Claude Code 2.1.243

登入驗證流程有了諸多貼心改善。現在執行 /login 前往 Anthropic Console 時，新增了免金鑰的「使用 Console 帳號登入」建議選項，方便禁止建立 API 金鑰的企業員工快速登入。透過 SSH 遠端連線登入時，登入網址會立即顯示，按下鍵盤 c 鍵會誠實回報網址如何複製而非一律回傳成功，全螢幕模式下也會附帶選取文字的提示。此外，登入網址等超出容器邊界的文字不會在畫面重繪時遺失前幾欄內容，登入後首度啟動工作階段（session）時由伺服器管理的企業公告 companyAnnouncements 也會正常顯示。

派遣子代理處理任務時的狀態掌握度更上一層樓。現在輸入 /tasks 或打開代理詳細對話框時，清單會明確列出每個子代理運行的模型名稱以及思考努力等級（effort level）。如果背景子代理正在等待它指派的最後一個背景 Bash 指令執行，系統修復了指令完成後子代理未被喚醒的狀況，現在會如期接續後續工作。在 VSCode 編輯器中，你在 Focus 視圖中手動展開的區塊，也不會再因為子代理頻繁呼叫工具而被強制自動摺疊收合。

連接外部工具時的穩定性大幅提升。在非互動 -p 模式或透過 SDK 呼叫時，遠端 MCP（讓 AI 連接外部工具的協定）伺服器若遇到網路斷線，現在會自動嘗試重新連線。重試失敗後系統會明確回報錯誤，工作階段（session）不再永久失聯。針對支援用戶端中繼文件之服務如 Linear，修正了從桌面版應用程式啟動 MCP 登入時跳出重新導向網址無效的錯誤。若使用者的 ~/.claude.json 設定檔中包含格式錯誤的 MCP 伺服器項目，輸入 /status 時會清楚印出出錯的檔案路徑，不再顯示缺少檔名的空白提示。

開發自訂外掛（plugin）時的除錯體驗更為流暢。當你透過 --plugin-dir 參數同時載入多個外掛時，修正了宣告 marketplace 欄位的外掛相依性永遠無法解析的問題，現在本機相依套件能正常關聯。執行 /reload-plugins 指令時，若停用了最後一個支援 LSP（語言伺服器協定，提供程式碼定義跳轉與補全的服務）的外掛，系統會確實卸除殘留的 LSP 工具。同時，若外掛變更會引發重新讀取整個對話紀錄，指令會在執行前主動跳出安全警告，避免讀取流程被無預警中斷。

讓 Claude 自行執行任務的自動模式（auto mode，讓 AI 不逐一詢問就自主執行工具的模式）修復了多項卡住問題。以前若伺服器端曾暫時停用該功能並被暫存，後續特徵旗標抓取失敗時會導致自動模式永久不可用，現已排除此問題。當 API 短暫過載並要求重試時，自動模式呼叫工具在等候約 1 分鐘後不會再被錯誤判定為暫時無法使用而直接拒絕。在 VSCode 擴充套件中，剛安裝完成或尚未抓取特徵旗標前啟動的工作階段（session），也會正確採用你自訂的預設模式或自動模式，不再強制退回預設權限模式。

接續過往對話的操作體驗更為順暢。輸入 /resume 挑選歷史對話時，清單不再受限於最近 50 個工作階段（session），現在只要向下捲動就會持續載入更多對話記錄。當你使用 claude --teleport <session> 指令傳送至其他工作階段時，若當前目錄有未提交的程式碼變更，系統會主動詢問是否將變更放入 stash 暫存並繼續，不再直接退出程式。針對雲端工作階段，在對話中途重啟後，若有未完成的掛鉤（hook，特定事件觸發的小程式）或背景任務通知，系統會以正常的接續訊息送出，不會再誤把通知內容當成使用者的提示詞。

命令列介面的各項互動細節獲得全面修復。輸入 /clear 清除對話畫面時，先前透過 /rename 設定的自訂工作階段（session）名稱會完整保留在提示列中，不再意外消失。當歷史檔 ~/.claude/history.jsonl 中混入損毀項目時，按 Ctrl+R 搜尋歷史或按方向鍵上鍵不會再崩潰中斷，能穩定瀏覽其他指令。在支援特殊按鍵編碼（例如 modifyOtherKeys 或 kitty 協定）的終端機裡，修正了按 Ctrl+[ 無法退出 Vim 插入模式的問題。此外，緊接在 Emoji 表情符號後方輸入的錯字，拼字檢查 spellcheck 現在也能正確標上紅色底線。

多對話協作的通訊管道現在更穩健了。在 2.1.232 版進行通訊端資料夾強化後，使用者命名空間與無 root 容器內的跨工作階段（session）傳訊功能曾被默默關閉，這次修復徹底解決了此問題。同時，跨工作階段收件匣的通訊端（socket）機制加入防護，若連線進入後在 30 秒內未傳送任何完整的行，系統會自動中斷該連線，因此發送資料的腳本應在資料準備完畢後再建立連線。

遠端操控對話時的連線韌性顯著提升。當伺服器在中途遺失環境時，執行 claude remote-control 不會再直接退出而丟下連線中的遠端工作階段（session），現在具備自動恢復機制。針對非管理員或擁有者角色的 Team 與 Enterprise 團隊成員，修正了停止並重啟 claude remote-control 後工作階段偶爾卡死的問題。此外，當你嘗試接續一個遠端遙控權限正被其他終端機佔用的對話時，提示訊息現在會清楚說明其他機器上的工作階段無法看見或連線至當前機器。

在受管網路環境下使用沙盒（sandbox）執行指令時，連線判斷更為靈活。沙盒環境中的 Bash 工具提示詞不再預先列出所有允許的網域清單，讓 Claude 能主動發起網路請求，方便你在執行當下核准新主機，避免 AI 預設未列出的主機皆被封鎖而直接放棄。當被攔截的指令依然回傳成功代碼 0 時（例如 curl 印出代理伺服器的 403 錯誤頁面），工具執行結果不會再遺漏沙盒的網路違規細節。此外，本機 IDE 連線現在同時相容大寫 NO_PROXY 與小寫 no_proxy 設定，避免本機連線誤走代理伺服器。

遭遇網路異常與額度重置時的體驗變得更加可靠。以前當 Anthropic API 完全沒有啟動回應時，工作階段（session）可能會陷入 10 分鐘以上的漫長無聲等待；現在請求會在約 3 分鐘後自動逾時並重試 1 次，若依然無回應則明確顯示 API Error: No response from API。驗證失敗、模型無法使用等用戶端產生的錯誤訊息，現在會以標準的錯誤行格式顯示，不再被誤渲染成 AI 的回覆文字。此外，當工作階段在閒置期間跨過限流視窗重置時間點，狀態列（status line，終端機最底部的即時狀態行）與 /usage 不會再殘留重置前的舊使用量百分比。

命令列參數與自動化規則的防呆驗證全面收緊。使用 --agents 參數啟動時，若傳入無效的 JSON 或格式錯誤的代理人定義，程式不再默默略過，而是會像 --mcp-config 一樣立即印出明確錯誤並退出。在掛鉤（hook，特定事件觸發的小程式）條件中，修正了類似 Bash(cat *) 的判斷條件在遇到包含 $() 或反引號指令替換時誤觸發不相干指令的問題。此外，若在關閉思考功能的狀態下設定了 xhigh 或 max 思考等級，錯誤訊息會明確指出該等級名稱、是哪項設定關閉了思考，並提示改用 /effort high 解決。

The authentication experience has received multiple refinements. Under /login → Anthropic Console, a keyless "Sign in with your Console account" option is now available (and recommended) alongside API key creation, allowing organizations that forbid manual API keys to sign in cleanly. Over SSH, the sign-in URL renders immediately, pressing c accurately reports how the URL was copied rather than always claiming success, and fullscreen hints explain how to select text. Long URLs overflowing container boundaries no longer lose their leading columns during repaints, and server-managed companyAnnouncements correctly display upon startup right after logging in.

Inspecting subagent tasks is now far clearer. Running /tasks or opening subagent detail dialogs explicitly displays the exact model and effort level assigned to each agent. When background subagents run long Bash jobs, an issue where agents failed to wake after their final background task completed has been resolved, ensuring uninterrupted workflow continuations. In VSCode, sections you manually expand inside the Focus view will no longer spontaneously collapse whenever subagents execute background tools.

Connecting to external tools via MCP is now significantly more dependable. In non-interactive -p runs and SDK sessions, dropped connections to remote MCP (Model Context Protocol) servers now automatically reconnect or report an explicit failure instead of silently hanging. For services supporting client ID metadata documents such as Linear, MCP server sign-in initiated from the desktop application no longer fails with an invalid redirect URI error. Additionally, when ~/.claude.json contains a malformed MCP server configuration, /status explicitly prints the offending filename instead of displaying an unhelpful dot without a path.

Plugin development and lifecycle management are significantly more dependable. When loading multiple local plugins together via --plugin-dir, dependencies declaring a marketplace field now resolve smoothly rather than failing indefinitely. Running /reload-plugins properly removes the LSP (Language Server Protocol) tool once the final LSP-providing plugin is disabled. Additionally, if an LSP plugin modification requires re-reading the entire conversation history, the command proactively surfaces a warning before proceeding.

Auto mode (where Claude acts without prompting for approval on every action) has received critical reliability fixes. If a temporary server-side disable was cached, subsequent feature flag fetch failures will no longer keep auto mode permanently disabled at startup. When the API experiences brief overload and requests client retries, tool calls in auto mode will no longer be rejected as "temporarily unavailable" after waiting about 1 minute. In VSCode, sessions started before feature flags finish their initial fetch (such as immediately after installation) now cleanly open in auto mode or your configured default mode rather than falling back to default permission prompts.

Picking up past conversations is now much smoother. Running /resume is no longer capped at displaying the 50 most recent sessions; scrolling down continuously fetches older conversations. When running claude --teleport <session> with uncommitted git changes, the CLI prompts to stash your changes and proceed rather than abruptly exiting. In cloud environments, sessions resuming after a mid-turn restart now send pending hook or background task notifications as regular continuation messages instead of mistaking them for user prompts.

Everyday terminal interactions are now significantly more reliable. Running /clear preserves your renamed session title in the prompt bar for the new session instead of dropping it. When ~/.claude/history.jsonl contains a malformed entry, searching history via Ctrl+R or pressing the up-arrow no longer breaks command browsing. Terminals that encode modified keys (such as modifyOtherKeys or kitty protocol) now properly exit Vim INSERT mode when pressing Ctrl+[. Furthermore, spellcheck accurately underlines misspelled words typed immediately following an emoji.

Inter-session communication is now significantly more dependable. Following directory security hardening in version 2.1.232, cross-session messaging that silently disabled itself inside user namespaces and rootless containers has been restored. Furthermore, the cross-session inbox socket now closes any connection that fails to send a complete line within 30 seconds, so custom scripts should connect only when payloads are ready to transmit.

Remote session control is significantly more resilient against network interruptions. When a server drops its environment mid-session, claude remote-control automatically recovers instead of exiting and stranding attached sessions. For Team and Enterprise members without admin or owner roles, sessions served by claude remote-control no longer get stuck after being stopped and restarted. When resuming a conversation currently locked by another terminal, the prompt explicitly informs you that sessions on other machines cannot be viewed from or reach your current host.

Network handling inside the execution sandbox is now much more flexible. The sandboxed Bash tool prompt no longer pre-lists allowed hosts, encouraging Claude to attempt requests so you can approve new domains on the fly instead of having the model assume unlisted hosts are blocked. When a blocked command exits with code 0 (such as curl outputting a proxy 403 page), the full network violation details are preserved in the tool results. Furthermore, local IDE connections now honor both uppercase NO_PROXY and lowercase no_proxy environment variables to prevent local traffic from routing through HTTP proxies.

Handling API dropouts and rate limits is now much more predictable. If the Anthropic API never begins streaming a response, requests no longer sit silent for 10+ minutes; they now time out after ~3 minutes, retry once, and surface API Error: No response from API if unresolved. Client-generated error messages (such as authentication or model availability errors) render as explicit error lines instead of normal model output. Furthermore, idle sessions accurately reset the status line rate_limits and /usage percentages once windows reset.

Command line inputs and automation rules now feature stricter validations. When launching with --agents, malformed JSON or invalid agent definitions will no longer be silently ignored; the CLI exits immediately with a descriptive error, matching --mcp-config. Hook if conditions like Bash(cat *) no longer mistakenly fire on unrelated commands containing $() or backtick substitutions. Additionally, setting effort levels to xhigh or max while thinking is disabled triggers a clear error naming the configured level, the offending setting, and /effort high as the fix.