發布日期：2026-09-22
沒特別指定模型的話，下次開 Claude Code 用的就是新一代旗艦 Claude Opus 5.5（claude-opus-5-5）。它是新的預設 Opus 模型，一次能讀 1M 上下文。價格是每百萬 token（AI 處理與計費的文字單位）輸入 4 美元、輸出 20 美元。讀取快取（把算過或查過的結果暫存起來直接用）每百萬只要 0.20 美元。Pro 與 Team Standard 方案的預設模型也從 Sonnet 改為 Opus，與 Max、Team Premium 及 Enterprise 看齊。思考力度也有調整：以前存下的設定不再套用到 Opus 5.5 這類新模型；Opus 4.7、Opus 4.8 與 Fable 5 也不再讓出廠預設值覆蓋 -p、SDK、專案或設定檔的 effortLevel。
讓 Claude 自己動手改檔、或開著自動模式的人，這批跟安全有關。透過符號連結（指向另一個位置的捷徑檔）寫入檔案時，提示詞會明確指出實際寫入的落點位置。如果真實目標落在專案目錄外，acceptEdits、允許規則與自動模式都不再放行。在自動模式下，如果安全檢查拒絕審查某項操作，該操作會直接被拒絕一次，並提示重試沒有幫助。若安全檢查沒有任何回應，Claude Code 會逐步拉長重試間隔，連續十次無回應就會停止該輪對話並顯示說明訊息。模型呼叫 Write 工具時如果傳入 path、file_text、file_content 或多餘的 description 等非標準參數，也能正確通過驗證。
在全螢幕裡捲清單、點選項，現在不必再跟對話框搶操作。你可以使用滑鼠滾輪捲動 /skills 清單，並點擊 /plugin 中的技能狀態選項。全螢幕下滾輪無法捲動含隱藏項目的清單（例如 /model 與 /permissions）已修復。在下列對話框中連按兩次 Ctrl+C 或 Ctrl+D 只會關閉對話框，不再退出程式：
/model, /effort, /config, /status, /usage, /plugin, /sandbox, /permissions,
/artifacts, /mobile, /login, /upgrade, /usage-credits, /install-github-app,
/setup-bedrock, /setup-vertex點擊視窗喚回前景時，不再誤觸游標底下的搜尋器、分頁列、代理與工作流程列、指令連結或選單。對話框取消按 y 或 n 鍵直接確認關閉，改用 Enter 確認與 Esc 取消（可在 keybindings.json 綁定 confirm:yes 與 confirm:no 還原）；文字輸入框也不再因快捷鍵遺失輸入的英數或空白。
送出前看不清自己打了什麼、聽寫按了停卻還在收音，這類小狀況這次一起修掉了。在 Windows 終端機中，按 Enter 清理不可見字元後會重新繪製畫面，讓你能看清楚送出的實際文字。不可見字元清理也不再誤刪波斯文與阿拉伯文用來連接字尾的零寬不連字元。語音聽寫時按下 Ctrl+C 會立即停止麥克風收音，在轉錄處理過程中按下 Esc 能順利取消，在轉錄檢視畫面與 vim NORMAL 模式下長按空白鍵也不會意外啟動聽寫。啟動掛鉤（特定事件發生時被觸發的小程式）執行時按 Esc，或是編輯、倒帶讓提示詞回到輸入框時，IDE 選取的內容不會再遺失。使用 Ctrl+S 暫存的 ! shell 模式提示詞，取回時依然維持 shell 模式，暫存後輸入 / 也能立即列出檔案路徑。
切換模型或接手子代理，最怕快取整個重算、白等一輪。當你在外部宿主應用程式中切換模型時，即使 Claude 正在執行任務，也不會再導致下一個提示詞發生快取未命中。外部宿主包含 Claude Desktop、VS Code 或 SDK。讀取快取（把算過或查過的結果暫存起來直接用）時，從分支復原的子代理會重新發送最初使用的工具清單，避免重建清單而破壞快取。/cost 指令說明快取沒命中的原因時，現在也會標出思考模式或思考顯示設定的變更。
裝了外掛或技能的人，這次更新比較不會出錯。背景外掛市集自動更新現在支援 Git 憑證庫（存登入證明的地方），私有儲存庫市集不再每次都重新複製或無法更新。當你從追蹤分支或標籤的 GitHub 儲存庫或 Git 網址更新外掛時，installed_plugins.json 會更新為最新修訂版本，不再保留安裝時的舊紀錄。執行 claude plugin update 時，若官方市集的快照檔案為連結或體積過大，也不會再將外掛版本誤標為未知。技能這邊，當 ~/.claude/skills/ 內包含列出技能名稱的 manifest.json 時，該目錄下的技能不再被意外移入 .trash 資料夾。新增外掛市集時，名稱如果模仿保留的官方市集名稱會直接被拒絕，已新增的也會停止載入。
這批是畫面上的小圖示和顏色修正，看狀態時比較不會誤會。在淺色與 ANSI 主題下，工作階段意見回饋調查現在會正常顯示游標停留時的高亮效果。執行 /workflows 時，若只有一筆執行紀錄，會直接開啟該紀錄，不再短暫閃過單行清單。在 /plugin 與 /skills 中，關閉的技能改以暗色圓圈 ◯ 顯示，不再與載入失敗外掛的紅色 ✘ 混淆。多選選項的說明文字縮排位置已校正至標籤文字下方，不再縮排在選項編號底下。全螢幕模式下 /plugin、/skills 與 /mcp 的搜尋框右邊框不再遺失。伺服器有警告時在 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）伺服器清單、詳細檢視畫面與 /plugin 中一律統一顯示為 ⚠ 圖示。
按 End 想跳到清單最底、按 Tab 想換欄位，結果不是沒反應就是偷改了設定？在 /config 設定清單以及 /model、/memory 和權限提示等選單中，按 Home 與 End 鍵現在能正確跳至最頂端與最底端。在 /skills 選單中按 PgUp 與 PgDn 鍵時，選取游標會停在第一個或最後一個技能，不再循環繞回另一端。在 /config 清單中按 Tab 鍵不再默默更改選定設定的值，改為不觸發任何變更。全螢幕模式下的 Ctrl+L 與 Cmd+K 恢復原本的重繪畫面功能，不再清除轉錄檢視內容。/autocompact 的頁尾提示已更新為使用 ← 與 → 調整數值，/fast 的頁尾也明確標示空白鍵為切換開關。
工作階段拖得越久，越怕一筆壞紀錄讓整段對話停擺。對話請求不再因為角色順序問題而拋出系統角色必須在助理訊息之前的 API 錯誤。開啟顧問模式時，若在不支援 advisor_20260301 標籤的代理伺服器或閘道後方遇到 API 400 錯誤，Claude Code 會自動拿掉該標籤重試。當工作階段儲存的歷史紀錄中包含無法載入的 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）工具錯誤通知時，對話與 /compact 不再每輪報錯。當儲存的轉錄紀錄包含格式損毀的系統訊息或缺少檔案清單的記憶體儲存通知時，復原工作階段不再發生崩潰。全螢幕長工作階段若發生損毀，損壞的快取（把算過或查過的結果暫存起來直接用）訊息清單現在會自動重建，避免跳出無法復原的介面錯誤而中斷。
讀設定檔讀到一半卡住不動，或一開 /config 就閃退？當設定檔或編輯後重新讀取的檔案在讀取過程中被具名管道替換時，Claude Code 不再無限期卡死等待。另一個是：某個設定項目已經移至 settings.json，但 ~/.claude.json 中仍然殘留 null 或 "false" 等舊值時，開啟 /config 不會再發生崩潰，各項開關偏好也能被正確讀取。
把工作丟給背景子代理（在背後替你跑任務的小助手）的人，這批修正讓交接不再掉東西。在無頭模式（沒有終端機視覺介面的執行模式）與 SDK 工作階段中，當背景子代理正在結束該輪對話時，傳送給它的訊息不會再默默遺失。當啟動子代理的主對話在讀取報告前執行了壓縮，已完成子代理的調查結果也不會再丟失。當 LSP 外掛啟用時，背景子代理現在能順利使用 LSP 工具進行程式碼語意分析。在背景子代理執行期間，按下兩次 Ctrl+C 即可順利退出，不再需要連續按三到四次。復原包含未完成背景代理、shell 或工作流程的工作階段時，不會再於你輸入前自行啟動模型對話。在非詳細模式下展開子代理交接訊息時，不再顯示內部的來源前言字串。
背景跑的指令明明正常結束，畫面卻顯示失敗，或 claude agents 開起來一片空白，這幾種狀況修好了。背景 shell 任務回傳良性的非零退出碼（例如未找到匹配項的 grep 指令）時，不再被誤報為執行失敗。當傳入工作階段的環境變數中包含空字元時，背景工作階段 claude --bg 現在依然能正常執行 Git、掛鉤（特定事件發生時被觸發的小程式）、外掛與其他輔助程式。當暫存目錄已滿、沒有寫入權限或屬於其他使用者時，執行 claude agents 會顯示明確的錯誤訊息，不再停留在無回應的空白畫面。跨工作階段通訊的啟動警告也做了說明改進，清楚指出可以使用 --debug-file 將除錯日誌寫入你指定的檔案路徑。
裝了很多 MCP 伺服器、或工具說明寫得特別長的人，這兩個改動跟你有關。你可以透過新增環境變數 CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH，自訂工作階段內所有 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）伺服器的工具說明與指示文字長度，打破原先 2,048 個字元的預設限制。另外，以前用 claude mcp remove 移除 MCP 伺服器後，若以相同名稱重新新增該伺服器，介面可能會停留在需要認證的錯誤狀態；現在會直接重新連線，不必再手動登入一次。
發布 Artifact 時，工具突然不見，或存取規則被悄悄重設？當組織政策因網路代理等因素無法載入時，Artifact 工具不再悄悄消失，Claude 現在會具體說明受阻原因。當重新發布 Artifact 時若遺漏了資料庫存取規則或檢視者個人檔案範圍，Claude 會直接拒絕發布，避免儲存的存取規則被意外重設或範圍遺失。當 Claude 無法讀取收到的 Artifact 連結時，會在繼續之前先告訴你。產出的 Artifact 網頁拿掉了列印按鈕、確認對話框以及檢視器阻擋的裝置功能，電子郵件與電話改以純文字呈現，深色模式也完整套用至表單控制項與捲軸。
審查明明還在跑卻顯示已停止、雲端 runner 認證失敗就斷線，這類意外中斷少了很多。使用 /ultrareview 上傳審查時，金鑰與設定檔複本（如 id_rsa copy 或 kubeconfig (1).yaml）留在本機不上傳。若工作階段被刪除或帳號變更，/ultrareview 不再誤報已停止的審查，也不會死等至逾時。在 Claude 應用程式中，遠端操控與雲端工作階段在 /compact 或 /clear 後上下文用量即時更新；當分支同時有已提交與未提交變更時，diff 檢視不再遺漏已提交檔案。伺服器過載且存取 token（授權令牌）輪替時，雲端與自架 runner 不再因認證失敗中斷。Cowork 編輯逾 10,800 字元記憶體衝突時，重試寫入不再丟失中段內容。
自己寫掛鉤（特定事件發生時被觸發的小程式）的人，出問題時更好查了。OpenTelemetry 的 hook_execution_complete 事件中新增了掛鉤輸出大小，以及因超出限制而存入檔案的超大輸出數量記錄。當 UserPromptSubmit 掛鉤執行逾時，逾時提示訊息與除錯日誌中現在會具體指出是哪一個掛鉤指令發生逾時。在 PermissionRequest 掛鉤中，代理型別（agent-type）的掛鉤因無法做出允許或拒絕的判定而不再允許執行，介面會顯示錯誤並引導你改用指令型或 HTTP 型掛鉤。
權限規則一多、清單一長，最容易迷失的是游標停在哪裡。在 /permissions 中檢視、新增或刪除規則後，介面焦點會自動回到規則清單中，刪除規則與移除目錄的確認對話框預設選項也改為 No 以防止手滑誤刪；在規則清單中按下 ←、→ 與 Tab 鍵可直接切換分頁，無需將焦點移至分頁列。執行 /install-github-app 進行 GitHub CLI 檢查與儲存庫選取步驟時，畫面現在會顯示「按 Esc 取消」的提示。/artifacts 與 /workflows 清單右側新增了捲軸，方便看清長清單的隱藏範圍與目前所在位置。工作流程進度樹中正在執行的代理與階段圖示從 ⟳ 改為暗色圓點。
打 @ 想找某個檔案，結果先跳出一堆只是資料夾名稱剛好對上的檔案？在全螢幕模式下，/plugin 的新增市集表單不再於窗格內額外繪製方框，文字與按鍵提示與其餘介面完全對齊；/workflows 詳細檢視畫面在全螢幕下也不會在窗格分隔線下方繪製多餘的第二條水平分隔線。未指定語言的程式碼區塊現在會採用類似行內程式碼的顏色標註，使指令能從周圍文字中清楚凸顯。當你在工具仍在執行時透過 /btw 提出旁支問題，Claude Code 現在認得這個呼叫還在跑，不會將其誤判為失敗呼叫。在輸入 @ 搜尋檔案時，檔案名稱直接符合搜尋詞的項目現在會優先排在僅資料夾名稱符合的項目之前。
用自己的機器替 Claude Code 跑工作（也就是自架執行器），或在 Windows 上把設定資料夾改指到別處的人，這批跟你有關。在 self-hosted runner 中，生命週期掛鉤（特定事件發生時被觸發的小程式）提交在 --configure-git 下無法簽章已修復。執行器達到 --retire-at 即將退役時，會等待該輪對話回報完畢再停止工作階段，避免遺失信號。生命週期掛鉤中的 Git 忽略執行器共用 Git 檔案指定的目錄與程式，本機路徑與 git:// 遠端現在需要設定 GIT_ALLOW_PROTOCOL。Windows 背景清理不再誤刪用於重定向 ~/.claude/session-env、image-cache 等目標的目錄符號連結（指向另一個位置的捷徑檔）或接合點。
VS Code 使用者這次拿到一批新對話框。你現在可以在對話框輸入 /status 開啟狀態對話框，檢視工作階段的版本、帳號、模型與伺服器詳細資訊；輸入 /sandbox 或從面板選單開啟沙盒對話框，管理沙盒模式、非沙盒備援與排除指令；輸入 /chrome 開啟 Claude in Chrome 對話框，查看擴充功能狀態、安裝、重新連線、權限頁面與預設啟用設定。輸入 /export 可將對話複製或儲存為純文字；輸入 /skills 開啟 Slash 指令對話框，能查看各個技能的來源、預估 token（AI 處理與計費的文字單位）用量與啟用狀態，並可直接點擊切換；輸入 /plan 則可切換至計畫模式、發送第一個計畫提示詞或查看目前計畫。
在 VS Code 貼上一大段文字時，Claude 現在分得出哪些是貼的、哪些是你打的。在聊天輸入框中貼上超過 800 個字元或包含超過 2 個換行符號的文字時，會自動標記成貼上內容。貼上文字時若包含不可見的 Unicode 格式化與標籤字元，清掉後會跳提示，其餘文字中的不可見字元也會在發送前自動移除。點擊「在新分頁開啟」時，Claude 視窗現在會直接開啟在你當前編輯器群組的旁邊，不再排在最後一個群組之後。重新開啟工作階段後，你發送的訊息中不再殘留貼上標記線。
VS Code 裡幾個會卡住或顯示錯的地方修好了。思考力度標籤不再顯示過期的儲存數值，改為即時呈現工作階段實際執行的力度等級。當 Python 擴充套件在啟用過程中卡住時，Claude Code 不再永久等待，而是在 60 秒後略過 Python 環境直接啟動。在計畫審批卡片中，當自動模式可用時，第一個選項現在會正常提供「是，並使用自動模式」，與終端機體驗保持一致。從鍵盤封存或取消封存工作階段後，工作階段清單中的方向鍵導覽不會再中斷失效。
管理員找不到例行任務（排好時間自動跑的工作）的開關了？它換了位置。管理員的例行任務開關設定已移至「Admin settings → Capabilities → Remote sessions」路徑下，Claude Code 管理頁面也附上了跳轉連結。在 GitHub Enterprise Server 儲存庫的雲端工作階段中執行 gh 指令與 GitHub API 呼叫時，存取 token（授權令牌）在約八小時後會自動續約，不再發生逾期中斷。若在排程執行前一刻編輯了例行任務，復原既有工作階段時會立即採用最新的提示詞與名稱，不再沿用舊設定執行。
點了檔案連結卻一直轉圈，或審批沒回就被記成你拒絕？在雲端工作階段轉錄紀錄中，指向工作目錄以外的檔案連結若無法載入，現在會直接停用並明確說明原因，不再打開永遠載入中的空白檔案卡片。在自動模式下，若審批提示逾期未回覆或被新訊息取代，不再記成你主動拒絕，工具呼叫能正常進行後續重試。在 Claude 應用程式中檢視雲端工作階段時，Claude 現在會將專供你使用的檔案儲存在應用程式可開啟的位置。在管理員關閉 GitHub 的組織中於自架環境啟動工作階段時，介面不再顯示空的儲存庫選擇器。
在 Slack 頻道把工作丟給 Claude 之後，你會想知道它做到哪了。Claude 在頻道討論串中回覆時，會顯示 Slack 原生的運作指示器、中斷按鈕與討論串標題；運作指示器會持續顯示直到任務完成，按下中斷按鈕可隨時中止執行。在訪客限制或僅限頻道設定下，當訪客加入或最後一名訪客離開而改變了 Claude 的回應方式時，頻道內會發出簡短通知說明變更。對於在加入 Enterprise Grid 之前就已連接 Claude 的 Slack 工作區，排程例行任務無聲失敗的問題已修復。
在 Slack 用 Claude 遇到錯誤時，現在比較知道是哪裡出問題。當遇到短暫的檔案掃描服務中斷時，Claude 不再要求你重新上傳檔案，而是會自動重試掃描並在掃描器離線時接收通知。在 Claude 的 Slack 回覆中，開頭帶有 +、- 或 * 符號的項目清單不再被錯誤渲染為空白項目符號與多餘的巢狀項目，文字內容與開頭符號完整保留為單一項目符號。當雲端環境設定指令碼執行失敗時，通知訊息不再僅顯示通用的重試提示，改為具體指出失敗的指令碼名稱並提示先修復問題。在管理員設定中，GitHub 連線橫幅現在會詳細說明未連線的具體原因，包括未登入、應用程式未連結或未安裝、登入逾期或 SSO 未授權。
審查規範寫得太長，有一部分可能根本沒被讀到。當儲存庫中的 REVIEW.md 審查指示檔案因為超過大小限制而在程式碼審查過程中遭到部分裁切或完全忽略時，Code Review 檢查執行會明確發出通知，具體指出檔案名稱與大小限制上限。團隊看到就知道規範沒有完全生效，可以回頭調整檔案長度。
If you haven't picked a model yourself, the next time you open Claude Code you'll be on the new flagship, Claude Opus 5.5 (claude-opus-5-5). It is now the default Opus model and offers a 1M context window. Pricing is $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million tokens. The default model for Pro and Team Standard plans has also changed from Sonnet to Opus, matching Max, Team Premium, and Enterprise. For thinking effort, settings saved in older versions no longer apply to new models like Opus 5.5. Opus 4.7, Opus 4.8, and Fable 5 also no longer let their launch defaults override effortLevel configured via -p, the SDK, projects, or settings files.
If you let Claude edit files on its own or run in auto mode, this batch is about safety. When writing to a file through a symlink, the prompt explicitly states where the write actually lands. If the real target sits outside the project directory, acceptEdits, allow rules, and auto mode will no longer approve it. In auto mode, if a safety check declines to review an action, the action is denied once with a notice that retrying will not help. If a safety check gives no response, retries now back off with increasing intervals, stopping the turn with an explanatory message after ten consecutive non-responses. In addition, Write tool calls that pass non-standard parameters such as path, file_text, file_content, or a stray description now pass validation properly.
Scrolling lists and clicking options in fullscreen no longer means fighting the dialogs. You can use the mouse wheel to scroll the /skills list and click on a skill's state options in /plugin. An issue where the wheel could not scroll lists with hidden items in fullscreen mode (such as /model and /permissions) has been fixed. Pressing Ctrl+C or Ctrl+D twice in the following dialogs now only closes the dialog instead of exiting the program:
/model, /effort, /config, /status, /usage, /plugin, /sandbox, /permissions,
/artifacts, /mobile, /login, /upgrade, /usage-credits, /install-github-app,
/setup-bedrock, /setup-vertexClicking a window to bring it to the foreground no longer accidentally triggers items under the pointer—such as search pickers, tab bars, agent and workflow rows, command links, or dropdowns. Dialogs no longer confirm or close on a stray y or n key, switching to Enter to accept and Esc to cancel (you can bind confirm:yes and confirm:no in keybindings.json to restore this behavior). In addition, text input fields in dialogs no longer lose typed letters, numbers, or spaces to matching keybindings.
Can't see what you're about to send, or dictation still listening after you hit stop? Those small snags are fixed. On Windows terminals, the screen repaints after invisible characters are removed on Enter, letting you clearly review the exact text being sent. Invisible character cleanup also no longer removes the zero-width non-joiner used in Persian and Arabic text to attach suffixes. For voice dictation, pressing Ctrl+C stops microphone recording immediately, Esc cleanly cancels while a transcript is processing, and holding Space in the transcript view or vim NORMAL mode will not accidentally trigger dictation. IDE selections are no longer lost when pressing Esc while startup hooks run, or when editing or rewinding brings a sent prompt back into the input box. A ! shell-mode prompt stashed with Ctrl+S remains in shell mode when restored, and typing / immediately lists file paths after stashing.
Switching models or taking over a subagent used to risk a full cache recompute and a wasted wait. When switching models from an external host application while Claude is working, the next prompt will no longer suffer a cache miss. Supported host applications include Claude Desktop, VS Code, and the SDK. When reading cache, resumed fork subagents re-send the tool list they originally used rather than rebuilding the list, preventing cache breakage. In addition, cache miss explanations in the /cost command have been improved to explicitly identify changes in thinking mode or thinking display settings.
If you use plugins or skills, updating them is less error-prone now. Background plugin marketplace auto-updates now support Git credential helpers, ensuring private repository marketplaces no longer re-clone every run or fail to update. When updating plugins from a GitHub repository or Git URL tracking a branch or tag, installed_plugins.json updates to the latest revision instead of keeping the install-time commit. Running claude plugin update also no longer misidentifies plugin versions as unknown when the official marketplace snapshot file is a link or too large. For skills, items in ~/.claude/skills/ are no longer accidentally moved to the .trash folder when a manifest.json in that folder lists their names. Furthermore, adding a plugin marketplace whose name imitates a reserved official marketplace name is rejected immediately, and existing ones stop loading.
This batch fixes small icons and colors on screen, so statuses are harder to misread. On light and ANSI themes, the session feedback survey now properly displays hover highlights. Running /workflows with only a single execution record directly opens that run instead of briefly flashing a one-row list. In /plugin and /skills, disabled skills are now represented with a dim circle ◯, avoiding confusion with the red ✘ used for failed plugin loads. Multi-select option descriptions are now properly indented under the label text instead of under the option number. The search box in /plugin, /skills, and /mcp no longer loses its right border in fullscreen mode. Finally, servers with warnings now consistently display a ⚠ icon across the MCP server list, detail views, and /plugin.
Pressed End to jump to the bottom of a list, or Tab to move on, only to get nothing or a silently changed setting? In the /config settings list, as well as selection lists such as /model, /memory, and permission prompts, pressing Home and End now jumps to the very top and bottom. In the /skills menu, pressing PgUp and PgDn stops the selection on the first or last skill rather than wrapping around. Pressing Tab in the /config list no longer silently changes the selected setting's value, instead doing nothing there. Ctrl+L and Cmd+K in fullscreen mode restore their original screen repainting behavior rather than clearing the transcript view. In addition, the footer hint for /autocompact now lists ← and → for adjusting values, while the footer for /fast explicitly identifies Space as the toggle key.
The longer a session runs, the more one bad record can threaten to stall the whole conversation. Conversations no longer fail on turn requests with an API error stating that role 'system' must precede an 'assistant' message. When advisor mode is enabled behind a proxy or gateway that does not support the advisor_20260301 tag, API Error 400 requests automatically retry without the tag. When saved session history contains a malformed notice about unloaded MCP tools, turns and /compact no longer fail on every turn. Resuming a session no longer crashes when its saved transcript holds a malformed system message or a memory-saved notice missing its file list. When long-running fullscreen sessions encounter corruption, damaged cached message lists are automatically rebuilt, preventing unrecoverable interface errors from aborting the run.
Stuck while reading a settings file, or /config crashing on open? When a settings file or a file being re-read after an edit is replaced by a named pipe mid-read, Claude Code no longer hangs indefinitely. In addition, when a preference has moved to settings.json while ~/.claude.json still holds a legacy value like null or "false", opening /config no longer crashes, and toggle preferences are parsed accurately.
If you hand work to background subagents, these fixes stop things from getting lost in the handoff. In headless mode and SDK sessions, messages sent to a background subagent are no longer silently lost when the subagent is finishing its turn. If the conversation that launched a subagent is compacted before reading the report, the finished subagent's findings will no longer be dropped. Background subagents can now use the LSP tool for semantic code analysis when an LSP plugin is active. Pressing Ctrl+C twice now cleanly exits while background subagents are running, removing the need for three or four presses. Resuming a session with unfinished background agents, shells, or workflows will not start a model turn on its own before you type anything. In addition, expanding subagent hand-back messages outside verbose mode no longer exposes internal provenance preambles.
A background command that finished fine but showed as failed, or claude agents opening to a blank screen: those are fixed. When background shell tasks return benign non-zero exit codes (such as a grep command finding no matches), they are no longer reported as failures. When environment variables passed into a session contain NUL characters, background sessions running claude --bg can still execute Git, hooks, plugins, and helper programs normally. When the temp directory is full, lacks write permissions, or is owned by another user, running claude agents displays a clear error message instead of remaining stuck on a blank, unresponsive screen. Startup warnings for cross-session messaging have also been improved to explain that --debug-file writes debug logs to a path you choose.
If you run lots of MCP servers or write long tool descriptions, these two changes are for you. You can now set the environment variable CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH to customize the length cap on MCP tool descriptions and server instructions across every MCP server in the session, overriding the default 2,048-character limit. In addition, when an MCP server was re-added under the same name after running claude mcp remove, the interface previously remained stuck in an error state requiring authentication; the system now reconnects directly without requiring repetitive manual sign-ins.
Publishing an Artifact and the tool suddenly vanished, or your access rules quietly reset? When organization policies fail to load due to factors like web proxies, the Artifact tool no longer disappears silently; Claude now explicitly explains what is blocking it. When republishing an artifact, if database access rules or viewer profile scopes are omitted, the system actively refuses publication, preventing stored access rules from being accidentally reset or scopes from being dropped. If Claude cannot read an artifact link it received, it notifies the user before continuing. For generated artifact web pages, Print buttons, confirmation dialogs, and device features blocked by the viewer have been removed, email and phone details appear as plain text, and dark mode extends across form controls and scrollbars.
Reviews shown as stopped while still running, or cloud runners dropping on auth failures: far fewer of these now. When uploading reviews with /ultrareview, renamed copies of key and config files (such as id_rsa copy or kubeconfig (1).yaml) remain on your local machine rather than being uploaded. If a session is deleted or the signed-in account changes, /ultrareview no longer misreports stopped reviews as completed or retriable errors, nor does it wait out the full timeout. In the Claude app, context usage figures for Remote Control and cloud sessions update immediately after /compact or /clear; when a branch has both committed and uncommitted changes, the diff view no longer drops committed files. When servers experience prolonged overloads during which access tokens rotate, cloud and self-hosted runner sessions no longer fail with authentication errors. When memory write conflicts arise in Cowork sessions for files exceeding about 10,800 characters, retried writes no longer drop the middle content.
If you write your own hooks, problems are now easier to track down. The hook_execution_complete OpenTelemetry event now includes hook output sizes and the count of oversized outputs saved to disk. When a UserPromptSubmit hook times out, the timeout notice and debug log now explicitly name which hook command timed out. In addition, agent-type hooks are no longer permitted in PermissionRequest hooks since their answers could never allow or deny requests; the interface now displays an error pointing to command or HTTP hooks.
Once permission rules pile up and lists grow long, the easiest thing to lose track of is where your cursor sits. After viewing, adding, or deleting a rule in /permissions, focus returns to the rule list automatically, and confirmation dialogs for deleting rules or removing directories now default to No to prevent accidental deletions. Pressing ←, →, and Tab within the rule list switches tabs directly without moving focus to the tab bar. Running /install-github-app now displays "Esc to cancel" during the GitHub CLI check and repository selection steps. Scrollbars at the right edge of /artifacts and /workflows lists make it easy to see how much of a long list is hidden and where you currently are. In addition, running agents and phases in the workflow progress tree now display a dim dot instead of ⟳.
Typed @ to find a file, only to see files that merely sit in a matching folder listed first? In fullscreen mode, the Add Marketplace form in /plugin no longer draws an extra box inside the pane, aligning its text and key hints with the rest of the interface. The /workflows detail view in fullscreen mode no longer draws an extra horizontal line beneath the pane divider. Code blocks without a specified language are now styled with inline-code coloring so commands stand out distinctly from surrounding text. Asking side questions via /btw while a tool is still running now recognizes the call as in progress rather than misinterpreting it as failed. When searching files with @, entries matching the query in their file name rank above those matching only across folder names.
If you run Claude Code work on your own machines (a self-hosted runner), or point settings folders elsewhere on Windows, this batch is for you. On self-hosted runners, lifecycle-hook commits failing to sign under --configure-git has been resolved. When a runner reaches --retire-at and is about to retire, it briefly waits for the turn to be reported before stopping the session, avoiding lost finish signals. Git inside lifecycle hooks now ignores hook directories and programs named in the runner's shared Git files, while local-path and git:// remotes now require GIT_ALLOW_PROTOCOL. On Windows, background cleanup no longer deletes directory symlinks or junctions used to relocate ~/.claude/session-env, image-cache, or other target directories.
VS Code users get a batch of new dialogs this time. You can type /status to open the Status dialog, inspecting session version, account, model, and server details. You can type /sandbox or open the Sandbox dialog from the panel menu to manage sandbox mode, unsandboxed fallbacks, and excluded commands. You can type /chrome or open the Claude in Chrome dialog from the panel menu to view extension status, installation, reconnection, permissions pages, and the enabled-by-default setting. In addition, typing /export copies or saves the conversation as plain text. Typing /skills opens the Slash commands dialog to check each skill's source, token estimate, and on/off state with click-to-toggle support. Typing /plan switches to plan mode, sends a first planning prompt, or views the current plan.
When you paste a big chunk of text in VS Code, Claude can now tell what you pasted from what you typed. When pasting text over 800 characters or containing more than 2 line breaks into the chat box, the text is automatically marked as pasted so Claude can distinguish it from what you typed. If pasted text contains invisible Unicode formatting and tag characters, the system notifies you after stripping them, and invisible characters in any other text are removed before sending. Clicking "Open in New Tab" now opens Claude right beside the editor group you are currently working in rather than after the last group. In addition, paste marker lines no longer linger in your own messages after reopening a session.
A few spots in VS Code that stalled or showed stale info are fixed. The effort chip no longer shows a stale saved effort level, instead displaying the actual level the session runs at in real time. When the Python extension hangs during activation, Claude Code no longer waits indefinitely, starting after 60 seconds without the Python environment. On the plan approval card, when auto mode is available, the first option now properly offers "Yes, and use auto mode", matching the terminal experience. In addition, arrow-key navigation in the session list no longer stops working after archiving or unarchiving a session from the keyboard.
Admins looking for the Routines switch: it moved. The admin Routines on/off setting has moved under "Admin settings → Capabilities → Remote sessions", with a direct link added to the Claude Code admin page. When executing gh commands and GitHub API calls inside a cloud session on a GitHub Enterprise Server repository, access tokens renew automatically after about eight hours without expiring. In addition, if a routine is edited moments before its scheduled execution, resuming an existing session immediately applies the updated prompt and name rather than running with old settings.
Clicked a file link that just kept loading, or had an unanswered approval logged as your denial? In cloud session transcripts, file links pointing outside the working directory that fail to load are now disabled with an explanation, rather than opening a blank file card that never loads. In auto mode, if an approval prompt expires unanswered or is superseded by a newer message, it is no longer recorded as an active user rejection, allowing tool calls to retry normally. When viewing cloud sessions in the Claude app, Claude now saves files intended for you where the app can open them. In addition, starting a session on a self-hosted environment in an organization where an admin turned GitHub off no longer shows an empty repository picker.
Once you hand work to Claude in a Slack channel, you want to know how far along it is. When Claude replies in channel threads, it displays Slack's native Working indicator, Stop button, and thread title; the indicator remains visible until Claude finishes, and pressing Stop interrupts the task at any time. Under Restrict or Channel only guest settings, when a guest joins or the last guest leaves and alters how Claude responds, a short notice is posted in the channel explaining the change. In addition, scheduled routines failing silently in Slack workspaces connected to Claude before joining Enterprise Grid have been fixed.
When something goes wrong using Claude in Slack, it's now clearer where the problem is. When a brief file-scanning service outage occurs, Claude no longer asks you to re-upload files, instead retrying the scan automatically and receiving a notice if the scanner is down. In Claude's Slack replies, list items starting with +, -, or * are no longer rendered as empty bullets with stray nested items, properly preserving the text and character as a single bullet. When cloud environment setup scripts fail, the notice no longer displays a generic retry prompt, instead naming the failed script and advising you to fix the issue first. In admin settings, the GitHub connection banner now explains exact disconnection reasons, including not signed in, app not linked or not installed, sign-in expired, or SSO not authorized.
If your review guidelines run long, part of them may never have been read at all. When a repository's REVIEW.md instruction file is partially cut or entirely omitted during code review because it exceeds size limits, the Code Review check run explicitly issues a notice naming the file and the exact limit. This allows teams to immediately understand why review guidelines did not take full effect and adjust instruction file lengths promptly.
Added Claude Opus 5.5 (`claude-opus-5-5`), now the default Opus model — 1M context, $4/$20 per Mtok with $0.20/Mtok cache reads
Changed the default model on Pro and Team Standard plans from Sonnet to Opus, matching Max, Team Premium, and Enterprise
Changed an effort level saved before `/effort` became per-model to no longer apply to newly released models such as Opus 5.5; they start at their default until you pick a level
Changed Opus 4.7, Opus 4.8 and Fable 5 to stop holding their launch-default effort over `/effort` in `-p` or the Agent SDK, a project, managed or `--settings` `effortLevel`, or a per-model levelFixed writes through a symlinked path being judged by their in-tree spelling: the prompt names where the write lands, and `acceptEdits`, allow rules and auto mode no longer approve one landing outside
Fixed auto mode retrying an action over and over when a safety check declined to review it; the action is now denied once, noting that retrying won't help
Fixed auto mode denying actions over and over without pause when a safety check gave no answer; retries now back off, and the turn stops with a message after ten in a row
Fixed Write calls failing validation when a model sends `path`, `file_text`, `file_content` or a stray `description` instead of `file_path` and `content`Added mouse support to more lists in fullscreen mode: the wheel scrolls the `/skills` list, and a skill's state options in `/plugin` can be clicked
Fixed Ctrl+C or Ctrl+D pressed twice in most dialogs (`/model`, `/effort`, `/config`, `/status`, `/usage`, `/plugin`, `/sandbox`, `/permissions`, `/artifacts`, `/mobile`, `/login`, `/upgrade`, `/usage-credits`, `/install-github-app`, `/setup-bedrock`, `/setup-vertex`) quitting Claude Code instead of closing the dialog
Fixed a click that only brought the terminal window to the front also triggering the item under the pointer — in search pickers, tab bars, agent/workflow rows, slash-command links and suggestion dropdowns
Fixed a stray `n` closing dialogs and a stray `y` confirming them; Enter and Esc accept and cancel (bind `y`/`n` to `confirm:yes`/`confirm:no` in `keybindings.json` to restore)
Fixed text fields in dialogs losing a typed letter, digit or Space to a keybinding on that key
Fixed the mouse wheel not scrolling selection lists with hidden options (such as `/model` and `/permissions`) in fullscreen modeFixed the prompt line staying scrambled on Windows terminals after invisible characters were removed on Enter; the screen is now repainted so you review the exact text that will be sent
Fixed the invisible-character cleanup removing the zero-width non-joiner that Persian and Arabic text uses to attach a suffix to a Latin word or number, such as the plural of "PDF"
Fixed voice dictation not stopping on Ctrl+C (the prompt cleared but the microphone kept recording), Esc not cancelling while a transcript was processing, and held Space starting dictation from the transcript view and vim NORMAL mode
Fixed IDE selection being dropped when a sent prompt comes back into the input, such as pressing Esc to edit it, rewinding to it, or pressing Esc while startup hooks run
Fixed a `!` shell-mode prompt stashed with Ctrl+S coming back as a plain prompt when restored, and `/` listing file paths right after stashing oneFixed a model switch made from a host app (Claude Desktop, VS Code, SDK) while Claude is working causing a prompt-cache miss on the next prompt
Fixed resumed fork subagents rebuilding their tool list instead of re-sending the one they first used, which broke prompt caching for that agent
Improved `/cost` cache-miss causes to name thinking mode and thinking display changesFixed `installed_plugins.json` keeping the install-time commit after updating a plugin from a GitHub repository or git URL that tracks a branch or tag
Fixed skills in `~/.claude/skills/` being moved to `~/.claude/skills/.trash/` when a `manifest.json` in that folder listed their names
Fixed background plugin marketplace auto-update ignoring git credential helpers, so private-repo marketplaces were re-cloned every run or never updated
Fixed `claude plugin update` clearing a plugin's recorded commit and moving it to version "unknown" when the official marketplace's snapshot file is a link or too large
Changed plugin marketplaces whose name imitates a reserved marketplace name to be refused when added, and to stop loading if one was already addedFixed the session feedback survey showing no hover highlight on light and ANSI themes
Fixed `/workflows` briefly showing a one-row list before opening the only run
Fixed a skill you switched off showing the same red ✘ as a plugin that failed to load in `/plugin` and `/skills`; off now shows a dim ◯
Fixed multi-select option descriptions being indented under the option number instead of under the label
Fixed the search box in `/plugin`, `/skills` and `/mcp` losing its right border in fullscreen mode
Fixed `/mcp` showing △ in the server list but ⚠ in the detail view for the same server; the list, detail views and `/plugin` now all show ⚠Fixed Home and End doing nothing in the `/config` settings list and in selection lists such as `/model`, `/memory` and permission prompts
Fixed PgUp/PgDn in the `/skills` menu wrapping past the first or last skill instead of stopping there
Fixed Tab silently changing the selected setting's value in the `/config` list; it now does nothing there
Reverted `ctrl+l` / `cmd+k` in fullscreen mode clearing the transcript view (added in 2.1.260); they redraw the screen again
Changed `/autocompact`'s footer hint to name ←/→, the keys that adjust other ordered values
Changed `/fast`'s footer to name Space as the toggle keyFixed conversations failing on every turn with a "role 'system' must precede an 'assistant' message" API error
Fixed conversations with the advisor on failing every turn with API Error 400 "Input tag 'advisor_20260301'" behind a proxy or gateway that doesn't support it; the request now retries without it
Fixed a session failing on every turn and `/compact` when its saved history held a malformed notice about MCP tools that could not be loaded
Fixed a crash when resuming a session whose saved transcript holds a malformed system message or a memory-saved notice without its file list
Fixed one cause of long-running fullscreen sessions exiting with "Claude Code exited after an unrecoverable interface error": a damaged cached message list is now rebuiltFixed Claude Code hanging when a settings file, or a file it re-reads after an edit, is replaced by a named pipe mid-read
Fixed `/config` crashing and some on/off preferences being misread when a preference that has moved to `settings.json` still holds a value like `null` or `"false"` in `~/.claude.json`Fixed subagent hand-back messages showing an internal provenance preamble when expanded outside verbose mode
Fixed resuming a session with unfinished background agents, shells or workflows starting a model turn on its own before you typed anything
Fixed messages sent to a background subagent being silently lost in headless and SDK sessions when the subagent was finishing its turn
Fixed a finished subagent's report being lost when the conversation that launched it was compacted before the report was read
Fixed background subagents being unable to use the LSP tool when an LSP plugin is active
Fixed Ctrl+C needing three or four presses to exit while background subagents are running; two presses now exitFixed background shell tasks reporting benign non-zero exits (e.g. grep with no matches) as failures
Fixed background sessions (`claude --bg`) being unable to run git, hooks, plugins and other helper programs when an environment variable handed to the session contained a NUL character
Fixed `claude agents` showing a blank, unresponsive screen instead of an error when the temp directory is full, not writable or owned by another user
Improved the cross-session messaging startup warning to explain that `--debug-file` writes a debug log to a path you chooseAdded `CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH` to change the 2,048-character cap on MCP tool descriptions and server instructions for every MCP server in the session
Fixed an MCP server re-added under the same name after `claude mcp remove` still showing as needing authentication instead of reconnectingFixed the Artifact tool silently disappearing when your organization's policy can't be loaded (for example behind a web proxy); Claude now says what's blocking it
Fixed artifact republishes silently resetting stored database access rules or dropping the viewer profile scope when that capability was re-sent without them; they are now refused
Improved the Artifact tool so that when Claude cannot read an artifact link it was given, it tells the user before continuing
Improved artifact pages: no Print buttons, confirm dialogs or device features the viewer blocks, email and phone details shown as text, and dark mode that reaches form controls and scrollbarsFixed `/ultrareview` reporting a stopped cloud review as completed or as an error to retry, and waiting out the full timeout when its session was deleted or the signed-in account changed
Fixed the Claude app showing a missing or stale context usage figure for Remote Control and cloud sessions right after /compact or /clear
Fixed the Claude app's diff view for Remote Control and cloud sessions dropping a branch's committed files whenever there are also uncommitted changes
Fixed cloud and self-hosted runner sessions failing with "Authentication failed" after waiting out a long overload during which the session's access token was rotated
Fixed memory write conflicts in Cowork sessions showing Claude only the start and end of a memory file over about 10,800 characters, so the retried write dropped the middle
Improved `/ultrareview` uploads: renamed copies of key files, such as `id_rsa copy` or `kubeconfig (1).yaml`, now also stay on your machineAdded hook output sizes and the number of oversized outputs saved to a file to the `hook_execution_complete` OpenTelemetry event
Improved the UserPromptSubmit hook timeout notice and the debug log to name which hook command timed out
Changed `PermissionRequest` hooks: an agent-type hook no longer runs there, since its answer could never allow or deny the request; it now shows an error pointing to command or http hooksImproved `/permissions`: focus returns to the rule list after viewing, adding or deleting a rule, and the delete-rule and remove-directory confirmations now default to No
Improved `/permissions` tab navigation: ←/→ and Tab pressed in a rule list now switch tabs without moving focus to the tab bar
Improved `/install-github-app`: the GitHub CLI check and repository selection steps now show "Esc to cancel"
Improved the `/artifacts` and `/workflows` lists: a scrollbar at the right edge shows how much of a long list is hidden and where you are in it
Improved the workflow progress tree: running agents and phases now show a dim dot instead of ⟳Improved `/plugin`'s Add Marketplace form in fullscreen: it no longer draws a box inside the pane, and its text and key hints line up with the rest of `/plugin`
Improved the `/workflows` detail view in fullscreen: it no longer draws a second horizontal rule under the pane's divider
Improved code blocks that don't name a language: they are now colored like inline code, so commands stand out from the surrounding text
Improved `/btw` asked while a tool is still running: the side question now knows that call is in progress instead of reading it as a failed one
Improved `@` file suggestions: a file whose name contains the query now ranks above one that only matches across its folder namesSelf-hosted runner: Fixed lifecycle-hook commits failing to sign under `--configure-git`
Windows: Fixed background cleanup deleting a directory symlink or junction used to relocate `~/.claude/session-env`, `image-cache` or another cleaned-up folder
Self-hosted runner: Fixed a turn that ended right at a `--retire-at` release losing its finished signal; the runner now briefly waits for the turn to be reported before stopping the session
Self-hosted runner: Changed git in lifecycle hooks to ignore hook folders and programs named in the runner's shared git files; local-path and `git://` remotes there now need `GIT_ALLOW_PROTOCOL`[VSCode] Added a Status dialog, with a typed `/status`, showing the session's version, account, model and server details
[VSCode] Added a Sandbox dialog for the sandbox mode, the unsandboxed fallback and excluded commands, opened from the panel menu or by typing `/sandbox`
[VSCode] Added a Claude in Chrome dialog (extension status, the install, reconnect and permissions pages, the enabled-by-default setting), opened from the panel menu or by typing `/chrome`
[VSCode] Added Export conversation, with a typed `/export`, to copy or save the conversation as plain text
[VSCode] Added each skill's source, token estimate and on/off state to the Slash commands dialog, with a click to change the state, and a typed `/skills` that opens it
[VSCode] Added a typed `/plan` that switches to plan mode, sends a first planning prompt, or shows the session's plan[VSCode] Improved pasted-text handling in the chat box: a paste over 800 characters or over 2 line breaks is now marked so Claude can tell it from what you typed
[VSCode] Improved prompt handling in the chat box: invisible Unicode formatting and tag characters are removed from pasted text with a notice, and from anything else before it is sent
[VSCode] Changed "Open in New Tab" to open Claude beside the editor group you are working in rather than after the last group
[VSCode] Fixed paste marker lines showing in your own messages after reopening a session[VSCode] Fixed the effort chip showing a stale saved effort level instead of the level the session runs at
[VSCode] Fixed Claude Code never starting when the Python extension hangs while activating; it now starts after 60 seconds without the Python environment
[VSCode] Fixed the plan approval card never offering auto mode: when auto mode is available, its first option is now "Yes, and use auto mode", as in the terminal
[VSCode] Fixed arrow-key navigation in the session list stopping after archiving or unarchiving a session from the keyboard[Claude Code on the web] Changed the admin Routines on/off setting to live under Admin settings → Capabilities → Remote sessions; the Claude Code admin page now links to it
[Claude Code on the web] Fixed `gh` and GitHub API calls inside a cloud session on a GitHub Enterprise Server repository failing after about eight hours; the token now renews automatically
[Claude Code on the web] Fixed a routine that resumes an existing session running with its old prompt and name when it was edited moments before the scheduled run started[Claude Code on the web] Fixed file links in a cloud session transcript that point outside the session's working directory opening a file card that never loads; they're now disabled and say why
[Claude Code on the web] Fixed auto mode refusing to retry a tool call because an approval prompt that expired unanswered, or was superseded by a newer message, had been recorded as your rejection
[Claude Code on the web] Improved cloud sessions viewed in the Claude app: Claude now saves files meant for you where the app can open them
[Claude Code on the web] Removed the empty repository picker shown when starting a session on a self-hosted environment in an organization where an admin has turned GitHub off[Claude Tag] Added Slack's native Working indicator, Stop button and thread title to Claude's threads in channels; the indicator stays up until Claude finishes, and Stop interrupts the task
[Claude Tag] Added a short notice in the Slack channel when a guest joining, or the last guest leaving, changes how Claude responds there under a Restrict or Channel only guest setting
[Claude Tag] Fixed scheduled routines silently failing to run in Slack workspaces that were connected to Claude before the workspace joined its Enterprise Grid[Claude Tag] Fixed Claude asking you to re-upload a Slack file when a brief file-scanning outage, not the file, was the problem; it now retries the scan and is told when the scanner is down
[Claude Tag] Fixed a bullet in Claude's Slack reply whose text starts with +, - or * rendering as an empty bullet with a stray nested item; it now shows as one bullet with the character kept
[Claude Tag] Fixed the Slack notice for a failed cloud environment setup script sometimes being a generic "mention me to retry"; it now names the setup script and says to fix it first
[Claude Tag] Improved the GitHub banner in Claude Tag admin settings to say why GitHub isn't connected: not signed in, app not linked or not installed, sign-in expired, or SSO not authorized[Code Review] Improved the Code Review check run to say when REVIEW.md instructions were cut or left out of a review for exceeding a size limit, naming the file and the limit