Claude Code 2.1.300

同時指派多位代理分頭處理工作，追蹤進度與回顧對話變得更加可靠。以前剛回覆完訊息的子代理在完成任務後會迅速從任務清單消失，讓你找不到入口回顧對話；現在會好好保留。

在檢視子代理時，已完成的父代理不再從頂端路徑脫落，階層關係清楚維持。快速輸入文字時若包含方向鍵單字，代理畫面以前會誤判為按鍵而移動預覽甚至清空草稿，這項問題已經修好。代理指令遇到個別請求資料不完整時，不會再卡住整個更新流程。技能中的代管代理工作階段（session）現在會完整保留，除非你主動要求封存或刪除。

背景跑著測試或編譯，最怕按退出時終端機卡死或殘留孤兒行程。

以前退出程式時如果還有背景指令在跑，工作階段結束掛鉤有時會漏跑；現在結束時若指令在收到停止通知後 1.5 秒內仍未關閉，系統會強制將其終止。停止常駐服務（daemon）時，不會再出現背景服務卡死卻顯示已停止的假象，若無法清除會回報錯誤代碼 1。關閉遙測時停止常駐服務，也不再默默無聲退出卻把工作階段留在清單中。執行背景指令時不再反覆查詢輸出檔或陷入休眠。背景工作階段接續工作流程時，改為透過工作流程工具的核可機制恢復執行。

大型專案常需開闢多個工作樹（worktree，Git 讓同一專案平行開多份副本的功能）處理不同功能，版本切換現在更加準確。

以前如果分支或標籤名稱長得像提交代碼，啟動工作樹可能會切到錯誤的版本，狀態中還會出現別的分支名稱；這個辨識問題已經修好。採用新型儲存機制的專案中，分支名稱顯示異常以及無法重新開啟工作樹的錯誤已獲修正。在無法取得遠端預設分支的裸儲存庫工作樹中，會改以當前簽出的版本作為備援基準。啟動工作樹時，外掛對應的建立與移除掛鉤現在會確實執行。

從網頁版或桌面版連回自己電腦操作，遠端控制的連線穩定度與檔案對照都獲得改善。

遇到短暫的登入憑證問題時，伺服器不會直接放棄更新連線，而是會自動重試並在需要時提醒你重新登入。當同一台電腦同時提供多個連線時，正在運作的工作階段不會再被誤標為已停止。網頁與桌面端的差異檢視現在能精確分辨被刪除與更名的檔案，不再全部標成修改，並能正確呈現衝突狀態。移除了過去完全不起作用的沙盒命令列參數，若需沙盒防護請在設定檔中將 sandbox.enabled 設為 true。

串接外部服務最怕網路短暫重連或格式錯誤，導致整場對話直接拋出中斷。開啟自動工具搜尋時，如果 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）伺服器在中途重新連線，這項搜尋功能先前會無故關閉，修復後會持續運作。

當外部工具規格被拒絕引發錯誤時，後續對話會改以不帶參數方式傳送該工具，避免每一輪對話都卡死在 400 錯誤。工作階段壓縮後才連線完成的伺服器，其工具與說明會立刻更新。恢復歷史對話時，也不會重複搜尋先前已經載入的延遲載入工具。

深度長對話或無頭程式執行最依賴定時摘要，避免上下文超出容量限制。在無頭模式（headless，沒有終端機視覺介面的執行模式）或 Agent SDK 中，若在回答達到字數上限或重試時觸發壓縮，重新連線可能會遺失近期待保留的訊息或摘要；這類情境已徹底修復。

程式異常中斷導致對話紀錄最後一行被截斷時，重新開啟後的第一條新紀錄不再丟失。含有特殊換行符號的對話不再導致分支指令遺失訊息或清理指令略過子專案。縮小自動壓縮視窗時，壓縮指令不會再誤發高額度用量警告。本地工作階段紀錄現在會完整記錄每次請求的系統提示詞與工具參數，方便精準重現。

雲端長任務最怕碰上伺服器重啟或政策讀取失敗，動輒卡住好幾分鐘。以前組織政策若無法讀取，雲端工作階段每輪會卡住 5 分鐘；現在會立刻結束該輪並清楚說明原因。

當工作階段自身的網路設定阻擋主機時，會誠實回報，不再誤稱是組織政策所封鎖。雲端容器被替換之後，曾恢復過一次的背景代理不會再被誤報為永久遺失，依舊能順利接續。修復了在長回覆期間重啟導致已回答訊息被重複執行第二次的問題。在雲端對話中送出第二則訊息後，斜線選單裡的指令說明不再集體消失。

在 Cowork 雲端環境指派任務，系統以前可能把本機未收到的檔案誤報為存檔成功。以前在雲端執行檔案寫入或編輯時，若指定本機路徑即使檔案根本沒傳到電腦也會顯示成功；這項假回報已修正。

網頁抓取工具現在遇到剛讀取失敗的網址時，會主動告知失敗原因，不再盲目重新抓取同一網頁。當組織停用永遠允許時，連接器工具的核准提示先前漏掉了允許本次任務選項，這項回退問題已經修好。在雲端掛鉤中核准請求時，掛鉤若嘗試更動模式會被直接略過，但其餘核准項目依然有效。

面對公司安全檢查與網關管制的企業環境，登入憑證與連線政策的控管現在更加嚴密。

以前登入若被強制指定組織的規則拒絕，被拒組織的授權令牌（token，用來證明身分的驗證憑證）仍會殘留；現在拒絕後會徹底清除。在透過作業系統憑證庫進行 TLS 檢驗的代理伺服器後方，新功能旗標與分階段推送能正常更新，不再卡在舊版狀態。託管環境的工作階段現在會忽略設定檔裡的自訂憑證與 TLS 變數，避免覆蓋公司規定。桌面版的 Code 分頁在指定網關時會套用託管設定，啟動時必須確保網關可連線。遇到桌面端無法使用的政策鍵時，網關指令會在啟動時直接指名該鍵並中止執行，不再僅發出警告。

各類終端機與編輯器內建面板的顯示細節，現在能看清暗淡提示且不再重複啟動檢查。在深色 ANSI 主題與 16 色終端機下，次要暗淡文字以前看起來和一般文字完全相同，難以辨識主次；現在已調好對比度。

在 VS Code、Cursor 或 Devin Desktop 的內建終端機啟動時，檢查擴充套件的指令以前會重複執行 2 次，現在精簡為只啟動 1 次。沙盒選單中的選項文字由原本容易誤解的名稱改為更直觀的指令名稱。修正了說明文件在介紹限制模式參數時，誤列出已移除工具的問題。

自己架設伺服器跑自動化任務，最怕日誌沒人讀取時整個執行器默默凍結。以前無人讀取輸出時，執行器與調度器會停止輪詢、健康檢查斷線且無法透過訊號關閉。

現在日誌會保留最新的 256 KiB 並記錄被捨棄的訊息數量，但 Windows 系統除外，讓服務始終保持運作。

在 VS Code 裡提問，現在能防止編輯器選取的文字被自動塞進輸入框。新增 claudeCode.attachSelection 設定項，設為 false 即可關閉自動夾帶。

在 Windows 下按住 Shift 拖曳檔案至輸入框時，不再插入難以辨認的磁碟機路徑，而是正確產生相對於工作目錄的參照。在 WSL、遠端 SSH 或容器環境中拖曳檔案也不再毫無反應。按向上鍵調出先前的斜線指令再按 Enter，不會再漏掉原本夾帶的參數。編輯曾標記的提示詞時，不會再被突然彈出的檔案選單遮擋。在 Windows 與 Linux 的檔案總管、其他樹狀檢視或偵錯主控台按 Ctrl+Alt+F，現在會開啟搜尋，不會再誤切換成專注檢視（Focus view）。

VS Code 裡切換不同模型或重新載入視窗，最怕額度用光後請求一直卡在重送死循環。快速模式在額度耗盡時以前會持續開啟導致請求不斷被拒重送，現在會附帶錯誤提示並自動關閉。

對於固定會進行深度思考的模型，介面不再顯示多餘的思考開關。對話字型大小變更現在會立即套用至差異檢視與計畫預覽，不再需要重新載入視窗，也不會被主題切換覆蓋。重新載入分頁或點擊繼續工作階段時，不會再遺失已移至工作樹的分支對話。系統在檢查其他視窗工作階段時改為直接呼叫工具，避免頻繁搜尋路徑產生額外行程。電腦在當機重開後，不會再因為行程識別碼被其他軟體重複使用而誤報對話仍在其他地方開啟。

在 Pull Request 中召喚 Claude 審查程式碼，非公開成員現在也能正常觸發審查流程。以前具有寫入權限的開發者如果將自己的 GitHub 組織成員身分設為非公開，在留言中輸入 @claude review 常會被忽視；現在系統能正確識別權限並啟動審查。

當儲存庫的審查功能處於暫停狀態時，留言不再引導管理員點進無關的設定首頁，而是會清楚指名有權限修復的人員，並直接附上 Git 提供者設定頁面的連結。

Assigning multiple agents to work in parallel makes tracking progress and reviewing conversations more reliable. Previously, subagents that had just replied would quickly vanish from the task list after completing their work, leaving you with no way to revisit the conversation; they are now properly preserved.

When inspecting subagents, completed parent agents no longer drop from the top breadcrumb path, keeping hierarchical relationships clear. When typing quickly, if text contained arrow-key names, the agent screen previously misinterpreted them as keypresses, shifting the preview or even clearing drafts; this issue has been resolved. Agent commands no longer stall the entire update flow when encountering individual requests with incomplete data. Managed agent sessions within skills are now fully preserved unless you explicitly request archiving or deletion.

Running tests or builds in the background makes you worry about frozen terminals or leftover orphaned processes when exiting.

Previously, if background commands were still running when exiting the program, session-end hooks were sometimes skipped. Now, if a command does not close within 1.5 seconds after receiving a stop notification upon exit, the system forcibly terminates it. Stopping a daemon no longer creates the illusion that a frozen background service has stopped, and it returns error code 1 if cleanup fails. Stopping daemons when telemetry is disabled also no longer exits silently while leaving sessions stranded on the list. Executing background commands no longer polls output files repeatedly or falls into sleep. When background sessions resume workflows, they now resume execution through the workflow tool approval mechanism.

Large projects often require creating multiple Git worktrees to handle different features, and switching versions is now far more accurate.

Previously, if a branch or tag name resembled a commit hash, launching a worktree could check out the wrong version and display another branch name in the status; this identification issue is now resolved. In projects adopting modern storage backends, bugs causing abnormal branch name display and failures when reopening worktrees have been fixed. In worktrees for bare repositories unable to fetch the remote default branch, the system falls back to using the currently checked-out version as the baseline. When launching a worktree, corresponding plugin creation and removal hooks now run reliably.

Connecting back to your computer from the web or desktop apps now offers better connection stability and file comparison for remote control.

When encountering brief credential issues, the server no longer abandons connection renewals immediately; it retries automatically and prompts you to log in again only when necessary. When a single computer hosts multiple connections simultaneously, active sessions are no longer mistakenly marked as stopped. Diff views across web and desktop clients now accurately distinguish between deleted and renamed files instead of labeling them all as modified, and conflict states display correctly. A previously ineffective sandbox command-line parameter has been removed; if you need sandbox protection, set sandbox.enabled to true in your configuration file.

Integrating external services poses risks when brief network reconnections or formatting errors interrupt an entire conversation. When auto tool search is enabled, if an MCP (Model Context Protocol) server reconnects halfway through, search previously shut down without warning; it now remains active after the fix.

When rejected external tool definitions trigger errors, subsequent turns transmit the tool without parameters, preventing every turn from getting stuck on 400 errors. For servers that finish connecting after session compaction, tools and descriptions update immediately. Resuming conversation history also avoids repeatedly searching for previously loaded deferred tools.

Deep, lengthy conversations and headless executions rely on scheduled summaries to prevent context from exceeding capacity limits. In headless mode or within the Agent SDK, if compaction triggered when answers hit length limits or during retries, reconnecting could drop recent pending messages or summaries; these scenarios have been thoroughly fixed.

When abnormal terminations truncate the final line of a conversation log, reopening no longer loses the very next log entry. Conversations containing unusual newline characters no longer cause branch commands to lose messages or cleanup commands to skip subprojects. Shrinking the auto-compaction window no longer causes compaction commands to emit false high-quota warnings. Local session logs now record system prompts and tool parameters for every request, allowing precise reproduction.

Long-running cloud tasks face risks when server reboots or policy fetch failures freeze progress for minutes at a time. Previously, if organization policies could not be retrieved, cloud sessions stalled for 5 minutes per turn; they now terminate the turn immediately and state the reason clearly.

When a session's own network configuration blocks the host, it reports this truthfully instead of misidentifying the cause as organization policy restrictions. After cloud containers are replaced, background agents that resumed once are no longer falsely reported as permanently lost, continuing execution normally. A bug where restarting during long responses caused completed messages to run a second time has also been resolved. Finally, command descriptions in the slash menu no longer vanish after sending a second message in a cloud conversation.

When delegating tasks in the Cowork cloud environment, the system previously could misreport unreceived local files as successfully saved. Writing or editing files in the cloud with a specified local path used to show success even if the file never reached the machine; this false confirmation is now fixed.

The web scraping tool now explicitly explains the failure reason when encountering a freshly failed URL, rather than blindly fetching the same page again. When organizations disable allow-always, connector tool approval prompts previously missed the allow-for-this-task option, a regression that has been resolved. In cloud hooks, when an approval attempt tries to alter the mode, that modification is skipped while keeping the remaining approved items intact.

For enterprise environments navigating corporate security inspections and gateway controls, login credential and connection policy management is now much tighter.

When a login was previously rejected by a rule enforcing specific organizations, authorization tokens for the rejected organization used to linger; they are now wiped clean upon rejection. Behind proxy servers that perform TLS inspection through the operating system credential store, feature flags and staged rollouts update normally instead of getting stuck on older versions. Sessions in managed environments now ignore custom certificates and TLS variables in configuration files, preventing corporate policies from being overridden. The desktop Code tab applies managed settings when a gateway is specified, requiring a reachable gateway at launch. When encountering policy keys unsupported on desktop, gateway commands now name the offending key and abort on startup rather than merely issuing a warning.

Display details across various terminals and integrated editor panels now render dim text clearly without duplicate startup checks. Under dark ANSI themes and 16-color terminals, secondary dim text previously appeared identical to standard text, making hierarchy hard to discern; contrast has now been adjusted.

When launching in the integrated terminal of VS Code, Cursor, or Devin Desktop, extension check commands previously ran 2 times, but are now streamlined to launch only 1 time. Option labels in the sandbox menu have been changed from misleading names to more intuitive command names. Documentation introducing restricted mode arguments was corrected after erroneously listing removed tools.

Running automated tasks on self-hosted servers risks silent runner freezes whenever logs go unread. Previously, when output remained unread, runners and dispatchers stopped polling, health checks disconnected, and processes could not be shut down via signals.

Logs now retain the latest 256 KiB and track the count of discarded messages, excluding Windows systems, keeping services continuously operational.

Asking questions in VS Code can now prevent editor-selected text from being automatically inserted into the input box. The new claudeCode.attachSelection setting disables automatic attachment when set to false.

Holding Shift while dragging files into the input box on Windows no longer inserts unreadable drive paths, correctly generating references relative to the working directory instead. Dragging files across WSL, remote SSH, or container environments is no longer unresponsive. Pressing the up arrow to recall previous slash commands and hitting Enter no longer drops original arguments. Editing previously bookmarked prompts is no longer obstructed by unexpected file menu popups. Pressing Ctrl+Alt+F across File Explorer, other tree views, or the debug console on Windows and Linux now opens search instead of erroneously toggling Focus view.

Switching models or reloading windows in VS Code risks requests getting trapped in retry loops once quotas run out. Fast mode previously stayed on when quota was exhausted, causing requests to be repeatedly rejected and retried, but now displays an error prompt and turns off automatically.

For models that always perform deep thinking, the interface no longer displays a redundant thinking toggle. Conversation font size changes now apply immediately to diff views and plan previews without requiring a window reload, and are no longer overridden by theme changes. Reloading tabs or clicking to resume a session no longer loses branch conversations moved to worktrees. The system now calls tools directly when checking sessions in other windows, avoiding redundant processes from frequent path searches. After a crash and reboot, the system no longer falsely reports conversations as open elsewhere when process IDs are reused by other software.

Summoning Claude to review code in pull requests now lets private members trigger the review workflow smoothly. Developers with write access whose GitHub organization membership was kept private previously found @claude review comments frequently ignored; the system now recognizes permissions accurately and launches reviews.

When repository review functionality is paused, comments no longer direct administrators to irrelevant settings landing pages, clearly naming authorized personnel instead and linking directly to the Git provider settings page.