Claude Code 2.1.247

在選單中快速切換選項時，系統不再因為按鍵反應落後而選錯上一行。過去在歷史搜尋、/config、/mcp、/skills、背景任務列表以及 /model 介面中，如果以極快速度連按方向鍵與 Enter 鍵，游標會誤觸剛才游標上方的那一列；現在按鍵時序已修正，確保選定的就是眼前停留的項目。

當子代理在首次呼叫模型時遭遇 404 找不到模型的錯誤，系統不再讓它直接死當。

現在子代理會依序使用當前工作階段設定的備援模型鏈繼續嘗試，而且回傳給父代理的錯誤訊息中，會完整包含錯誤類型、狀態碼、請求 ID（request id）以及目標模型名稱，方便排查問題。

掛鉤（hook，特定事件發生時被觸發的小程式）或背景代理若噴出數 MB 的海量錯誤訊息，過去會直接塞爆對話紀錄，導致工作階段卡死在「Prompt is too long」而無法繼續輸入。現在系統已修正溢位處理機制，防止過量的錯誤輸出癱瘓對話。

在支援 kitty 協定的終端機中，使用非拉丁鍵盤配置（例如西里爾字母）的使用者，現在按下 Ctrl 快捷鍵組合也能正常觸發功能。過去在切換為非拉丁輸入法時，Ctrl 按鍵訊號無法被正確辨識；修復後無論目前的輸入語言為何，組合鍵都能精確送達。

滑鼠回報資料若在跳脫字元前綴剛送出時被切成兩段讀取，過去會導致像 <35;150;7M 這樣的殘留字串被當成文字硬生生塞進提示輸入框裡。

現在系統修正了終端機輸入串流的解析邏輯，即便滑鼠封包被拆成多次讀取，也能完整辨識並過濾掉控制碼。

使用 nix/home-manager 或 GNU stow 等工具管理設定檔的使用者，不再需要擔心個人設定檔被意外刪除。過去當 Bash 沙盒在指令執行完畢進行清理時，若指向外部的符號連結（symlink，指向另一個位置的捷徑檔）~/.claude/settings.json 被重導至沙盒可寫入區域之外，清理程序會將該連結檔案直接刪除；現在沙盒已修正保護邏輯。

在 Zed 編輯器設定終端機快捷鍵時，執行 /terminal-setup 指令不再粗暴地洗掉你既有的全部按鍵綁定。

過去該指令會直接覆寫整個 keymap.json 檔案；現在它懂得將所需的快捷鍵合併進現有設定中，完整保留你原先自訂的其他按鍵組合。

使用 /rename 指令重新命名當前對話時，若內部工作階段註冊表（session registry）因故無法同步更新，系統不再假裝一切成功並默默確認。現在它會明確顯示提示訊息，告知其他工作階段可能暫時仍會顯示舊名稱，讓資訊維持透明。

以 --agent 參數啟動自訂代理工作階段時，執行 /compact 壓縮指令或點選「Summarize from here」整理脈絡時，系統過去會錯誤套用預設的系統提示詞來產生摘要。現在摘要程序已修正為嚴格依循該對話專屬的自訂系統提示詞，確保代理的角色定位與特殊規則在濃縮後依然完整維持。

在 claude agents 列表中檢視背景工作階段時，若終端機宿主處理程序意外終止，該列過去會永遠卡在「opening…」狀態。

現在系統會在數秒內將該列標記為失敗並清楚顯示具體原因，你只需按下 Enter 鍵就能直接重新啟動該工作階段。

當掛鉤（hook，特定事件發生時被觸發的小程式）或背景任務的輸出檔案因權限等原因無法寫入磁碟時，系統過去會在記憶體中無限累積暫存資料導致記憶體暴增。現在記憶體限制已獲得控制，且日後輸出檔案恢復可寫時，檔案內會明確加註哪一段輸出內容已遺失。

在遠端 SSH 連線環境下執行 /install-github-app 指令時，登入體驗變得更加可靠。複製快捷鍵現在會如實說明登入網址是如何複製的，不再盲目宣稱複製成功；當遠端伺服器無法自動開啟本機瀏覽器時，登入網址也會第一時間直接印在畫面上供你手動點擊。

由前景轉移至背景工作階段執行的 Shell 指令，在執行完畢時不再跳出內部錯誤紀錄或誤導人的 [exited with code -1] 退出碼。

現在指令在背景順利收尾時能正確回報真實狀態，避免造成執行失敗的假象。

當你在不同範圍重複安裝同一個未標註版號的市集外掛時，系統過去會直接刪除並重建既有的快取目錄。這往往會干擾到正在使用該外掛的工作階段；現在快取（把算過或查過的結果暫存起來直接用）處理邏輯已獲得修正，第二個範圍的安裝動作不會再中斷運作中的程式。

透過 /remote-control 指令開啟遠端控制工作階段時，已連線的用戶端現在能正確收到目前工作樹的修改差異（working-tree diff）。

過去這些變更資訊未被同步傳遞給連線端；修正後遠端操作者能即時檢視檔案的具體改動，大幅提升遠端協作的透明度。

自託管執行器（self-hosted runner）的工作階段過去在 Claude Code 尚未正式啟動前，就會提早回報為「running」運作中狀態，導致 Claude 桌面應用程式過早發出「Claude is waiting for your input」的提示通知。現在狀態回報時機已修正，只有在主程式真正啟動並就緒後才會發送通知。

在由管理員統一派發的受管設定（managed settings）中，若指定透過 Claude 應用程式閘道登入，而環境對外無法連線至 Anthropic 預設端點，初次啟動設定過去會直接報錯並退出。

現在系統已修正連線判斷邏輯，優先依循閘道設定順利完成初次設定，保障封閉網路環境的可用性。

在網頁版、桌面版或行動應用程式中使用雲端工作階段時，若在剛送出訊息的瞬間立即切換權限模式，介面過去有時會因為非同步傳輸的順序衝突，而跳回先前顯示的舊模式。現在狀態同步順序已獲得修復，無論你送出訊息後多快切換權限，模式切換都能精確反映當下的最新選擇，不再遺失設定。

當雲端工作階段在回合之間發生容器重啟，而背後的代理程式、Shell 指令或監控程序仍在執行時，系統過去在重啟後會陷入無聲無息的沉默狀態，讓你誤以為任務仍在背景順暢運作。現在工作階段在恢復連線後，會主動向你回報受影響的中斷工作項目與遺失進度，確保執行狀況一目了然。

Rapidly navigating interactive menus no longer triggers the wrong item due to timing latency. Previously, typing a rapid sequence of arrow keys and Enter in history search, /config, /mcp, /skills, background tasks, or /model could accidentally act on the row directly above your target; input event processing is now fixed to reliably target your current selection.

Sub-agents no longer crash outright when their initial API call encounters a 404 model not found error.

Instead, they now transition smoothly through the session's fallback model chain, and if the operation still fails, the error sent back to the parent agent includes the error type, status, request id, and model name.

Hooks or background agents that dump megabytes of error output can no longer flood the conversation history and wedge the session on "Prompt is too long". The system now prevents excessive diagnostic output from overwhelming the context window, keeping your terminal responsive.

Using Ctrl keyboard shortcuts in terminals supporting the kitty keyboard protocol now works reliably under non-Latin keyboard layouts such as Cyrillic. Previously, switching input layouts prevented modifier key combinations from firing; keyboard event translation is now fixed across all keyboard configurations.

Mouse tracking reports that arrive split across stream reads directly after an escape prefix will no longer dump stray text fragments like <35;150;7M straight into your prompt input.

Input stream buffering now properly reconstructs split control sequences.

Users managing dotfiles with tools like nix/home-manager or GNU stow no longer risk having their configuration link erased. Previously, when the Bash sandbox cleaned up after running a command, if the ~/.claude/settings.json symlink was repointed outside the sandbox's writable area, the cleanup routine deleted the symlink; the sandbox now safely preserves these links.

Running /terminal-setup to configure keybindings for the Zed editor no longer wipes out your entire keymap configuration.

Instead of completely overwriting keymap.json, the command now cleanly merges its required keybindings into your existing file, preserving all your personal custom shortcuts.

When running /rename to rename your session, the interface no longer gives a silent confirmation if the session registry cannot be updated. It now explicitly alerts you that other sessions might still display the old name, ensuring you are aware of the registry state.

Sessions launched with --agent now maintain their custom system prompt during context compaction. Previously, running /compact or using "Summarize from here" caused the summary to be generated under the generic default system prompt; it now strictly preserves the conversation's own prompt instructions.

In the claude agents list, background sessions whose terminal host process died will no longer remain stuck displaying "opening…" forever.

The row now marks itself as failed within seconds along with the failure reason, and pressing Enter immediately restarts the session.

If a hook or background task cannot write to its output file, the system no longer buffers the data indefinitely in memory, preventing unbounded memory growth. Additionally, the log file now records an explicit note indicating where output was lost.

Running /install-github-app over an SSH connection now handles clipboard and browser limitations gracefully. The copy shortcut explicitly reports how the sign-in URL was copied rather than falsely claiming universal success, and if no browser can be opened remotely, the URL prints immediately to the screen.

Shell commands handed off from the foreground to background sessions no longer log internal errors or display a misleading [exited with code -1] banner upon completion.

Background task status is now reported cleanly upon normal termination.

Installing a version-less marketplace plugin across a second scope no longer purges and recreates its active cache directory. This eliminates the risk of disrupting existing, active sessions that are currently executing the plugin.

Remote Control sessions launched via /remote-control now reliably broadcast the working-tree diff to connected clients.

Previously, attached clients could not see pending file modifications; this sync issue is resolved so remote collaborators see current changes in real time.

Self-hosted runner sessions no longer report a "running" status before Claude Code has actually booted up. This eliminates premature "Claude is waiting for your input" notifications fired by the Claude desktop app while initialization is still underway.

First-run setup no longer abruptly aborts with "Unable to connect to Anthropic services" when managed settings configure gateway sign-in for Claude apps while standard Anthropic endpoints are blocked.

Initial configuration proceeds smoothly over the designated gateway channel.

In cloud sessions across web, desktop, and mobile apps, rapidly toggling your permission mode right after dispatching a message no longer causes the UI to flicker back to the prior mode. State synchronization now ensures your updated mode remains active regardless of network transmission timing.

If a cloud session's container restarts between turns while background agents, shells, or monitors are active, the resumed session no longer silently drops into limbo. It now explicitly reports any lost work and background processes that were interrupted.