Claude Code 2.1.293

在開啟多個背景工作階段時，代理清單與底欄資訊現在能正確反映現況。以前系統短暫無法讀取工作階段資料夾（例如開啟太多檔案）時，底欄代理數量會消失，清單也會被佔位列洗掉。現在遇到讀取異常不會再弄亂介面。命名為 worker 的自訂代理在啟動時不再被誤標為 Agent，結束後詳細對話框標題也會保留原本類型。claude agents 以前提供跳過權限檢查的選項，但同意若只存在 .claude/settings.local.json 或 --settings 檔案中，背景工作階段會忽略該選項；現在會先詢問同意，仍忽略該選項的工作階段會持續顯示簡短提示。

在 Claude 運作途中按左箭頭 ← 鍵將工作階段切到背景時，未送出的文字與等待回答的狀態不會再被意外打斷。以前只要輸入框留有未送出的字，或畫面正在等待你回答問題，按了 ← 在 10 秒後仍會切入背景。現在遇到這些情況會自動取消背景切換。當排隊中的訊息無法轉移至背景時，系統會停在原地並提示原因，不再遺失訊息。剛按 ← 切換時若遇到權限確認，按 Esc，或選擇 No 且不附回饋，也能如期停止該輪對話。

開發自訂模組或外掛時，工具註冊與測試流程更可靠。$.tool.register 支援 isDeferred，設為 false 會直接在提示詞列出工具結構，不必透過工具搜尋。外掛掛鉤（特定事件發生時被觸發的小程式）工作程序重啟時，不再漏掉模組的 classic.* 事件掛鉤，避免只剩設定掛鉤單獨回應。呼叫 $.session.append 的模組曾讓 claude plugin test 失敗，現在測試能透過 mock.session 讀回追加紀錄。裝有 Docker Desktop 的 Mac 上，~/.docker/bin 下的符號連結（指向另一位置的捷徑檔）曾讓 claude plugin eval 拒絕所有授予 Bash 權限的執行。這項錯誤已修好，拒絕訊息現在會指出憑證庫（存登入證明的地方）哪個部分持有該連結。

透過遠端操控或雲端環境進行對話時，長篇回覆的顯示與背景狀態更正常。以前在長對話中文字偶爾會整塊卡住才突然跳出，現在能順暢逐字顯示。憑證復原之後，系統不會再重複上傳工作階段一開始的歷史紀錄。透過 claude remote-control 指令開啟時，推播通知也不會再誤報遠端操控未啟用。

使用程式碼審查或 /ultrareview 上傳專案時，檔案檢查與錯誤指引更清晰。以前在 Linux 上，沙盒指令執行中可能因設定檔「無法解析」而誤拒部分儲存庫，例如儲存庫放在另一份已取出的專案副本裡。現在這項讀取問題已經修正。遇到分割索引檔案遭拒時，系統不再給出可能讓 Git 讀不到索引的錯誤修復建議。在 Code Review 的新增儲存庫對話框中，現在會逐一列出無法加入的儲存庫，並寫明各自的原因（例如缺少 GitHub 寫入權限）。

剛在網頁版修改了自訂技能的說明文字時，終端機裡的 Claude 曾有讀不到變更的問題。編輯同步自 claude.ai 的技能描述後，不再有時要開新對話或輸入 /clear，模型才讀得到修改後的說明。當沒有工作階段在使用時，背景同步檢查技能變更的間隔從每 10 分鐘調整為約每 40 分鐘一次。

在終端機貼上文字、設定快速鍵或挑選模型參數時，各項輸入判定更精準。開頭與結尾單字相同的貼上文字，曾被當成手動輸入。貼上技能名稱後，緊接著輸入或貼上的重音符號若併入最後一個字母，曾讓技能名稱被當成手動輸入；兩項都已修好。在 /model 選單用左右箭頭調整思考深度時，不再越過最高或最低值循環跳轉，避免誤存 Low 為預設值。以前透過 /feedback 傳送報告時，按 Ctrl+O 或 Ctrl+Z 會跳回草稿清單，導致無法取消傳送；這個問題現在已修好。keybindings.json 接受單一空白鍵 " "，對 "ctrl+ k" 這類多餘空格則顯示警告。

習慣在終端機開啟 Vim 模式編輯指令的讀者，縮排與刪除操作現在更符合直覺。以前在純空白行按 >> 或 << 縮排時，游標會漂移至行尾之外，導致接著按 x 無法刪除字元。現在游標會留在正確位置。在 Visual 模式以 V 接 d 刪除整行後，游標會停在首個非空白字元，且按下 . 鍵也能正確在該行重複操作。

執行日常維護或清理檔案時，行程管理與清理指令更穩定。執行 claude logs、claude stop、claude kill、claude rm 與 claude daemon status、claude daemon stop、claude daemon uninstall 時，不再因登入即將或已經過期而突然被登出。claude purge 以前遇到無法刪除的檔案或資料夾，可能靜默結束或卡在終端機；現在會刪完其餘項目、列出失敗清單，並以 exit 1 結束。Windows 停止狀態列、掛鉤或指令時，也不再誤殺拿到相同行程 ID 的無關程式。

在長時間對話或自動執行任務時，Claude 對歷史進度的判斷與權限通知更穩定。以前觸發上下文壓縮（將過長記錄精簡以釋放空間的機制）時，Claude 可能把壓縮前做完的事誤判為剛做完，導致重複做工或收回成果。現在這項時序錯亂問題已經修復。自動模式拒絕訊息的改動已撤回，不再告訴 Claude 拒絕涵蓋整體結果而非單一指令。雲端容器重啟若遺失待執行的 /loop 喚醒或排程任務，不再通知 Claude，工作階段會維持休眠。

專案連接多個外部工具或在自託管環境大量運作時，記憶體佔用與伺服器負載更平衡。以前 HTTP MCP（讓 AI 連接外部工具的協定）連線會把每個已送出的請求保留到連線關閉。這會造成記憶體洩漏（不再需要的資料仍占著記憶體），現已修好。提供給模型的代理與 MCP 伺服器清單中，非 ASCII 字元名稱會排在 ASCII 名稱之後。OpenTelemetry（蒐集程式運作資料的工具）的 claude_code.at_mention 記錄項目，每次讀取提示詞最多送出 100 個代理事件及 100 個 MCP 資源事件。自託管環境將固定 5 秒輪詢改為間隔 4～6 秒，避免同一環境的多個副本同時發出請求。

在 Slack 頻道中與 Claude 協同作業或執行例行任務時，訊息指派與權限變更更平順。在跨工作區的 Enterprise Grid 共享頻道中，遇到未連線的工作區標籤時，Claude 不會再誤報工作區未設定。管理員在任務途中修改連接器、外掛、技能或規則時，Claude 不會中途停下，改為完成當前任務後才套用。在討論串帶備註觸發例行任務時，能正確在該串內接續執行與回報。當 Claude 加入僅限管理員發言的公告頻道時，會安靜加入而不主動發送自我介紹。

When running multiple background sessions, the agent list and bottom-bar information now correctly reflect the current state. Previously, when the system briefly couldn't read the session folder (for example, because too many files were open), the agent count in the bottom bar would vanish and the list would be wiped out by placeholders. Now read errors no longer scramble the interface. Custom agents named worker are no longer mislabeled as Agent at startup, and the detail dialog title keeps the original type after they finish. claude agents previously offered an option to skip permission checks, but if the consent existed only in .claude/settings.local.json or a --settings file, background sessions would ignore that option; now consent is asked for first, and sessions that still ignore the option keep showing a short notice.

When you press the left arrow ← key mid-run to switch a session to the background, unsent text and pending-question states are no longer accidentally interrupted. Previously, as long as the input box held unsent text, or the screen was waiting for you to answer a question, pressing ← would still switch to the background after 10 seconds. Now the background switch is automatically canceled in these situations. When a queued message cannot be transferred to the background, the system stays in place and explains why, instead of losing the message. If a permission confirmation appears right after you press ←, pressing Esc, or choosing No without feedback, also stops that turn as expected.

When developing custom modules or plugins, tool registration and testing flows are more reliable. $.tool.register supports isDeferred; setting it to false lists the tool schema directly in the prompt without going through tool search. When the plugin hook (a small program triggered by specific events) worker process restarts, a module's classic.* event hooks are no longer dropped, avoiding a situation where only settings hooks respond alone. Modules calling $.session.append used to make claude plugin test fail; tests can now read back the appended records via mock.session. On Macs with Docker Desktop installed, symlinks (shortcut files pointing to another location) under ~/.docker/bin used to make claude plugin eval reject every execution granted Bash permission. This bug is fixed, and the rejection message now indicates which part of the credential store (where login proofs are kept) holds the link.

When conversing through remote control or cloud environments, long replies display and background status behave more normally. Previously, in long conversations text would occasionally freeze in a block and then suddenly pop out; it now displays smoothly word by word. After credentials are restored, the system no longer re-uploads the history from the start of the session. When launched via the claude remote-control command, push notifications no longer falsely report that remote control is not enabled.

When uploading projects with code review or /ultrareview, file checks and error guidance are clearer. Previously on Linux, sandboxed command execution could wrongly reject some repositories because a config file "couldn't be parsed"—for example, when the repository sat inside another checked-out copy of the project. This read problem is now fixed. When a split index file is rejected, the system no longer suggests a fix that could leave Git unable to read the index. In Code Review's add-repository dialog, repositories that can't be added are now listed one by one, each with its reason (such as missing GitHub write permission).

When you had just edited a custom skill's description on the web, Claude in the terminal sometimes couldn't read the change. After editing a skill description synced from claude.ai, you no longer sometimes need to start a new conversation or type /clear for the model to pick up the revised description. When no session is in use, the background sync that checks for skill changes now runs about every 40 minutes instead of every 10 minutes.

When pasting text in the terminal, setting shortcuts, or picking model parameters, input detection is more accurate. Pasted text whose first and last words matched used to be treated as typed by hand. After pasting a skill name, an accent mark typed or pasted right after it that merged into the last letter could make the skill name count as manual input; both issues are fixed. In the /model menu, using the left and right arrows to adjust thinking depth no longer wraps past the highest or lowest value, avoiding accidentally saving Low as the default. Previously, when sending a report via /feedback, pressing Ctrl+O or Ctrl+Z jumped back to the draft list, making it impossible to cancel sending; this is now fixed. keybindings.json accepts a single space key " ", and shows a warning for extra spaces like "ctrl+ k".

For users accustomed to editing commands with Vim mode enabled in the terminal, indentation and deletion operations are now more intuitive. Previously, pressing >> or << to indent on blank lines caused the cursor to drift past the end of the line, preventing a subsequent x from deleting characters. The cursor now remains in the correct position. After deleting an entire line in Visual mode using V followed by d, the cursor lands on the first non-blank character, and pressing . correctly repeats the operation on that line.

Process management and cleanup commands are now more reliable during routine maintenance and file cleanup. Running claude logs, claude stop, claude kill, claude rm, claude daemon status, claude daemon stop, or claude daemon uninstall no longer causes unexpected logouts when login credentials are near expiration or already expired. When encountering files or directories that cannot be deleted, claude purge previously exited silently or hung in the terminal; it now finishes deleting the remaining items, lists failures, and exits with code 1. On Windows, stopping the status line, hooks, or commands no longer terminates unrelated programs that happen to share the same process ID.

Claude tracks historical progress and permission notifications more reliably during long conversations or automated tasks. Previously, triggering context compaction (a mechanism that condenses long transcripts to free up space) could cause Claude to misjudge tasks finished before compaction as just completed, leading to duplicate work or reverted progress. This timeline confusion has now been fixed. A change to auto-mode rejection messages has been reverted, no longer telling Claude that a rejection covers the entire outcome rather than a single command. When cloud container restarts lose pending /loop wakeups or scheduled tasks, notifications are no longer sent to Claude, allowing the session to stay dormant.

Memory consumption and server load are more balanced when projects connect multiple external tools or run at scale in self-hosted environments. Previously, HTTP MCP (a protocol connecting AI to external tools) connections retained every sent request until closed. This caused a memory leak (where data no longer needed continues occupying memory), which is now resolved. In the list of agents and MCP servers provided to the model, names with non-ASCII characters are sorted after ASCII names. OpenTelemetry (a tool for collecting application runtime telemetry) claude_code.at_mention log entries now emit at most 100 agent events and 100 MCP resource events per prompt read. Self-hosted environments replaced fixed 5-second polling with a 4–6 second jitter interval, preventing multiple instances in the same environment from sending requests simultaneously.

Message routing and permission updates are smoother when collaborating with Claude or running routine tasks in Slack channels. In Enterprise Grid shared channels across multiple workspaces, Claude no longer erroneously reports that a workspace is unconfigured upon encountering an unconnected workspace tag. When administrators update connectors, plugins, skills, or rules mid-task, Claude avoids abrupt interruptions and instead applies changes after finishing the current task. Triggering routine tasks in threads with notes now reliably executes and reports back within the thread. When Claude joins announcement channels where only administrators can post, it joins quietly without posting an unprompted introduction.