發布日期：2026-10-16
當你開啟自動模式（auto mode，讓 Claude 不逐一詢問就自主執行工具的模式）處理繁重複雜的任務時，現在能自訂本機安全分類器使用的模型。透過環境變數 CLAUDE_CODE_AUTO_MODE_MODEL，可指定用戶端分類器的模型，伺服器端模型不變。對話太長而超出伺服器端安全檢查範圍時，現在會先壓縮，不再一律拒絕工具呼叫。
多數自動模式分類器阻擋事件漏記 classifier_category 已修復；OTEL_LOGRECORD_ATTRIBUTE_COUNT_LIMIT 較低時，新增屬性擠掉 tool_use_id、tool_source、tool_parameters 的問題也已修復。
平時讓對話視窗閒置太久時，你可能會發現 Claude 自動整理了先前的發言紀錄。現在 /config 設定選單新增了「Idle compaction」選項，方便你自由開啟或關閉閒置自動壓縮；整理結束後的提示也會主動說明關閉方式。以前若在對話正在整理壓縮時中途停止回答，畫面會跳出提示詞過長（Prompt is too long）的錯誤；現在這個問題已經修好。
若你的閘道使用了帶有前綴的 Opus 4 或 4.1 名稱（例如 anthropic/claude-opus-4-1），執行 /autocompact 與 /effort 也已修正為儲存至正確模型。
當你開啟新的對話時，Claude in Chrome 的操作指引現在能從共用提示詞快取讀取。Claude 調整了這些指引在系統提示詞中的擺放位置，讓新對話能共用既有的快取（cache，把算過或查過的結果暫存起來直接用）。
在工作階段中途動態加入的工具，即使經過 ToolSearch 再次載入或對話壓縮，也不會再發生快取未命中的問題。
若你平時習慣在終端機使用介面設計指令，現在執行相關操作時會看見更準確的時程與導引提示。當你輸入 /design-sync 或 /design-login 時，畫面會新增提示，說明獨立網頁版 Claude Design 預計於 2026 年 12 月 14 日關閉服務。
在雲端工作階段執行 /design consent 與 /design revoke 失敗時，提示訊息也已改為指引至 claude.ai/design/settings，不再建議無法使用的登入指令。
當你為專案安裝了大量外掛程式時，現在啟動 Claude Code 會感覺到等待時間明顯縮短。系統減少了讀取外掛指令與代理檔案時的檔案系統呼叫次數，這項改進在 Windows 環境尤其有感。以前執行 claude plugin 等子指令時，程式偶爾會在設定檔寫入完成前提早退出，導致變更遺失或殘留 .tmp 暫存檔，現在已確保寫入完成才結束。當 Git 預設採用 reftable 引用格式時，市集不再每次重整都重複下載整個倉庫，新裝外掛也不會顯示版本未知。PowerShell 工具現在能順利呼叫外掛 bin/ 目錄下的指令，搜尋為空時串接格式化指令不再噴錯，查詢內建外掛失敗時也正確指引至 /plugin。
當你在桌面應用程式中使用外部工具，若需要重新授權或連線，現在會得到完全符合當前介面的操作指引。以前桌面版 Claude 在需要重新連接 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）伺服器時，會誤導你去翻找網頁版設定或打 /mcp 指令，即使該環境根本沒有這些選項；現在這項引導已修正。
同時，當外部伺服器協商使用 2026-07-28 協定版本，但未提供 resultType、ttlMs 或 cacheScope 等非必要欄位時，工具清單不再一片空白，呼叫工具也不會全數失敗。
在桌面應用程式、VS Code 或是透過程式呼叫 Claude 時，分頭處理多項任務的規則現在與終端機統一。分叉子代理（subagent_type: "fork"）在 claude -p、Agent SDK、桌面版與 VS Code 改為預設開啟，若要關閉可設 CLAUDE_CODE_FORK_SUBAGENT=0，claude mcp serve 則不提供。桌面版與 VS Code 主對話的子代理現在固定在背景執行，不再由模型決定；子代理發起的代理呼叫則預設在前景執行。檢視畫面中捲動長清單或摺疊分組後內容反白空行的問題已經修復，底部的 ← for agents 提示文字也會在工作階段開啟 10 秒後自動隱藏，僅保留等待處理的代理數量計數。
平時在企業雲端平台透過指令或程式庫呼叫 Claude 時，中途切換供應商通道不再卡關。以前在工作階段進行到一半時切換至 Vertex AI 或 Bedrock，請求容易因輸入參數不符（diagnostics: Extra inputs are not permitted）而失敗，這項相容性問題現已修復。在第三方供應商上，若未指定權限模式，且關閉遙測或屬於全新安裝，claude -p 與 Agent SDK 工作階段會以 default 啟動，若要維持自動授權需明確加上 --permission-mode auto。從其他資料夾重新開啟背景工作階段若缺少模型閘道，系統現在會具體說明無法連線的原因，不再含糊顯示未登入。
當你的自動化工作流替同一事件設定了許多自訂攔截指令時，可以留意哪些正在執行、哪些仍在等待。提示詞與代理掛鉤（hook，特定事件發生時被觸發的小程式）現在有了並行上限：同一事件最多同時執行 12 個，其餘掛鉤等待輪到自己。這個數量是按每個事件計算。
當你在不同作業系統或特殊磁碟格式下編寫專案時，現在檔案修改與版本控制會順暢許多。在 Windows 環境中，若專案位於 exFAT 或 FAT32 格式的磁碟槽上，包含加上 --plugin-dir 參數啟動的工作階段，以前每次寫入檔案都會遭到系統拒絕或頻繁彈出確認，現在這項權限誤判已經修好。在 Git 2.56 或更新版本中，若目錄殘留了指向不存在路徑的孤立 .git 檔案，工作樹（worktree，Git 讓同一專案平行開多份副本的功能）隔離機制不再拒絕建立或復原工作樹。對於自架執行器（self-hosted runner），若重複使用 reftable 格式的簽出目錄，git fetch 抓取資料時不再因物件錯誤而中斷。
當你在終端機瀏覽 Claude 輸出的長網址或包含圖示的文字時，現在視覺呈現與點擊體驗都更加俐落。以前包裹在反引號內的超長網址若跨行折行，在支援超連結的終端機裡往往只有第一行可以點擊；現在整段網址已整合成一個完整的點擊連結。連結起點落在表情符號（emoji）內時，折行寬度會多算一個字元欄位，造成該表情符號在問題預覽中被隱藏，並讓表格變成清單；這些問題已修復。
當你關注帳戶的額度消耗或在雲端持續進行專案時，現在用量與成本統計變得更加精準。如果你的訂閱方案針對特定模型僅能以額外點數計費，Claude 在回覆該次提問並扣除點數後，會自動跳過後續的提示建議，避免在背景悄悄消耗你的用量點數。同時，雲端工作階段在新的行程中恢復時，先前累積的費用與用量不再歸零。
當你請 Claude 讀取本機圖片時，來源大小會影響它是否載入。以前讀取工具（Read tool，用來查看檔案內容的工具）會把任何大小的圖片檔，或名稱像圖片的管道、裝置，整份載入記憶體。現在超過 128 MB 的圖片來源會被拒絕載入。
當你讓 Claude 在對話中動態生成或即時修改展示成果物時，發布過程不再因為小意外而中斷。成果物（artifact，獨立生成的展示面板或程式碼檔案）在發布時，若 Claude 傳送了空的修改清單，系統不再判定失敗。同時，若某份成果物在送出更新的過程中已由其他地方產生了更新的版本，遭到退件後 Claude 也不會再以錯誤的方式盲目重試，而是會遵循正確的重試機制更新內容。
當你在公司或企業受管的團隊環境下使用終端機，若想對工具功能提出建言，可以留意回饋指令顯示的提示。當所屬組織的管理政策關閉了 /feedback 指令時，畫面上的訊息現在會引導你前往 GitHub issues（用來提交問題的頁面），不再要求你聯絡內部組織管理員。
當你透過遠端操控（Remote Control）跨裝置管理工作階段時，現在連線狀態一目了然。如果你的電腦在遠端連線期間離線，介面中的子代理面板不再無止境轉圈等待，而是會明確顯示 Remote Control 連線已中斷。
同時，若你在尚未登入帳號的狀態下執行 claude remote-control --help，系統現在會正常印出完整的指令用法並順利退出，不再異常無反應。
當你在 VS Code 編輯器中一邊檢視 Claude 的回答一邊連接周邊工具時，現在畫面顯示與連線狀態都更為精準。以前即使根本無法連線到任何安裝了 Claude 擴充功能的瀏覽器，頂部的 @browser 橫幅仍會誤報「Browser connected」；現在連線狀態已校正為真實情況。當 MCP 伺服器的網址中帶有查詢字串（query string）時，從面板進行登入授權不再開啟損壞的驗證網頁。在清單項目中若包含尚未閉合的程式碼區塊，答案串流生成時後續文字一度先顯示為純文字再突兀跳成程式碼的渲染問題，現在也已徹底修正。
當你在雲端設定例行排程任務（routine）讓系統自動跑報表或分析時，現在任務的執行履歷更加可靠。在 routine 的管理頁面上，若上次執行未能順利啟動，頂部現在會以醒目橫幅具體說明原因，例如組織已達到同時並行雲端工作階段的上限。以前如果雲端工作階段發生非預期中斷，任務狀態會被誤記為已完成。若在啟動前就遭封存，則會錯誤顯示為已取消。現在這些異常情境均已正確改標記為失敗。同時，當排程任務中途失敗時，系統不會再於一小時或更久之後自作主張從頭重跑，避免重複寄送報告或造成重覆扣款。
當你透過瀏覽器進入雲端工作階段與團隊協作時，現在多項操作與介面同步修復得更加平穩。在工作階段中點擊歷史倒轉（Rewind to here）時，以前只有最新一則訊息有效、其餘皆會失敗，現在所有歷史節點都能順暢倒轉；若使用自架執行器則需搭配 Claude Code v2.1.246 或更新版本。當 Pull Request 發生合併衝突時，狀態列不再於 GitHub 重新複查期間短暫跳回準備就緒狀態。當 GitHub 儲存庫被重新命名或移動位置後，差異檢視（diff view）遺漏修改或載入失敗的問題已排除。同時，從另一台裝置傳送的訊息現在會即時出現在已開啟的頁面上，不必再手動重新整理網頁。若工作階段啟動失敗，錯誤訊息改為明確顯示連線中斷（Connection lost）或伺服器錯誤，不再含糊籠統。
當你在 Slack 頻道或私訊中標記 Claude 協助工作時，現在對話格式與傳檔體驗都更加健全。在 Slack 中上傳 HEIC、SVG、TIFF 與 BMP 等格式的圖片，現在會作為工作檔案正常傳遞給 Claude，不再遺失。當 Claude 透過編輯舊訊息回答問題後，標記下方的「is thinking…」提示會正確消除，不再永久殘留。程式碼區塊內附在清單步驟下的網址不再被多餘的角括號包裹，複製指令可直接貼上使用。回覆中的超連結也不會再把後方的粗體、斜體符號或全形標點符號吃進網址內而造成破版。私訊遇到單次 Slack 查詢延遲時，系統現在會先主動重試，不再直接回覆私訊暫不可用並丟棄訊息。在 Slack Enterprise Grid 中，私訊會在你最近開啟 Claude 的那個已連接工作區執行。
管理者在後台調整 Claude 的授權範圍與連線通道時，各項權限現在被整理得井井有條。Claude Tag 後台管理介面重新梳理為四大分頁：Bundles、Connectors、Skills and plugins 以及 Channels，各項權限歸類一目了然。儲存庫搜尋效能明顯提升，現在最多可涵蓋 100 個已連結的 GitHub 帳號，並能搜尋超過 GitHub App 安裝前 5,000 個以外的私人儲存庫。在新增位置功能中，手動輸入或貼上頻道名稱時，系統能立即顯示該頻道，不必再苦等組織的完整頻道清單全數載入。工作區或頻道頁面中的「Add > Domain」選項，只要當前頁面載入完成便立即可用，不必苦等所有頻道載入完畢。
當企業針對不同團隊角色嚴格管控 Claude 的使用權限時，管理員身分與外部工具的設定規則現在徹底釐清。在限制角色權限的組織中，新任命的頻道管理員先前可能會失去在 Slack 使用 Claude 的權限；現在管理員角色已自動涵蓋 Claude Tag 的存取授權。組織設定中的角色頁面也進行了整理，為頻道管理員建立的角色現在會集中收納在組織既有角色下方的摺疊區塊中，版面乾淨許多。若多個 MCP 伺服器共用同一個主機位址，後台不再將其錯誤合併為單一連接器，改由每個伺服器各自獨立顯示。當頻道達到每月支出上限時，系統提示不再誤稱頻道管理員能調高額度，而是明確標明組織擁有者並附上用量頁面連結。
在後台設定程式碼審查（Code Review）流程時，中途跳轉授權不再弄丟原本的進度。以前在設定流程中如果中途離開頁面去連接 GitHub 帳號，返回管理後台時系統無法自動恢復進度，常常需要重新開始；現在這項跳轉中斷問題已經修復，返回後會直接重新開啟在儲存庫挑選畫面，讓你順利接續完成設定。
When you enable auto mode (a mode that lets Claude execute tools autonomously without asking each time) to handle heavy and complex tasks, you can now customize the model used by the local safety classifier. Through the environment variable CLAUDE_CODE_AUTO_MODE_MODEL, you can specify the model for the client-side classifier while the server-side model remains unchanged. When a conversation is too long and exceeds the scope of server-side safety checks, it is now compacted first instead of rejecting tool calls outright.
An issue where most auto mode classifier block events failed to record classifier_category has been fixed. Another issue where newly added attributes crowded out tool_use_id, tool_source, and tool_parameters when OTEL_LOGRECORD_ATTRIBUTE_COUNT_LIMIT was set low has also been resolved.
When you leave a conversation window idle for too long, you might notice that Claude automatically reorganizes earlier message history. The /config settings menu now includes an "Idle compaction" option so you can freely enable or disable automatic idle compaction, and the notice displayed after reorganization finishes now explains how to turn it off. Previously, if you stopped generation midway while a conversation was being compacted, the screen would throw a "Prompt is too long" error; this issue has now been resolved.
If your gateway uses prefixed Opus 4 or 4.1 names (such as anthropic/claude-opus-4-1), running /autocompact and /effort has also been corrected to save to the proper model.
When you start a new conversation, operating instructions for Claude in Chrome can now be read from a shared prompt cache. Claude adjusted the placement of these instructions within the system prompt so that new conversations can share the existing cache.
Tools added dynamically midway through a session will no longer suffer cache misses, even after being reloaded via ToolSearch or undergoing conversation compaction.
If you regularly use UI design commands in your terminal, you will now see more accurate timelines and guidance notices during related operations. When you enter /design-sync or /design-login, the screen adds a notice explaining that the standalone web version of Claude Design is scheduled to shut down on December 14, 2026.
When executing /design consent and /design revoke fails in a cloud session, the error message now points you to claude.ai/design/settings instead of recommending unusable login commands.
When you have many plugins installed for a project, launching Claude Code now comes with noticeably shorter wait times. The system has reduced filesystem calls when reading plugin commands and agent files, an improvement that is especially noticeable on Windows. Previously, when executing subcommands like claude plugin, the process would occasionally exit before configuration file writes finished, causing lost changes or leaving behind .tmp temporary files; it now ensures writes complete before exiting. When Git defaults to the reftable reference format, the marketplace no longer repeatedly re-downloads entire repositories on every refresh, nor do newly installed plugins display an unknown version. PowerShell tools can now smoothly invoke commands under a plugin's bin/ directory, piping into formatting commands when search results are empty no longer throws errors, and failures when querying builtin plugins correctly point to /plugin.
When using external tools in the desktop app, you will now receive guidance that perfectly matches your current interface if re-authentication or reconnection is required. Previously, when desktop Claude needed to reconnect to an MCP (Model Context Protocol, a protocol that connects AI to external tools) server, it misdirected you to check web settings or run the /mcp command, even though neither option existed in that environment; this guidance has now been corrected.
Meanwhile, when an external server negotiates protocol version 2026-07-28 without providing optional fields such as resultType, ttlMs, or cacheScope, the tool list no longer goes blank, and tool calls no longer fail completely.
When using Claude in the desktop app, VS Code, or programmatically, rules for handling multiple tasks in parallel are now unified with the terminal. Forked subagents (subagent_type: "fork") are now enabled by default in claude -p, the Agent SDK, desktop, and VS Code; to disable them, set CLAUDE_CODE_FORK_SUBAGENT=0, while claude mcp serve does not offer them. Subagents in main conversations on desktop and VS Code now always run in the background rather than being decided by the model, whereas agent calls initiated by subagents run in the foreground by default. An issue in the viewer where scrolling long lists or collapsing groups resulted in highlighted blank lines has been fixed, and the bottom ← for agents hint text now automatically hides 10 seconds after a session opens, leaving only a count of pending agents.
When calling Claude on enterprise cloud platforms via the CLI or SDKs, switching provider channels midway through a session will no longer stall. Previously, switching to Vertex AI or Bedrock in the middle of a session frequently failed due to mismatched input parameters (diagnostics: Extra inputs are not permitted); this compatibility issue is now resolved. On third-party providers, if no permission mode is specified and telemetry is disabled or the setup is a fresh install, claude -p and Agent SDK sessions will start with default, requiring an explicit --permission-mode auto to maintain automatic approvals. When reopening background sessions from other folders that lack a model gateway, the system now specifically explains why the connection failed instead of vaguely reporting that you are not logged in.
When your automation workflow sets up many custom interception commands for the same event, you can keep track of which are currently executing and which are still waiting. Prompt and agent hooks (small programs triggered by specific events) now have a concurrency limit: at most 12 can run concurrently for the same event, with remaining hooks waiting their turn. This limit is calculated on a per-event basis.
When working on projects across different operating systems or specialized disk formats, file modifications and version control will now run much more smoothly. On Windows, if a project is located on an exFAT or FAT32 partition—including sessions started with the --plugin-dir flag—file writes that were previously rejected by the system or triggered frequent confirmation prompts have had their permission misjudgments fixed. In Git 2.56 or newer, if a directory retains an orphaned .git file pointing to a nonexistent path, the worktree (a Git feature for maintaining multiple copies of the same project in parallel) isolation mechanism no longer refuses to create or restore worktrees. For self-hosted runners that reuse checkout directories formatted with reftable, running git fetch to retrieve data will no longer fail due to object errors.
When you browse long URLs or text containing icons output by Claude in the terminal, visual presentation and click experiences are now much cleaner. Previously, when an extra-long URL wrapped in backticks wrapped across lines, only the first line was clickable in terminals that support hyperlinks. Now, the entire URL is consolidated into a single clickable link. When a link started inside an emoji, the wrap width counted an extra character cell, causing that emoji to be hidden in question previews and turning tables into lists. These issues are now fixed.
When you monitor account quota consumption or continue projects in the cloud, usage and cost tracking are now more precise. If your subscription plan charges for specific models solely via extra credits, Claude automatically skips follow-up prompt suggestions after answering the query and deducting credits, preventing silent credit consumption in the background. Meanwhile, when a cloud session resumes in a new process, previously accumulated costs and usage no longer reset to zero.
When you ask Claude to read local images, the source size determines whether it loads. Previously, the Read tool (a utility for inspecting file contents) loaded image files of any size, or pipes and devices named like images, entirely into memory. Now, image sources exceeding 128 MB are rejected from loading.
When you have Claude dynamically generate or modify display artifacts in conversation, the publishing flow is no longer interrupted by minor glitches. When publishing an artifact (a standalone display panel or code file), the system no longer fails if Claude transmits an empty list of modifications. Meanwhile, if a newer version of an artifact was already created elsewhere while an update was being submitted, Claude no longer blindly retries in an erroneous manner upon rejection, but follows the proper retry mechanism to update content.
When using the terminal in a corporate or enterprise-managed team environment and looking to share suggestions about tool features, notice the prompt shown by the feedback command. When your organization's administrative policy disables the /feedback command, the on-screen message now directs you to GitHub issues (a page used to submit bug reports), instead of asking you to contact an internal organization admin.
When managing sessions across devices through Remote Control, connection status is now clear at a glance. If your computer goes offline during a remote connection, the subagent panel in the interface no longer spins indefinitely, but clearly shows that the Remote Control connection has dropped.
Meanwhile, if you run claude remote-control --help without being logged in to an account, the system now prints complete command usage and exits cleanly, instead of hanging unresponsive.
When viewing Claude's responses while connecting peripheral tools inside the VS Code editor, visual display and connection status are now more accurate. Previously, even when unable to connect to any browser with the Claude extension installed, the @browser banner at the top mistakenly reported "Browser connected". Now, connection status is calibrated to reflect actual conditions. When an MCP server URL contains a query string, completing login authorization from the panel no longer opens a broken authentication webpage. If a list item contained an unclosed code block, a rendering glitch where subsequent text briefly displayed as plain text before abruptly snapping into code during streamed generation has now been completely fixed.
When configuring automated routines in the cloud to run reports or analyses, task execution histories are now more dependable. On the routine management page, if the previous run failed to start properly, a prominent banner at the top now details the specific reason, such as the organization reaching its concurrent cloud session limit. Previously, unexpected interruptions in a cloud session caused the task status to be misrecorded as completed. If a task was archived before starting, it was erroneously displayed as canceled. Now, these edge cases are properly marked as failed. Meanwhile, when a scheduled task fails midway, the system no longer takes it upon itself to restart from scratch an hour or more later, preventing duplicate reports or repeated billing charges.
When collaborating with teams in cloud sessions through a browser, multiple operations and interface sync mechanisms are now smoother. When clicking "Rewind to here" inside a session, previously only the latest message worked while all others failed, but now every historical checkpoint rewinds smoothly. Self-hosted runners require Claude Code v2.1.246 or later. When a pull request encounters merge conflicts, the status line no longer briefly flashes back to a ready state during GitHub re-reviews. When a GitHub repository is renamed or moved, issues where the diff view missed modifications or failed to load have been resolved. Meanwhile, messages sent from another device now appear instantly on open pages without requiring a manual refresh. If a session fails to launch, error messages now explicitly indicate "Connection lost" or a server error instead of staying vague.
When tagging Claude for assistance in Slack channels or direct messages, conversation formatting and file-sharing experiences are now much more reliable. Images uploaded to Slack in formats such as HEIC, SVG, TIFF, and BMP now pass properly to Claude as working files instead of getting lost. When Claude answers a question by editing an earlier message, the "is thinking…" indicator below the tag clears properly rather than lingering permanently. URLs inside code blocks attached under list steps are no longer wrapped in superfluous angle brackets, allowing copied commands to be pasted directly. Hyperlinks in responses no longer absorb trailing bold markers, italics, or full-width punctuation into the URL and break formatting. When a direct message encounters a delay in a single Slack lookup, the system now proactively retries first, rather than immediately replying that direct messages are unavailable and discarding the request. In Slack Enterprise Grid, direct messages run in whichever connected workspace where you most recently opened Claude.
When administrators adjust Claude authorization scopes and connection channels in the admin console, permissions are now neatly organized. The Claude Tag admin interface has been reorganized into four tabs: Bundles, Connectors, Skills and plugins, and Channels, making permission categories clear at a glance. Repository search performance has noticeably improved, now covering up to 100 connected GitHub accounts and searching private repositories beyond the first 5,000 indexed by GitHub App installations. In the add location feature, manually entering or pasting a channel name immediately displays that channel without waiting for the organization's full channel list to load. The "Add > Domain" option on workspace or channel pages is ready as soon as the current page finishes loading, without waiting for all channels to load.
When enterprises strictly control Claude usage permissions across different team roles, configuration rules for administrator roles and external tools are now thoroughly clarified. In organizations with restricted role permissions, newly appointed channel managers previously could lose access to Claude in Slack. The manager role now automatically includes access authorization for Claude Tag. The roles page in organization settings has also been tidied up, consolidating roles created for channel managers into a collapsible section beneath existing organization roles for a much cleaner layout. If multiple MCP servers share the same host address, the admin console no longer mistakenly merges them into a single connector, displaying each server independently instead. When a channel reaches its monthly spend cap, the system notice no longer mistakenly claims that channel managers can increase quotas, instead clearly identifying organization owners with a link to the usage page.
When configuring Code Review workflows in the admin console, redirecting mid-process for authorization no longer loses your initial progress. Previously, leaving the setup workflow to connect a GitHub account prevented the admin console from automatically restoring progress upon return, often requiring a fresh start. This redirection interruption is now resolved, reopening directly to the repository selection screen upon return so you can smoothly resume setup.
Added `CLAUDE_CODE_AUTO_MODE_MODEL` to choose the model the client-side auto mode classifier uses; it does not change the model for server-side auto mode
Fixed `classifier_category` missing from OpenTelemetry `tool_decision` events for most auto mode classifier blocks, and new `tool_decision` attributes pushing out `tool_use_id`, `tool_source` and `tool_parameters` when `OTEL_LOGRECORD_ATTRIBUTE_COUNT_LIMIT` is low
Improved auto mode with the server-side classifier: a conversation too long for the safety check is now compacted, instead of every checked tool call being deniedAdded an `Idle compaction` row to `/config` for turning idle compaction off or back on; the notices shown after an idle compaction now say how to turn it off
Fixed a "Prompt is too long" error when stopping a turn while the conversation was being compacted
Fixed `/autocompact` and `/effort` saving to the wrong model when a gateway names Opus 4 or 4.1 with a prefix, such as `anthropic/claude-opus-4-1`Fixed a prompt cache miss when a tool added mid-session was loaded again through ToolSearch or after compaction
Changed where the Claude in Chrome instructions sit in the system prompt so that new conversations can read them from the shared prompt cacheAdded a notice to `/design-sync` and `/design-login` that standalone Claude Design (claude.ai/design) closes by December 14, 2026
Fixed `/design consent` and `/design revoke` suggesting `/design-login` when they fail in a cloud session, where it isn't available; the message now points to claude.ai/design/settingsFixed `claude plugin` and other subcommands sometimes exiting before a queued settings or config write finished, which could lose the change or leave a `.tmp` file behind
Fixed plugin marketplaces re-cloning on every refresh, and newly installed git plugins showing version "unknown", when git defaults to the reftable ref format
Fixed the PowerShell tool not finding commands in a plugin's `bin/` folder, and reporting an error when a search that finds nothing (rg, grep, findstr, git grep) is piped into `Select-Object` or another formatting cmdlet
Fixed `claude plugin details <name>@builtin` suggesting `claude plugin list`, which never shows built-in plugins, when the plugin isn't found; it now points to `/plugin`
Improved startup time with many plugins enabled: plugin command and agent files now take far fewer file system calls to load, most of all on WindowsFixed Claude in the desktop app pointing you to claude.ai settings or `/mcp` to authorize or reconnect an MCP server where that option isn't available
Fixed MCP servers showing no tools, or every tool call failing, when the server negotiates protocol version 2026-07-28 but omits `resultType`, `ttlMs` or `cacheScope`Fixed rows in the agents view staying blank after moving down a long list or folding a section
Changed fork subagents (`subagent_type: "fork"`) to be on by default in `claude -p`, the Agent SDK, the desktop app and VS Code, as they already are in the terminal; `CLAUDE_CODE_FORK_SUBAGENT=0` turns them off, and `claude mcp serve` never offers them
Changed the desktop app and VS Code to always run the main conversation's subagents in the background, as the terminal does, rather than let Claude choose; a subagent's own Agent calls in `claude -p`, the Agent SDK, the desktop app and VS Code now run in the foreground by default
Changed the `← for agents` footer hint to hide 10 seconds after a session opens; the count of agents waiting on you still showsFixed requests failing with "diagnostics: Extra inputs are not permitted" after switching to Vertex or Bedrock mid-session
Improved the message when a background session reopened from another folder has no LLM gateway: it now says why, instead of "Not logged in"
Changed `claude -p` and Agent SDK sessions with no permission mode to start in `default` mode on third-party providers, with telemetry off or on a fresh install; `--permission-mode auto` keeps autoChanged prompt and agent hooks so that at most 12 of one event run at a time; the rest wait their turnFixed worktree isolation refusing to create or resume worktrees under git 2.56 or later when a stray `.git` file points at a directory that doesn't exist
Self-hosted runner: Fixed git fetches failing with "bad object refs/heads" in sessions that reuse a checkout stored in git's reftable format
Windows: Fixed every file write being refused or asked about for projects on exFAT or FAT32 drives, including in sessions started with `--plugin-dir`Fixed URLs written in backticks being clickable only on their first row when they wrap; the whole URL is now one link in terminals that support hyperlinks
Fixed an emoji that a link starts inside wrapping one column too wide, which hid it in question previews and turned tables into listsFixed prompt suggestions spending usage credits on models your plan pays for only with credits: they are now skipped after a reply billed to credits
Fixed a cloud session's cost and usage restarting from zero after the session resumed in a new processFixed the Read tool loading an image file of any size, or a pipe or device named like an image, entirely into memory; image sources over 128 MB are now refusedFixed artifact publishes sometimes failing when Claude sent an empty list of edits, and Claude retrying an artifact update the wrong way after a newer version of the artifact turned it awayChanged the message shown when your organization's policy turns off `/feedback`: it now points to GitHub issues instead of an organization adminFixed `claude remote-control --help` not printing its usage and exiting when not signed in
[Remote Control] Fixed a subagent's pane spinning forever in a Remote Control session after your computer goes offline; it now says Remote Control is disconnected[VSCode] Fixed text after an unclosed code block in a list showing as plain text while a reply streams, then turning into code
[VSCode] Fixed MCP server sign-in from the panel opening a broken authorization page when the server URL contains a query string
[VSCode] Fixed the `@browser` banner saying "Browser connected" when no browser with the Claude extension can be reached[Cloud sessions] Added a banner on a routine's page that says why its last run didn't start, such as your organization reaching its limit on concurrent cloud sessions
[Cloud sessions] Fixed routine runs showing as Completed after their cloud session stopped unexpectedly, and as Cancelled when archived before they started, instead of failed
[Cloud sessions] Fixed a scheduled routine run that failed partway being run again from the start an hour or more later, which could repeat work such as posting the same report twice[Cloud sessions] Fixed "Rewind to here" failing on every message except the newest in cloud sessions; self-hosted runners need Claude Code v2.1.246 or later
[Cloud sessions] Fixed the pull request status bar in cloud sessions flipping a pull request with merge conflicts back to ready while GitHub rechecked it
[Cloud sessions] Fixed the diff view in cloud sessions missing changes or failing to load after the session's GitHub repository was renamed or moved
[Cloud sessions] Fixed an open cloud session sometimes not showing messages sent from another device until you reloaded the page or clicked away and back
[Cloud sessions] Improved the error shown when a cloud session fails to start: it now says "Connection lost" or "Server error" instead of "Something went wrong"[Claude Tag] Fixed "is thinking…" staying up under a mention in a Slack channel or DM after Claude answered it by editing one of its earlier messages
[Claude Tag] Fixed URLs in code blocks under a numbered or bulleted step of Claude's Slack replies being wrapped in angle brackets, which broke copied commands
[Claude Tag] Fixed HEIC, SVG, TIFF and BMP images attached in Slack never reaching Claude; they now arrive as files in its working files, like other file types
[Claude Tag] Fixed Claude in Slack sometimes answering a direct message with "DMs are temporarily unavailable" and dropping it when a single Slack lookup stalled; it now retries first
[Claude Tag] Fixed direct messages with Claude on a Slack Enterprise Grid ignoring Slack's workspace dropdown; they now run under the connected workspace you last opened Claude in
[Claude Tag] Fixed links in Claude's Slack replies absorbing the closing bold or italic marks, or full-width punctuation, right after them, which broke the link and the formatting[Claude Tag] Improved repository search in Claude Tag admin settings to cover up to 100 connected GitHub accounts and find private repositories past the first 5,000 of a GitHub App install
[Claude Tag] Improved Add place in Claude Tag admin settings to list a channel you type or paste right away, instead of only after your organization's full channel list has loaded
[Claude Tag] Improved Add > Domain on a workspace's or channel's page in Claude Tag admin settings to be available once that page loads, not only after every channel has loaded
[Claude Tag] Changed Claude Tag admin settings to organize Claude's access into four tabs: Bundles, Connectors, Skills and plugins, and Channels[Claude Tag] Fixed new channel managers losing access to Claude in Slack in organizations that restrict Claude Tag by role; the manager role now also grants Claude Tag in Slack
[Claude Tag] Fixed custom connectors whose MCP servers share one host showing as a single connector in Claude Tag admin settings; each server is now its own connector
[Claude Tag] Fixed Claude in Slack's notice for a channel's monthly spend limit saying a channel manager can raise it; it now names an organization owner and links the usage page
[Claude Tag] Changed the Roles page in organization settings to group the roles created for Claude Tag channel managers in a collapsed section below your organization's own roles[Code Review] Fixed Code Review setup not resuming after you leave to connect GitHub; back in admin settings, it now reopens at the repository picker