時，Claude 現在可隨時讀取「問題」面板中的錯誤與警告，不必等到剛編輯完檔案。管理外掛時，安裝有自訂選項的外掛會詢問尚未填寫的欄位，之後可按齒輪圖示修改。操作若失敗，彈出視窗會說明原因；有可行修法時才會顯示修法。移除市集，或停用專案共用 .claude/settings.json 已啟用的外掛前，也會先詢問。重新載入視窗若中斷對話，恢復的分頁末端會標明不會再有後續回覆。

新功能

VS Code 多視窗開啟主動警示，對話復原不再一片空白

在編輯器中開啟多重視窗或翻閱過往歷史紀錄時，工作階段的狀態保護更加周全。當你試圖開啟已在其他視窗或應用程式中運行的對話時，系統會先主動詢問確認，不再無預警直接開啟重複副本。點擊歷史對話清單時，也不會意外把眼前正在進行的即時對話覆蓋掉。擴充套件重啟後重新載入的分頁，不再只留一片空白，而會說明如何復原。若對話卡在無法恢復的狀態，錯誤訊息會說明原因，並提供開啟新對話的選項。

新功能

VS Code 訊息同步不遺失，標籤名稱隨重新命名即時更新

在 VS Code 中，Claude 正在回覆時送出的訊息，重新開啟工作階段後不再消失。用 /rename、SessionStart 掛鉤（工作階段開始時執行的小程式），或在 claude.ai 重新命名對話後，編輯器分頁會跟著顯示新名稱。切換帳號後，工作階段清單的「Web」分頁會顯示目前帳號的工作階段，不再顯示前一個帳號的紀錄。登入時若返回上一步，再選相同的登入方式，也能離開授權碼畫面繼續操作。

問題修復

VS Code 對話框操作修復，指令選單與快捷鍵不再衝突

在 VS Code 對話面板輸入斜線指令後按 Enter，現在會執行選定指令，不再被模糊搜尋帶到無關項目或毫無反應。點過指令選單中的項目後按 Esc，也只會關閉選單，不會中止正在執行的回合。訊息若引用 Claude Code 或 IDE 標籤，後面的文字不再被截掉；長工作階段修剪最舊紀錄時，面板也不再卡住。視窗重載後，掛鉤（特定事件發生時執行的小程式）阻擋或停止提問的理由會保留。若啟動 VS Code 時設定了 CLAUDE_CODE_RESUME_INTERRUPTED_TURN，但 Continue After Reload 已關閉，恢復的分頁也不會誤重跑中斷的回合。

新功能

VS Code 代理人地圖資訊更精準，檔案編輯不再卡住十分鐘

在 VS Code 中調度多個代理人處理複雜任務時，代理人地圖與檔案操作的幾個問題已修正。代理人地圖現在會正確標註各個子代理人實際運行的模型名稱，不再誤標成工作階段的模型；即時發送給運行中代理人的訊息與最新紀錄也會完整顯示，多代理人同時運作時對話也不會遺失。當編輯器在檔案工具執行前因自動存檔無回應時，系統修復了每次讀寫與編輯操作卡死 10 分鐘才跳過的問題。此外，在外掛管理介面中解除安裝專案外掛時，系統能準確移除指定外掛，不會再誤刪其他安裝或解除安裝失敗。

新功能

企業方案開放一對一私訊，Slack 備援模型註記不再消失

在企業通訊軟體中使用 Claude Tag 時，持有企業方案 Standard Chat 或 Usage-Based Chat 席位、且具備 Cowork 的成員，現在可與 Claude 一對一私訊，不必另有包含 Claude Code 的席位。管理員的 Default model 設定與頻道的 Configure 頁面，也只會列出組織有權使用的模型，避免儲存或建立工作階段失敗。Claude 在 Slack 以備援（原模型不可用時改用的模型）回覆時，訊息下方會註明所用模型和原因；Claude 之後編輯訊息，註記也不再消失。

English

FEATURE

Launch Desktop App Directly from Terminal to Resume Existing Sessions

If you want to switch to a windowed interface in the middle of working on the command line, you can now do so with a single command. When using Claude Code in the terminal, entering claude --desktop launches the desktop application directly in the current directory. If you want to continue a session you were just having, simply add --continue or specify an ID with --resume <id>, and the system brings the entire existing conversation into the window.

FEATURE

Preconfigure Plugins on Install for Ready-to-Use MCP Servers

Having to manually hunt through menus to fill in parameters after installing an extension plugin often interrupts your workflow. You can now use the claude plugin configure <plugin-name> command to directly view all configuration items and unfilled fields for that plugin, or pass values all at once from standard input using --values-stdin. When installing plugins that package MCP servers, you can pass parameters directly after claude plugin install --config with <server>.<key>=<value>. Previously, if a plugin's built-in server was missing configuration, the system would often silently skip running it; now, both install prompts and the plugin list will proactively remind you to configure it.

FEATURE

Stricter Plugin Identification with Clearer SSH and Marketplace URL Validation

Subtle name differences when managing community plugins could cause installations or toggles to affect the wrong plugin. Downloading plugins or marketplace listings over SSH previously ignored the programs specified by the GIT_SSH environment variable and the Git config core.sshCommand, but they are now applied. If two plugin identifiers differ only by ., -, or @, or only by casing on macOS and Windows, installation previously risked misusing an installed plugin's cache or data directory; installation is now rejected. Enabling or disabling plugins using the full name@marketplace identifier will also no longer modify settings for a differently cased variant. Git URLs for plugins, marketplaces, and the current repository remotes undergo more thorough checks, and the interface explains the reason when a marketplace URL is rejected.

FEATURE

Centralized Enterprise Policy Controls Block Project Overrides of Security Sandboxes

When a company needs to restrict which API providers Claude Code can use, administrators can specify the scope in managed settings using allowedProviders. Project settings cannot relax or disable admin-mandated sandboxes, replace proxy servers protected by managed blocklists, expand strict allowlists, or reopen files barred from reading. Sessions authenticated to the Anthropic API via ANTHROPIC_AUTH_TOKEN now load organization policies. Team, Enterprise, and sessions with unresolvable subscription tiers pause WebFetch if policies fail to load at startup. If the operating system denies read access to managed configuration files, Claude Code warns and starts without policies from that file; other read errors or unparseable files will halt the session.

FEATURE

Support Disabling Web Fetch and Clarify URL Security Check Errors

When you do not want Claude Code to fetch web data on its own, you can disable WebFetch using the CLAUDE_CODE_DISABLE_WEB_FETCH environment variable. Error messages for URL security checks have also been corrected: previously, when domain checks hit rate limits, they could be misreported as network issues or enterprise policy blocks; now, rate limiting is accurately displayed.

FEATURE

Subagents Fully Inherit Parent Permissions and Streamline Background Reporting

When using subagents to assist with tasks, managing results and permissions is now easier. Running claude -p with CLAUDE_CODE_FORK_SUBAGENT=1 executes further Agent calls from subagents in the foreground, allowing them to capture results from nested agents. With claude -p --permission-prompt-tool, permission requests from background subagents are routed to the specified prompt tool instead of being automatically rejected. Forked subagents inherit the parent's permission mode in the session, including plan mode and dontAsk, and cannot exit plan mode on their own. In auto mode, background subagents no longer prompt for a second turn after reporting, and subagents finish immediately once returning reports to their callers.

FEATURE

Auto Mode Expands to Third-Party Providers with Streamlined Command Defaults

If you connect other model providers or disable telemetry, you will now also receive a one-time prompt to switch permissions. This prompt appears when the user is configured in other modes, offering auto mode as an optional new default. Running claude -p or Python Agent SDK sessions in these environments without a specified permission mode will start directly in auto mode just like interactive sessions, which can still be overridden using --permission-mode.

FEATURE

Enforce Background Command Timeouts and Improve Session Management

Background Bash or PowerShell commands now stop after a 30-minute default timeout, configurable up to 2 hours, and Claude receives a notification if a command halts due to timeout. /tasks groups Claude Code's own background work under a "System tasks" row that expands when pressing Enter. Background sessions can now be opened with /resume or claude --resume, and accompanying text is submitted as the next conversation turn only when using claude --resume <id> "prompt". Background sessions and sessions launched by Claude Code tools cannot enable auto-memory, though it can still be turned off. Responding to a background session awaiting permissions from claude agents will also avoid accidentally approving pending commands.

FEATURE

Reliable Remote Control Messaging and Mid-Turn Visibility for Side Questions

When sending messages and attachments from other devices via remote control, file downloads that fail due to network errors, timeouts, or server errors are now retried up to twice. Queued messages are marked as read only when Claude starts processing them; if the terminal closes early, unprocessed messages are delivered upon resuming the session next time. In sessions hosted by applications such as Claude Desktop, asking side questions with /btw now shows the turn currently in progress instead of only the previous turn. Remote control clients are also prohibited from running /claude-api.

FEATURE

Identify Hostnames for Chrome Connections with Clearer Configuration Guidance

Managing browser automation across computers is now easier to identify. The local client proactively reports its computer hostname to the browser, replacing generic numeric labels when multiple devices connect. Entering /config chrome=true in the terminal guides you to open the settings panel to complete configuration, while entering chrome=false disables the connection directly. Additionally, documentation for remote control clarifies default flag behavior: remote spawned sessions do not enable browser control by default unless started with the explicit --chrome flag.

BUGFIX

Cloud Sessions Receive Full Fixes, Smoothing Environment Loading and Tool Caching

When resuming a compacted conversation in the cloud, you can now update artifact files previously read or published during that session. Creating cloud sessions or running /remote-env is no longer restricted to viewing only the 20 most recent environments. If MCP_DISCOVERY_CACHE=1 is set in cloud environment variables, restarts reuse the connector tool list, while other MCP servers are no longer cached and connect upon startup. Brief delays on opening responses and false warnings about missing upload folders before files are uploaded have also been resolved. When clicking "Run now" on a schedule and getting rejected before startup, the screen displays the same explanation found in schedule failure notifications rather than exposing raw internal error text.

BUGFIX

MCP CLI Output Filters Escape Characters and Safely Conceals Plugin Values

When managing MCP servers in the terminal, newlines and terminal escape characters within names or configuration values no longer scramble the output of claude mcp list and claude mcp get, nor do they garble not-found error messages for remove, login, and logout. Running claude mcp list now displays URLs and connection states for WebSocket servers. If type is omitted in stdio server configurations, claude mcp get still displays Type, Command, Args, and Environment. For stdio servers provided by plugins, commands, arguments, and environment variable values are concealed, revealing only variable names.

FEATURE

MCP Tool Removal Takes Immediate Effect with Accurate Lazy Loading

In SDK or -p sessions, if an MCP server is added mid-session and subsequently closed, its associated tools are now removed immediately rather than lingering for the model to use. If a tool sets _meta['anthropic/alwaysLoad'] to false, it still lazy-loads even if its parent --mcp-config, Agent SDK, or plugin server is configured with alwaysLoad. Cloud sessions and self-hosted runners also reserve widgets and related names like widgets_ for system use, requiring custom servers using those names to be renamed before loading.

FEATURE

Artifact Publishing Streamlines into One Step with Rewind Overwrite Protection

When asking AI to write documents, web pages, or generate display artifacts, creation and publishing are now faster and more secure. When editing or creating pages, the system proactively suggests publishing in the same step and streamlines subsequent prompt text to save tokens. To prevent file contents from being accidentally overwritten, the system forces a re-read of the latest source file before allowing publication whenever a prior read was interrupted, the file changed externally, or the user rewound the conversation by double-pressing Esc. A bug where transient server retries misidentified successful publications as session conflicts has also been resolved.

BUGFIX

Closing Artifact Items Preserves Links and Controls External File Publishing

Permission safeguards and version associations are now more reliable when managing lists of generated display artifacts. Clicking the close icon on an individual entry in the /artifacts popover no longer disassociates the source file from the artifact, ensuring that publishing the same file again properly updates its version instead of creating a duplicate. Even if permission prompts were bypassed beforehand, the system still asks for confirmation if an artifact file resides outside the working directory, unless that directory was explicitly allowlisted via --add-dir. Furthermore, reading artifacts created by others or uploaded resources in auto mode strictly enforces security reviews without skipping checks just because approval was granted in another mode.

FEATURE

Ultrareview Supports Partial Clone Snapshots and Requires Git 2.31 or Later

Uploading a local repository with /ultrareview on macOS or Linux now requires Git 2.31 or later. Projects using partial clones upload as worktree snapshots when meeting the version requirement rather than falling back to legacy methods. Older Git versions encountering partial clones missing worktree files reject the upload with an explanation instead of fetching missing files first, while repositories cloned without --filter can still upload. Symbolic references are excluded during uploads. If the current branch is a symbolic reference, the upload is rejected with an explanation; repositories that separate their Git directory via --separate-git-dir are also rejected instead of being uploaded with the older method.

BUGFIX

Ultrareview Fixes Cross-Platform Uploads and Bolsters Sensitive Credential Protection

Several special filenames and paths no longer cause errors when uploading repositories using /ultrareview. On WSL Linux volumes, modified filenames containing colons or ending with dots or spaces no longer trigger upload rejections. Slow uploads caused by certain special filenames on macOS and Linux have been resolved, and credential file checks no longer overlook files or directories containing multiple backup or editor markers. Uncommitted changes to credential files with colons before extensions, such as server:8443.key, are no longer mistakenly included in uploads. False core.worktree is set errors when project folders are temporarily unreadable, upload failures on macOS and Linux when Git worktrees configure core.longpaths, and erroneous uploads of linked worktrees when repositories reside in user profile folders on Windows have all been fixed.

FEATURE

Code Review Workflow Relaxes Constraints as GitHub Organization Lists Add Infinite Scroll

GitHub integration and rule application feedback are clearer when configuring automated pull request fixes or scheduled reviews. Code reviews and /ultrareview still run even with disableWorkflows enabled, unless the machine executing the review was configured with this key by an administrator. In the "Add repository" dialog, the organization list now supports scrolling down to load more entries instead of only showing the initial few. When custom project REVIEW.md rules cannot be applied because a pull request is too large or REVIEW.md is a symlink, the check report proactively adds an explanatory note. In addition, /autofix-pr and /schedule no longer falsely report that the GitHub App is not installed for repositories with in-progress status checks.

FEATURE

API Requests Share Retry Quotas as Non-Streaming Fallbacks Send Periodic Keepalives

When unstable network conditions cause repeated streaming failures, requests switch to a non-streaming fallback. Previously, fallbacks maintained a separate retry pool that could re-send the same request up to 21 times, but they now share retry counts with the original request. The environment variable CLAUDE_CODE_NONSTREAMING_TIMEOUT_RETRIES can limit retries when fallback requests time out. If an API output content filter blocks a response, the screen immediately displays the error instead of endlessly retrying for minutes. When partial messages are enabled, non-streaming fallbacks across the Anthropic API, Claude Platform on AWS, and gateways emit a ping event every 30 seconds.

FEATURE

Cloud Models Support Same-Tier Automatic Downgrades and Cache Startup States for a Day

When using Claude Code via Bedrock or Vertex AI, if an administrator removes access to the default model, sessions, conversation titles, and summaries automatically fall back to an older available model in the same tier instead of failing outright. Startup checks for both platforms also remember inaccessible models for up to a day, avoiding repeated lookups on every launch. Startup checks for Bedrock, Vertex AI, and Mantle attach headers consistent with standard requests and identify the traffic as originating from Claude Code. When using non-default ports, Bedrock, Mantle, and Claude Platform on AWS include the port in the signed Host header. If Bedrock streaming encounters model timeouts or service unavailability, the screen displays a clean error message rather than a raw JSON dump.

BUGFIX

Custom Servers Enable Million-Token Context, Mid-Session Model Switches Apply Limits Instantly

When connecting through a custom ANTHROPIC_BASE_URL, models supporting million-token context—such as Opus 4.7 and above, Sonnet 5 and above, and Fable—now use a 1M context window. If a gateway only supports 200K, you can run /autocompact 200k to adjust the compression threshold. When switching models mid-session via a set_model request, such as setModel in the Agent SDK, the new model's output token limits and automatic compact window take effect immediately. If the server responds with a rejection using an unexpected fallback model, /cost and the SDK's modelUsage attribute the usage to the model that actually responded.

FEATURE

Long Conversations Stay Responsive Without Minor Stutter, Expanding and Exiting History Runs Smoother

Terminal interactions remain responsive even when conversations accumulate numerous file searches and code modifications. In long conversations, the occasional one-second stutter in non-fullscreen mode has been fixed and no longer recurs after running /clear or compacting history. Per-turn computational performance is improved when configuring large numbers of permission denial rules and external tools. In fullscreen mode, pressing Ctrl+O to open a history containing hundreds of file reads or searches no longer freezes the screen momentarily, and exiting history in non-fullscreen mode is also smoother. Tools that are still running when history opens append their results in real time upon completion. Additionally, HTML whitespace entities used in table indentation render properly as spaces instead of printing raw entity strings on screen.

BUGFIX

URL Passwords Masked Reliably, Malformed History Records Fixed

When viewing logs or transcripts, URL passwords containing @ or encoded as %40 are now fully masked without exposing partial or full passwords. After editing in an external editor via Ctrl+G in Vim mode, returning to NORMAL mode and pressing x or r will not corrupt paste markers at the end of the prompt text. Saved session transcripts containing missing fields or malformed compact markers and resurrection entries will no longer fail compaction or session resumption, drop history, or crash outright.

BUGFIX

Hooks No Longer Hang on Background Subprocesses, Cancellation Intercepts Commands Promptly

When extending workflows via hooks, if a background process launched by a synchronous hook keeps its output pipe open, Claude Code now proceeds shortly after the hook itself exits instead of waiting indefinitely. Commands or hooks canceled while preparing to launch will no longer run to completion. Upon exiting plan mode, both hooks and SDK permission callbacks can read the plan written in the same turn. If agent(), parallel(), or pipeline() calls in a workflow fail, even if awaited later or unawaited entirely, they are no longer treated as unhandled Promise rejections that terminate background sessions.

BUGFIX

Smarter Sandbox Allow Rules, Upgraded Permission Checks and System Variable Protections

When executing inline scripts like python3 -c or node -e, the sandbox no longer prompts for approval simply because content contains an = character. Allow rules in .claude/settings.local.json are no longer blocked when Git trace2 is configured outside a repository. If the PowerShell command parser fails to start, permission checks do not skip deny and ask rules, nor do they cache the failure. When approving an Edit from the IDE diff view or modifying the edit during approval, edits targeting devices like /dev/null now complete properly. Windows project and local settings can no longer use env to configure variables related to ALLUSERSPROFILE, SystemDrive, or CommonProgramFiles, requiring user or managed settings instead.

BUGFIX

Fixes for Authentication and CLI Commands, Expanded Image Preview and Conversion Support

During daily login and session resumption, the terminal no longer hangs after the browser indicates a successful login, and two sessions simultaneously claiming a stale lock from a crash will no longer occasionally fail authentication. CLAUDE_CODE_RESUME_INTERRUPTED_TURN will not rerun turns that ended due to --max-turns. When aliases place options before claude attach, logs, stop, respawn, or rm, command names are no longer mistaken for new prompt text. Fetching remote data via worktrees or /teleport fails fast instead of hanging the terminal when encountering SSH key passphrase or new host confirmation prompts. When Claude sends BMP, HEIC, HEIF, AVIF, or TIFF images, the Claude app now displays previews if Claude Code can convert them.

BUGFIX

Evaluation Tools Prevent Symlink Traversal, Truncated Responses Scored Separately

When evaluating models with /claude-api, the evaluation runner and report generator now prevent writes from traversing pre-existing symlinks or hard links placed at output file paths. Responses truncated upon hitting the max_tokens limit, which were previously blended into average scores, are now flagged as truncated and computed separately. This distinction makes it easy to separate complete responses from unfinished outputs when reviewing results.

FEATURE

VS Code Adds Proactive Problem Diagnostics, Visual Configuration Forms for Plugin Management

When working on projects in the VS Code sidebar, Claude can now inspect errors and warnings in the Problems panel at any time without waiting for a file edit to finish. When managing plugins, installing plugins with custom options prompts for unconfigured fields, which can later be edited via the gear icon. If an operation fails, popups explain why and only present actionable fixes when available. Prompts now ask for confirmation before removing marketplaces or disabling plugins enabled in shared project .claude/settings.json files. If reloading the window interrupts a conversation, the restored tab indicates at the bottom that no further responses will follow.

FEATURE

VS Code Warns on Multi-Window Collisions, Restored Conversations No Longer Open Blank

Session state is now better protected when opening multiple editor windows or browsing past history. When attempting to open a conversation already running in another window or application, the system asks for confirmation instead of opening duplicate copies without warning. Clicking entries in the conversation history list will not overwrite an ongoing live conversation. Tabs reloaded after an extension restart no longer remain blank, displaying recovery instructions instead. If a conversation becomes permanently unrecoverable, an error message explains the cause and provides an option to start a new chat.

FEATURE

VS Code Prevents Dropped Messages During Replies, Tab Names Update on Conversation Renames

In VS Code, messages sent while Claude is replying no longer disappear after reopening a session. Renaming a conversation via /rename, a SessionStart hook, or claude.ai updates the editor tab label immediately with the new title. After switching accounts, the "Web" tab in the session list displays sessions for the current account rather than stale records from the previous account. Going back during login and reselecting the same sign-in method allows you to leave the authorization code screen and proceed normally.

BUGFIX

VS Code Chat Fixes Eliminate Command Menu and Shortcut Conflicts

Pressing Enter after typing a slash command in the VS Code chat panel now runs the selected command instead of drifting to unrelated fuzzy search results or becoming unresponsive. Pressing Esc after clicking an item in the command menu now closes only the menu without aborting the active turn. Text following references to Claude Code or IDE tags is no longer truncated, and the panel no longer hangs when long sessions prune the oldest history entries. After a window reload, reasons for hook blocks or stopped questions are preserved. Restored tabs will not mistakenly rerun interrupted turns if VS Code launched with CLAUDE_CODE_RESUME_INTERRUPTED_TURN set while Continue After Reload is disabled.

FEATURE

VS Code Subagent Map Delivers Accurate Info and Eliminates 10-Minute File Edit Freezes

Several issues with the agent map and file operations are fixed when orchestrating multiple agents for complex tasks in VS Code. The agent map now accurately displays the actual model running for each subagent instead of mislabeling it with the session model. Real-time messages sent to running agents and latest logs appear fully, preventing lost conversations during concurrent multi-agent runs. When the editor becomes unresponsive due to autosave before file tools execute, operations no longer hang for 10 minutes on each read, write, or edit before skipping. In addition, uninstalling project plugins from the plugin management interface accurately removes only the targeted plugin without accidentally deleting other installations or failing.

FEATURE

Enterprise Plans Enable Direct Messaging and Preserve Slack Fallback Model Notes

When using Claude Tag in enterprise messaging software, members on Enterprise Standard Chat or Usage-Based Chat seats with Cowork can now message Claude directly one-on-one without needing a separate Claude Code seat. The admin Default model setting and channel Configure page now list only the models your organization has permission to use, preventing failed saves or session creation errors. When Claude responds in Slack using a fallback model, the message displays a note below indicating the active model and reason, and editing the message later no longer erases this annotation.

原始 CHANGELOG

Original changelog終端機指令直開桌面版，直接接續既有對話

Added `claude --desktop` to open the Claude desktop app on the current directory, or on a session with `--continue` / `--resume <id>`

Original changelog外掛安裝支援預先設定，MCP 伺服器即裝即用

Added `claude plugin configure <plugin>` to show a plugin's options and which are unset, or save new values read from stdin with `--values-stdin`
Added `<server>.<key>=<value>` to `claude plugin install --config`, so a bundled `.mcpb` MCP server's own settings can be set at install time and it starts without visiting `/plugin` → Configure
Fixed plugins silently skipping a bundled `.mcpb` MCP server that still needs configuration: `/plugin`, the install message and `claude plugin install` now say so and point to Configure

Original changelog外掛辨識更加嚴謹，SSH 與市集網址驗證更清楚

Fixed plugin and marketplace installs and updates over SSH ignoring the ssh program set in `GIT_SSH` or in your git config's `core.sshCommand`
Fixed `claude plugin disable` and `enable` with a full `name@marketplace` id changing a settings entry in another letter case instead of the installed plugin's own
Fixed installing a plugin with `claude plugin install` or `/plugin` putting it into an installed plugin's cache or data folder when their ids differ only in `.`, `-`, `@` or (macOS, Windows) capitals; the install is now refused
Improved plugin marketplace errors to name why a git address is refused instead of citing enterprise policy
Improved validation of git URLs for plugins, marketplaces and the current repository's remote

Original changelog企業政策集中管控，專案設定不得覆寫安全沙盒

Added `allowedProviders` managed setting to limit which API providers a machine may use (Anthropic API, a custom endpoint, Bedrock, Mantle, Vertex AI, Foundry, Claude Platform on AWS, or a Cloud gateway)
Fixed Claude Code refusing to start when the OS denies reading the managed settings file; it now warns and starts without that file's policies. Other read errors and unparseable files stop every session
Fixed sessions that authenticate with `ANTHROPIC_AUTH_TOKEN` against the Anthropic API never loading the organization's policy
Changed Team and Enterprise sessions, and sessions whose sign-in plan Claude Code can't determine, to withhold WebFetch until the organization policy loads if it couldn't be loaded at startup
Changed sandbox settings so project settings cannot widen or turn off an admin-required sandbox, replace the proxy behind a managed deny list, extend a strict allowlist, or reopen managed read-denies

Original changelog支援停用網頁擷取，網址安全檢查錯誤回報更精準

Added `CLAUDE_CODE_DISABLE_WEB_FETCH` environment variable to turn off the WebFetch tool
Fixed WebFetch reporting a rate-limited domain safety check as a network or enterprise policy block

Original changelog子代理完整繼承父層權限，背景執行回報不灌水

Fixed `claude -p` with `CLAUDE_CODE_FORK_SUBAGENT=1`: a subagent's own Agent call now runs in the foreground, so the subagent gets the child's result
Fixed `claude -p --permission-prompt-tool`: a background subagent's permission request now goes to the prompt tool instead of being auto-denied
Fixed fork subagents not keeping the session's plan mode or `dontAsk` mode: a fork now runs under its parent's permission mode and cannot exit plan mode
Fixed background subagents in auto mode prompting a second, redundant reply after each report
Improved subagents in auto mode: a subagent's run now ends as soon as it hands its report back to its caller, instead of taking extra turns that reach no one

Original changelog自動模式支援第三方平台，指令執行預設更俐落

Changed the one-time offer to make auto mode your default permission mode to also show on third-party providers and with telemetry off, when your user settings default to another mode
Changed `claude -p` and Python Agent SDK sessions on third-party providers or with telemetry off to start in auto mode when no permission mode is configured, like interactive sessions; `--permission-mode` still overrides it

Original changelog背景命令設定逾時上限，工作階段管理與切換更順手

Fixed a reply sent from `claude agents` to a background session waiting on a permission prompt sometimes approving the pending command
Improved `/resume` and `claude --resume` on a session that is running in the background: they now open that session instead of refusing, and a prompt given with `claude --resume <id> "prompt"` is sent to it as its next turn
Changed background Bash and PowerShell commands to stop after a time limit (their `timeout` with `run_in_background`, default 30 min, max 2 h); Claude is notified when one is stopped
Changed /memory so that Auto-memory can no longer be turned on from a background session or from a session one of Claude Code's own tools started; turning it off there still works
Changed `/tasks` to fold background work Claude Code runs for itself under one "System tasks" row; press Enter on it to show those tasks

Original changelog遠端控制訊息不遺漏，插話提問能掌握進行中進度

Fixed files attached to a message sent over Remote Control being left out after a single failed download; a network error, timeout or server error is now retried up to twice
Fixed Remote Control marking a message as read as soon as it arrived instead of when Claude started on it, and losing a message still queued when the terminal quit (it now arrives on the next resume)
Improved Remote Control: a `/btw` side question asked of a session hosted by an app such as Claude Desktop now sees the turn in progress, not only the last finished one
Changed `/claude-api` so it can no longer be run from Remote Control clients

Original changelogChrome 連線辨識主機名稱，設定指引更明確

Fixed `claude remote-control --help` saying `--[no-]chrome` defaults to the machine's `/chrome` setting; spawned sessions keep Claude in Chrome off unless `--chrome` is passed
Improved Claude in Chrome: the native host now reports your computer's name, so connected browsers can be labeled by computer instead of "Browser 1" / "Browser 2"
Changed `/config chrome=true` to direct you to the /config panel instead of enabling Claude in Chrome by default; `/config chrome=false` still turns it off when it was on

Original changelog雲端工作階段全面修復，環境載入與工具快取更順暢

Fixed cloud sessions that restarted after their conversation was compacted refusing the next update to an artifact the session had already read or published
Fixed cloud session creation and `/remote-env` reading only the newest 20 of an account's environments
Fixed the first reply in cloud sessions arriving tens of milliseconds late, a regression in 2.1.283
Fixed cloud sessions reporting the uploads folder as missing before any file had been uploaded
[Cloud sessions] Fixed Run now on a routine showing internal error text when the run is refused before it starts; it now shows the same explanation as the routine's failure notification
[Cloud sessions] Changed `MCP_DISCOVERY_CACHE=1`, when set in your cloud environment's variables rather than a settings file, to reuse your connectors' tool lists after a session restart; other MCP servers are no longer cached and connect at startup

Original changelogMCP 指令列輸出過濾跳脫字元，安全隱藏外掛設定值

Fixed `claude mcp list` and `claude mcp get`, and the not-found error of `claude mcp remove`, `login` and `logout`, printing line breaks and terminal escape sequences from MCP server names and values
Fixed `claude mcp list` leaving out WebSocket (`ws`) MCP servers; each is now listed with its URL and health status
Fixed `claude mcp get` showing no Type, Command, Args, or Environment for stdio servers whose config entry omits the `type` field
Changed `claude mcp get` to hide the command, arguments, and environment values of stdio MCP servers provided by plugins; variable names are still shown

Original changelogMCP 工具關閉即刻生效，自訂工具延遲載入更精準

Fixed switching off an MCP server added mid-session in SDK and `-p` sessions leaving its tools available
Changed MCP tools so a tool that sets its own `_meta['anthropic/alwaysLoad']` to false stays deferred when its `--mcp-config`, Agent SDK or plugin server is set to `alwaysLoad`
Changed the MCP server name `widgets` to be reserved in cloud sessions and on self-hosted runners: your own server under it, or a close spelling such as `widgets_`, no longer loads, so rename it

Original changelog工件編輯發布一步到位，對話倒帶嚴防誤覆寫

Fixed Artifact tool publishes after a conversation rewind (Esc Esc) overwriting a file's newer content that Claude had read only in the rewound turns; the publish is now refused until Claude re-reads the file
Fixed the Artifact tool so that publishing a page no longer lets Claude overwrite its source file without re-reading it when Claude's earlier read was cut short or the file had changed since
Fixed the Artifact tool sometimes reporting a publish as a conflict with another session after retrying a temporary server error, when the first attempt had actually succeeded
Improved Artifact tool results: they now suggest publishing in the same step as writing or editing the page, which can save a round trip
Improved Artifact tool publish results to use fewer tokens: the note on updating an artifact is shorter, and where to find your artifacts is no longer repeated after every publish

Original changelog工件清單關閉不中斷連結，目錄外檔案發布更受控

Fixed dismissing a row (x) in `/artifacts` unlinking its file from the artifact, so publishing the same file again created a new artifact instead of updating it
Fixed an `Artifact` allow rule ("don't ask again") letting the Artifact tool publish a file outside the working directories without asking; add the file's folder with `--add-dir` for the rule to cover it
Fixed auto mode skipping its classifier for Artifact tool asset uploads and reads of someone else's artifact when you had approved that artifact earlier in another permission mode

Original changelogUltrareview 支援部分複製快照，要求 Git 2.31 以上版本

Changed `/ultrareview` on macOS and Linux to leave symbolic refs out when uploading a local checkout; a checkout whose current branch is a symbolic ref is now refused with an explanation
Changed `/ultrareview` on macOS and Linux to require git 2.31 or newer to upload a local repository; checkouts made with `--separate-git-dir` are now refused instead of being uploaded with an older method
Changed `/ultrareview` uploads on macOS and Linux to send a partial clone as a working-tree snapshot on git 2.31 or newer, instead of falling back or refusing when git's version looked too old
Changed `/ultrareview` uploads on macOS and Linux to refuse, instead of fetching, a partial clone missing some of its working tree's files on older git versions; a clone made without `--filter` uploads

Original changelogUltrareview 跨平台檔案上傳修復，敏感憑證防護更周全

Fixed a misleading "core.worktree is set" error from `/ultrareview` when the project folder briefly could not be read
Fixed `/ultrareview` on macOS and Linux failing to upload the working tree from a git worktree whose per-worktree config sets `core.longpaths`
Fixed `/ultrareview` uploads including uncommitted changes to credential files whose name has a colon before the extension, such as `server:8443.key`
Fixed `/ultrareview` uploads on macOS and Linux running slowly on some unusual file names, and their credential-file check missing file or folder names with many backup or editor marks
WSL: Fixed `/ultrareview` refusing to upload a checkout on a Linux volume when a changed file's name has a colon or ends in a dot or space
Windows: Fixed `/ultrareview` uploading a linked worktree of a repository rooted at your home folder in some cases

Original changelog程式碼審查流程放寬，GitHub 組織選單滾動載入

Fixed `/autofix-pr` and `/schedule` saying the Claude GitHub App is not installed on a repository whose install status had not been checked yet
Changed Code Review's pull request reviews and `/ultrareview` to run when `disableWorkflows` is on, unless the machine running the review has it set by its own administrator (MDM or the managed-settings file)
[Code Review] Fixed the organization menu in Code Review's "Add a repository" dialog showing only a few of your GitHub organizations; it now loads more as you scroll
[Code Review] Improved the Code Review check run to say when your repository's REVIEW.md wasn't applied, for example on a very large pull request or when REVIEW.md is a symbolic link

Original changelogAPI 請求共享重試額度，非串流備援定時發送存活訊號

Added `CLAUDE_CODE_NONSTREAMING_TIMEOUT_RETRIES` environment variable to cap re-sends of a non-streaming fallback request that timed out
Fixed responses blocked by the API's output content filter being re-sent and retried, sometimes for minutes, instead of showing the filter's error right away
Fixed a failing API request being retried up to 21 times when streaming kept failing; the non-streaming fallback now shares the request's retry budget instead of getting a fresh set of retries
Improved SDK liveness during a non-streaming fallback request: with partial messages on, a `ping` stream event is now sent every 30 seconds on the Anthropic API, Claude Platform on AWS and gateways

Original changelog雲端模型支援同級自動降級，啟動狀態快取一整天

Fixed Amazon Bedrock mid-stream `modelTimeoutException` and `serviceUnavailableException` errors showing a raw JSON body instead of the error message
Improved Bedrock and Vertex AI sessions to switch to an older available model of the same tier, instead of failing, when an admin removes access to the default model; session titles and summaries now fall back with it
Improved Bedrock and Vertex start-up model checks: models your account cannot use are now remembered for up to a day instead of being re-checked on every launch
Improved Bedrock, Vertex and Mantle start-up model checks to send the same User-Agent, x-app and session ID headers as regular requests
Changed Bedrock, Mantle and Claude Platform on AWS requests to a base URL with a non-default port to include the port in the SigV4-signed Host header
Changed Bedrock, Vertex and Mantle start-up model checks to identify themselves as Claude Code, like other Claude Code requests

Original changelog自訂伺服器啟用百萬視野，對話中途換模型限制即時生效

Fixed switching models mid-session with a `set_model` request (such as the Agent SDK's `setModel`) leaving the new model on the built-in output-token limit and auto-compact window until restart
Fixed `/cost` and SDK `modelUsage` reporting a turn under the wrong model when the server answered a refusal with a different fallback model than the client expected
Changed sessions behind a custom `ANTHROPIC_BASE_URL` to use the 1M context window of models that have one (Opus 4.7+, Sonnet 5+, Fable); run `/autocompact 200k` if your gateway stops at 200K

Original changelog長對話畫面不再微幅卡頓，歷史紀錄展開與退出更滑順

Fixed the fullscreen ctrl+o transcript freezing briefly when opened on turns with hundreds of file reads or searches; tool calls still running when the transcript opens now show their results when they finish
Fixed `&nbsp;` showing as literal text in the terminal when a reply uses it to indent text, such as row labels in a markdown table
Fixed a brief freeze (up to a second) partway through long sessions outside fullscreen mode, which came back after `/clear` or `/compact`
Improved per-turn performance when many permission deny rules and MCP tools are configured
Improved responsiveness when leaving the ctrl+o transcript view in long sessions when fullscreen rendering is off

Original changelog網址密碼遮蔽滴水不漏，修復歷史紀錄格式異常

Fixed redacted logs and transcripts showing part of a URL password that contains `@`, or all of it when the URL writes its `@` as `%40`
Fixed vim mode: after editing in the external editor (Ctrl+G), `x` or `r` in NORMAL mode no longer breaks a pasted-text placeholder at the end of the prompt
Fixed compacting or resuming a session failing, opening without its history, or crashing when its saved transcript holds a compaction marker or loop wakeup entry with missing or malformed fields

Original changelog掛鉤不再因背景子行程卡死，取消指令即刻攔截

Fixed hooks and SDK permission callbacks seeing a missing or outdated plan on ExitPlanMode when the plan was written in the same response
Fixed a failed `agent()`, `parallel()` or `pipeline()` call that a workflow script awaits later, or not at all, being treated as an unhandled promise rejection, which could end a background session
Fixed synchronous hooks hanging Claude Code while a background process the hook started (for example `some-daemon &`) kept its output open; the hook now finishes shortly after its own process exits
Fixed a cancelled shell command or hook still starting, and running to its end, when the cancel arrived while it was being set up

Original changelog沙盒放行規則更智慧，權限檢查與系統變數防護升級

Fixed sandbox auto-allow asking for approval on every run of many inline scripts (`python3 -c`, `node -e`) just because they contain `=`
Fixed the PowerShell tool's permission check skipping deny and ask rules, and caching that failure for later checks, when its command parser failed to start (for example when the machine was out of memory)
Fixed `.claude/settings.local.json` allow rules being held back outside a git repository when git's trace2 output is configured
Fixed an approved Edit never going through when its target is a device, such as a file symlinked to /dev/null, and the approval came from the IDE diff view or changed the edit
Windows: Changed project and local settings `env` to no longer set `ALLUSERSPROFILE`, `SystemDrive`, or the `CommonProgramFiles` variables; set them in user or managed settings instead

Original changelog登入驗證與指令列操作修復，支援多種圖片預覽轉檔

Fixed SSH passphrase and new-host prompts from worktree and `/teleport` fetches taking over the terminal; these fetches now fail fast instead of asking
Fixed a rare auth failure when two sessions recover a login refresh lock left by a crashed process at the same time
Fixed `CLAUDE_CODE_RESUME_INTERRUPTED_TURN` re-running a turn that had ended at `--max-turns`
Fixed sign-in that could wait forever after the browser showed success
Fixed `claude attach`, `logs`, `stop`, `respawn` and `rm` starting a new session with the command name as its prompt when options came before it, such as from a shell alias
Improved pictures Claude sends as BMP, HEIC, HEIF, AVIF or TIFF files: the Claude apps now show a preview where Claude Code can convert them

Original changelog評測工具防止符號連結穿透，截斷回覆獨立計分

Fixed the `/claude-api` eval runner scaffold and report builder writing through a symlink or hard link planted at an output file
Fixed the `/claude-api` eval runner scaffold counting responses cut off at `max_tokens` in the score averages; they are now marked truncated and counted separately

Original changelogVS Code 支援主動診斷問題，外掛管理提供視覺化設定表單

[VSCode] Added a note under a restored tab's last message when a window reload interrupted it and no reply will follow
[VSCode] Added a plugin options form to Manage plugins: installing a plugin that has options asks for the unset ones, and a gear on its row changes them later
[VSCode] Added an on-demand diagnostics tool so Claude in the panel can read the Problems panel's current errors and warnings at any time, not only right after it edits a file
[VSCode] Improved the Manage plugins dialog: a failed plugin action now opens a popup that explains it and, where there is one, offers the fix
[VSCode] Changed the Manage plugins dialog to ask before removing a marketplace or turning off a plugin that your project's shared `.claude/settings.json` turns on

Original changelogVS Code 多視窗開啟主動警示，對話復原不再一片空白

[VSCode] Fixed opening Past conversations replacing a live conversation with its saved copy
[VSCode] Fixed a Claude tab reloaded after an extension restart staying blank instead of saying how to recover
[VSCode] Fixed opening a conversation that is already open in another window or app starting a second copy of it without warning; it now asks first
[VSCode] Fixed tabs stuck on a conversation that can't be resumed: the error now says so and offers to start a new conversation

Original changelogVS Code 訊息同步不遺失，標籤名稱隨重新命名即時更新

[VSCode] Fixed a message sent while Claude was working disappearing from the conversation after the session was reopened
[VSCode] Fixed the session list's Web tab showing the previous account's sessions after an account switch
[VSCode] Fixed sign-in staying on the authorization-code step after going back and choosing the same sign-in method again
[VSCode] Fixed the editor tab keeping an old name after a session was renamed with /rename, by a SessionStart hook, or on claude.ai

Original changelogVS Code 對話框操作修復，指令選單與快捷鍵不再衝突

[VSCode] Fixed pressing Enter after typing a slash command running an unrelated menu item picked by fuzzy match, or doing nothing
[VSCode] Fixed a message that quotes a Claude Code or IDE tag losing the rest of its text in the chat
[VSCode] Fixed restored tabs re-running an interrupted turn when VS Code was started with `CLAUDE_CODE_RESUME_INTERRUPTED_TURN` set, even with Continue After Reload off
[VSCode] Fixed Escape stopping the running turn instead of closing the command menu after clicking one of its rows
[VSCode] Fixed a hook's reason for blocking or stopping a prompt disappearing after a window reload
[VSCode] Fixed the chat panel stalling when a long session trims its oldest rows

Original changelogVS Code 代理人地圖資訊更精準，檔案編輯不再卡住十分鐘

[VSCode] Fixed an open agent transcript losing the agent's newer messages during a long session
[VSCode] Fixed every file Read, Write and Edit stalling for ten minutes and then being skipped when the editor stops responding to the extension's automatic save before the tool runs
[VSCode] Fixed the agent map labeling a sub-agent with the session's model instead of the model it actually ran on (e.g. under `CLAUDE_CODE_SUBAGENT_MODEL_FORCE` or an agent's own `model`)
[VSCode] Fixed uninstalling a plugin from the Manage plugins dialog, which removed the wrong installation or failed for a plugin installed for the project
[VSCode] Fixed the conversation disappearing from a session when many agents run
[VSCode] Fixed the agent map's transcript view leaving out messages sent to a running agent

Original changelog企業方案開放一對一私訊，Slack 備援模型註記不再消失

[Claude Tag] Added direct messages with Claude for members on an Enterprise plan Standard or Usage-Based Chat seat who also have Cowork; a seat that includes Claude Code is no longer required
[Claude Tag] Fixed the Default model setting in admin settings and a channel's Configure page offering models your organization can't use, which made saves or new sessions fail
[Claude Tag] Fixed the note under Claude's Slack messages saying it answered on a fallback model, and why, disappearing when Claude later edited that message