Claude Code 2.1.275：即時插話快捷鍵送出佇列，雲端問答與 Vim 游標精準對齊

繁體中文

新功能

即時插話快捷鍵送出佇列，雲端問答與 Vim 游標精準對齊

當 Claude 正在輸出長篇回答時，你隨時能按下快捷鍵 ctrl+enter（或 ctrl+x ctrl+s）立即中斷當前輪次，並一次性送出所有排隊中的佇列訊息。這些送出與排隊中的文字會以灰色顯示，直到模型正式接收處理。在雲端工作階段（session）恢復執行時，被後續訊息取代的過期提問不再殘留在對話記錄中。在 Vim 模式下透過快速鍵或點號重複指令切換至 Shell 模式時，游標位置也不會再往右偏移一格。

新功能

應用程式網關支援登入確認，憑證安全註銷不留連線殘留

當你透過 Claude 應用程式網關（gateway，負責轉發與管理企業連線的中繼伺服器）登入時，若網關指名了帳號名稱，系統會在儲存登入憑證前要求你確認，輸入 /status 也能隨時查看目前帳號。

執行 /logout 登出時，若網關支援權杖註銷，系統會同步結束該工作階段（session）。此外，若企業網路網關在拒絕測試版標頭時改寫了 API 錯誤回應，終端機也不會再於每個輪次卡在 400 錯誤。

新功能

網頁帳號技能與外掛同步終端機，本機修改明確提示儲存方式

你在網頁版 claude.ai 啟用的技能與外掛，現在登入同帳號的終端機工作階段（session）時會自動同步啟用。若想關閉此功能，可在設定檔中指定以下開關：

{
  "syncClaudeAiSkills": false,
  "syncClaudeAiPlugins": false
}

ListPlugins 工具說明已更新，讓模型清楚知道這是列出帳號外掛而非本機外掛。修改同步技能資料夾中的檔案時，編輯結果會明確提醒變更尚未存回雲端帳號並附上儲存指引。內建 claude-api 技能即時來源表中的四個失效文件網址也已修正。

新功能

外掛安裝支援來源市集提示，敏感密碼遮蔽與套件腳本防護

安裝外掛現在能直接指定來源市集，指令格式為 /plugin install <plugin> --marketplace <source>，系統會在安裝前主動詢問是否先加入該市集。市集更新失敗時不會再誤刪本地副本，記錄與清單中包含在網址內的密碼或授權權杖已全面隱藏。某些 Git 位址顯示錯誤伺服器的問題已修復。從封存檔載入的外掛在預覽重載時，不再覆寫執行中工作階段（session）的解壓縮檔案。從 npm 來源安裝的外掛改採 npm pack --ignore-scripts 並校驗完整性，徹底防止安裝腳本暗中執行。

新功能

Artifact 發布共享指引強化，跨裝置接續與例行任務自動更新

當工作階段（session）換機恢復或暫存區被清空時，更新失敗的問題已修復。系統會自動還原該頁面最後發布的版本。

發布與讀取結果現在會清楚標明瀏覽權限與分享選單選項，初次發布時改為要求單字分頁圖示，並在頁面包含空位元組時發出警告。被賦予編輯權限的共享頁面會直接在原處更新，不再另行建立複本。例行任務在執行時能自動儲存並重新發布你可編輯的頁面，省去反覆詢問的步驟。

新功能

提示詞快取命中率提升，遙測設定示警與方案用量查詢共用

當你在 --system-prompt 中加入 __SYSTEM_PROMPT_DYNAMIC_BOUNDARY__ 標記時，分界線以上的文字現在會進行全域快取（cache，把計算過的結果暫存起來直接取用）。在對話壓縮或恢復執行後，記憶檔案的時間標註不再於每次請求間變動，有效解決了快取失效的問題。回覆中偶爾出現的多餘標籤已修正。同一台機器上的編輯器視窗與非互動工作階段（session）現在會共用一分鐘內查詢過的方案用量，避免重複呼叫端點。啟動時若設定的遙測小幫手執行失敗，系統會主動發出警告。

問題修復

檔案搜尋與讀取不再卡住，沙盒指令狀態碼精準回報

使用 Grep、Glob 或輸入 @ 搜尋檔案時，若搜尋結果超出上限，程式不再發生記憶體不足或卡死，ripgrep 遇到大量警告時也會正確回報錯誤。自訂檔案建議指令或輸入 @. 時，檔案建議不會再被 MCP（讓 AI 連接外部工具的協定）資源排在下方。在記憶體吃緊導致大檔案無法解碼時，讀取工具現在會明確回報錯誤而非無回應。Linux 沙盒環境下使用 Zsh 時，失敗的 Bash 指令不再誤報成功結束代碼，且沙盒環境已支援寫入名為 hooks/ 或 config/ 的專案目錄。

問題修復

工作階段還原容錯強化，損毀記錄與啟動異常自動修復

恢復工作階段（session）時的穩定性現在更加健全。

以前若儲存的歷史紀錄中包含格式異常的任務提醒、附件項目或對話內容區塊，常導致 --resume、預覽選單或背景代理還原失敗，現在均已能妥善處理。全螢幕檢視此類對話時，在向上捲動狀態下接收新訊息也不會再引發當機。使用 --resume 或 --continue 接續任務時，即使啟動時包含的內建工具已被伺服器端旗標關閉，對話先前的思考歷程依然完整保留。此外，設定檔 ~/.claude.json 包含異常數值導致啟動當機的問題也已修復。

問題修復

全螢幕檢視流暢度升級，背景通知定位與選取複製修復

在全螢幕工作階段（session）中，背景任務的完成通知現在會即時關閉展開的工具列，並精準顯示在抵達位置，不再被壓在摺疊行底下。向上捲動瀏覽長篇檔案修改比對時，畫面不再短暫凍結或反白數秒。在斜線指令下拉選單開啟時快速打字或按住按鍵，不再觸發無法復原的介面錯誤而退出程式。在 claude --resume 選擇清單中用滑鼠選取的文字，現在能順利複製至剪貼簿。當終端機運作緩慢或暫停時，系統反應已獲得改善，輸出內容不會在等待期間持續落後。

問題修復

子代理訊息轉發完整保留，背景任務與復原指令穩定執行

當你在分支技能（forked skill）中啟動子代理時，--forward-subagent-text 輸出與 SDK 不再遺漏轉發訊息。在背景工作階段（session）中執行 /rewind 倒轉時，即使歷史備份未完全複製，也不會再還原出全零或遭截斷的損毀檔案。在檔案描述元耗盡的機器上透過標準輸入（stdin）執行指令時，背景工作單元不再異常崩潰重啟。自架執行器在收到關機訊號時會短暫等待當前輪次回報，避免丟失最終結果。帶有特定比對條件的 SubagentStop 掛鉤（hook，在特定事件發生時被觸發的小程式），也不會再對代理類型為空的子代理錯誤觸發。

新功能

權限設定規則精確套用，託管環境重啟與圖片存取順暢

執行 /update-config 更新權限時，系統現在會正確寫入 Edit(path) 規則而非無法比對的寫入規則。貼上與附加的圖片現在會存放在安全位置，讓模型直接開啟為檔案而無需反覆彈出權限提示。在雲端託管的工作階段（session）中，容器重啟後未回答的授權提示會繼續保留，不必重新詢問。Chrome 擴充套件在自動模式下改為跳過逐站檢查，修復了重新導向後的批次瀏覽權限問題。無法開啟桌面版時，錯誤訊息會清楚說明原因與後續步驟。此外，單次排程例行任務執行完畢後的啟動提示通知已移除。

新功能

［VSCode］記憶對話框支援編修，檔案修改分項審查與無字送圖

在 VS Code 的記憶管理對話框中，你現在可以直接檢視、修改或刪除已儲存的偏好紀錄。

修改比對頁籤中為每處變更增設了接受與拒絕按鈕，讓你能逐項審查修改內容。附加圖片現在支援直接送出而無需輸入額外文字。若工作階段（session）偏好位置設定為側邊欄，開啟新工作階段時會直接在側邊欄顯示而非開啟新分頁。在高對比淺色佈景主題下，計畫預覽與權限規則對話框的文字對比度已修復，且在擴充套件無回應時誤送被忽略檔案中選取文字的罕見問題也已解決。

新功能

［VSCode］代理地圖視覺化升級，背景任務崩潰復原狀態更精準

VS Code 中的代理地圖現在獲得顯著強化：狀態標籤能即時顯示執行中的代理數量並在失敗時轉為紅燈，捲動時主代理會持續固定在畫面上，所有代理清單依狀態與結束時間整齊排序。重新開啟工作階段（session）後，先前因崩潰而未完成的代理會正確標記為失敗而非停止。在背景任務執行期間若發生崩潰並重載視窗，接續步驟的提示通知能正常出現，且接續次數上限不會被異常重設。背景代理在工具執行期間完成時的結果通知與地圖狀態，在視窗重新整理後也能完整保留。

問題修復

［VSCode］對話捲動體驗改善，斜線指令排隊與發送設定自訂

當 Claude 正在處理任務時，新發送的訊息現在會整齊停留在對話最底部排隊，直到模型開始處理它。在此期間輸入的斜線指令，會在當前回答結束後依序執行，不再被當成純文字誤送給模型。當你向上捲動閱讀歷史訊息時，串流回覆不再強行將畫面拉回底部，若想徹底關閉發送時自動捲動至底部的行為，可新增以下設定：

{
  "claudeCode.scrollToBottomOnSend": false
}

發送訊息時捲動未對齊新回覆，以及權限卡片等待確認時畫面微幅跳動的問題均已修復。

問題修復

［VSCode］工作階段清單時間精準，分支對話權限與重新命名修復

在 VS Code 的工作階段（session）清單中，排序時間現在忠實依據最後一則訊息的發送時間顯示，不再因重新開啟視窗而更新。

重新命名執行中的工作階段時，名稱不會再無故跳回系統預設名稱。執行「從此處分支對話」或倒轉對話時，新對話會完整沿用原先設定的授權權限模式。修復了在 Claude 運作期間發送訊息後接續分支會失敗的問題，且重新開啟工作階段後提示詞快取（cache，暫存計算結果）的有效計時器已能正確顯示剩餘時間。部分雲端對話在 VS Code 開啟時呈現空白畫面的問題也已順利修復。

新功能

［VSCode］MCP 伺服器支援重試，遠端遙控開關與外掛設定安全

當 MCP（讓 AI 連接外部工具的協定）伺服器清單載入失敗時，對話框中現在會主動顯示「重試」連結供你快速重新連線。在遠端遙控（Remote Control）連線中若想取消，再次輸入 /remote-control 即可立即關閉連線，且連線關閉失敗後狀態能正確顯示為已關閉。外掛管理對話框中不再外洩市集網址內輸入的密碼或權杖，並修正了無效安裝連結的彈窗問題。此外，若在環境變數設定中留下空的 CLAUDE_CONFIG_DIR，系統不會再將設定檔案錯誤存放在工作區目錄中。

新功能

［Web］例行任務頁面新增捷徑，通知完整度與雲端網域設定防護

當例行任務（Routine）連結失效時，錯誤頁面現在新增了「新增例行任務」按鈕，方便你快速建立新任務。

任務暫停或保留的通知字句不再被中途截斷，訂閱暫停時會清楚提示你在恢復後手動重啟任務。若請 Claude 操作非它建立的任務，它會直接提供該任務的設定頁連結。雲端工作階段（session）設定過長的允許網域清單時，儲存當下便會主動阻擋並提示應縮減的長度，避免日後啟動失敗。個人帳號遇到 GitHub 存取受阻時，系統指引會直接附上正確的帳號連結頁面。

問題修復

［Claude Tag］存取套件支援外部頻道，雲端監控預設與 S3 連線修復

在 Claude Tag 管理設定中，擁有者現在可以為存取套件（access bundle）設定附加條件，允許該套件套用至包含訪客的頻道或 Slack Connect 跨企業頻道，不再侷限於僅限成員的私密頻道。在憑證標籤頁中，新增了 Amazon CloudWatch、CloudWatch Logs、Amazon SNS、Google Cloud Monitoring 與 Cloud Logging 等預設範本。Datadog 連線新增了多個站點預設並限制為唯讀與查詢路由。透過 AWS 連線使用新版 CLI 與 SDK 上傳至 S3 儲存庫時觸發 502 錯誤的問題已修復。

新功能

［Claude Tag］Slack 討論串身分維持，頻道活躍度判斷與企業網格提示

當 Claude 仍在透過例行任務或討論串在 Slack 頻道中發言時，不會再將該頻道誤判為閒置而略過未標記的訊息。

在對話工作階段（session）重新整理或重啟後，討論串中的專屬任務名稱不再跳回預設名稱，手動切換的語言模型也不會被默默重設為頻道預設值。當其他應用程式或機器人在頻道頂層提及 Claude 時，偶爾重複回覆兩次的問題已修復。在 Enterprise Grid 跨工作區共用頻道中，若尚未設定工作區，系統通知會清楚說明原因。

問題修復

［Code Review］多代理審查分析完整呈現，大量審查 PR 合併複審輕量化

當多個審查代理協同分析程式碼時，即使個別代理以非預期的格式回傳發現，審查報告也不會再遺失部分分析結論。在累積超過一百次 Claude 審查的 Pull Request（合併請求）中，從基礎分支進行乾淨合併時，系統現在會精確執行專注於合併內容的輕量複審，不再耗費資源觸發通篇完整重審。以前每次合併都會強制全面重審，如今省去了重複審查的等待時間。

English

FEATURE

A Send-Now Key Flushes Queued Messages, Cloud Session Resume and Vim Cursor Get Precise

While Claude is streaming a long reply, you can now press ctrl+enter (or ctrl+x ctrl+s) at any time to interrupt the current turn and send every queued message at once. Sent and queued text shows in gray until the model actually receives it. When a cloud session resumes, a stale question that a later message had already superseded no longer lingers in the transcript. And in vim mode, switching to shell mode with a dot-repeated shortcut or command no longer nudges the cursor one character to the right.

FEATURE

Apps Gateway Sign-In Adds a Confirmation Step, Logout Now Clears the Connection Cleanly

When you sign in through the Claude apps gateway (a relay server that forwards and manages enterprise connections), if the gateway names an account, the system now asks you to confirm it before the credential is saved, and /status shows the current account at any time.

Running /logout now also ends the session on gateways that advertise token revocation. And when a corporate network gateway rewrites API error responses while rejecting a beta header, the terminal no longer gets stuck with a 400 error on every turn.

FEATURE

Web Account Skills and Plugins Now Sync to the Terminal, Local Edits Say Clearly How to Save

Skills and plugins you've enabled on claude.ai on the web now sync automatically into terminal sessions signed in with the same account. To opt out, set these in your config:

{
  "syncClaudeAiSkills": false,
  "syncClaudeAiPlugins": false
}

The ListPlugins tool description has been updated so Claude knows it lists plugins enabled on your account, not ones installed locally. When you edit a file inside the synced skills folder, the edit result now says clearly that the change isn't saved to your cloud account, with instructions on how to save it. Four dead documentation URLs in the bundled claude-api skill's live-sources table have also been fixed.

FEATURE

Plugin Install Can Name a Source Marketplace, Passwords Get Masked and Install Scripts Blocked

Installing a plugin can now name a source marketplace directly with /plugin install <plugin> --marketplace <source>, and the system offers to add that marketplace first if it isn't already added. A failed marketplace update no longer deletes the local copy by mistake, and passwords or tokens embedded in a URL are now hidden everywhere they'd show up in messages and lists. A bug that showed the wrong server for some git addresses has been fixed. Reloading a plugin loaded from an archive no longer overwrites the extracted files of a running session. Plugins installed from an npm source are now fetched with npm pack --ignore-scripts and integrity-verified, so install scripts can no longer run silently.

FEATURE

Artifact Publish Guidance Gets Sharper, Cross-Machine Resume and Routines Update in Place

A bug where Artifact updates failed after a session resumed on another machine or its scratchpad was cleared has been fixed; the page's last published version is now restored automatically.

Publish and read results now say clearly who can open the page and what the owner's Share menu offers, a first publish now asks for a one-word tab icon instead, and a warning appears when the page contains a NUL byte. A shared page you've been given edit access to now updates in place instead of publishing a separate copy. Scheduled and Run now routine runs can automatically save data to, and republish, a page you can edit, without asking each time.

FEATURE

Prompt Cache Hit Rate Goes Up, Telemetry Setup Gets a Warning and Usage Reads Get Shared

When your --system-prompt contains a __SYSTEM_PROMPT_DYNAMIC_BOUNDARY__ line, the text above it is now cached globally. After a compaction or resume, a restored memory file's age note no longer changes between requests, which had been causing cache misses. A stray closing tag that occasionally appeared in replies has been fixed. Editor windows and non-interactive sessions on the same machine now share a plan-usage read made within the last minute instead of each calling the usage endpoint separately. A startup warning now appears when a configured telemetry headers helper fails to run.

BUGFIX

File Search and Reads No Longer Hang, Sandboxed Commands Report Status Correctly

When Grep, Glob, or an @ search returns more results than the cap, the tool no longer hangs or runs out of memory, and system ripgrep now reports an actual error instead of "no matches" after a flood of warnings. With a custom file-suggestion command or when typing @., file suggestions are no longer buried below MCP (a protocol that connects AI to external tools) resources. When memory pressure keeps a large file from being decoded, the Read tool now reports an error instead of hanging. On Linux sandboxes running zsh, a failed Bash command no longer reports exit code 0, and sandboxes can now write to project directories named hooks/ or config/.

BUGFIX

Session Resume Gets More Fault-Tolerant, Corrupted Records and Startup Crashes Now Recover

Resuming a session is now noticeably more stable.

A saved history containing a malformed task-reminder, @-file attachment entry, or message content block used to make --resume, the resume picker preview, or a resumed background agent fail; all of these now recover properly. Viewing such a conversation in fullscreen no longer crashes when new messages arrive while you're scrolled up. Continuing with --resume or --continue now keeps a conversation's earlier thinking intact, even when a built-in tool it started with has since been switched off by a server-side flag. A crash at launch caused by a malformed value in ~/.claude.json has also been fixed.

BUGFIX

Fullscreen Mode Gets Smoother, Background Notices and Text Selection Get Fixed

In a fullscreen session, a background task's completion notice now closes the open tool row right away and shows up exactly where it arrived, instead of getting buried beneath a long turn's collapsed tool row. Scrolling up through a large file diff no longer freezes or blanks the screen for several seconds. Typing fast or holding a key while the slash-command dropdown is open no longer triggers an unrecoverable interface error that exits the program. Text selected with the mouse in the claude --resume session picker now actually reaches the clipboard. Responsiveness has also improved when the terminal is slow or paused, so output no longer keeps falling further behind while it waits.

BUGFIX

Subagent Message Forwarding Stays Complete, Background Tasks and Rewind Run Stably

When you spawn a subagent inside a forked skill, --forward-subagent-text stream-json output and the SDK no longer drop its forwarded messages. Running /rewind in a background session no longer restores a zero-filled or truncated file when the session's file-history backups couldn't be fully copied. A command run through stdin on a machine that's run out of file descriptors no longer crashes and restarts the background worker. A self-hosted runner receiving a shutdown signal now waits briefly for the current turn to be reported before it loses the final result. A SubagentStop hook with a specific matcher no longer fires incorrectly for every stopping subagent whose agent type is empty.

FEATURE

Permission Rules Now Apply Correctly, Hosted Restarts and Image Access Go Smoother

Running /update-config now correctly writes Edit(path) permission rules instead of Write(path) rules that file permission checks don't match. Pasted and attached images are now saved somewhere Claude can open them directly as files without repeated permission prompts. In a cloud-hosted session, an unanswered permission prompt now stays up after a container restart instead of asking again. Claude in Chrome in auto mode now skips the extension's per-site check for classifier-approved calls, fixing a permission issue with batch browsing after a redirect. When Claude Desktop won't open, the error message now says clearly why and what to do next. The startup notice about a one-off scheduled routine having run since your last session has also been removed.

FEATURE

[VSCode] Memory Dialog Gets Editing, Diff Review Goes Change-by-Change and Images Send Without Text

In VS Code's Memory dialog, you can now view, edit, or delete a saved memory directly.

The proposed-change diff tab now has accept and reject buttons under each change, so you can review an edit change by change. An attached image can now be sent without typing any text. When Preferred Location is set to Sidebar, a new session now opens in the sidebar instead of always opening another tab. Text contrast in the plan preview and the Hooks and Permission rules dialogs has been fixed under the High Contrast Light theme, and a rare case where text selected in a git-ignored file could be sent after the extension was unresponsive has also been fixed.

FEATURE

[VSCode] Agent Map Visualization Improves, Crash Recovery Status Gets More Accurate

The agent map in VS Code has gotten a solid upgrade: the pill now counts running agents live and turns red after a failure, the main agent stays pinned in view while the map scrolls, and the agent list sorts by state then end time. After reopening a session, an agent a crash left unfinished is now correctly marked as failed instead of stopped. If a crash happens during a background task and the window reloads, the "Continuing the step" notice now appears correctly and the continue limit no longer resets by mistake. A background agent's completion notice and its status on the map, when it finishes while a tool is running, now also survive a window reload.

BUGFIX

[VSCode] Conversation Scrolling Improves, Slash Commands Queue and Send Behavior Is Configurable

While Claude is working, a newly sent message now waits neatly at the bottom of the conversation until Claude starts on it. A slash command typed during that time now runs in order once the current reply finishes, instead of being sent to the model as plain text. When you scroll up to read earlier messages, a streaming reply no longer forces the view back to the bottom; to turn off the jump-to-bottom-on-send behavior entirely, add this setting:

{
  "claudeCode.scrollToBottomOnSend": false
}

A bug where scrolling on send didn't line up with the new reply, and a small jump in the screen while a permission card waits for confirmation, have both been fixed.

BUGFIX

[VSCode] Session List Timestamps Get Accurate, Forked Conversations Keep Permissions and Renaming Gets Fixed

In VS Code's session list, the sort time now faithfully reflects when the last message was sent, instead of updating just because the window was reopened.

Renaming a running session no longer reverts to the generated name for no reason. Using "Fork conversation from here" or rewinding a conversation now keeps the permission mode you had originally picked. A bug where forking failed on the message right after one sent while Claude was working has been fixed, and the prompt cache clock now shows the correct remaining time after reopening a session with a message sent while Claude was working. Some cloud conversations opening in VS Code as a blank screen has also been fixed.

FEATURE

[VSCode] MCP Server List Gets a Retry, Remote Control Toggle and Plugin Settings Get Safer

When the MCP (a protocol that connects AI to external tools) server list fails to load, the dialog now shows a Retry link so you can reconnect quickly. If you want to cancel a Remote Control connection, running /remote-control again now turns it off immediately, and it now shows as off correctly after a failed turn-off. The Manage plugins dialog no longer leaks a password or token typed into a marketplace URL, and a bug with invalid install links opening the dialog has been fixed. An empty CLAUDE_CONFIG_DIR left in the environment variables setting no longer makes Claude Code keep its files in the workspace by mistake.

FEATURE

[Web] Routine Pages Get a New Shortcut, Notifications and Cloud Domain Settings Get Safer

When a routine link no longer resolves, the error page now shows a "New routine" button so you can create one quickly.

A routine's "paused" and "on hold" notifications are no longer cut off mid-sentence, and the paused-subscription notice now says clearly that you need to turn the routine back on yourself once it resumes. When you ask Claude to edit, delete, or run a routine it didn't create, it now links you straight to that routine's page. A cloud environment with a very long allowed-domains list now fails to save up front with a message on how much to trim, instead of saving fine and failing every session start later. Claude's guidance when a personal account's GitHub access is denied now links directly to the correct account page.

BUGFIX

[Claude Tag] Access Bundles Can Reach Outside Channels, Cloud Monitoring Presets and S3 Get Fixed

In Claude Tag admin settings, an Owner can now set attach conditions for an access bundle, letting it also apply in channels with guests or Slack Connect channels, not just member-only ones. New presets have been added to the Credentials tab, including Amazon CloudWatch, CloudWatch Logs, Amazon SNS, Google Cloud Monitoring, and Cloud Logging. Datadog connections now add presets for several sites and are limited to Datadog's read and query routes. A bug where S3 uploads from recent AWS CLI and SDK versions failed with a 502 error over an AWS connection has been fixed.

FEATURE

[Claude Tag] Slack Threads Keep Their Identity, Channel Activity Detection and Enterprise Grid Notices Improve

Claude no longer mistakes a channel for inactive and skips untagged messages there while it's still posting in that channel from a routine or a thread.

After a session refreshes or restarts, a thread's dedicated task name no longer reverts to the default, and a manually switched model no longer silently resets to the channel default. A bug where Claude sometimes replied twice when another app or bot mentioned it in a top-level channel message has been fixed. In Enterprise Grid channels shared across workspaces, the notice now says clearly when no workspace has been set up yet.

BUGFIX

[Code Review] Multi-Agent Analysis Stops Dropping Findings, Heavy-Review PRs Get Lighter Merge Checks

When multiple reviewing agents analyze code together, the review report no longer loses part of its analysis just because one agent returned its findings in an unexpected format. On a pull request with more than 100 Claude reviews, a clean merge from the base branch now correctly triggers the lighter, merge-focused review instead of a full re-review. Every merge used to force a complete re-review, which wasted time; that repeated wait is now gone.

原始 CHANGELOG

Original changelog即時插話快捷鍵送出佇列，雲端問答與 Vim 游標精準對齊

Added a send-now key (ctrl+enter, or ctrl+x ctrl+s) that interrupts the current turn and sends all queued messages at once; sent and queued messages show in gray until the model receives them
Fixed a resumed cloud session leaving an unanswered question open in the transcript after a queued message superseded it
Fixed vim mode placing the cursor one character right after a dot-repeated "!" or a fast-typed "i!" switched a non-empty prompt into shell mode

Original changelog應用程式網關支援登入確認，憑證安全註銷不留連線殘留

Added the signed-in account to Claude apps gateway sign-in: when the gateway names it, you confirm it before the credential is saved, and `/status` shows it
Fixed a terminal `API Error: 400` on every turn for users behind a network gateway that rewrites API error responses when a beta request header is rejected
Updated `/logout` for Claude apps gateway sign-ins to also end the session on gateways that advertise token revocation

Original changelog網頁帳號技能與外掛同步終端機，本機修改明確提示儲存方式

Added syncing of the skills and plugins enabled on your claude.ai account to terminal sessions signed in with it; opt out with `syncClaudeAiSkills: false` or `syncClaudeAiPlugins: false`
Fixed four dead documentation URLs (Pricing, Computer Use, Skills, CLI) in the bundled claude-api skill's live-sources table
Improved the `ListPlugins` tool description so Claude knows it lists plugins enabled on your claude.ai account, not plugins installed locally with `/plugin`
Improved Write and Edit results for files in the synced account-skills folder: they now say the change is not saved to your account and how to save it

Original changelog外掛安裝支援來源市集提示，敏感密碼遮蔽與套件腳本防護

Added `/plugin install <plugin> --marketplace <source>`, which offers to add the marketplace before installing the plugin
Fixed `claude plugin marketplace update` deleting a GitHub marketplace's local copy when the fetch failed and the marketplace was named after its repository
Fixed plugin and marketplace messages, logs and `claude plugin marketplace list` showing a password or token stored in a git, ssh or marketplace URL
Fixed plugin messages, logs and the VS Code plugin dialog showing the wrong server for some git addresses
Fixed plugin reload previews replacing a running session's extracted plugin files when the plugin was loaded from a `--plugin-dir` or `--plugin-url` archive
Changed plugins installed from an npm source to be fetched with `npm pack --ignore-scripts` and integrity-verified, so a package's install scripts no longer run

Original changelogArtifact 發布共享指引強化，跨裝置接續與例行任務自動更新

Fixed Artifact updates failing with "File not found" after a session resumes on another machine or its scratchpad is cleared: the page's last published version is restored
Improved the Artifact tool's publish and read results: they now say who can open the page and what the owner's Share menu offers
Improved artifact publish results: they name the tab icon sent, warn when the page contains a NUL byte, and retry a flaky fetch of the newer page to merge after a stale publish
Improved the Artifact tool's guidance so Claude updates a shared artifact in place when you were given edit access to it, instead of publishing a separate copy
Changed the Artifact tool to ask for a one-word tab icon on a first publish instead of an emoji favicon
Changed scheduled and Run now routine runs to save data to, and republish the page of, an artifact you can edit without asking; public artifacts, first publishes and deletes still ask

Original changelog提示詞快取命中率提升，遙測設定示警與方案用量查詢共用

Added a startup warning when a configured `otelHeadersHelper` fails, so sessions that silently export no telemetry are noticed
Fixed a restored memory file's age note changing between requests after a compaction or resume, which caused prompt cache misses
Fixed a stray `</ccmemory>`-style closing tag occasionally appearing in responses
Improved prompt caching for a `--system-prompt` that contains a `__SYSTEM_PROMPT_DYNAMIC_BOUNDARY__` line: the text above it is now cached globally, as the SDK's array form already is
Improved plan-usage reads: editor windows and non-interactive sessions on one machine now share a read made in the last minute instead of each calling the usage endpoint

Original changelog檔案搜尋與讀取不再卡住，沙盒指令狀態碼精準回報

Fixed @-mention file suggestions being buried below MCP resources when using a custom `fileSuggestion` command or typing `@.`/`@./`
Fixed sandboxed Bash commands on Linux reporting exit code 0 for failed commands when the shell is zsh
Fixed the Read tool hanging instead of reporting an error when part of a large file could not be decoded under memory pressure
Fixed Grep, Glob and @-file suggestions hanging or running out of memory on searches over the 20MB output cap, and system ripgrep reporting "no matches" instead of an error after a flood of warnings
Fixed sandboxed Bash commands being unable to write to project directories named `hooks/` or `config/`

Original changelog工作階段還原容錯強化，損毀記錄與啟動異常自動修復

Fixed `--resume`, the resume picker preview, resumed background agents and the transcript view failing on a session whose saved history contains a malformed task-reminder or @-file attachment entry
Fixed a crash when resuming a conversation whose transcript contains a malformed message entry, and a fullscreen crash when such a conversation received new messages while scrolled up
Fixed sessions failing to resume or start when their saved transcript contains a malformed message content block
Fixed a crash at launch when `~/.claude.json` holds a malformed `mcpNeedsAuthNoticed` value
Fixed `--resume` and `--continue` dropping a conversation's earlier thinking when a built-in tool it started with has since been switched off by a server-side flag

Original changelog全螢幕檢視流暢度升級，背景通知定位與選取複製修復

Fixed fullscreen mode placing background-task completion notices beneath a long turn's collapsed tool row instead of where they arrived; each notice now closes the open row
Fixed fullscreen mode freezing or blanking for several seconds when scrolling up past a large file diff
Fixed fullscreen sessions sometimes exiting with "Claude Code exited after an unrecoverable interface error" when typing fast or holding a key with the slash-command dropdown open
Fixed text selected with the mouse in the fullscreen `claude --resume` session picker never reaching the clipboard
Improved responsiveness when the terminal is slow or paused: output no longer falls further behind while the terminal catches up

Original changelog子代理訊息轉發完整保留，背景任務與復原指令穩定執行

Fixed `--forward-subagent-text` stream-json and SDK output dropping the messages of subagents spawned by a `context: fork` skill, and of forked skills invoked by a subagent or another forked skill
Fixed `/rewind` in a forked or background session restoring a zero-filled or truncated file when the session's file-history backups could not be fully copied
Fixed background sessions crashing and restarting their worker when a command fed through stdin ran on a machine that had run out of file descriptors
Fixed self-hosted runners with `--drain-wait-sec` losing the final result of a turn that finished during a SIGTERM drain; the runner now waits briefly for the turn to be reported
Fixed `SubagentStop` hooks with a specific `matcher` firing for every stopping subagent whose agent type was empty

Original changelog權限設定規則精確套用，託管環境重啟與圖片存取順暢

Fixed `/update-config` writing `Write(path)` permission rules, which file permission checks don't match, instead of `Edit(path)` rules
Improved the `/desktop` error when Claude Desktop does not open: it now says why and what to do next
Improved pasted and attached images: they are now saved where Claude can open them as files without a permission prompt, including in Desktop and VS Code
Changed hosted sessions to keep an unanswered permission prompt up after a container restart, instead of asking again
Changed Claude in Chrome in auto mode to skip the extension's per-site check for classifier-approved calls, as bypass mode does, fixing `browser_batch` "Permission denied" after a redirect
Removed the startup notice that told you a one-off scheduled routine had run since your last session

Original changelog［VSCode］記憶對話框支援編修，檔案修改分項審查與無字送圖

[VSCode] Added viewing, editing and deleting a saved memory inside the Memory dialog
[VSCode] Added sending an attached image without typing any text
[VSCode] Added accept and reject buttons under each change in the proposed-change diff tab, so an edit can be reviewed change by change
[VSCode] Fixed a rare case where text selected in a git-ignored file could be sent to Claude after the extension was unresponsive for several seconds
[VSCode] Fixed unreadable code in the plan preview and the Hooks and Permission rules dialogs with the High Contrast Light theme
[VSCode] Changed New session in a Claude editor tab to open in the sidebar when Preferred Location is set to Sidebar, instead of always opening another tab

Original changelog［VSCode］代理地圖視覺化升級，背景任務崩潰復原狀態更精準

[VSCode] Fixed the agent map showing agents a crash left unfinished as stopped instead of failed once the session is reopened
[VSCode] Fixed the "Continuing the step" notice not appearing, and the continue limit resetting, after a reload that follows a crash with background tasks still running
[VSCode] Fixed a background agent that finished while Claude was running a tool losing its completion notice, and its result on the agent map, after a window reload
[VSCode] Improved the agent map: the pill counts running agents and turns red after a failure, the main agent stays in view while the map scrolls, and agents sort by state then end time

Original changelog［VSCode］對話捲動體驗改善，斜線指令排隊與發送設定自訂

[VSCode] Fixed the transcript creeping toward the bottom in small steps while a permission card waits and content keeps arriving
[VSCode] Fixed the scroll to the bottom on send stopping short of the reply when the reply starts arriving during the scroll
[VSCode] Fixed slash commands typed while Claude is responding being sent to the model as text instead of running once the response finishes
[VSCode] Fixed the conversation pulling you back to the bottom while a reply streams after you scroll up, and added a `claudeCode.scrollToBottomOnSend` setting to turn off the jump on send
[VSCode] Changed a message sent while Claude is working to wait at the bottom of the conversation until Claude starts on it

Original changelog［VSCode］工作階段清單時間精準，分支對話權限與重新命名修復

[VSCode] Fixed rewound and forked conversations not keeping the permission mode you had picked for the original conversation
[VSCode] Fixed the session list showing when a session was last reopened, such as after a window reload, instead of when its last message was sent
[VSCode] Fixed "Fork conversation from here" failing on the message right after one sent while Claude was working
[VSCode] Fixed the prompt cache clock showing too few minutes after reopening a session with a message sent while Claude was working
[VSCode] Fixed renaming a running session reverting to the generated name (regression in 2.1.269)
[VSCode] Fixed some claude.ai/code sessions opening in VS Code as an empty conversation with no messages

Original changelog［VSCode］MCP 伺服器支援重試，遠端遙控開關與外掛設定安全

[VSCode] Added a Retry link to the MCP servers dialog when the server list fails to load
[VSCode] Fixed an empty `CLAUDE_CONFIG_DIR` entry in the `environmentVariables` setting making Claude Code keep its files in the workspace
[VSCode] Fixed plugin install links opening the Manage plugins dialog for plugin names and marketplace addresses that can't be used in a link
[VSCode] Fixed Remote Control staying shown as connected after a turn-off that Claude Code reported as failed; it now shows as off
[VSCode] Fixed `/remote-control` being ignored while Remote Control is still connecting: running it again now turns Remote Control off immediately
[VSCode] Fixed the Manage plugins dialog showing a password or token that was typed into a marketplace URL

Original changelog［Web］例行任務頁面新增捷徑，通知完整度與雲端網域設定防護

[Claude Code on the web] Added a "New routine" button to the page shown when a routine link no longer resolves, next to the link back to your routines list
[Claude Code on the web] Fixed routine "paused" and "on hold" notifications being cut off mid-sentence; the paused-subscription notice now says to turn the routine back on yourself
[Claude Code on the web] Fixed cloud environments with a very long allowed-domains list saving fine and then failing every session start; saving now fails up front and says how much to trim
[Claude Code on the web] Fixed Claude's guidance when a cloud session on a personal account is denied GitHub access: it now links to claude.ai/connect-github instead of an admin settings page
[Claude Code on the web] Improved what Claude tells you when asked to edit, delete or run a routine it didn't create: it now links to the routine's page so you can do it yourself

Original changelog［Claude Tag］存取套件支援外部頻道，雲端監控預設與 S3 連線修復

[Claude Tag] Added attach conditions for access bundles in Claude Tag settings: an Owner can let a bundle also apply in channels with guests or Slack Connect channels, not just member-only
[Claude Tag] Added Amazon CloudWatch, CloudWatch Logs, Amazon SNS, Google Cloud Monitoring and Cloud Logging presets to an access bundle's Credentials tab in Claude Tag admin settings
[Claude Tag] Added Datadog presets for the US3, AP1, AP2 and US1-FED sites; new Datadog connections are now limited to Datadog's read and query API routes
[Claude Tag] Fixed S3 uploads from recent AWS CLI and SDK versions failing with a 502 error when sent through an AWS connection

Original changelog［Claude Tag］Slack 討論串身分維持，頻道活躍度判斷與企業網格提示

[Claude Tag] Fixed Claude treating a channel as inactive, and skipping untagged messages there, while it was still posting in that channel from a routine or a thread
[Claude Tag] Fixed a thread's "Claude [task]" display name reverting to plain "Claude" after the session behind that thread was refreshed or restarted
[Claude Tag] Fixed the model you switched to in a Slack thread silently reverting to the channel's default after that thread's session was restarted or refreshed
[Claude Tag] Fixed Claude sometimes replying twice when another app or bot @mentioned it in a top-level channel message
[Claude Tag] Improved Claude's notices in Enterprise Grid channels shared across workspaces: they now say when no workspace is set up yet, or why only organization defaults apply

Original changelog［Code Review］多代理審查分析完整呈現，大量審查 PR 合併複審輕量化

[Code Review] Fixed reviews occasionally dropping part of their analysis when one of the reviewing agents returned its findings in an unexpected format
[Code Review] Fixed pull requests with more than 100 Claude reviews getting a full re-review on every clean merge from the base branch instead of the lighter merge-focused review