發布日期：2026-10-08
當你加上自訂掛鉤（hook，特定事件發生時被觸發的小程式）時，啟動失敗也能阻擋動作。命令與 HTTP 掛鉤支援 onFailure: "block"，啟動失敗、逾時或傳回非預期結束碼就阻擋後續動作。深層巢狀工具輸入不再被截斷而漏掉防護。外掛掛鉤背景工作單元重新啟動、模組重新載入期間，原本會繞過另一模組帶有 .catch 的防護掛鉤的呼叫，現在會被拒絕。非同步掛鉤分多行印出的 JSON 也能正確讀取；雲端工作階段不再因 session.receive 掛鉤先請求權限而卡住。工具執行後若遭外掛拒絕，Claude 與你會得知工具已執行，但結果被攔截。
管理擴充工具時最怕設定檔載入失敗卻毫無動靜。執行 claude plugin install、enable、disable 或新增市集時，若要寫入的設定檔無法載入，會發出警告。在 /plugin 錯誤分頁按 Enter，不再未經確認就移除市集及其外掛。新增市集若名稱無法用於安裝外掛，會直接拒絕；新增或重新整理含大型 Git 子模組的市集時，只抓取含外掛檔案的子模組。子模組是儲存庫內連結的另一份儲存庫。claude plugin validate 遇到 README 缺少安裝指令，會印出可貼上的建議，不影響結束碼，即使使用 --strict 也一樣。掛鉤模組因頂層 var 重新綁定遭拒時，錯誤會指出行號、原因與修法。
開發或使用自訂介面模組時，資訊排版與提醒層級現在更有彈性。模組現在可透過 $.ui.notify 發送符合系統設定的原生桌面通知，並明確標記發送來源。按鈕元件支援放入文字子節點，方便在單一可點擊按鈕內呈現標籤或次要細節。由所屬組織發布的彈出提示（toast）會排在其他模組前面優先顯示。內建的外掛編寫技能在沒有終端機的工作階段中不再要求執行指令，並僅在被詢問時才解說分享流程；Workflow（工作流程工具）拒絕 scriptPath 時，會提示改用 script 行內傳入腳本，這種方式可在沒有 Read 讀取工具的工作階段使用。
在修改擴充模組並即時查看效果時，漏接存檔可能讓你仍看到舊版本。熱重載（hot reload，修改程式碼後不重啟即時生效的技術）已修好兩種漏接：/reload-plugins 或工作階段啟動正在讀取模組檔案時存檔，以及上次重載後幾毫秒內再次存檔。重載詢問直接放行但未選答案後，不再每回合重複出現；SDK 宿主若在回答前關閉詢問，詢問最多會出現三次，不再整個工作階段都關閉熱重載。外掛選項命名為 constructor 或 prototype 時，不再一直讀到預設值，編輯後也能重新載入。加入大型介面的外掛，在外掛掛鉤背景工作單元停滯時，不再使另一個外掛被卸載。
透過外掛管控對話設定或改寫輸入時，設定與記錄應該反映外掛的決定。使用 /model、/fast、/output-style 或在 /config 切換自動更新管道，會先詢問 config.set 掛鉤（特定事件觸發的小程式），再儲存設定。prompt.submit 掛鉤改寫或捨棄輸入後，提示詞歷程與對話記錄中的待發送提示詞紀錄，不再保存改寫或捨棄前的原始輸入。claude plugin test 不再錯誤放行會移除工具呼叫、工具結果或思考區塊的 session.append 掛鉤，這些區塊在真實工作階段會保留。-p 與 SDK 工作階段不再把產生的型別檔寫入 --plugin-dir 外掛資料夾，只寫到正在開發模組的位置。掛鉤模組程式碼巢狀數千層時，載入或驗證失敗並只顯示堆疊溢位訊息的問題也已修好。
在 Windows 環境或多個連線環境間使用外掛時，可能遇到路徑與管理歸屬判斷錯誤。VS Code 在 Windows 上因磁碟代號大小寫不同而無法載入專案範圍外掛的問題已修好。Windows 的 plugin install、uninstall 在專案或本機範圍操作時，也不再因 C: 與 c: 這類差異漏掉安裝紀錄或新增重複紀錄。長時間運作的工作階段，在另一個工作階段同步 claude.ai 外掛更新後，掛鉤不再錯報「Plugin directory does not exist」。託管設定會列出或啟用外掛的機器上，遭竄改的伺服器託管設定快取（把查過的資料暫存起來直接用）將個人外掛誤判為組織管理外掛的問題也已修好。
企業透過閘道轉送雲端模型請求時，可以限制等待與模型去向。timeouts.upstream_ttfb_ms 可設定 Bedrock、Vertex、Foundry 等上游的串流開始等待毫秒數，逾時就切換備援（主通道失敗時接手的替代方案）或回傳 502。每個上游可設 models 清單，只轉送列出的模型，備援轉送也受限制；每個項目中的一個 * 可作萬用字元。Amazon Bedrock 位於 AWS 美國地理區域以外時，未列於 models: 的模型請求原本會被 AWS 拒絕；閘道現在會嘗試你自己的區域，若仍遭拒，會指出需加入的模型。背景請求優先使用 Haiku 4.5，閘道未提供時才退回工作階段原本的模型。
如果企業需要集中管控員工的登入管道與使用軌跡，可以設定登入去向並比對請求紀錄。沒有託管設定的機器，可在使用者設定加入 forceLoginMethod: "gateway" 與 forceLoginGatewayUrl，讓 /login 開啟指定閘道。模型推論的審計紀錄新增 upstream_request_id，成功回應標頭加入 request-id，方便對照雲端上游、閘道日誌與 Claude Code 遙測。機器已設定 OpenTelemetry（收集監控資料的標準），或工作階段已登入閘道時，登入驗證事件能正確匯出。若遇到內附 Claude Desktop 結構定義尚不認識的 desktop 原則鍵，閘道會帶著警告啟動並提供該設定，不必為這類新桌面設定升級閘道。
管理員查看閘道用量時，工作階段移出群組後不再殘留舊群組與花費上限。Amazon Bedrock 閘道改用 AWS CountTokens API 計算權杖（token，AI 處理與計費的文字單位），取代向模型請求單一權杖的方式，供 /context 與大檔讀取使用；需授予 bedrock:CountTokens 權限。PostgreSQL 資料庫處於唯讀狀態時，每 30 秒記錄警告，說明哪些操作失敗及如何復原。CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC 不再隱藏 /usage 與狀態列的閘道花費上限，花費上限請求會送往目前工作階段登入的閘道。已登入閘道的工作階段，availableModels 中的 Fable 模型現在會顯示於 /model 選單，以內建列取代同模型的 modelPicker 列。
讓 AI 執行命令或讀取檔案時，參數小差異不再直接卡住工具。Bash（執行終端機命令的工具）收到 command_description 而非 description 時不再失敗；Read 收到空的 pages 參數會視為未填。Grep 搜尋支援附帶 -l、-c、-r 旗標，不再因此呼叫失敗。--tools 與 --restricted 的限制會套用到啟動後才註冊的內建工具，已棄用的工具名稱也不再繞過呼叫者的工具集合。Bash 對萬用字元迴圈的權限檢查已修正。無法取得功能旗標（控制功能是否開啟的開關）時，Claude 3 Opus、Claude 3.x Sonnet 工作階段不再被提供它們會拒絕的工具搜尋。
當處理龐大程式碼庫或超長上下文時，系統與模型之間的溝通更加穩健。使用標記百萬上下文（[1m]）的模型時，若雲端平台拒絕該測試旗標，系統會自動移除旗標並重新發送請求，避免整批請求直接失敗。在檔案讀寫方面，即使檔案修改時間未更新但內容已變動，或是終端機指令未印出內容，編輯工具都不會誤以為檔案已經完整讀取過。大型檔案上傳失敗且未提供原因時，Claude 得到的說明已改善，會回報失敗而不縮小檔案。xhigh 與 max 思考力度下，網路搜尋及 agent 掛鉤（特定事件觸發的小程式）評估耗時過長的問題已修好。
在無頭模式（headless，沒有終端機視窗的執行模式）或 SDK 程式整合中連接遠端工具時，斷線後不再一直離線。遠端 MCP（讓 AI 連接外部工具的協定）伺服器中斷超過 15 秒後無法重連的問題已修好；反覆連上就斷線時，重連間隔會逐步拉長，最多 30 秒。伺服器錯誤回覆含網路錯誤名稱，不再使連線被誤切斷；分頁指標重複也不再反覆索取同頁。等待資源清單時按中斷鍵，能結束回合而無需等逾時。WebSocket 傳輸收到超過 16 MiB 的訊息，不再解析並會關閉連線，與其他傳輸方式的上限一致。
透過外部工具回傳程式碼、樣式或圖示檔案時，檔案格式現在能正確保留。以前 MCP（讓 AI 連接外部工具的協定）工具回傳的 CSS、JavaScript、XML 容易被存成 .bin，使 Read 拒絕讀取；現在這些檔案及字型、圖示都有對應副檔名。工具搜尋載入的說明上限由 2,048 放寬至 16,384 字元。claude mcp serve 的背景 Bash 結果會指出輸出檔名，工具說明也修正了原本承諾會收到、實際卻永遠不會到達的通知。不取得功能旗標的安裝中，claude.ai 連接器預設協商 MCP 協定版本 2026-07-28；功能旗標是控制功能是否開啟的開關，可用 MCP_PROTOCOL_NEGOTIATION=legacy 退出這項協商變更。
當你將耗時較長的 MCP（讓 AI 連接外部工具的協定）工具呼叫放到背景執行時，任務動態的同步變得更加清晰。如果你在任務面板或連線客戶端中手動中止了執行中的 MCP 工具，Claude 會得知工具已停止。工具轉移到背景時所跳出的系統通知，現在會顯示完整的任務識別碼以利辨識。在自動化無人值守的非互動工作階段中，若任務完全不需要使用特定 MCP 伺服器，不再跳出多餘的登入驗證提示；無頭模式（沒有終端機視窗的執行模式）與 SDK 工作階段中，延遲完成的登入或重連，不再把 setMcpServers() 設定的 MCP 伺服器誤列為使用者自己的伺服器。
讓子代理（分頭執行子任務的 AI 助理）在獨立工作樹工作時，不再混入主工作階段的 Git 分支、狀態與近期提交。工作樹是同專案的平行副本。Workflow 子代理的技能結果回給呼叫者；代理夥伴首次 SendMessage 不再因工具未載入而失敗。工作階段遭終止後第二次恢復時，也會提醒 Claude 可用 SendMessage 恢復背景子代理。名為 worker 的自訂代理，不再於自動模式拒絕通知及任務詳情活動清單中誤顯示為「Agent」。Skill 工具先於回應串流結束時，allowed-tools 與 effort 不再遺失；兩欄指定可用工具與思考力度，避免 -p 的 Bash 指令遭拒。子代理從 skills 最多預載 32 個技能、各一次，有 Skill 工具仍可呼叫其餘技能。
按 ← 將工作移到背景、還在等工具完成時，停止與繼續工作的操作有不同結果。停止回合會顯示「Backgrounding cancelled」；輸入文字取消背景化，Claude 則繼續在前景工作。CLAUDE_AUTO_BACKGROUND_TASKS 不再於子代理後方仍有編輯或命令排隊時搶先背景化，避免後方呼叫過早開始。macOS 或未安裝服務的 Linux 上，claude agents 與背景服務一起停止時，執行中的工作階段約一分鐘後停止並顯示通知；期間再次執行 claude agents 可避免停止。兩者退出時不再意外啟動新服務，拖延關機或重啟中斷的工作階段。恢復時也不再誤告知 Claude 未完成的背景工作流程可能遭 TaskStop 停止。
使用 /loop 定時檢查時，離開連線畫面與停止循環是兩件事。連接背景工作階段，在閒置提示字元按兩次 Ctrl+C 只會離開連線，待執行的喚醒仍保留；按 Esc 才停止循環，取消待執行喚醒也會顯示通知。用 ← 移到背景的自定步調 /loop，Esc 無法停止的問題已修好。若下次喚醒在程序停止期間到期，循環會停止；現在工作階段會顯示通知，Claude 也會得知。無人看守的重試模式 CLAUDE_CODE_RETRY_WATCHDOG 可用 CLAUDE_CODE_RETRY_WATCHDOG_MAX_WAIT_MS 環境變數（啟動程式前提供的設定），限制遇到 429 用量限制或 529 伺服器超載時的最長等待毫秒數。
使用 /rewind 倒帶對話或分支出新主題時，不會再遺失提示詞或誤重啟已停止的任務。倒帶後送出的提示詞，在工作階段移到背景或遭終止後恢復時不再遺失，被刪除的回合也不再重現。已停止的排程不再於倒帶後恢復時重啟；對話壓縮後按 Esc 或倒帶，再恢復工作階段時遺失排程的問題也已修好。剛按 ← 後，用 ↑ 或 Esc 將待發送訊息拉回輸入框，提示詞歷程會保留它。執行 /resume 或 /branch 後，SessionStart 掛鉤（啟動工作階段時觸發的小程式）寫入 CLAUDE_ENV_FILE 環境檔案的變數能傳給 Bash 工具；未變更的非同步啟動掛鉤上下文，也不再於每次恢復時重複加入對話。
與 Claude 互動或將輸出交給其他程式時，狀態與回覆不該漏掉。支援程式狀態協定 OSC 7501 的終端機，能在分頁或狀態列顯示 Claude 正在工作、等你輸入或已完成。/copy 可選取引言文字，複製時不帶 >。管線（將輸出接給另一個程式）中緊接淡色字的粗體字，不再被畫成淡色。背景工作階段的命令與掛鉤不再繼承 FORCE_COLOR=3，避免 Claude 讀到顏色跳脫碼。claude -p 若因背景工作開啟新回合，每回合回覆會在回合結束時印出，不再丟失早先回覆。最後回合結束後仍保持開啟時，只有標準錯誤輸出 stderr 連到終端機，才會印出正在等待什麼的說明。
在終端機打字、貼上特殊字元或點擊超連結時，游標與連結顯示不該誤導你。從 macOS 剪貼簿將帶重音標記的文字貼到提示字元中間，游標不再多跳一格。vim（以按鍵指令編輯文字的模式）的 ~ 不再把游標移過行尾，使後續 x 刪除指令無效；3~ 也不再跨入下一行。回覆或代理夥伴訊息中的原始終端機超連結位元組，不再被畫成隱藏位址的可點連結；文字網址也不再被殘留的顏色或隱藏樣式遮住，包括表格換行與問題預覽。/model 的 Max 思考力度會正確標示只適用目前工作階段。/tui 無法重新啟動 Claude Code 時，無訊息退出或顯示原始系統錯誤的問題也已修好。
當 Claude 產生數萬行回覆或逐漸加深的巢狀引言時，終端機曾卡到難以操作。數萬行回覆造成畫面凍結、Ctrl+C 失效，以及引言每隔幾行就加深而造成數秒停頓與數 GB 記憶體用量的問題已修好。語法高亮（用顏色區分程式碼）遇到超長行、大量連續空白行、未閉合字串或 heredoc 時，停頓數秒至數分鐘的問題也已修好；heredoc 是標記一整段文字起訖的寫法。含 Tab 或雙向控制字元的文字，不再於螢幕邊緣被截斷或覆蓋鄰近行。Tab 定位從文字起點計算；Windows 換行格式 CRLF 的短輸出不再擠成一行。
使用遠端操控（Remote Control，從其他裝置操作工作階段）或在 macOS 執行背景任務時，可能遇到不該出現的拒絕或權限提示。遠端一次傳送超過 256 個檔案遭「could not be vouched for」拒絕的問題已修好。claude remote-control 剛收到手機訊息、未用 /config 開關，或從 Claude Code 內啟動時，推播已送出卻誤報未送出的問題也已修好。遠端操控、claude.ai 與桌面介面等待權限確認時，會顯示 MCP 工具的伺服器與好讀名稱，取代內部識別碼。在 macOS 家目錄或其上層啟動時，檔案計數誤觸其他應用程式資料存取提示的問題已修好；更新後一起重啟的背景工作階段，也不再偶爾跑到穩定的應用程式包裝外而再次引發資料夾權限提示。
設定網址封鎖規則、自動模式或清理磁碟時，先前幾種不合預期的行為已修好。Chrome 擴充功能的 host:80 封鎖規則，能套用到該主機的一般 HTTP 網頁。移除 disableAutoMode 後，桌面端與 SDK 程式整合不用重新啟動即可切回自動模式。中斷 /ultrareview 上傳後留在 ~/.claude/seed-admin 的暫存檔，會由保留期限清理機制移除。/advisor 保存的顧問模型已不可用時，面板會開在「No advisor」；登入重新整理後，啟動時偶爾載入託管設定失敗的問題也已修好。無頭模式（沒有終端機視窗的執行模式）的 rate_limit_event 用量上限警告，會說明帳號是否開啟額外用量。
瀏覽 Claude 產出的互動式成果（Artifact，產出的獨立網頁或程式成果預覽）時，視覺與操作體驗更加流暢。未自訂背景顏色的成果頁面現在改以純白底色呈現，取代以往偏灰的米白色。在關閉診斷遙測傳輸的環境下，成果工具的權限確認對話簡化為與一般環境相同的五個核心問題，不再於每次讀取他人成果或修改資料時反覆打斷；在關閉遙測的安裝中，排程與 Run now（立即執行）例行工作可免詢問發布新的私有成果，或更新自己的成果。
在 VS Code 編輯器內與 Claude 協作時，檔案查閱與多視窗切換不再互相干擾。聊天室會將 Claude 傳送的檔案獨立列出一行，點擊檔名即可在編輯器開啟，下方附有說明，專注檢視中也保持可見。在背景代理活動或面板自身狀態列之後，使用「從此處分岔對話」或倒帶而恢復失敗、顯示「Message not found in session」的問題已修好。輸入框的 Ctrl+Shift+Tab、Cmd+Shift+Tab、Alt+Shift+Tab 不再誤切換權限模式，Ctrl+Tab 也不再誤接受建議提示詞。兩個 Claude 視窗同時可見時，鍵盤焦點不再來回跳動，導致文字送到錯的對話。
在雲端工作階段中運行長時任務或排程時，斷線重連與設定體驗均有感提升。解除封存後立即傳送訊息，偶爾因無法啟動容器（執行工作的隔離環境）而沒有回覆的問題已修好，舊版例行排程也會如實帶入預存的提示詞。當網路連線短暫中斷後重新連上時，漏掉的對話紀錄會整批一次性呈現在畫面上，不再逐行緩慢爬出。雲端環境設定表單中的安裝腳本輸入框預設高度增加，不僅能隨內容自動擴展，也支援自由拖曳調整大小。
在 Slack 團隊頻道用 @Claude 協作時，確認卡片與通知有幾項調整。設定確認卡片從 10 分鐘延長至 30 分鐘。Enterprise Grid（Slack 的多工作區企業方案）跨工作區共享頻道使用組織預設設定的通知，只在有人標記 Claude 時發送，每月至多一次。分岔討論串承接卡片的頻道連結與成員標記，不再顯示原始 Slack 碼。@Claude !restart 確認重啟後，運算中旋轉圖示不再持續轉動；忙碌頻道停止讀取未標記訊息後，其他應用程式或機器人的 @Claude 不再漏回覆。管理員新增頻道管理者時，若會將成員移至 Custom roles（自訂權限角色），可能移除其他存取權，操作前會先確認。
在企業自託管環境執行任務或自動程式碼審查時，大型專案曾因等待限制或規則檔太大而出問題。大型 Git 儲存庫抓取期間，伺服器仍回報進度卻被中斷的問題已修好；環境變數（啟動程式前提供的設定）CLAUDE_RUNNER_FETCH_SERVER_PROGRESS_CAP_MS 可調整或關閉等待限制。執行器不再將 CCR_AUTO_MODE_ALLOW、CCR_AUTO_MODE_ENVIRONMENT、CCR_AUTO_MODE_SOFT_DENY 傳入工作階段環境。重新審查時，先前發現的問題仍未解決，評論會指出剩餘數量，不再誤報零問題。合併請求修改大型 CLAUDE.md 專案規則檔時，程式碼審查仍跳過其中規則的問題也已修好。
When you add custom hooks (small programs triggered by specific events), startup failures can now block actions. Command and HTTP hooks support onFailure: "block", blocking subsequent actions if they fail to start, time out, or return an unexpected exit code. Deeply nested tool inputs are no longer truncated to bypass protections. During plugin hook worker restarts and module reloads, calls that previously bypassed another module's shielding hook with .catch are now rejected. Multi-line JSON printed by async hooks is parsed correctly, and cloud sessions no longer hang when a session.receive hook requests permissions upfront. If a tool execution is rejected by a plugin after running, both Claude and you are informed that the tool executed but its output was intercepted.
When managing extensions, silent configuration load failures can cause unexpected issues. Running claude plugin install, enable, disable, or adding a marketplace now issues a warning if the target configuration file fails to load. Pressing Enter on the /plugin error tab no longer removes the marketplace and its plugins without confirmation. Adding a marketplace is rejected if its name cannot be used to install plugins; adding or refreshing marketplaces with large Git submodules now fetches only submodules containing plugin files. Submodules are separate repositories nested inside a repository. When claude plugin validate finds a README missing install instructions, it prints pasteable suggestions without affecting the exit code, even when using --strict. When a hook module is rejected due to top-level var rebinding, the error message indicates the line number, reason, and remediation.
When building or using custom UI modules, layout and alerting options are now more flexible. Modules can now deliver native desktop notifications via $.ui.notify respecting system preferences and clearly attributing the source. Button components support text child nodes, allowing labels or secondary details to be rendered within a single clickable button. Toast popups published by your organization are prioritized ahead of other modules. The built-in plugin authoring skill no longer prompts to run commands in sessions without a terminal and only explains sharing flows when asked. When Workflow (a workflow tool) rejects scriptPath, it prompts you to use script to pass inline scripts instead, which works in sessions without Read tools.
When modifying extension modules and inspecting changes in real time, missed file saves can leave you looking at outdated code. Hot reload (applying code changes immediately without restarting) has fixed two missed save cases: saving files while /reload-plugins or session startup is reading module files, and saving again within milliseconds of the previous reload. Reload prompts allowed without an explicit selection no longer recur every turn. If an SDK host dismisses prompts before answering, the prompt appears up to three times rather than disabling hot reload for the entire session. Plugin options named constructor or prototype no longer continuously read default values and reload cleanly after edits. Plugins with large UI surfaces no longer cause another plugin to unload when a plugin hook worker stalls.
When managing session settings or rewriting inputs via plugins, configs and histories should reflect plugin decisions. Using /model, /fast, /output-style, or toggling auto-update channels in /config now consults config.set hooks (small programs triggered by specific events) before persisting settings. After a prompt.submit hook rewrites or discards input, prompt history and session pending prompt logs no longer retain the original input prior to rewriting or discarding. claude plugin test no longer erroneously passes session.append hooks that strip tool calls, tool results, or thinking blocks, which are preserved in real sessions. Headless -p and SDK sessions no longer write generated type definitions to the --plugin-dir folder, writing only to the module location under active development. An issue where hook module code nested thousands of levels deep failed loading or validation with only a stack overflow error has also been fixed.
When using plugins on Windows or across multiple connection environments, path and governance misclassifications can arise. An issue where VS Code on Windows failed to load project-scoped plugins due to drive letter casing mismatches has been fixed. Windows operations with plugin install and uninstall in project or local scopes no longer drop install records or add duplicates because of C: versus c: differences. In long-running sessions, hooks no longer falsely report "Plugin directory does not exist" after another session syncs claude.ai plugin updates. On machines where managed configurations list or enable plugins, a bug where a tampered server-managed config cache misclassified personal plugins as organization-managed plugins has also been resolved.
When enterprises route cloud model requests through a gateway, they can constrain wait times and model destinations. timeouts.upstream_ttfb_ms configures the time-to-first-byte waiting threshold in milliseconds for upstreams like Bedrock, Vertex, and Foundry, switching to fallbacks or returning a 502 on timeout. Each upstream can define a models list to route only listed models, which also constrains fallback routing; a single * in each entry acts as a wildcard. When Amazon Bedrock is located outside AWS US regions, model requests not listed under models: were previously rejected by AWS; the gateway now tries your own region and, if still rejected, identifies which model needs to be added. Background requests prioritize Haiku 4.5, falling back to the session's original model only when the gateway does not provide it.
Enterprises needing centralized governance over employee login channels and usage trails can configure login destinations and correlate request logs. Machines without managed configurations can add forceLoginMethod: "gateway" and forceLoginGatewayUrl in user settings to direct /login to the designated gateway. Model inference audit logs now include upstream_request_id, and successful response headers append request-id to facilitate cross-referencing between cloud upstreams, gateway logs, and Claude Code telemetry. When a machine has OpenTelemetry (a standard for collecting monitoring data) configured or a session is logged into the gateway, login authentication events are properly exported. If the gateway encounters desktop policy keys not yet recognized by the bundled Claude Desktop schema, it starts with a warning and serves the configuration without requiring a gateway upgrade for new desktop settings.
When administrators review gateway usage, sessions moved out of a group no longer retain stale groups or spend limits. The Amazon Bedrock gateway now uses the AWS CountTokens API to calculate tokens (text units used for AI processing and billing) instead of requesting a single token from the model, serving /context and large file reads; this requires granting bedrock:CountTokens permissions. When the PostgreSQL database enters a read-only state, warnings are logged every 30 seconds detailing failed operations and remediation steps. CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC no longer hides gateway spend limits from /usage and the status line, and spend limit requests are routed to the gateway logged into by the current session. For sessions logged into a gateway, Fable models in availableModels are now displayed in the /model menu, replacing modelPicker rows with built-in rows for the same model.
When having AI execute commands or read files, minor parameter differences no longer break tool invocations. Bash (a tool for running terminal commands) no longer fails when receiving command_description instead of description; Read treats an empty pages parameter as unset. Grep searches now support appending -l, -c, and -r flags without causing call failures. Restrictions from --tools and --restricted apply to built-in tools registered after startup, and deprecated tool names no longer bypass caller toolsets. Bash permission checks for wildcard loops have been fixed. When feature flags (switches controlling whether features are enabled) cannot be retrieved, Claude 3 Opus and Claude 3.x Sonnet sessions are no longer offered tool searches they would otherwise reject.
Communication between the system and models is more resilient when handling massive codebases or very long contexts. When using models tagged with 1-million context ([1m]), if the cloud platform rejects the test flag, the system automatically removes the flag and resends the request so the entire batch does not immediately fail. For file reading and writing, editor tools no longer assume a file has been read in its entirety if the modified timestamp remains unchanged despite content updates, or if a terminal command produces no output. When large file uploads fail without a specified reason, the explanation given to Claude is improved, reporting failure rather than shrinking the file. Issues where web searches and agent hooks (small programs triggered by specific events) took excessively long to evaluate under xhigh and max thinking efforts have been fixed.
Connecting to remote tools in headless mode (an execution mode without a terminal window) or SDK integrations no longer remains permanently offline after a disconnection. An issue where remote MCP (a protocol connecting AI to external tools) servers failed to reconnect after disconnecting for over 15 seconds has been fixed. When connections drop repeatedly right after connecting, reconnection intervals now back off progressively up to 30 seconds. Server error responses containing network error names no longer mistakenly terminate connections, and duplicate pagination cursors no longer fetch the same page repeatedly. Pressing interrupt while awaiting resource lists now terminates the turn without waiting for timeouts. WebSocket transports receiving messages over 16 MiB now skip parsing and close the connection, matching limits on other transports.
File formats are now preserved correctly when returning code, style, or icon files through external tools. Previously, CSS, JavaScript, and XML returned by MCP (a protocol connecting AI to external tools) tools were often saved as .bin, causing Read to reject them; now these files as well as fonts and icons receive matching extensions. The description limit loaded by tool search has been expanded from 2,048 to 16,384 characters. Background Bash results in claude mcp serve now state the output filename, and tool descriptions have been corrected to remove promises of notifications that never arrive. In installations that do not fetch feature flags, the claude.ai connector negotiates MCP protocol version 2026-07-28 by default. Feature flags are switches controlling whether capabilities are enabled, and you can opt out of this negotiation change using MCP_PROTOCOL_NEGOTIATION=legacy.
Synchronization of task activity is clearer when running long-running MCP (a protocol connecting AI to external tools) tool calls in the background. If you manually abort a running MCP tool from the task popover or connected client, Claude learns that the tool has stopped. System notifications displayed when tools move to the background now show the full task ID for easy identification. In automated, unattended non-interactive sessions, unnecessary authentication prompts no longer appear when tasks do not require specific MCP servers. In headless mode (an execution mode without a terminal window) and SDK sessions, delayed logins or reconnections no longer mistakenly list MCP servers configured via setMcpServers() as user-provided servers.
When subagents (AI assistants running subtasks in parallel) work in separate worktrees, they no longer mix in Git branches, status, and recent commits from the primary session. A worktree is a parallel copy of the same project. Workflow subagent skill results now return to callers, and initial SendMessage calls between peer agents no longer fail due to unloaded tools. When resuming a terminated session for a second time, Claude is also reminded to use SendMessage to resume background subagents. Custom agents named worker no longer mistakenly appear as "Agent" in auto-mode rejection notices and task detail activity lists. When the Skill tool finishes before the response stream ends, allowed-tools and effort are no longer lost; these fields specify allowed tools and thinking effort, preventing -p Bash commands from being rejected. Subagents preload up to 32 skills once each from skills, and can still invoke remaining skills when the Skill tool is available.
When pressing ← to background work while waiting for tools to finish, stopping versus continuing work now yields distinct outcomes. Stopping a turn displays "Backgrounding cancelled", while typing input cancels backgrounding and lets Claude continue in the foreground. CLAUDE_AUTO_BACKGROUND_TASKS no longer eagerly backgrounds tasks when edits or commands remain queued behind a subagent, preventing subsequent calls from starting prematurely. On macOS or Linux without installed services, when claude agents stops alongside the background service, active sessions stop after roughly one minute with a notification; running claude agents again during this window prevents the shutdown. Exiting both no longer inadvertently starts a new service, which previously delayed shutdown or restarted interrupted sessions. Resuming also no longer mistakenly informs Claude that incomplete background workflows may be stopped by TaskStop.
When using /loop for scheduled checks, disconnecting from the session and stopping the loop are separate actions. In a connected background session, pressing Ctrl+C twice at an idle prompt only disconnects the client while preserving pending wakeups; pressing Esc stops the loop and notifies you of cancelled wakeups. An issue where Esc could not stop self-paced /loop tasks moved to the background with ← has been fixed. If the next wakeup expires while the process is stopped, the loop terminates; sessions now display a notification and Claude is informed. Unattended retry mode CLAUDE_CODE_RETRY_WATCHDOG can use the CLAUDE_CODE_RETRY_WATCHDOG_MAX_WAIT_MS environment variable (configuration provided before starting a program) to cap maximum wait milliseconds upon encountering 429 rate limits or 529 server overloads.
Prompts are no longer lost and stopped tasks are not accidentally restarted when using /rewind to rewind conversations or branch into new topics. Prompts submitted after rewinding are preserved when sessions move to the background or resume after termination, and deleted turns no longer reappear. Stopped schedules no longer restart upon resuming after a rewind, and an issue where schedules were lost when resuming after conversation compaction followed by Esc or rewind has been fixed. Using ↑ or Esc immediately after pressing ← to pull a pending message back into the input box preserves it in prompt history. After running /resume or /branch, variables written to the CLAUDE_ENV_FILE environment file by SessionStart hooks (small programs triggered on session start) are passed to the Bash tool. Unchanged async startup hook context is also no longer repeatedly added to the conversation across resumes.
Status and responses are no longer missed when interacting with Claude or piping output to other programs. Terminals supporting the application status protocol OSC 7501 can display in tabs or status lines whether Claude is working, awaiting input, or finished. /copy allows selecting blockquote text, copying it without >. Bold text directly following dim text in pipelines (passing output to another program) is no longer rendered dim. Commands and hooks in background sessions no longer inherit FORCE_COLOR=3, preventing Claude from reading ANSI color escape codes. When claude -p opens new turns due to background tasks, each turn's response prints when the turn ends, preventing earlier responses from being lost. When remaining open after the final turn ends, explanations of what is being awaited only print if standard error stderr is connected to a terminal.
Cursor and link displays no longer mislead you when typing, pasting special characters, or clicking hyperlinks in the terminal. Pasting text with accented characters from the macOS clipboard into the middle of the prompt no longer moves the cursor an extra cell. In vim (a mode for editing text using keyboard commands), ~ no longer advances the cursor past the end of the line to break subsequent x delete commands, and 3~ no longer spans across to the next line. Raw terminal hyperlink bytes in responses or peer agent messages are no longer rendered as clickable links with hidden URLs, and plaintext URLs are no longer obscured by lingering colors or hidden styling, including table line wraps and question previews. The Max thinking effort in /model is properly labeled as applying only to the current session. Silent exits or raw system errors when /tui fails to restart Claude Code have also been fixed.
When Claude generated tens of thousands of lines of output or progressively deepening nested blockquotes, the terminal could become completely unresponsive. Freezes caused by tens of thousands of lines of output, Ctrl+C failures, and multi-second pauses with gigabytes of memory usage from blockquotes deepening every few lines have been fixed. Syntax highlighting (using colors to distinguish code) stalling for seconds to minutes on ultra-long lines, large blocks of consecutive blank lines, unclosed strings, or heredocs is also fixed; a heredoc is a syntax for defining multi-line text blocks. Text containing tabs or bidirectional control characters is no longer clipped at screen edges or overlapping adjacent lines. Tab stops are now calculated from the start of the text, and short outputs with Windows CRLF line endings are no longer collapsed into a single line.
When using Remote Control to manage sessions from other devices or running background tasks on macOS, unexpected rejections and permission prompts could occur. An issue where sending more than 256 files at once remotely was rejected with 'could not be vouched for' has been fixed. False warnings reporting that push notifications were not sent when claude remote-control just received a mobile message, had not used the /config toggle, or was started from inside Claude Code are now resolved. When Remote Control, claude.ai, and desktop interfaces await permission confirmation, they display the MCP tool's server and human-readable name instead of internal identifiers. File counting when launched in the macOS home directory or any directory above it no longer triggers false data access prompts for other applications. Background sessions restarted alongside updates no longer occasionally fall outside the stable application bundle and re-trigger folder permission prompts.
Several unexpected behaviors when configuring URL block rules, Auto-Mode, or performing disk cleanup have been resolved. The Chrome extension's host:80 blocking rule now correctly applies to standard HTTP pages on that host. After removing disableAutoMode, desktop and SDK programmatic integrations can switch back to Auto-Mode without restarting. Temporary files left in ~/.claude/seed-admin after interrupted /ultrareview uploads are now removed by retention-based cleanup. When an advisor model saved by /advisor is no longer available, the panel opens to 'No advisor', and occasional startup failures loading managed settings after login refreshes are fixed. The rate_limit_event warning in headless mode (an execution mode without a terminal window) clarifies whether the account has additional usage enabled.
Viewing interactive Artifacts (standalone web pages or previews of generated programs) generated by Claude offers a smoother visual and operational experience. Artifact pages without custom background colors now render on a pure white background instead of the previous grayish off-white. In environments with diagnostic telemetry disabled, Artifact tool permission prompts are streamlined to the same five core questions as standard setups, avoiding repetitive interruptions when reading others' artifacts or modifying data. Scheduled and Run now routine jobs in telemetry-disabled installations can now publish new private artifacts or update your own artifacts without prompting.
When collaborating with Claude in VS Code, file navigation and multi-window switching no longer interfere with each other. Files sent by Claude are listed on dedicated lines in the chat; clicking a filename opens it in the editor with a description beneath it, remaining visible in focused views. An issue where using 'Fork conversation from here' or rewind failed to restore after background agent activity or the panel's own status bar, returning 'Message not found in session', has been resolved. In the input box, Ctrl+Shift+Tab, Cmd+Shift+Tab, and Alt+Shift+Tab no longer accidentally toggle permission modes, and Ctrl+Tab no longer unintentionally accepts suggested prompts. When two Claude windows are visible simultaneously, keyboard focus no longer jumps back and forth to send text to the wrong conversation.
Running long-running tasks or schedules in Cloud sessions brings noticeable improvements to reconnection and configuration. An issue where sending messages immediately after unarchiving sometimes produced no response due to container (an isolated environment for running tasks) startup failures is resolved, and legacy routine schedules reliably load pre-saved prompts. When reconnecting after brief network drops, missed conversation history renders on screen as a single batch rather than slowly crawling line by line. The default height of the install script input field in cloud environment settings has been increased, expanding automatically with content while supporting manual drag-resizing.
When collaborating with @Claude in Slack team channels, several adjustments have been made to confirmation cards and notifications. Setting confirmation cards now remain active for 30 minutes, up from 10 minutes. Notifications for Enterprise Grid (Slack's enterprise plan for multiple workspaces) cross-workspace shared channels applying organization defaults now trigger at most once a month, only when someone mentions Claude. Forked thread continuation cards now render proper channel links and member mentions instead of raw Slack markup. The spinning indicator no longer persists after confirming an @Claude !restart, and @Claude mentions from other apps or bots are no longer dropped after a busy channel stops reading unmentioned messages. When adding channel managers moves members into Custom roles (custom permission roles) and potentially removes other permissions, admins are prompted for confirmation beforehand.
When running tasks or automated code reviews in self-hosted enterprise environments, large projects previously encountered issues with timeout limits or oversized rules files. An issue where large Git repository fetches aborted while the server was actively reporting progress has been resolved; the CLAUDE_RUNNER_FETCH_SERVER_PROGRESS_CAP_MS environment variable (configuration provided before starting a program) allows adjusting or disabling this timeout. The runner no longer passes CCR_AUTO_MODE_ALLOW, CCR_AUTO_MODE_ENVIRONMENT, or CCR_AUTO_MODE_SOFT_DENY into the session environment. During re-reviews where previously detected issues remain unresolved, review comments accurately report the remaining count rather than falsely claiming zero issues. An issue where code review skipped rules when a pull request modified a large CLAUDE.md project rules file has also been fixed.
Added `onFailure: "block"` for command and HTTP hooks: a hook that can't start, times out, or exits with an unexpected code blocks the action instead of letting it through
Fixed a mod's hook being handed a deeply nested tool input cut short with no error, so a guard could pass content it never saw
Fixed calls a mod makes while it reloads during a plugin hooks worker restart getting past another mod's guard hook that has a `.catch`; such calls are now refused
Fixed an async hook's JSON output being ignored when it is printed over several lines
Fixed cloud sessions hanging when a mod's `session.receive` hook asked for permission before passing a message on
Changed what Claude and you read when a mod denies a tool call after the tool ran: it now says the tool ran and a plugin withheld its resultAdded a warning to `claude plugin install`, `enable`, `disable` and `marketplace add` when the settings file they write to does not load
Added advice to `claude plugin validate` when a plugin's README has no install line: it prints the line to paste and never changes the exit code, even with `--strict`
Fixed `/plugin`'s Errors tab removing a marketplace that failed to load, and uninstalling its plugins, on Enter without asking first
Fixed `claude plugin marketplace add` reporting success for a marketplace whose name no plugin can be installed under; such an add is now refused
Fixed plugin marketplaces whose repository has large git submodules failing to add or refresh: only the submodules that hold plugin files are now fetched
Improved `claude plugin validate` and plugin loading: when a hooks module is refused over a rebound top-level `var`, the error now names the line that rebinds it, the cause, and a fixAdded `$.ui.notify` for mods: raises a native notification through your own notification setting and says which channel sent it
Added children to a mod's `Button`: strings and `Text`, so a row of a list is one pressable with a chip or a dim detail inside
Improved the built-in `plugin-authoring` skill: it no longer tells a session with no terminal to run terminal commands, and explains sharing a mod only when asked
Improved the Workflow tool's `scriptPath` refusal: it now says to pass the script inline via `script`, the route that works in sessions without a Read tool
Changed toasts from your organization's mods to show ahead of other mods' toasts instead of waiting behind themFixed plugin options named `constructor` or `prototype` always reading as their default and never reloading the plugin when edited
Fixed a mod's hot reload missing a save made while `/reload-plugins` or session start was reading the mod's files
Fixed the mod hot-reloading question being asked again every turn for the whole session when it is approved without picking an answer
Fixed a mod's hot reload missing a save made within a few milliseconds of its last reload, which left the mod on its old version until the next save
Fixed mod hot-reloading turning off for the whole session when an SDK host closed its enable question before anyone answered; the question is now asked up to three times
Fixed a plugin that adds very large interfaces getting another plugin unloaded when the plugin hooks worker stallsFixed plugins whose hooks module nests code thousands of levels deep failing to load or validate with a bare stack-overflow message
Fixed `claude plugin test` passing a `session.append` hook that removes tool call, tool result or thinking blocks that a real session keeps
Fixed `/model`, `/fast` and `/output-style` saving their setting without asking a plugin's `config.set` hook
Fixed generated type files being written into a `--plugin-dir` plugin's folder in `-p` and SDK sessions; they are now written only where a mod is being developed
Fixed a prompt that a mod's `prompt.submit` hook rewrote or dropped still being saved as typed to prompt history and to the transcript's queued-prompt records
Fixed changing the auto-update channel in `/config` saving before a plugin's `config.set` hook was askedFixed a tampered cache of server-managed settings making a person's own plugin count as organization-managed on machines whose managed settings list or enable plugins
Fixed project-scope plugins not loading in VS Code on Windows due to drive-letter casing (anthropics/claude-code#74612)
Fixed `plugin install` and `uninstall` at project or local scope on Windows missing the install record, or adding a second one, when only the drive letter's case differs
Fixed claude.ai-synced plugin hooks failing with "Plugin directory does not exist" in long-running sessions after another session synced a plugin updateAdded support for `timeouts.upstream_ttfb_ms` on the Claude apps gateway's Bedrock, Vertex, Foundry and other cloud upstreams: a value you set now limits how long a stream may take to start there, after which it fails over or gets a 502
Added an optional `models` list to every Claude apps gateway upstream: only the listed models are sent there, on failover too, and one `*` in an entry is a wildcard
Improved background requests behind a Claude apps gateway to use Haiku 4.5 instead of the session's model, falling back to the session's model where the gateway does not serve Haiku 4.5
Improved how the Claude apps gateway handles models not listed under `models:` in Amazon Bedrock regions outside AWS's US geography. AWS always refused those requests; the gateway now tries the model in your own region, and a refusal names the model to addAdded support for `forceLoginMethod: "gateway"` and `forceLoginGatewayUrl` in your own user settings on machines with no managed settings, so `/login` opens on that Claude apps gateway
Added `upstream_request_id` to the Claude apps gateway's `inference` audit event: the request ID from Amazon Bedrock, the Anthropic API or another upstream, for support cases
Added a `request-id` header to the Claude apps gateway's successful inference responses, so `request_id` in Claude Code telemetry matches the gateway's audit log
Fixed the `claude_code.auth` OpenTelemetry login event not being emitted on Claude apps gateway sign-in. The event is exported when OpenTelemetry is configured on the machine or the session is already signed in to the gateway
Changed the Claude apps gateway to start with a warning and serve a `desktop` policy key its bundled Claude Desktop schema doesn't know, so new Desktop settings need no gateway upgradeFixed Fable models listed in `availableModels` not appearing in the `/model` picker for sessions signed in to a Claude apps gateway; such a model now shows its built-in row in place of a `modelPicker` row added for the same model
Fixed the Claude apps gateway's admin spend view showing a developer's old groups and group cap after their session stopped carrying groups
Fixed `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC` hiding the Claude apps gateway spend limit in `/usage` and the status line. The spend limit request goes to the gateway the session is signed in to
Improved token counts behind a Claude apps gateway on Amazon Bedrock, which `/context` and large file reads use: the gateway now gets them from AWS's CountTokens API instead of a one-token model request. Grant `bedrock:CountTokens` to use it
Improved the Claude apps gateway to log a warning every 30 seconds while its PostgreSQL database is read-only, saying what fails and how to recoverFixed Bash tool calls failing when the model passes `command_description` instead of `description`
Fixed `--tools` and `--restricted` not applying to built-in tools that register after launch, and deprecated tool names reaching tools outside the caller's tool set
Fixed Bash permission checks for for-loops over glob patterns, improving permission-check accuracy
Fixed Claude 3 Opus and Claude 3.x Sonnet sessions being offered tool search, which those models reject, where feature flags are unavailable
Fixed the Read tool rejecting calls with an empty `pages` parameter instead of treating it as omitted
Improved Grep input handling: a search sent with grep's `-l`, `-c` or `-r` flag now runs instead of failing the callFixed every request failing on a `[1m]` model when a gateway, Bedrock, Vertex or Foundry refuses the context-1m beta; Claude Code now resends without it
Fixed a file being treated as already read after a Bash command such as `cat` ran without printing it
Fixed Edit treating a file as fully read when its contents changed without its modification time advancing
Improved what Claude is told when a large file it sends fails to upload with no reason given, so it reports the failure instead of shrinking the file
Fixed web searches and `agent` hook evaluations taking much longer at `xhigh` and `max` effortFixed remote MCP servers in headless and SDK sessions staying disconnected after an outage longer than 15 seconds, or reconnecting in a tight loop to a server that drops each connection right after it connects; repeated drops now back off, up to 30s
Fixed remote MCP connections being dropped when a server's error reply happened to contain a network error name
Fixed MCP servers that repeat a pagination cursor being asked for the same page up to 20 times at every connect
Fixed an interrupt not ending the turn until a timeout while Claude waited for an MCP server's resource list
Changed WebSocket (`ws`) MCP servers: a message over 16 MiB is no longer parsed and closes the connection, the limit the other transports already haveFixed CSS, JavaScript and XML files returned by MCP tools being saved as .bin, which the Read tool refuses; font and icon files now get their own extension too
Fixed `claude mcp serve` background Bash results not naming the output file, and its tool description promising a notification that never arrives
Changed MCP tool descriptions the model loads through tool search to be cut at 16,384 characters instead of 2,048
Changed claude.ai connectors to negotiate MCP protocol version 2026-07-28 by default on installs that fetch no flags; `MCP_PROTOCOL_NEGOTIATION=legacy` opts outFixed Claude not being told when you stop a long-running MCP tool call from the tasks panel or a connected client
Fixed the notification for an MCP tool call that moved to the background showing a shortened task id
Fixed non-interactive sessions telling you about MCP servers that need authentication when your request doesn't use them
Fixed an MCP sign-in or reconnect that finishes late in headless and SDK sessions relisting a server set with `setMcpServers()` as your ownFixed subagents in their own linked worktree being shown the parent session's git branch, status and recent commits
Fixed a custom agent named `worker` being shown as "Agent" in auto-mode denial notices and in the activity lists of task detail views
Fixed forked skills invoked from Workflow subagents delivering their results to the main conversation instead of the invoking agent
Fixed in-process agent teammates failing their first SendMessage call until they loaded the tool through tool search
Fixed Claude not being told it can resume background subagents with SendMessage after a session was killed and resumed a second time
Fixed a skill's `allowed-tools` and `effort` being dropped when the Skill tool finished before the response stream ended, which denied the skill's Bash commands in `-p` runs
Changed subagents to preload at most 32 skills from the `skills` field, each once; a subagent with the Skill tool can still invoke the restAdded a "Backgrounding cancelled" message when you stop the turn while `←` is waiting for the current tool to finish
Fixed `CLAUDE_AUTO_BACKGROUND_TASKS` moving a subagent to the background while an edit or shell command waited behind it, which started that call before the subagent finished
Fixed Claude stopping at the end of the current tool when text typed in the prompt cancelled backgrounding with `←`; Claude now keeps working in the foreground
Fixed `claude agents` starting a new background service as it exited when both were stopped with a session open (for example at a reboot), which could delay shutdown and restart an interrupted session
Fixed Claude being told on resume that an unfinished background workflow may have been stopped by `TaskStop`, which can never be the cause
Changed `claude agents` stopped with its background service (macOS, or Linux without the service installed): running sessions now stop in about a minute unless it is run again, and a notice says soAdded `CLAUDE_CODE_RETRY_WATCHDOG_MAX_WAIT_MS` to limit how long unattended retry mode (`CLAUDE_CODE_RETRY_WATCHDOG`) waits out 429 and 529 errors
Fixed Esc not stopping a self-paced `/loop` that was moved to the background with `←`; cancelling a pending wakeup now shows a notice
Fixed a background session's `/loop` stopping without notice when its next wakeup came due while the session's process was down; the session now says so and Claude is told
Changed Ctrl+C at the idle prompt of an attached background session to leave a pending `/loop` wakeup alone, so pressing it twice detaches and the loop keeps running; press Esc to stop itFixed an async SessionStart hook's unchanged context being added to the conversation again on every resume
Fixed variables a SessionStart hook writes to `CLAUDE_ENV_FILE` not reaching the Bash tool after an in-app `/resume` or `/branch`
Fixed a queued message being lost when pulled back into the prompt with `↑` or `Esc` just after `←`: prompt history now keeps it
Fixed a stopped scheduled task coming back on resume after a rewind, and a scheduled task being lost on resume after Esc or a rewind following a compaction
Fixed the prompt sent after a `/rewind` being lost, and the removed turns coming back, when the session was moved to the background or resumed after being killedAdded Program Status Protocol (OSC 7501) support: terminals that implement it can show whether Claude Code is working, waiting on you, or done
Added quoted text to the `/copy` picker, so a drafted message copies without its `>` markers
Added a line on stderr, when it is a terminal, that says what a `claude -p` run is waiting for when it stays open after its last turn
Fixed `claude -p` text output dropping earlier responses when background work started another turn; each turn's response now prints when the turn ends
Fixed commands and hooks in background sessions inheriting `FORCE_COLOR=3`, which put color escape codes into the output Claude reads
Fixed bold text being drawn faint in piped output when it directly follows dim textFixed `/tui` exiting without a message, or with a raw system error, when Claude Code could not be started again
Fixed a link address shown as text being hidden by a colour or conceal style a reply left on, including where it wraps in a table or a question preview
Fixed raw terminal hyperlink bytes in a reply or a teammate's message being drawn as a clickable link with a hidden address
Fixed vim mode's `~` moving the cursor past the last character of a line, so `x` after it did nothing, and `3~` running on into the next line
Fixed the cursor landing one character too far after pasting accented text copied from macOS (such as a file name) into the middle of the prompt
Fixed `/model` saying a Max effort pick was saved as your default for new sessions; Max applies to the current session onlyFixed the terminal freezing, with ctrl+c ignored, when a response ran to tens of thousands of lines
Fixed replies whose quotes nest deeper every few lines freezing the terminal for seconds and using gigabytes of memory
Fixed text with tabs or bidirectional control characters losing its end at the edge of the screen or drawing over nearby rows
Fixed short command output with Windows line endings (CRLF) being drawn on a single line
Fixed the terminal freezing for seconds or minutes while syntax highlighting code with very long lines, long runs of blank lines, or unclosed strings and heredocs
Changed tab stops to count from where a text starts instead of from the screen's left edge: an answer indented by 2 has its first stop 8 cells in, where it was 6Fixed PushNotification saying a push was not sent in `claude remote-control` sessions right after a phone message, without the `/config` toggle, or when started from inside Claude Code
Fixed files sent from remote sessions being refused as "could not be vouched for" when more than 256 were sent at once
Fixed macOS background sessions that restart together after an update sometimes running outside the stable app wrapper, so folder-access prompts could reappear
macOS: Fixed a file count at session start that could reach other apps' data and raise the "access data from other apps" prompt when a session starts in the home folder or above it
Improved the status shown in Remote Control, claude.ai and the desktop app for a session waiting on a permission prompt to name an MCP tool by its server and readable name instead of its `mcp__server__tool` identifierFixed Claude in Chrome not applying a site deny rule written with port 80 (`host:80`) to plain `http://` pages on that host
Fixed a rare failure to load managed settings at startup right after a login refresh
Fixed the `/advisor` dialog showing a checkmark on a saved advisor model that is no longer available; it now opens on "No advisor"
Fixed temporary files left under `~/.claude/seed-admin` by an interrupted `/ultrareview` upload never being removed by the retention cleanup
Fixed Desktop and SDK sessions refusing to switch back to auto mode until restart after `disableAutoMode` was removed from settings
Improved headless `rate_limit_event` usage-limit warnings to say whether the account has extra usage turned onChanged Artifact pages with no background of their own to show on white instead of off-white
Changed the Artifact tool's permission prompts with telemetry disabled to the same five questions other installs get, instead of asking before each read of someone else's artifact and each data edit
Changed scheduled and Run now routine runs in installs with telemetry disabled to publish new private artifacts, and update your own, without asking, as in other installs[VSCode] Added a chat row for files Claude sends you: click a file's name to open it in the editor; Claude's caption shows underneath, and the row stays visible in Focus view
[VSCode] Fixed "Fork conversation from here" and Rewind's conversation restore failing with "Message not found in session" after a background agent's activity or one of the panel's own status lines
[VSCode] Fixed Ctrl+Shift+Tab, Cmd+Shift+Tab and Alt+Shift+Tab in the chat input also switching the permission mode, and Ctrl+Tab accepting the suggested prompt
[VSCode] Fixed keyboard focus bouncing between two Claude views visible at once, which could send typed text to the wrong conversation[Cloud sessions] Fixed a message sent right after unarchiving a cloud session sometimes getting no reply because no container could start for it
[Cloud sessions] Fixed some older routines starting their runs without ever giving Claude their saved prompt; those routines now run their prompt again
[Cloud sessions] Improved how quickly a cloud session catches up after its connection drops: activity you missed now appears all at once instead of row by row
[Cloud sessions] Improved the setup script box in the cloud environment form: it starts taller, grows with your script, and can be resized by dragging[Claude Tag] Changed the settings confirmation cards Claude posts in a Slack channel to stay open for 30 minutes instead of 10
[Claude Tag] Changed Claude's notice that a channel shared across Enterprise Grid workspaces uses organization defaults to post only when someone mentions Claude, at most once a month
[Claude Tag] Fixed channel links and @-mentions in a forked thread's "Continued from" card showing as raw Slack codes
[Claude Tag] Fixed Slack's Working status spinning on under a thread after `@Claude !restart` had already confirmed the restart
[Claude Tag] Fixed another app's or bot's @Claude going unanswered in a busy channel where Claude had stopped reading untagged messages
[Claude Tag] Added a confirmation in Claude Tag admin settings before adding a channel manager moves that member to Custom roles, which can remove their other accessSelf-hosted runner: Changed the runner to stop passing `CCR_AUTO_MODE_ALLOW`, `CCR_AUTO_MODE_ENVIRONMENT` and `CCR_AUTO_MODE_SOFT_DENY` into a session's environment
Self-hosted runner: Fixed git fetches being cut off while the git server was still reporting progress on a large repository; `CLAUDE_RUNNER_FETCH_SERVER_PROGRESS_CAP_MS` tunes or turns off the wait
[Code Review] Fixed Code Review re-reviews saying no issues were found while earlier findings were still open; the comment now says how many remain open
[Code Review] Fixed Code Review still skipping a large CLAUDE.md's rules when the pull request edits that file