Claude Code 2.1.296

當你指派子代理處理大量繁重的研究分支時，背景助理的對話紀錄常常比主線更快塞滿記憶空間。

現在你在子代理的 frontmatter 或 --agents 參數定義中，可以加入 autoCompactWindow 設定，讓子代理比主對話更早觸發自動壓縮。同時，當工作階段關閉導致工具呼叫中斷時，重新啟動的子代理不會再誤以為是使用者按下了拒絕。如果你在自訂代理檔案中打錯了標頭欄位名稱，開啟 --debug 時主控台也會直接標示無法識別的欄位，並附上可能拼錯的文字提示。

企業 IT 人員統一部署安全連線時，能透過中央政策同時控管命令列與桌面端工具。Claude apps gateway 受控政策新增 code 鍵值，套用與 cli 相同設定並在桌面端 Code 標籤頁生效，搭配 desktop 還能啟用桌面端閘道模式。若閘道許可提供者包含 "gateway"，先前在設定指定該閘道的筆電被鎖在門外的問題已修復。受控設定強制以 gateway 登入卻未指定 forceLoginGatewayUrl 時，先前登入狀態被忽略的問題也已修正。針對 Haiku 5.5 等自適應思考模型，部分閘道計算 token（AI 處理與計費的文字單位）失敗或誤計為預算思考的情況已修復。桌面端工作階段若無法啟動，原因現在會直接以回覆顯示。

當你依賴外部工具擴充功能時，工具本身的參數說明現在可以寫得更詳盡。系統一開始傳遞給模型的前置 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）工具說明與伺服器指引，預設長度上限已從原本的 2,048 個字元調高至 4,096 個字元。在無頭模式（沒有終端機視窗的執行模式）的工作階段中，切換目錄或重新載入外掛後，以前會錯誤啟動該資料夾原本已關閉的 .mcp.json 或外掛 MCP 伺服器，現在已修正。在環境變數啟用中斷回溯時，如果中斷前是由 MCP 工具回傳結果結束輪次，重新啟動後不會再錯誤重跑已完成的輪次。當 MCP 伺服器需要驗證時，雲端工作階段也不會再被誤判為非互動式而跳出無效的本機設定建議。

讓 Claude 自主執行一連串任務時，操作介面對工具執行狀態的呈現現在更友善。在自動模式下，如果某次工具呼叫因為檢查程序沒有可用答案而未被執行，畫面不再直接亮起鮮紅色的錯誤訊息，而是改以灰暗的「Not run」列呈現。而在雲端工作階段中，若功能旗標尚未載入完成，即使已儲存允許權限，自動模式有時仍會跳過對 Chrome 瀏覽器操作的檢查，這項漏洞現已補正。而編輯自己產出的工件時，偶爾會被系統誤指分享狀態改變而遭到拒絕的問題，也已徹底修復。

當你將除錯日誌分享給團隊或執行自動修改程式碼時，系統對敏感資訊與特殊編碼檔案的保護更加周全。

在分享的對話紀錄與除錯日誌中，機密遮蔽功能過去若遇到沒有數值的鍵名，接在其後的值或寫在 Shell 字串內的 JSON 可能會漏未隱藏，現在都會遮蔽。安全防護上，若 Bash 指令對 BASH_ARGV0 變數賦值隨後呼叫，以前權限檢查可能自動放行，這類指令現在會跳出提示請求使用者確認。而 Edit 與 NotebookEdit 工具遇到 Windows-1252、Shift-JIS 或 GBK 等非有效 UTF-8 的檔案時，以前會把非 ASCII 字元全數置換，現在會直接拒絕編輯，避免毀損原始內容。

在包含大量程式碼的長篇對話中按快捷鍵切換紀錄時，視窗畫面的反應速度現在變得更快。透過快取（把算過的結果暫存起來直接用）語法醒目標示的程式碼區塊，按 Ctrl+O 切換對話紀錄時不再感到延遲。本地垃圾回收機制遇到大型紀錄檔時，若重寫所能釋放的空間低於一成，將不再耗費資源重新寫入檔案。另外，先前當文字內容包含 Unicode 行或段落分隔符號 U+2028 或 U+2029 時，本地垃圾清理會遺失該則訊息，而執行 claude purge 也會殘留該提示詞，這兩個問題皆已修復。最後，使用 --teleport 時若無法讀取工作階段紀錄，也不再出現開啟空白對話的異常現象。

當你使用方向鍵將當前對話推到背景執行時，背景服務的處理行為現在更加可預測。以前按下左方向鍵 ← 將工作階段切換至背景時，剛啟動的輪次或以 ! 開頭的 Shell 指令會在背後默默執行完畢，現在則會立即停止執行。針對輸入時機的異常也做了嚴格防護：若在按 ← 後立刻送出提示詞且背景服務回應緩慢，過去可能在前景與背景各跑一次，現在不再重複送出。而在啟動階段的 hook（掛鉤，特定事件發生時被觸發的小程式）仍在執行時按 ← 導致輸入內容消失的問題，現在也能透過上方向鍵 ↑ 找回。在介面顯示方面，全螢幕模式下游標懸停的連結會加上底線提示。開啟 /diff 面板時不會再卡住通知訊息。在 MATE 等終端機中複製文字時印出 52;c;… 亂碼跳脫字元的瑕疵也已修正。

當你從雲端同步外掛或跨作業系統共用設定時，外掛安裝流程現在更加穩健。

過去若市集名稱命名為 constructor，執行市集新增、更新或外掛安裝時會引發內部錯誤，現在新增指令會明確拒絕該名稱。同時，名稱為 constructor 或 prototype 的外掛金鑰，也不會再因為儲存選項而被誤刪。在多個工作階段共用設定目錄的情境下，從 claude.ai 同步的技能過去會在每次同步時重新下載、始終裝不完，此問題已獲得解決。而且宣告了市集相依性的外掛在兩者皆從雲端同步時不會再被誤停用。從 Linux 複製外掛路徑至 macOS 時因指向 /home 導致啟動卡住的狀況也已修正。

企業管理者設定安全規則防範危險工具時，攔截工具後的流程現在能正常收尾。

在受控設定中，若 PreToolUse 掛鉤以 "continue": false 拒絕工具，或受控 prompt 掛鉤攔截請求，回絕後卻未結束該輪對話的問題已修正。部分工作階段中 PostToolUse 掛鉤未套用 updatedMCPToolOutput 的問題也已解決。提示詞送出掛鉤執行時按 Esc 或中斷，不再導致無頭模式工作階段結束、清空輸入或放行未檢查提示詞。只要 CLAUDE_ENV_FILE 裡只有單純的變數賦值，PowerShell 指令現在就能讀到掛鉤寫入的環境變數。

當你撰寫自訂模組擴充 Claude 的內建能力時，模組提供的輔助方法現在能精準對應呼叫者的環境。外掛所提供的 $ 方法先前會在主工作階段的工作目錄下運作，而非呼叫該方法的代理所在目錄，並且會忽略呼叫掛鉤所持有的輪次，現在已修正為在正確的代理目錄下執行並尊重輪次。在模組管理方面，若模組已經重新載入或遭到移除，但其殘留的掛鉤仍在執行，呼叫 $.agent.register 過去仍會成功註冊，現在系統會明確拒絕該呼叫。同時，模組內的網路請求工具 $.http.fetch 過去在發送 HEAD 請求時，若伺服器回傳的標頭宣告內容長度超過 4 MiB 上限，即便該請求根本不下載主體內容也會被回絕，這項不合理的限制如今已正式排除。

閱讀自動審查意見或請 Claude 繪製統計圖表時，產出的視覺呈現現在更賞心悅目。

在雲端工作階段、Agent SDK 以及 IDE 整合環境中執行 /code-review 指令時，審查結果以前會停留在未格式化的原始 JSON 陣列，現在會整齊排版成帶有編號的重點清單，方便逐條檢視修訂。同時，內建隨附的資料視覺化（dataviz）技能也進行了介面調優：淺色模式下第七組數列採用更輕盈的淺紫色、深色模式下主要文字顏色變得更柔和易讀，垂直 Y 軸標籤也改採如 1K 這種緊湊簡潔的縮寫格式。

在 Windows 環境下執行終端機操作或安裝工具時，多項權限與程序管理的瑕疵獲得改善。過去在 PowerShell 中執行長度超過約 1 KB 的指令時，系統總是一律彈出確認視窗；現在已放寬為長度達 32 KB 以內的指令都能正常套用允許規則與唯讀偵測，減少頻繁確認的干擾。在 MCP 伺服器管理方面，以前透過標準輸入輸出運作的伺服器在關閉時會被強制中斷，現在會先關閉標準輸入（stdin），若超過 300 毫秒仍未退出才會終止整個行程樹。安裝外掛方面，未設定 GitHub SSH 金鑰的電腦執行 claude plugin install 複製儲存庫失敗時，會自動改用 HTTPS 重試。同時，修正了在略過權限模式下於 Git Bash 執行刪除使用者家目錄指令卻未跳出確認提示的危險問題。

在 VS Code 編輯器中使用 Claude 側邊欄時，瀏覽器動作確認與介面設定現在更加嚴謹。透過 Claude in Chrome 執行網頁操作時，如同命令列終端機的規範，現在在所有工作階段（包含使用 @browser 連線時）執行動作前都會主動詢問使用者。只要在該工作階段內允許特定網站，後續就不會重複詢問。設定檔方面，擴充套件設定中補上了 claudeCode.spinnerVerbs 鍵值，讓 settings.json 能提供自動補全與格式驗證，格式錯誤的數值也不會再造成聊天面板當機。在 Windows 上點擊對話中完整路徑連結無法開啟檔案的問題已獲修復。當系統或設定啟用減少動態效果時，進度指示動畫會依約定停止播放。而螢幕關閉或遠端視窗重新連線後，面板延遲收到訊息，會讓快取時鐘（判斷提示詞快取是否仍有效的計時）誤判快取仍有效，這也已修正。

在 Slack 頻道中與 Claude 連續對話時，討論串的版面閱讀體驗變得更加清爽。現在 Claude 僅會在對話串中最新的一則回覆保留帶有工作階段連結、模型名稱與費用資訊的頁尾，較早的回覆會自動撤下頁尾，避免畫面充斥重複資訊。此版本亦修復了多項格式細節。網址內包含括號時不會再破壞超連結或開啟錯誤網頁。文字中出現如六位數 PR 編號等類似十六進位色碼的字串時，不再於旁邊誤畫色塊。外掛確認卡片會正確顯示外掛名稱而非內部 ID。頻道未包含的外掛被提及時，會改為說明外掛來源而非誤貼移除卡片。而在自託管環境中任務中途更換執行器導致回覆遺失的問題也已修正。

送出 Pull Request 等候程式碼審查時，上面的 Code Review 檢查標記現在能即時反映審查完畢的真實狀態。

過去在審查程序完成後，Pull Request 上的 Code Review 檢查狀態偶爾會停滯在「進行中」而未更新，現在已修復為在完成時立刻顯示審查的具體判定結果。同時，管理者在調整 Code Review 的後台設定時，若因為組織政策限制等原因而無法儲存，介面不再只丟出含糊的失敗訊息，而是會明確說明無法儲存的具體原因。

When you dispatch subagents to handle heavy research branches, the background assistant's conversation history often fills up memory space faster than the main conversation.

You can now add an autoCompactWindow setting in a subagent's frontmatter or --agents flag definition, triggering automatic compaction for the subagent earlier than in the primary session. Additionally, when a tool call is interrupted by closing the session, restarting the subagent will no longer mistake the interruption for a user rejection. If you mistype a header field name in a custom agent file, the console will now directly flag unrecognized fields when running with --debug, complete with suggestions for potential typos.

When enterprise IT administrators deploy secure connections centrally, managed policies allow them to govern command-line and desktop tools simultaneously. The Claude apps gateway managed policy adds a code key that applies the same settings as cli to take effect in the desktop Code tab, and pairing it with desktop enables desktop gateway mode. When permitted gateway providers include "gateway", an issue where laptops specifying that gateway in settings were locked out has been fixed. When managed configurations mandate logging in via gateway without defining forceLoginGatewayUrl, an issue where login status was ignored has also been resolved. For adaptive thinking models such as Haiku 5.5, cases where certain gateways failed to calculate tokens (units of text used by AI for processing and billing) or miscounted them as budgeted thinking have been fixed. If a desktop session fails to launch, the exact reason is now directly shown as a response.

When you rely on external tools for feature extensions, parameter descriptions for tools can now be written with much greater detail. The default maximum length for initial MCP (Model Context Protocol, a standard connecting AI to external tools) tool descriptions and server instructions passed to the model has been increased from 2,048 characters to 4,096 characters. In headless sessions, an issue where switching directories or reloading plugins erroneously started disabled .mcp.json or plugin MCP servers from that directory has been resolved. When interrupt rewind is enabled via environment variable, restarting will no longer mistakenly rerun completed turns if the turn previously ended with an MCP tool result. When an MCP server requires authentication, cloud sessions will no longer be misidentified as non-interactive and trigger invalid local setup suggestions.

When letting Claude autonomously execute sequences of tasks, the interface presentation of tool execution states is now much clearer. Under auto mode, if a tool call remains unexecuted because checks have no available answer, the screen no longer flashes a bright red error message, displaying a muted "Not run" row instead. In cloud sessions, auto mode previously bypassed checks on Chrome browser actions when feature flags had not finished loading even after permissions were saved, a loophole that is now plugged. Additionally, an issue where editing your own generated artifacts occasionally met with rejections due to the system mistakenly claiming a change in share state has been completely resolved.

When you share debug logs with teammates or run automated code edits, protections for sensitive information and non-standard encoded files are now more comprehensive.

In shared conversation histories and debug logs, secret redaction previously missed values following empty keys or JSON embedded in shell strings, all of which are now masked. For security enforcement, Bash commands that assign to BASH_ARGV0 right before invocation previously bypassed permission checks automatically, but now prompt the user for confirmation. Meanwhile, when the Edit and NotebookEdit tools encounter files with invalid UTF-8 encodings such as Windows-1252, Shift-JIS, or GBK, they now outright reject edits rather than replacing non-ASCII characters and corrupting original contents.

When using shortcut keys to switch conversation histories in long sessions packed with code snippets, window rendering is now significantly faster. By caching syntax-highlighted code blocks, pressing Ctrl+O to cycle through history no longer feels laggy. When inspecting large log files, local garbage collection will skip rewrites if the reclaimable space is less than 10%, avoiding unnecessary disk writes. Furthermore, an issue where local garbage collection dropped messages containing Unicode line or paragraph separators U+2028 or U+2029, as well as an issue where running claude purge left such prompts behind, have both been resolved. Finally, sessions failing to load history when using --teleport will no longer anomalously open blank conversations.

When using arrow keys to push an active conversation to the background, background handling behavior is now far more predictable. Previously, pressing the left arrow ← to background a session allowed newly started turns or shell commands prefixed with ! to continue running silently, but they are now aborted immediately. Strict safeguards are also in place for input timing: submitting a prompt immediately after pressing ← while the background service is slow to respond no longer duplicates runs across foreground and background. Input dropped after pressing ← while startup hooks are still executing can now be recovered using the up arrow ↑. On the visual side, hovered links in full-screen mode now display an underline. Opening the /diff panel no longer blocks notifications. An issue causing garbage escape sequences like 52;c;… when copying text in terminals such as MATE has also been fixed.

When synchronizing plugins from the cloud or sharing settings across operating systems, the installation pipeline is now significantly more reliable.

In the past, naming a marketplace constructor triggered internal errors during marketplace addition, updates, or plugin installation; the add command now explicitly disallows this name. At the same time, plugin keys named constructor or prototype will no longer be inadvertently wiped out when saving settings. In scenarios where multiple sessions share a settings directory, skills synced from claude.ai no longer get downloaded again at every sync without ever finishing their install.ai repeatedly downloaded and never finished installing has been resolved. Plugins declaring marketplace dependencies are also no longer mistakenly disabled when both items are synced from the cloud. A startup freeze caused by copying plugin paths pointing to /home from Linux to macOS has also been fixed.

When enterprise administrators set security rules to guard against dangerous tools, the workflow following tool interception now finishes cleanly.

In managed environments, an issue where the turn did not terminate after a PreToolUse hook denied a tool with "continue": false or after a managed prompt hook blocked a request has been fixed. An issue where PostToolUse hooks failed to apply updatedMCPToolOutput in certain sessions has also been resolved. Pressing Esc or interrupting while prompt submission hooks are running no longer causes headless sessions to exit, clears the input, or permits unreviewed prompts. Furthermore, as long as CLAUDE_ENV_FILE contains only basic variable assignments, PowerShell commands can now read environment variables written by hooks.

When you write custom modules to extend Claude's built-in features, helper methods provided by modules now target the caller environment accurately. The $ method provided by plugins previously executed in the working directory of the main session rather than that of the agent calling it, while ignoring the turn held by the calling hook; it now runs in the correct agent directory and respects the turn. For module lifecycle management, calling $.agent.register previously succeeded even if the module was reloaded or removed while its lingering hooks remained running, but the system now explicitly rejects such calls. Meanwhile, the network request utility $.http.fetch previously rejected HEAD requests whenever server headers declared a content length over the 4 MiB limit even though no body content was downloaded, and this restriction has now been removed.

When reviewing automated feedback or asking Claude to chart statistics, visual output is now noticeably cleaner.

In cloud sessions, the Agent SDK, and IDE integrations, running /code-review previously rendered an unformatted raw JSON array, but it now neatly structures results into a numbered list for straightforward line-by-line inspection. At the same time, the bundled data visualization (dataviz) skill received interface adjustments: light mode adopts a softer light purple for the seventh series, dark mode uses a gentler, more legible tone for primary text, and vertical Y-axis labels now use compact abbreviations such as 1K.

When running terminal tasks or installing tools on Windows, several permission and process handling quirks have been smoothed out. PowerShell previously prompted for confirmation on every command longer than roughly 1 KB; it now applies allow rules and read-only detection to commands up to 32 KB, cutting down on frequent interruption prompts. In MCP server lifecycle management, servers running over standard input and output were previously killed outright upon shutdown, but the system now closes standard input (stdin) first and waits 300 ms before terminating the full process tree. For plugin installations, running claude plugin install on machines without configured GitHub SSH keys now falls back to HTTPS automatically when repository cloning fails. In addition, a dangerous issue where running commands in Git Bash under bypass permissions to delete the user home directory failed to trigger a confirmation prompt has been fixed.

When using the Claude sidebar inside VS Code, browser action checks and interface preferences are now handled with greater rigor. Web operations executed via Claude in Chrome follow command-line terminal conventions by prompting the user before actions across all sessions, including connections via @browser. Once a site is approved within a session, repeated prompts for it are suppressed. In configuration handling, extension settings now include the claudeCode.spinnerVerbs key to provide autocomplete and schema validation in settings.json, preventing malformed values from crashing the chat panel. An issue where clicking full file path links inside chat on Windows failed to open files has been resolved. Progress spinner animations honor system or user settings to reduce motion by stopping playback. Furthermore, delayed message delivery after the display turned off or a remote window reconnected no longer makes the prompt cache clock wrongly read as warm (still valid).

When holding continuous conversations with Claude in Slack channels, thread readability is now significantly cleaner. Claude now retains the footer containing session links, model names, and cost details only on the most recent reply in a thread, while earlier replies automatically drop their footers to prevent duplicate clutter. This release also resolves multiple formatting details. URLs containing parentheses will no longer break hyperlinks or open incorrect web pages. Strings that resemble hex color codes, such as six-digit PR numbers, will no longer trigger accidental color swatches next to them. Plugin confirmation cards now properly display plugin names instead of internal IDs. When plugins not included in the channel are mentioned, the system explains the plugin source instead of mistakenly posting removal cards. Additionally, an issue in self-hosted environments where replies were lost when runners were swapped mid-task has been resolved.

When submitting a pull request for review, the Code Review check badge now reflects the real status as soon as the review finishes.

In the past, after a review finished, the Code Review check status on the pull request occasionally lingered on "in progress" without updating; it now immediately displays the specific determination result upon completion. Meanwhile, when administrators adjust Code Review configuration in settings and cannot save due to organization policy restrictions, the interface no longer outputs a vague failure message, but explicitly explains the specific reason why saving failed.