發布日期：2026-09-11
外掛開發者現在能直接評估擴充套件的表現，透過全新指令可對 Claude Code 執行評測套件，產出可重現評分報告（包含 JSON 與 HTML 報告）：
claude plugin eval --help # 查看評測指令使用說明多項外掛安全與路徑問題也已修復。工作階段（session）的外掛封存檔，修復了先前其他本機使用者也能讀取、解壓縮檔案保留全域可寫入權限，以及舊檔案在重新解壓縮後仍殘留的狀況。組織透過受管設定啟用的外掛，修復了在無頭模式（沒有視覺介面的執行模式）未正常載入的問題。Claude Desktop 需等桌面端打包這版 CLI 後，才會在下個工作階段載入。此外，headersHelper 同意提示不再顯示容易誤讀為主機名稱的網址路徑，Windows 上包含 @ 資料夾的相對路徑也不再被誤判為遮蔽網址。
當你在雲端工作階段（session）使用從 claude.ai 同步的技能時，套件名稱已改為 anthropic-skills:<name> 格式以對齊桌面端行為；若沒有其他同名項目衝突，輸入簡稱依然可以正常呼叫。若 Skill 工具呼叫到未註冊的簡稱，只要剛好唯一符合某個外掛技能，錯誤訊息現在會主動提示完整名稱。背景執行的外掛 LSP 語言伺服器（例如 rust-analyzer），過去在工作階段結束時若拒絕 shutdown 參數會持續殘留，現在即便關閉請求失敗也會補送 exit 指令確實終止行程。此外，將工件資料庫讀取並儲存到工作階段暫記板的操作，不再停下來詢問工作目錄的核准權限。
終端機現在新增了樣式切換指令：
/output-style # 列出所有可用的輸出風格
/output-style [name] # 切換至指定的輸出風格這項功能也支援遠端控制、雲端環境與各類無頭模式（沒有終端機畫面的執行模式）工作階段。等待回覆的旋轉動畫現在會隨機推薦 /focus 指令，讓你切換到僅保留提示詞、一行工作摘要與最終成果的精簡畫面。檢視檔案變更的 /diff 面板現在改為一次完整渲染呈現，不再先顯示載入狀態。此外，過去夾帶印刷用 CMYK 色彩格式的 JPEG 圖片時跳出解碼失敗的問題已經修復，現在會像一般圖片一樣正常縮放並附加到對話中。
當終端機透過 Bash 執行指令修改檔案時，現在工具的輸出結果會直接附帶檔案變更差異對照。此功能由受管設定 bashEditDiffEnabled 提供支援。安全檢查也針對管線寫入做了補強。過去透過 Bash 的 tee 指令寫入檔案時，Edit() 拒絕規則與寫入路徑檢查可能遭到略過，現在防護已經修好，且 Bash(tee:*) 允許規則不再涵蓋工作目錄以外的目的地。此外，在對話壓縮後系統通知 Claude 的 Git 狀態，現在會即時更新為當下的最新狀態，不再沿用工作階段（session）剛開始時的舊記錄。
遙測與權限設定迎來多項調整：
# 標記 vcs.* 程式庫屬性；提交事件附帶 vcs.ref.head.*
export OTEL_METRICS_INCLUDE_REPOSITORY=1
export OTEL_LOG_TOOL_DETAILS=1以驚嘆號 ! 開頭的拒絕或詢問規則，現在僅於撰寫該設定的來源生效而不擴散，單獨的 ! 符號會被忽略。路徑拒絕規則所阻擋的 Read、Edit 與 Write 呼叫，先前在 --output-format stream-json 的 permission_denials 中被漏列的問題已修復。此外，署名提示不再覆蓋 CLAUDE.md 或記憶中反對 commit 與 PR 署名的規則；但受管設定寫入的署名行仍會套用，未帶協定的 gRPC 端點在核准視窗中也會正確標記收集器名稱。
使用企業模型閘道器或第三方雲端平台的使用者，連線彈性與多工穩定性迎來改善。若閘道器的 /v1/models 端點回應較慢，現在可透過環境變數自訂等候時間（預設為 3 秒）。
export CLAUDE_CODE_GATEWAY_MODEL_DISCOVERY_TIMEOUT_MS=10000在 Bedrock、Vertex、Foundry 與閘道環境中，若帳號無法存取預設的 Opus 模型，執行 /insights 指令時改以當前工作階段（session）使用的模型分析，解決了過去直接回傳失敗的問題。當其他本機行程剛好在同一瞬間重新整理登入時，工作階段也能完整載入組織原則上限。此外，在桌面版搭配 Bedrock、Vertex 或閘道環境時，回合結束的通知訊息現在會附上情境文字說明 Claude 需要什麼。
面對大量推論並行需求，現在可透過環境變數調高 Workflow 工具每次執行的背景代理人併發上限（可設定為 1 至 256）：
export CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS=32多項背景任務與顯示問題也已改善。在遠端或無頭模式（沒有終端機畫面的執行模式）工作階段中，修復了背景代理仍在運作時卻誤報等待輸入的問題，若想恢復舊行為可設定 CLAUDE_CODE_BG_TASKS_REPORT_RUNNING=0。連續點擊 /fork 產生的收據不再卡住背景化流程，SDK 或桌面端啟動的工作階段在代理清單中不再顯示未知狀態。此外，恢復背景任務時，磁碟記錄裡的跳脫碼、換行與過長文字不再進到工作清單與任務通知，輸入框頂部邊界也不再被換行撐裂。
長時間對話的流暢度與暫存效能顯著提升。當回答因輸出上限中斷並自動恢復時，下一個回合的提示快取（把算過結果暫存起來直接用）不再部分失效。在 Claude 思考中途打斷後恢復工作階段（session），修復了重新傳送脈絡改變而損害快取重用的問題。雲端工作階段在首個請求前會短暫等待伺服器設定，避免了快取未命中。自動壓縮在缺乏完整對話歷史可總結時，不再卡死在提示過長狀態。更新對話紀錄不再重新處理整個對話來建立摺疊摘要。此外，CLAUDE_CODE_RESUME_INTERRUPTED_TURN 不再重跑超過 6 小時前因 API 錯誤失敗的回合，亦可自訂 CLAUDE_CODE_RESUME_INTERRUPTED_TURN_MAX_AGE_MS 上限。
鍵盤操作與不同終端機的相容性獲得全面校準。過去在 kitty 協定終端機中 F1、F2 與 F4 無效、st 中的 Delete 鍵失效、rxvt-unicode 中 Alt 加方向鍵被誤判為 Escape，以及 WezTerm 中 Shift 加標點打出未換檔按鍵的問題均已修復。部分終端機啟動時不再把 capability query 回覆顯示成雜字。慢速連線如 SSH 或瀏覽器終端機中，22c、顏色或版本回覆不再被打進輸入框。此外，透過 SSH 或在未知終端機中，只要軟體能回應 kitty 鍵盤查詢（如 foot 與 Alacritty 0.16 以上版本），現在都能正常使用 Shift+Enter 與 Ctrl+Shift 快速鍵。
終端機介面的畫面排版與游標顯示修復了多項瑕疵。在全螢幕模式下調整終端機視窗大小後，對話紀錄最頂部或底部的列數先前會異常變成空白，現在已經修復。當啟用終端機的原生游標時，權限規則、自動模式規則、新增目錄、工作階段（session）重新命名與反饋檢閱的文字欄位，修復了游標消失看不見打字位置的問題。此外，GNOME Terminal 與 Konsole 若屬於不支援同步輸出的版本，Claude Code 不再單憑終端機名稱就預設啟用同步輸出，避免了畫面更新異常。
習慣呼叫外部編輯器編輯長文的使用者，切換視窗後的畫面渲染更加穩定。在非全螢幕模式下按 Ctrl+G 從外部編輯器返回後，畫面先前會被重複繪製兩次，這項重繪問題已經修正。在 Konsole 終端機中，從外部編輯器返回時介面被重複繪製兩次的問題也一併排除。針對 rxvt-unicode 終端機，離開或重新進入全螢幕時終端機方塊游標在介面下方穿透顯示的問題已修好，在全螢幕模式下從外部編輯器返回時方塊游標持續殘留的問題也已解決。
遠端與無頭模式（沒有終端機畫面的執行模式）工作階段的通訊穩定度迎來重要修復。恢復遠端工作階段（session）時，外掛連線已修好。同步外掛的 MCP（讓 AI 連接外部工具的協定）伺服器先前無法連線的問題已獲解決。在恢復的無頭工作階段中，若在回合中途切換模型或重試請求，不再遺失該回合的回覆。當更新的設定僅改變伺服器網址的查詢參數順序時，MCP 伺服器不再不必要地重新連線。關閉遙測的第一方工作階段中，對話中途連線完成的 alwaysLoad 伺服器在下一回合即可直接使用。在 Windows 上，送入背景執行的 PowerShell 工具指令在 Claude Code 結束時不再被迫中斷。
當你使用進階指令執行長效任務時，流程控制變得更加可靠。過去 /goal 遇到網路中斷、API 錯誤或達到 token（AI 處理與計費的文字單位）上限時會靜默卡住。現在它會自動進行退避重試，或暫停並清楚說明原因，包含等候用量重置。/ultrareview --post 現在審查結果出爐時會直接發布 PR 評論並印出連結，不必另開第二個雲端工作階段（session）。在輸入列中，日文、中文、泰文等詞間無空格的文字提示建議遺失問題已修復。中日韓文字的提示建議現在會保留單詞與混合語系建議，並比照英文剔除後設評估字眼。此外，/btw 旁支提問先前可能捏造工具呼叫與輸出，現在已明確要求不撰寫虛構呼叫，若有出現也會標記為未執行。
VSCode 擴充套件為多代理人協同帶來視覺化新工具。點擊頁尾的代理人標籤即可開啟地圖，以專屬卡片檢視當前工作階段（session）的所有子代理人，並支援停止個別代理人與閱讀唯讀紀錄。在 Focus 專注檢視中，工具呼叫群組下方新增了執行中子代理人的即時進度列。排程任務等純文字提示觸發的新回合，先前在專注檢視被誤歸入上一回合的問題已經修復。計畫預覽方面，較晚載入完成時隱藏評論框或顯示舊計畫的狀況已獲排除，分頁關閉後也不再接受無處送出的留言。此外，當 Claude 撰寫給非使用者本人的文件時，會針對該受眾撰寫並於回覆頂部指名。
VSCode 指令選單新增了掛鉤（特定事件發生時被觸發的小程式）與權限規則兩組全新對話框。使用者能在視覺面板中瀏覽使用者、專案與本機設定中的規則。掛鉤可新增、編輯或移除；權限規則可新增或移除。掛鉤的受管、外掛與工作階段（session）規則唯讀；權限規則的 startup-option、session-only 與受管規則唯讀。輸入框的目前檔案標籤改由叉叉圖示直接移除，取代先前的隱藏開關。右鍵選單與編輯器頂部無法對所選分頁生效的項目已全數清理移除。此外，斜線指令選單、標註選單、樣式選擇器、傳送停止按鈕、問題卡片與入門清單的螢幕閱讀器與鍵盤無障礙支援均全面升級，新聊天視窗啟動時按快捷鍵插入標註也不再延遲滯後。
工作階段清單修復了多項設定與視窗重載後的狀態不同步。在設定檔配置 CLAUDE_CONFIG_DIR 時，清單不再錯誤殘留預設目錄的工作階段（session）。開啟遠端控制時在擴充套件重新命名，新名稱現在會正確同步至網頁端，長回合中重新命名在視窗重載後也不會被自動產生的名稱覆蓋。在終端機將遠端控制重置為預設後，全域開關按鈕不再停留在舊狀態。視窗重載後還原的分頁不再被清單視為未開啟而重複打開第二個分頁。此外，側邊欄帳號標頭出現時工作階段清單瞬間下跳的跳動問題也已解決。
帳號管理與用量資訊的介面交互更加貼心。在切換帳號畫面新增了「取消」按鈕，點擊即可直接返回原帳號的進行中工作階段（session）。當你重新登入相同帳號時，分頁不會再忘記你先前已經關閉的用量上限警告。側邊欄用量計量表在帳號失去特定模型的每週配額限制後，不會再錯誤保留過期的每週限制列。頁尾的提示快取（把算過結果暫存起來直接用）時鐘在面板較窄時不再隱藏分鐘數字。此外，在最後一次回覆後壓縮並關閉的工作階段，重新打開時的重開通知與快取時鐘現在會正確標示為冷啟動。
在網頁版雲端工作階段（session）中送出訊息後，只要 Claude 尚未讀取，你現在隨時能反悔抽回。你可以直接從排隊佇列中移除它，或是按下 Esc 或向上方向鍵 ↑，未送出的文字就會完整返回輸入框讓你重新編輯。自動化例行排程的執行邏輯也經過修復。單次排程的定時例行任務，修復了在伺服器遭遇短暫錯誤後偶爾重複執行第二次的問題。此外，使用子代理人的例行排程任務過去偶爾會被過早判定為已完成，進而導致真實失敗時跳過重試或意外啟動重複執行，這項狀態判定漏洞現在已經完全修好。
網頁版管理介面與存取指引迎來友善改動。雲端環境管理頁面現在會直接列出所有環境，不再把表格限制在五列並隱藏於可能無法點擊的展開按鈕後面。針對免費方案使用者，前往 claude.ai/code 時改為開啟方案頁面並提供升級路徑，取代先前無路可走的「已被組織管理員停用」畫面。在限制可用模型的組織中，於雲端工作階段（session）執行 /model default 導致後續每則訊息皆發送失敗的問題已經排除。此外，當 Claude 在程式碼庫的子資料夾中作業時，對話紀錄中的檔案連結在 GitHub 上開出 404 找不到網頁的錯誤也已修復。
管理後台的防護與效能獲得顯著強化。在管理設定中針對 GitHub 安裝點擊全部連線或中斷連線時，現在會先跳出確認對話框以防止組織層級的誤觸。當使用者嘗試切換至組織尚未啟用的模型時，Claude 不再默默接受並以備援（fallback）模型私下回答，而是會直接拒絕並提示可由管理員啟用。Slack 存取設定中的外掛橫列不再只顯示未標籤的原始 ID，改為顯示外掛名稱並連結至管理套件。此外，從 Slack 啟動的工作階段（session），共享工作階段橫幅與分享對話框都會指名 Slack 頻道受眾，管理頁面與頻道選擇器的載入速度也獲得改善。
在 Slack 頻道中與 Claude 協同作業時，通知與排版呈現更加穩定。當回合失敗遇到 Slack 短暫限流時，失敗通知不再被遺棄而導致討論串陷入沉默，現在會自動重試發送。當同一則訊息中包含附件檔案時，表格不再被降級為未排版的直槓原始文字，改為正常渲染表格並將附加檔案以簡明說明接續呈現。在未啟動 Claude 的頻道頂層輸入 @Claude !restart 時，不再開啟無關對話，而是以私人訊息回覆目前沒有可重啟的工作。此外，頻道中的定時例行任務現在能直接回覆在既有的討論串內，避免在頂層不斷發布新訊息洗版；即時進度檢查清單的時間戳記也改為顯示讀者本機時間與更新距今時間，取代原本固定的 UTC 時間。
Plugin developers can now directly evaluate extension performance, using a new command to run an eval suite against Claude Code and generate scored, reproducible results (including JSON and HTML reports):
claude plugin eval --help # View plugin evaluation command helpSeveral plugin security and path issues have also been fixed. For plugin archives in a session, fixes resolve issues where other local users could read extracted files, extracted files retained world-writable permissions from archives, and stale files survived re-extraction. For plugins enabled by organizations through managed settings, an issue where they failed to load in headless mode has been resolved; Claude Desktop will load them from the next session once Desktop bundles this CLI version. Additionally, headersHelper consent prompts no longer display URL paths that could be misread as hostnames, and relative paths on Windows containing folders starting with @ are no longer misidentified as redacted URLs.
When using skills synced from claude.ai in cloud sessions, package names are now formatted as anthropic-skills:<name> to match Claude Desktop behavior; bare names can still be used when there are no conflicting items. If the Skill tool is called with an unregistered bare name that uniquely matches a plugin skill, the error message now actively suggests the full name. For plugin LSP language servers running in the background (such as rust-analyzer) that previously remained running at session end if they rejected shutdown parameters, an exit command is now sent even if the shutdown request fails to ensure processes are terminated. Additionally, reading the artifact database and saving it into the session scratchpad no longer stops to ask for working-directory approval.
The terminal now adds style-switching commands:
/output-style # List all available output styles
/output-style [name] # Switch to the specified output styleThis feature also supports Remote Control, cloud environments, and various headless sessions. The spinner animation while waiting for responses now suggests /focus to switch to a streamlined view retaining only your prompt, a one-line work summary, and the response. The /diff panel for reviewing file changes now opens fully rendered in one step instead of showing a loading state first. Additionally, an issue where attaching CMYK JPEG images caused decoding failures has been fixed; they are now converted, resized, and attached like regular JPEGs.
When commands executed through Bash modify files in the terminal, tool results now directly include a diff of the changed files. This feature is supported by the managed setting bashEditDiffEnabled. Security checks have also been reinforced for pipeline writes. Previously, when writing files via Bash's tee command, Edit() deny rules and write-path checks could be bypassed; this protection is now fixed, and Bash(tee:*) allow rules no longer cover destinations outside the working directories. Additionally, the Git status told to Claude after compaction is now refreshed to the current status instead of using outdated records from the start of the session.
Telemetry and permission settings receive several adjustments:
# Tag vcs.* repository attributes; commit events include vcs.ref.head.*
export OTEL_METRICS_INCLUDE_REPOSITORY=1
export OTEL_LOG_TOOL_DETAILS=1Deny or ask permission rules starting with an exclamation mark ! now apply only within the settings source that wrote them without spreading, and a bare ! negation is ignored. An issue where Read, Edit, and Write calls blocked by path-scoped deny rules were omitted from permission_denials in --output-format stream-json results has been fixed. Additionally, attribution reminders no longer override CLAUDE.md or memory rules against commit and PR attribution; lines set by managed settings still apply, and gRPC endpoints without a scheme correctly name the collector in the approval dialog.
Connection flexibility and multitasking stability have improved for users on enterprise model gateways or third-party cloud platforms. If a gateway's /v1/models endpoint responds slowly, the wait timeout can now be customized via an environment variable (defaulting to 3 seconds).
export CLAUDE_CODE_GATEWAY_MODEL_DISCOVERY_TIMEOUT_MS=10000Across Bedrock, Vertex, Foundry, and gateway environments where an account cannot access the default Opus model, running /insights now uses the current session model instead, resolving previous direct failures. When another local process happens to refresh login at the exact same moment, the session still fully loads organization policy limits. Additionally, for desktop sessions using Bedrock, Vertex, or a gateway, turn-end notifications now include contextual text explaining what Claude needs.
For high-volume concurrent inference demands, an environment variable can now raise the Workflow tool's per-run concurrent background agent limit (configurable from 1 to 256):
export CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS=32Several background task and display issues have also been resolved. In remote and headless sessions, an issue where background agents running still mistakenly reported waiting for input has been fixed (set CLAUDE_CODE_BG_TASKS_REPORT_RUNNING=0 to restore the old behavior). Rapidly clicking /fork receipts no longer blocks immediate backgrounding, and sessions started via the SDK or desktop app no longer show unknown statuses in agent lists. Additionally, when resuming background tasks, escape codes, line breaks, and oversized text from on-disk records no longer reach task lists or notifications, and prompt box top borders no longer split into extra lines from line breaks.
Responsiveness and caching performance in long conversations have improved significantly. When a response is cut off at the output limit and automatically resumed, prompt caching on the next turn no longer partially invalidates. Resuming a session after interrupting Claude mid-thought fixes an issue where changed re-sent context hurt cache reuse. Cloud sessions briefly wait for server configuration before the first request, preventing cache misses. Auto-compaction no longer gets permanently stuck on prompt-too-long states when lacking complete earlier exchanges to summarize. Transcript updates no longer re-process the entire conversation to build collapsed tool-use summaries. Additionally, CLAUDE_CODE_RESUME_INTERRUPTED_TURN no longer re-runs turns that failed with API errors over 6 hours ago, and CLAUDE_CODE_RESUME_INTERRUPTED_TURN_MAX_AGE_MS can customize this limit.
Keyboard interactions and compatibility across terminals have been thoroughly calibrated. Issues where F1, F2, and F4 failed in kitty-protocol terminals, Delete failed in st, Alt+arrows acted as Escape in rxvt-unicode, and Shift+punctuation typed unshifted keys in WezTerm have all been resolved. Replies to capability queries no longer appear as stray text at startup in certain terminals. Over slow connections like SSH or browser terminals, 22c, color, or version replies are no longer typed into the prompt. Additionally, over SSH or in unrecognized terminals that answer kitty keyboard queries (such as foot and Alacritty 0.16+), Shift+Enter and Ctrl+Shift shortcuts are now fully supported.
Terminal interface layout and cursor rendering have received multiple fixes. After resizing terminal windows in fullscreen mode, an issue where rows at the top or bottom of transcripts went blank has been fixed. When the terminal's native cursor is enabled, text fields for permission rules, auto-mode rules, add-directory, session renames, and feedback reviews fix missing cursors that obscured typing positions. Additionally, for GNOME Terminal and Konsole versions that do not support synchronized output, Claude Code no longer assumes synchronized output based merely on terminal names, preventing display update anomalies.
Window switching and rendering are now more stable for users accustomed to launching external editors for long-form edits. Returning from an external editor outside fullscreen mode via Ctrl+G previously caused the interface to be drawn twice; this redraw issue has been corrected. An issue where returning from an external editor in Konsole drew the interface twice has also been resolved. For rxvt-unicode terminals, issues where terminal block cursors showed through beneath the interface after leaving or re-entering fullscreen, or stayed visible after returning from an external editor in fullscreen mode, are now resolved.
Communication stability in remote and headless sessions receives important fixes. Resuming remote sessions now connects synced plugin MCP servers properly. In resumed headless sessions, switching models or retrying requests mid-turn no longer loses that turn's replies. When updated configurations only change query parameter ordering in server URLs, MCP servers no longer reconnect unnecessarily. In first-party sessions with telemetry disabled, alwaysLoad servers that finish connecting mid-conversation are immediately usable on the next turn. On Windows, PowerShell tool commands sent to the background no longer terminate when Claude Code exits.
Process control is now more dependable when executing long-running tasks with advanced commands. Previously, /goal silently stalled after network drops, API errors, or reaching token limits. It now automatically retries with backoff, or pauses and clearly states why, including waiting until usage limits reset. /ultrareview --post now posts PR comments directly when findings arrive and prints comment links, without needing a second cloud session. In the input prompt, prompt suggestions dropped for text without spaces between words (such as Japanese, Chinese, and Thai) have been restored. Prompt suggestions for CJK text now retain single-word and mixed-script suggestions, dropping meta or evaluative text just like English. Additionally, /btw side questions previously could fabricate tool calls and output; Claude is now instructed not to write them, and any that appear are flagged as unexecuted.
The VSCode extension brings visual tools for multi-agent collaboration. Clicking the agent pill in the footer opens an agent map with dedicated cards for all subagents in the current session, supporting stopping individual agents and reading read-only transcripts. In Focus view, live progress rows for running subagents have been added beneath tool-call groups. An issue where new turns triggered by plain-text prompts (such as scheduled tasks) were mistakenly grouped into the previous turn in Focus view has been fixed. For plan previews, issues where late-loading previews hid comment boxes or displayed older plans have been eliminated, and closed tabs no longer accept orphaned comments. Additionally, when Claude writes documents or messages for someone other than the user, it addresses that audience and names it at the top of the reply.
The VSCode command menu adds two new dialogs: Hooks and Permission rules. Users can browse rules across user, project, and local settings in visual panels. Hooks can be added, edited, or removed; permission rules can be added or removed. Managed, plugin, and session hooks remain read-only; startup-option, session-only, and managed permission rules remain read-only. The current-file chip in the input box is now removed directly via an X icon, replacing the previous Hide toggle. Context menu and editor title bar items that could not act on the selected tab have been cleaned up and removed. Additionally, screen reader and keyboard accessibility have been upgraded across slash-command menus, @-mention menus, output-style pickers, Send/Stop buttons, permission and question cards, and onboarding checklists, while inserting @-mentions via shortcuts when starting new chat views no longer experiences lag.
The session list resolves multiple state synchronization issues after configuration changes and window reloads. When CLAUDE_CONFIG_DIR is set in settings, the list no longer incorrectly retains sessions from the default folder. Renaming a session in the extension while Remote Control is active now properly syncs the new name to the web interface, and renames during long turns are no longer overwritten by generated names after window reloads. Resetting Remote Control to default from a terminal no longer leaves the global toggle stuck in its previous state. Restored tabs after window reloads are no longer treated as closed by the list to open duplicate tabs. Additionally, an issue where the session list jumped downward when the account header appeared in the sidebar has been resolved.
Interface interactions for account management and usage tracking are now smoother. A Cancel button on the Switch account screen returns directly to your active session under the current account. Signing back in to the same account no longer causes tabs to forget previously dismissed usage-limit warnings. The sidebar usage meter no longer retains stale weekly limit rows after an account loses per-model weekly limits. The footer's prompt cache clock no longer hides minute digits when panels are narrow. Additionally, sessions compacted after their last reply and then closed now correctly display cold states in reopen notices and cache clocks upon reopening.
After sending messages in web cloud sessions, queued messages can now be taken back at any time before Claude reads them. You can remove them directly from the queue, or press Esc or Up arrow ↑ to return unsent text to the input box for editing. Execution logic for automated routines has also been fixed. One-off scheduled routines resolve issues where runs occasionally executed a second time after transient server errors. Additionally, routine runs using subagents that were previously treated as finished too early—which could skip retries after real failures or trigger duplicate runs—now have their status determination fully fixed.
The web management interface and access guidance receive welcoming improvements. The Cloud environments admin page now lists every environment directly, no longer capping tables at five rows behind potentially unreachable Show more controls. For Free-plan users, navigating to claude.ai/code now opens the plans page with upgrade paths instead of a dead-end "Disabled by org admin" page. In organizations restricting available models, an issue where running /model default in cloud sessions caused all subsequent messages to fail has been eliminated. Additionally, when Claude works from repository subfolders, file links in transcripts opening GitHub 404 errors have been fixed.
Admin console protection and performance have been significantly reinforced. Clicking Connect all or Disconnect on GitHub installations in admin settings now prompts a confirmation dialog to guard against accidental organization-wide changes. When attempting to switch to a model not enabled by the organization, Claude no longer quietly accepts and answers with a fallback model, but instead declines and indicates an admin can enable it. Plugin rows in Slack access settings no longer show only unlabeled raw IDs, but now display plugin names and link to the managing bundle. Additionally, for sessions started from Slack, shared session banners and Share dialogs name the Slack channel audience, while load times for admin settings and the channel picker have improved.
When collaborating with Claude in Slack channels, notifications and formatting render more reliably. When a failed turn hits brief Slack rate limits, failure notices are no longer dropped to leave threads silent, but are now automatically retried. When tables are posted alongside attached files in the same message, tables are no longer downgraded to raw pipe text, instead posting as normally rendered tables followed by plain file captions. Entering @Claude !restart at the top level of a channel where Claude is not active no longer starts unrelated conversations, replying privately instead that there is nothing to restart. Additionally, scheduled routines in channels can now reply directly in existing threads rather than spamming new top-level messages; timestamps on live progress checklists now display each reader's local time and relative update elapsed time instead of fixed UTC times.
Added `claude plugin eval`: run a plugin's eval suite against Claude Code and get scored, reproducible results (JSON + HTML report); see `claude plugin eval --help`
Fixed plugin `headersHelper` consent prompts showing a URL path that could be misread as a different host
Fixed plugin errors showing `[redacted URL]` in place of a relative Windows path with a folder name that starts with `@`
Fixed organization plugins enabled through managed settings not loading in headless sessions and on Claude Desktop (once Desktop bundles this CLI version); they load from the next session
Fixed plugin archives extracted for a session being readable by other local users, extracted files keeping world-writable bits from the archive, and stale files surviving re-extractionFixed plugin LSP servers that reject `shutdown` params (e.g. rust-analyzer) being left running at session end; `exit` is now sent even if `shutdown` fails
Improved the Skill tool's "Unknown skill" error to name the plugin skill's full name when a bare name matches exactly one plugin skill
Changed artifact database reads that save into the session scratchpad so they no longer stop for working-folder approval
Changed skills synced from claude.ai in cloud sessions to be named `anthropic-skills:<name>`, matching Claude Desktop; the bare name still works when nothing else uses itAdded `/output-style [name]` to list and switch output styles, including over Remote Control and in cloud and other headless sessions
Added a spinner tip suggesting `/focus` for a view with just your prompt, a one-line work summary, and the response
Fixed CMYK JPEG images failing to attach with "cannot decode"; they are now converted and resized like other JPEGs
Improved the `/diff` panel to open fully rendered in one step instead of showing a loading state firstAdded a diff of the files a Bash command changed to the Bash tool result when the Bash tool handles file edits (setting `bashEditDiffEnabled`)
Fixed the git status Claude is told after a compaction: it is now the current status, not the one from the start of the session
Fixed `Edit()` deny rules and the write-path check not applying to the file a Bash `tee` command writes; a `Bash(tee:*)` allow rule no longer covers destinations outside the working directoriesAdded `OTEL_METRICS_INCLUDE_REPOSITORY` to tag OpenTelemetry metrics and events with `vcs.*` repository attributes; commit events get `vcs.ref.head.*` with `OTEL_LOG_TOOL_DETAILS`
Fixed a deny or ask permission rule starting with `!` applying beyond the settings source that wrote it; such a rule now applies only within its own source, and a bare `!` negation is ignored
Fixed the managed settings approval dialog not naming the collector for a gRPC telemetry endpoint set without a scheme
Fixed the attribution reminder overriding a CLAUDE.md or memory rule against commit and pull request attribution; lines set by managed settings still apply
Fixed `permission_denials` in `--output-format stream-json` results omitting Read, Edit and Write calls blocked by a path-scoped deny ruleAdded `CLAUDE_CODE_GATEWAY_MODEL_DISCOVERY_TIMEOUT_MS` to extend the LLM gateway `/v1/models` discovery timeout (default 3s)
Fixed `/insights` failing on Bedrock, Vertex, Foundry, and gateway deployments whose account can't reach the default Opus model by using the session model there instead
Fixed organization policy limits not loading for the session when another Claude Code process refreshed the login at the same moment
Fixed Claude Desktop sessions using Bedrock, Vertex, or a gateway not getting the contextual "what Claude needs" turn-end notification textAdded `CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS` (1–256) to raise the Workflow tool's per-run concurrent agent limit for inference-bound fan-outs
Fixed remote and headless sessions reporting "waiting for your input" while background agents were still running (set `CLAUDE_CODE_BG_TASKS_REPORT_RUNNING=0` to restore the old behavior)
Fixed terminal escape codes, line breaks and oversized text from a background task's on-disk record reaching the task list and task notifications when work is resumed
Fixed repeated clicks on a `/fork` receipt, each under a second apart, never backgrounding the session right away while it waited for the current tool to finish
Fixed sessions run through the SDK or the desktop app showing an unknown status in other sessions' agent list
Fixed the prompt box's top border splitting into extra lines when viewing a background agent whose name or description has line breaks or is wider than the terminalFixed the prompt cache being partially invalidated on the turn after a response was cut off at the output-token limit and automatically resumed
Fixed a case where resuming a session after interrupting Claude mid-thought could change how earlier context was re-sent, hurting prompt-cache reuse
Fixed sessions getting permanently stuck on "Prompt is too long" when auto-compaction had no complete earlier exchange to summarize (mostly Agent SDK sessions with very large prompts)
Fixed prompt cache misses in cloud sessions by waiting briefly for server configuration before the first request
Fixed `CLAUDE_CODE_RESUME_INTERRUPTED_TURN` re-running a turn that had failed with an API error over 6 hours earlier, or longer ago than `CLAUDE_CODE_RESUME_INTERRUPTED_TURN_MAX_AGE_MS` when set
Improved responsiveness in long sessions: transcript updates no longer re-process the whole conversation to build the collapsed tool-use summariesFixed F1/F2/F4 not working in kitty-protocol terminals and Delete in st, Alt+arrows acting as Escape in rxvt-unicode, and Shift+punctuation typing the unshifted key in WezTerm (regression in 2.1.247)
Fixed the terminal's replies to capability queries (`^[[?1;2c`) appearing as stray text at startup in some terminals
Fixed stray characters like `22c`, or a terminal's color or version reply, being typed into the prompt at startup over slow connections (ssh, browser terminals)
Improved keyboard support over SSH and in unrecognized terminals: terminals that answer the kitty keyboard query (such as foot and Alacritty 0.16+) now get Shift+Enter and Ctrl+Shift shortcutsFixed rows at the top or bottom of the transcript going blank in fullscreen after resizing the terminal
Fixed missing cursor in the permission-rule, auto-mode-rule, add-directory, session-rename and feedback-review text fields when the terminal's native cursor is enabled
Fixed synchronized output being assumed from the terminal's name in GNOME Terminal and Konsole versions that do not support itFixed the terminal's block cursor showing under the interface in rxvt-unicode after leaving or re-entering fullscreen
Fixed the cursor block staying visible after returning from an external editor in fullscreen mode on rxvt-unicode
Fixed the interface being drawn twice after returning from an external editor (Ctrl+G) outside fullscreen mode
Fixed the interface being drawn twice in Konsole after returning from an external editorFixed synced plugin MCP servers not connecting when a remote session resumes
Fixed resumed headless sessions losing a turn's replies when the model was switched or a request was retried mid-turn
Fixed MCP servers reconnecting when an updated config only changed the order of the server URL's query parameters
Windows: Fixed PowerShell tool commands sent to the background stopping when Claude Code exits
Improved first-party sessions with telemetry disabled: an `alwaysLoad` MCP server that finishes connecting mid-conversation is usable on the next turn without a tool-search round tripFixed prompt suggestions being dropped for text in Japanese, Chinese, Thai and other languages written without spaces between words
Fixed `/goal` runs silently stalling after API errors, network drops, or token limits: the goal now retries with backoff, or pauses and says why, including until a usage limit resets
Fixed `/btw` answers that contained made-up tool calls and output: the side question is now told not to write them, and any that appear are flagged as not executed
Improved prompt suggestion filtering for Japanese, Chinese and Korean text: mixed-script and single-word suggestions are kept, and meta or evaluative text is dropped as it is for English
Changed `/ultrareview --post` to post the PR comment directly when the findings arrive and print the comment link, instead of starting a second cloud session to post it[VSCode] Added an agent map: an "N agents" footer pill opens a map of the session's sub-agents with per-agent cards, Stop agent, and read-only transcripts
[VSCode] Added live progress rows for running subagents under the tool-call groups in Focus view
[VSCode] Fixed Focus view showing a turn started by a delivered plain-text prompt, such as a scheduled task's, as part of the previous turn
[VSCode] Fixed a plan preview that finished loading late sometimes hiding its comment box or showing an older plan
[VSCode] Fixed a plan preview accepting comments that went nowhere after its Claude tab closed
[VSCode] Improved documents and messages written for someone other than the user: Claude now writes them for that audience and names it at the top of its reply[VSCode] Added a Hooks dialog to the command menu for viewing hooks and adding, editing, or removing them in user, project, and local settings; managed, plugin, and session hooks stay read-only
[VSCode] Added a Permission rules dialog that lists permission rules and adds or removes them in user, project, and local settings; startup-option, session-only, and managed rules stay read-only
[VSCode] Fixed a rare case where an @-mention sent with the keyboard shortcut while a new chat view was still starting could be inserted into the input long after the keystroke
[VSCode] Improved screen reader and keyboard accessibility in the slash-command menu, @-mention menu, output-style picker, Send/Stop button, permission and question cards, and onboarding checklist
[VSCode] Changed the current-file chip in the message box: an X now removes it, replacing the Hide toggle
[VSCode] Removed the Claude Code items from a session tab's right-click menu and the editor title bar's "..." menu; they could not act on the tab the menu was opened on[VSCode] Fixed the session list keeping sessions from the default folder when `CLAUDE_CONFIG_DIR` is set in a settings file or the `environmentVariables` setting
[VSCode] Fixed a session renamed in the extension while Remote Control is on keeping its old name on claude.ai/code
[VSCode] Fixed the "Enable Remote Control for all sessions" toggle keeping its last position after the setting was reset to default from a terminal
[VSCode] Fixed restored Claude tabs not counting as open in the session list after a window reload until clicked, and their row opening a second tab
[VSCode] Fixed a session rename being replaced by the generated name after a window reload when the session was renamed during a long turn
[VSCode] Fixed the session list jumping down when the Account & usage header appeared a moment after opening the Claude side bar[VSCode] Added a Cancel button to the Switch account screen that returns to your session as the current account
[VSCode] Fixed the footer's prompt cache clock hiding its minutes when the panel is narrow
[VSCode] Fixed the prompt cache clock and reopen notice for a session compacted after its last reply and then closed, which now reads as cold when reopened
[VSCode] Fixed Switch account making a tab forget its dismissed usage-limit warnings when you sign back in as the same account
[VSCode] Fixed the sidebar usage meter keeping a stale per-model weekly limit row after the account loses that limit[Claude Code on the web] Added taking back a queued message in a cloud session before Claude reads it: remove it from the queue, or press Esc or Up, and the text returns to the message box
[Claude Code on the web] Fixed one-off scheduled routines occasionally running a second time after a transient server error
[Claude Code on the web] Fixed routine runs that use subagents sometimes being treated as finished too early, which could skip the retry after a real failure or start a duplicate run[Claude Code on the web] Fixed `/model default` in a cloud session leaving every later message failing in organizations that restrict which models Claude Code can use
[Claude Code on the web] Fixed file links in cloud session transcripts opening a GitHub 404 when Claude was working from a subfolder of the repository
[Claude Code on the web] Changed the Cloud environments admin page to list every environment instead of capping each table at five rows behind a Show more control that could be unreachable
[Claude Code on the web] Changed claude.ai/code for Free-plan users to open the plans page with a path to upgrade, instead of a "Disabled by org admin" page with no way forward[Claude Tag] Added a confirmation dialog before Connect all or Disconnect on a GitHub installation in admin settings, to guard against accidental organization-wide changes
[Claude Tag] Fixed Claude accepting a switch to a model your organization hasn't enabled and then quietly answering with a fallback model; it now declines and says an admin can enable it
[Claude Tag] Fixed plugin rows in Slack access settings showing an unlabeled raw ID with no way to turn the plugin off; they now show its name and link to the bundle that manages it
[Claude Tag] Fixed the shared-session banner and Share dialog on sessions started from Slack claiming the whole organization could open the link; they now name the Slack channel's audience
[Claude Tag] Improved load time of the admin settings page and its Slack channel picker, most noticeably for organizations with many channels or several connected workspaces[Claude Tag] Fixed threads occasionally going silent after a failed turn because the failure notice was dropped when Slack briefly rate-limited it; the notice is now retried
[Claude Tag] Fixed a table posting as raw pipe text when Claude attached files to the same message; the table now posts as a normal reply and the files follow with a plain caption
[Claude Tag] Fixed `@Claude !restart` at the top level of a channel where Claude isn't active starting an unrelated conversation; it now privately says there is nothing to restart
[Claude Tag] Improved scheduled routines in Slack channels: a routine run can now reply in an existing thread instead of always posting a new top-level channel message
[Claude Tag] Improved the timestamp on Claude's live progress checklists to show each reader's local time and how long ago it was updated, instead of a fixed UTC time