Claude Code 2.1.282

處理長篇對話或延續舊工作階段時，多種可能導致思考脈絡遺失與請求失敗的問題已修復。使用 --continue 或 --resume 恢復工作階段（session，使用者與 AI 互動的對話週期）時，早期訊息不會再因改寫重送而遺失先前的擴展思考。恢復對話時若透過 --tools 拿掉先前提過的內建工具，也不會再丟失思考紀錄。在 Claude 工作時執行 /model、/rename、/artifacts 或其他立即執行的斜線指令，也不再中斷思考紀錄。對話紀錄若含有 API 無法解密的搜尋結果不再引發 400 錯誤，這可能出現在經第三方閘道回答的回合。遇到 redacted_thinking 思考區塊資料無效時，Claude Code 會丟棄對話中的思考區塊後重試一次。關閉思考但強度設為極高時遇到安全切換也不再報錯。

切換模型與處理長對話摘要時，多項備援與提示機制已修正。長對話壓縮在摘要請求遭到拒絕時，現在會自動改用備援（fallback，主要方案失效時接手的替代方案）模型重新嘗試。在 Claude Desktop 等由 SDK 託管的工作階段中，未回應的額度購買提示不再擅自切換模型，而是直接結束該輪；遠端控制用戶端也會看到模型切換通知。用 /model 輸入 Fable 完整模型 ID，且方案需要尚未開啟的使用額度時，會正常彈出額度購買提示而非停在 API 錯誤。Bedrock 與 Bedrock Mantle 的安全攔截訊息現在會補上請求識別碼與訊息識別碼。Vertex AI 上遇到尚未登記的新模型也能正常提供網路搜尋，桌面版遇到未知模型時也會提示改選其他模型。

在電腦上同時開啟多個編輯器視窗進行登入時，憑證更新不再互相衝突卡住。如果前一個處理序在重新整理憑證（credentials，驗證使用者身分的登入證明）途中被關閉或強制結束，其他視窗不會再卡在另一處理序正在重新整理的登入錯誤長達一分鐘。同時啟動的各個工作階段（session，使用者與 AI 互動的對話週期）若剛好遇上登入重整，現在也會自動重試抓取組織安全政策。

企業管理員在派發組織安全性原則時，設定解析邏輯變得更加穩固。管理鎖定鍵如 disableClaudeAiConnectors 或 allowManagedPermissionRulesOnly 若被不小心打錯布林值，安全性鎖定依然強制生效並在啟動時點名該鍵。permissions、autoMode、工作樹（worktree，Git 讓同一專案平行開多份副本的功能）與 attribution 區塊若包含無效欄位，其餘有效設定仍會正常套用。下列任一情況，sandbox.excludedCommands 會忽略專案及本機設定的項目：

受管設定或 --settings 設 allowUnsandboxedCommands: false
受管設定設 allowManagedDomainsOnly: true

Windows 與 WSL 的管理員政策（HKLM、managed-settings.json）若存在但無效或無法讀取，可由使用者修改的 HKCU 與 WSL /etc/claude-code 設定就不會套用。

啟動專案與執行命令時的權限檢查更加安全嚴謹。啟動時不再經由儲存庫符號連結（symlink，指向另一個位置的捷徑檔）以 .. 或 /.vol 類型路徑讀取 macOS 的 /Network 中的 CLAUDE.md 與規則，也不再列出指向 /home 的規則連結。設定檔中含有中段萬用字元 :* 的 Bash 權限規則現在全面生效，並會在啟動時提示比對規則。受管 allowManagedPermissionRulesOnly 下，儲存庫、使用者及 --add-dir 來源的技能與指令，還有技能目錄中的外掛 manifest，都不能再藉 allowed-tools 自行預先核准工具。遠端工作階段（session，使用者與 AI 互動的對話週期）的背景工作單元重啟後，在恢復的權限提示上核准的指令不會再執行兩次，授權提示中的特殊字元顯示也已改善。

終端機介面在各種視窗尺寸與輸出情境下的畫面渲染問題已全面修復。以全螢幕啟動時不再有一瞬間的空白畫面閃爍。在一般視窗模式下，執行指令輸出時若同時刪除提示行，導致畫面縮短後仍高於終端機高度，輸出內容不會再錯位或亂成一團。程式碼差異比較中若中日韓文字元或表情符號換行重繪，最後一欄不再留下殘留的舊字元。設定了 SessionStart 掛鉤（hook，特定事件發生時被觸發的小程式）的專案，啟動時範例提示文字不會再閃爍消失。螢幕報讀模式、引述中的清單與長清單中，若程式碼區塊在清單項目開頭，或在引述內開頭，區塊頂部空白行也不會再被誤刪。

終端機的輸入操作與工具執行錯誤回報變得更加清楚準確。Bash 與 PowerShell 遇到硬碟配額（quota，儲存空間上限）已滿時，不再只回傳含糊的代碼 Exit code 1 並在暫存資料夾留下龐大檔案，而是明確說明空間不足。工具參數驗證失敗時，現在會一次列出該次呼叫的所有無效參數，不再每次只報一個。在工作階段（session，使用者與 AI 互動的對話週期）貼上多行文字時，若貼上模式中途重置，不會再變成一行行被當作指令送出。含有 Tab 字元的歷史提示詞被叫出時，游標不再停在非末尾位置。Windows Terminal 舊版發送提示改為顯示 ctrl+x ctrl+s。讀取 PDF 失敗時，包含特殊字元的路徑能正常顯示，資料夾名稱若類似 password 或 invalid，也不會再誤導錯誤原因。

執行各項指令與外掛管理時，取消操作與錯誤防護機制已獲修正。/install-github-app 在使用者取消後，不再繼續推送分支或儲存密鑰，而是立刻中斷並回報已完成的進度。在 Bedrock 與 Vertex 等平台執行 /feedback、/bug 或 /share 時，若在儲存途中取消，不會再強制寫入回報檔案。卸載外掛時若設定檔仍啟用該外掛或無法讀取，系統會停止解除安裝並指出該檔案，不再謊報成功或刪除設定。卸載後若讀不到外掛清單，已儲存的選項與機密資訊仍會保留，解除安裝也會告知使用者已保留這些資料。claude remote-control --debug 不再因為參數未識別而報錯。

終端機選單與清單視窗的互動體驗修復了多處細節。全螢幕下 /feedback 草稿清單在滑鼠懸停時會顯示捲軸。在 /skills 中，緊接著 / 輸入的按鍵不會再移動技能清單，而會進入搜尋框。繼續輸入關鍵字時，打字游標也不再跳回清單本體，避免游標隱藏與輸入法輸入位置錯亂。非全螢幕下 /skills 與 /mcp 清單不再窄兩欄。代理面板底部按鍵說明在自訂長快捷鍵時不會再折成兩行，且收合提示會遵守自訂鍵設定。若在 keybindings.json 解除綁定停止所有代理的快捷鍵，/tasks 對話框底部不再出現多餘的分隔符號。

在終端機輸入框使用 Vim 鍵盤模式時，多項編輯指令已校正為標準行為。>> 縮排指令不再對空白行增加縮排，r 帶數字時超出長度不會破壞文字，2J 不會多合併一行，最後一行的重複刪除與縮排（如 2dd 或 2>>）也不會位移異常或誤刪末行。使用 dd、dj、dG 刪除或用 p/P 整行貼上後，游標會落在第一個非空白字元，yy 複製不再意外移動游標，表情符號後按 Esc 也不會卡在字元內部。使用 . 重複 x、s、p、d 或 c 時能正確識別前置次數；上一行折行時，整行 p/P、o、O、J、>>、<< 等編輯指令也不再作用到錯誤行。從歷史紀錄叫出或從佇列拉回提示詞時，一般模式下的游標能精準落在結尾字元供 x 直接刪除。

VS Code 擴充套件在長篇對話與視窗整合上的多個問題已修復。生成長回覆時，面板不再於每次字元更新時重新解析全部文字，改善了回覆內容跟不上串流速度的延遲卡頓。訊息輸入框增高出現捲軸時，語音輸入麥克風按鈕不再遮擋住捲軸。在本機發起的遠端控制工作階段（session，使用者與 AI 互動的對話週期），現在能從 Web 清單項目點開本機對話，除非工作階段正在別處執行。擴充套件主機重啟後，編輯器分頁中的登入畫面若卡住，會顯示「停止回應」通知。

跨部門使用 Slack 多工作區架構協作時，團隊常會遇到機器人沒有自動加入新頻道或無法正常回覆的情況。企業全組織安裝時，單一工作區設定的頻道自動加入規則不再被忽略，Claude 會正常加入符合條件的新頻道。在同組織不同工作區共享的頻道中，Claude 不會再因為版本設定來自另一工作區而沉默不語。若頻道共享給尚未安裝 Claude 的工作區，提示訊息會明確指出是哪一個工作區需要安裝，不再只回報無法檢查頻道。中斷 GitHub 組織連線後，從權限套裝中移除該授權能順利儲存。舊版 Slack 應用程式連接 GitHub Enterprise Server 時，進度卡片現在會正確顯示儲存庫名稱並提供有效的建立 PR 按鈕。

Slack 內部對話的模型切換與訊息呈現進行了多項優化。當討論串原本使用的舊模型下線時，討論串會直接遷移至可用模型，不再於每次回覆時以較慢的備援（fallback，主要方案失效時接手的替代方案）模型重試並跳出提示。Claude 上傳至 Slack 的檔案說明現在支援表格排版，格式與一般回覆一致。Agents & tools 清單不再把討論串首則訊息誤當名稱，後續重新命名也會同步更新。工作階段的雲端工作單元重啟後，回覆底部的費用與 token（AI 處理與計費的文字單位）計量不再膨脹數倍。回覆中的計畫、表格與細節外框卡片預設改用寬版呈現，新連接的 Slack 工作區也會跟隨目前預設模型的變動。

Several issues that could cause lost thinking context and failed requests during long conversations or when continuing old sessions have been fixed. When resuming a session (the conversational cycle of interaction between you and the AI) with --continue or --resume, earlier messages no longer lose their prior extended thinking due to rewriting and resending. When resuming a conversation, removing previously mentioned built-in tools via --tools no longer drops thinking records either. Running /model, /rename, /artifacts, or other immediately executed slash commands while Claude is working no longer interrupts thinking records. Conversation history containing search results the API cannot decrypt no longer triggers a 400 error, which could occur on turns answered through a third-party gateway. When a redacted_thinking thinking block contains invalid data, Claude Code discards the thinking blocks in the conversation and retries once. Hitting a safety switch while thinking is off but effort is set to maximum no longer errors either.

Several fallback and prompt mechanisms for switching models and summarizing long conversations have been corrected. When a summarization request is rejected during long-conversation compaction, it now automatically retries with a fallback model (the alternative that takes over when the primary option fails). In SDK-hosted sessions such as Claude Desktop, an unanswered usage-purchase prompt no longer switches models on its own and instead simply ends that turn; remote-control clients also see the model-switch notification. Entering a full Fable model ID with /model when the plan requires usage that hasn't been enabled now properly pops up the usage-purchase prompt instead of stalling on an API error. Safety-interception messages from Bedrock and Bedrock Mantle now include the request ID and message ID. Unregistered new models on Vertex AI can still provide web search, and the desktop app now prompts you to pick a different model when it encounters an unknown one.

When signing in with multiple editor windows open on the same computer, credential updates no longer conflict and jam. If a previous process is closed or force-killed partway through refreshing credentials (the sign-in proof that verifies your identity), other windows no longer get stuck for up to a minute on a sign-in error saying another process is refreshing. Sessions that start at the same time and happen to hit a sign-in refresh now also automatically retry fetching the organization's security policy.

When enterprise admins distribute organizational security policies, the settings-parsing logic has become more reliable. If a managed lockdown key such as disableClaudeAiConnectors or allowManagedPermissionRulesOnly is given a mistyped boolean value, the security lockdown is still enforced and the key is named at startup. If the permissions, autoMode, worktree (a Git feature that lets one project have multiple parallel copies), or attribution blocks contain invalid fields, the remaining valid settings still apply normally. Under any of the following conditions, sandbox.excludedCommands ignores entries from project and local settings:

Managed settings or --settings set allowUnsandboxedCommands: false
Managed settings set allowManagedDomainsOnly: true

If Windows or WSL administrator policies (HKLM, managed-settings.json) exist but are invalid or unreadable, the user-modifiable HKCU and WSL /etc/claude-code settings will not be applied.

Permission checks at project startup and command execution are now safer and stricter. At startup, CLAUDE.md and rules in macOS /Network are no longer read via repository symlinks (shortcut files that point to another location) using .. or /.vol-style paths, and rule links pointing to /home are no longer listed. Bash permission rules containing a mid-pattern wildcard :* in config files now fully take effect, with a matching-rule notice at startup. Under managed allowManagedPermissionRulesOnly, skills and commands from repository, user, and --add-dir sources, as well as plugin manifests in skill directories, can no longer pre-approve tools on their own via allowed-tools. After a background work unit of a remote session restarts, commands approved on the restored permission prompt no longer run twice, and the display of special characters in authorization prompts has also been improved.

Rendering issues in the terminal interface across various window sizes and output situations have been fixed across the board. Launching in fullscreen no longer shows a brief flash of a blank screen. In normal window mode, when running a command deletes the prompt line at the same time and the screen shrinks but is still taller than the terminal height, output no longer shifts out of place or scrambles. In code diffs, when CJK characters or emoji wrap and redraw, the last column no longer leaves behind stale old characters. For projects with a SessionStart hook (a small program triggered when a specific event happens), the example hint text no longer flickers away at startup. In screen-reader mode, in lists inside quotes, and in long lists, when a code block sits at the start of a list item or at the start inside a quote, the blank line at the top of the block is no longer wrongly deleted.

Input handling and tool error reporting in the terminal are now clearer and more accurate. When Bash and PowerShell hit a full disk quota (the storage space limit), they no longer return only the vague Exit code 1 and leave huge files in the temp folder; they now state clearly that space ran out. When tool argument validation fails, all invalid arguments for that call are now listed at once instead of reporting just one each time. When pasting multi-line text during a session (the conversation cycle between you and the AI), if paste mode resets midway, the lines are no longer sent one by one as commands. When a history prompt containing Tab characters is recalled, the cursor no longer stops at a non-final position. The send hint on older Windows Terminal versions now shows ctrl+x ctrl+s. When reading a PDF fails, paths with special characters display properly, and folder names resembling password or invalid no longer mislead about the cause of the error.

Cancel handling and error safeguards when running commands and managing plugins have been corrected. After you cancel, /install-github-app no longer keeps pushing the branch or saving the key; it stops immediately and reports the progress already completed. When running /feedback, /bug, or /share on platforms like Bedrock and Vertex, canceling midway through saving no longer force-writes the report file. When uninstalling a plugin, if the config file still enables that plugin or can't be read, the system stops the uninstall and points out that file, instead of falsely reporting success or deleting settings. If the plugin list can't be read after uninstalling, your saved options and secret information are still kept, and the uninstall tells you this data was preserved. claude remote-control --debug no longer errors out over an unrecognized argument.

Several details of interacting with terminal menus and list windows have been fixed. In fullscreen, the /feedback draft list now shows a scrollbar on mouse hover. In /skills, keys typed right after / no longer move the skills list and instead go into the search box. As you keep typing keywords, the typing cursor no longer jumps back to the list body, avoiding a hidden cursor and misplaced input-method text. Outside fullscreen, the /skills and /mcp lists are no longer two columns narrower. The key hints at the bottom of the agents panel no longer wrap onto two lines with custom long shortcuts, and the collapse hint respects your custom key settings. If you unbind the shortcut for stopping all agents in keybindings.json, the extra separator no longer appears at the bottom of the /tasks dialog.

When using Vim keyboard mode in the terminal input box, several editing commands have been corrected to standard behavior. The >> indent command no longer adds indentation to blank lines, r with a count no longer corrupts text when it exceeds the line length, 2J no longer merges one extra line, and repeated deletes and indents on the last line (like 2dd or 2>>) no longer shift oddly or delete the final line by mistake. After deleting with dd, dj, or dG, or pasting whole lines with p/P, the cursor lands on the first non-blank character; yy yanking no longer moves the cursor unexpectedly, and pressing Esc after an emoji no longer gets stuck inside the character. Using . to repeat x, s, p, d, or c now correctly recognizes a preceding count; when the previous line is wrapped, whole-line editing commands like p/P, o, O, J, >>, and << no longer act on the wrong line. When recalling a prompt from history or pulling it back from the queue, the cursor in normal mode lands precisely on the final character so x can delete it directly.

Several issues in the VS Code extension around long conversations and window integration have been fixed. When generating long replies, the panel no longer re-parses the entire text on every character update, improving the laggy stutter where reply content couldn't keep up with the stream. When the message input box grows tall enough to show a scrollbar, the voice input microphone button no longer covers the scrollbar. For a remote-control session (the conversation cycle between you and the AI) started locally, you can now open the local conversation from the Web list item, unless the session is running elsewhere. After the extension host restarts, if the sign-in screen in an editor tab gets stuck, an "unresponsive" notification is shown.

When teams collaborate across departments using Slack's multi-workspace architecture, they often run into situations where the bot fails to auto-join new channels or cannot reply properly. With an org-wide enterprise installation, channel auto-join rules configured for a single workspace are no longer ignored, and Claude joins qualifying new channels as expected. In channels shared between different workspaces of the same organization, Claude no longer stays silent just because its version setting comes from another workspace. If a channel is shared with a workspace that has not installed Claude yet, the prompt message now clearly states which workspace needs the installation, instead of only reporting that the channel cannot be checked. After disconnecting a GitHub organization, removing that authorization from a permission bundle now saves successfully. When a legacy Slack app connects to GitHub Enterprise Server, the progress card now correctly shows the repository name and provides a working Create PR button.

Model switching and message rendering in Slack internal conversations received several improvements. When the older model a thread was using goes offline, the thread migrates directly to an available model, instead of retrying with the slower fallback model on every reply and popping up a notice. File descriptions Claude uploads to Slack now support table formatting, consistent with regular replies. The Agents & tools list no longer mistakes a thread's first message for its name, and later renames sync as well. After a session's cloud work unit restarts, the cost and token (the text unit AI processes and bills by) counts at the bottom of replies no longer inflate severalfold. Plan, table, and detail border cards in replies now default to a wide layout, and newly connected Slack workspaces also follow changes to the current default model.