Claude Code 2.1.239

你設定了預算上限，卻老是抓不準何時該停手或何時額度重設。現在輸入 /cost、查看狀態列或指定 --max-budget-usd 10.0 旗標時，針對有資料落地限制的工作區，系統會將僅限美國推論的 1.1 倍加成自動計入花費預估。當月度花費額度耗盡時，用量上限提示現在會同步標明你的工作階段（session）或每週額度何時重置。若你開啟了持續重試模式 CLAUDE_CODE_RETRY_WATCHDOG=1，一旦遇到組織花費上限或額度用光，程式會立即失敗結束，不會再無休止地空等重設。VSCode 橫幅裡的查看用量按鈕也調整至警示文字同一行。

在公司內部網路透過代理伺服器使用 Claude 時，連線與計費更加可靠。以前在 Bedrock 平台後方遇到移除回應標頭的代理伺服器，系統會誤以為串流（streaming）失敗而私下以非串流重跑，造成 API 計費莫名翻倍；現已修正。同時，若在 Bedrock 搭配 SSO 與 awsAuthRefresh，啟動前的憑證（credentials）預先檢查現在會遵循 export HTTPS_PROXY="http://proxy.corp.com:8080" 代理變數，解決卡死問題。在網頁版工作階段中，來自 Bash 等工具前往 non-API 網域的請求也會正確走代理連線，落實網域白名單。

現在多代理人（agent）協作的通訊機制全面完善。執行 ListAgents 或輸入 /list-agents 時，清單中會列出所有在線的活隊友，不再像以往只顯示子代理與其他工作階段而漏掉身旁隊友。工作階段（session）現在能透過 ListAgents 獲知同儕用來聯絡自己的名稱；若誤向自己發送 SendMessage，系統會明確提醒那是你自己的名字，而非回報找不到該對象。此外，標題以 / 開頭的工作階段不再顯示為無標題且可正常通訊，Windows 平台也正式支援跨工作階段傳訊，跨機器協作與 macOS 及 Linux 同步。

把耗時的任務交給背景執行，你不再需要被頻繁的打擾打斷工作節奏。現在使用 /goal 執行長任務時，簽到間隔會逐步拉長，從 30 分鐘延長至 1 小時，之後每 2 小時一次，不再每半小時重複打擾。若你透過 claude --resume 挑選並接續工作階段（session），系統會自動還原該對話中進行中的活躍目標。針對遠端與雲端工作階段，在執行耗時的 SessionStart 或 Setup 掛鉤（hook，特定事件觸發的小程式）期間會持續發送保活訊號，防止容器在掛鉤執行途中被當成閒置而回收。此外，閒置背景工作者重啟後也能穩定維持原有的計畫模式。

辛辛苦苦為對話取了自訂名稱，卻在聊了幾輪後突然在清單裡找不到。這次修復了多項接續工作階段（session）的歷史紀錄問題。以前為對話改名後只要後續對話超過約 64 KB，自訂標題就會在 /resume 選單中消失，即使後續對話超過約 64 KB，標題仍會留在 /resume 清單中。使用 claude -c 或接續功能時，路徑若僅有底線、連字號或點號等符號差異，系統不再錯誤抓取其他資料夾的對話。只是重新開啟對話或檔案被外部觸碰，不會再被誤判為最近更新而擾亂排序。在全專案模式下，若對話原本的工作目錄（例如已被刪除的工作樹副本）已不存在，系統會直接在當前目錄接續，不再叫你切換進不存在的資料夾。

在全螢幕下叫出長表單或設定面板，常遇到按鈕被邊界切斷或擋住最新對話。現在全螢幕模式下的視窗排版與彈跳對話框變得更清晰。當 MCP（讓 AI 連接外部工具的協定）的互動表單高度超出終端機視窗時，表單會自動縮放貼合視窗，隱藏欄位可透過捲動瀏覽，而接受與拒絕按鈕會固定在底部始終可見。輸入 /config 或 /model 等斜線指令叫出設定面板時，對話內容會保持固定在面板上方，不再被面板直接遮蔽。此外，在 Claude 正在輸出回覆時打開 /workflows 工作流詳細對話框，視窗也不會再超出終端機而導致標頭被擠出螢幕外。

習慣終端機快速鍵的開發者，現在游標移動與編輯更順手。在 ~/.claude/settings.json 加入 {"keybindingFlavor": "readline"} 後，單字跳轉與刪除（Alt+F、Ctrl/Option+→、Alt+D，可用 Ctrl+Y 貼回）對齊 Bash 並以標點分詞。搜尋框按 Ctrl+Backspace 改為刪除單字。Shift 延伸選取後 selection:copy 不再漏字。游標在 [Pasted text #N] 內以 Ctrl+W、Ctrl+U、Ctrl+K、Option+Backspace、Option+D 或 vim df/dt 刪除不留損毀片段。代理人視圖 vim 模式按 Escape 切回 NORMAL 模式並保留輸入。

在終端機輸入登入驗證碼或移動滑鼠時，敏感字串與雜訊字元容易意外洩漏或干擾輸入欄位。以前在密碼遮罩輸入框中輸入的敏感文字，若按兩次 Esc 清除，可能會殘留在歷史紀錄或被 Ctrl+Y 意外貼出；現在機密文字已被嚴格隔離。在瀏覽器型終端機中，滑鼠移動回報若分段寫入，曾會將 "35;150;7M" 等座標字串誤塞進提示欄，現已全面過濾乾淨。

當你在命令列前加上驚嘆號切換至 Shell 模式（以 ! 開頭），並按下 Tab 鍵補全本機腳本路徑時，指令能直接正確執行。以前輸入 ./script 並按 Tab 鍵進行自動補全時，系統會自動剔除前方的 ./ 前綴，導致補全後的指令因為缺少當前目錄路徑而無法被 Shell 辨識與執行；現在補全時會完整保留 ./ 標記。

在雲端與本機同步外掛時，套件管理變得井然有序。在雲端工作階段（session）中，由 claude.ai 同步的外掛會明確標記為 name@synced，你可以透過 claude plugin enable <name>@synced 或 claude plugin disable <name>@synced 獨立開關，且永不覆蓋本機自行安裝的同名外掛。此外，在外掛市集設定 {"metadata": {"pluginRoot": "packages"}} 後，純外掛來源名稱會如說明文件所述正確解析至該子目錄下。

遠端 MCP（讓 AI 連接外部工具的協定）伺服器的容錯與連線穩定度大幅提升。以前在雲端工作階段中途重連，或是透過 SDK 的 setMcpServers() 動態更新伺服器清單時，若遠端 MCP 伺服器遭遇短暫的 5xx 伺服器錯誤，連線會永久卡死在失敗狀態；現在連線重試修復後能順暢復原並繼續提供服務。

清理暫存目錄或設定 Git 排除規則時，路徑問題不再使程式崩潰。從已刪除目錄啟動 Claude Code 會印出清晰診斷。工作階段（session）目錄若途中遭刪除，系統不再當機。掛鉤（hook，觸發的小程式）改由專案根目錄或家目錄執行，避免 ENOENT 錯誤。在 Git 工作樹（worktree，平行副本）中，.worktreeinclude 以 **/ 開頭的規則能正常比對已忽略資料夾內的檔案；在設定 extensions.worktreeConfig 的儲存庫中，Linux 沙盒不再因缺少 .git/config.worktree 使 Git 指令失效。若在設定檔加入：

{"claudeMdExcludes": [".claude/rules"]}

無論指向規則目錄或符號連結（symlink，捷徑檔）本身皆能排除。

針對採用 musl 標準函式庫的 Alpine Linux 環境，Claude Code 改為載入專屬 musl 編譯二進位模組，徹底解決以往因 glibc 格式不相容而被執行環境拒絕載入的問題。現在輕量容器與 Alpine 系統下，原生圖片貼上、系統剪貼簿讀寫與音訊擷取擴充功能皆能隨開即用。

在 Windows 或特定文字編輯器中建立說明文件時，常因隱藏的檔頭標記導致自訂功能無故失效。以前定義代理人（agent）、技能與自訂指令的 .md 檔案若開頭包含 UTF-8 BOM，會被系統默默忽略而形同虛設。現在系統能正確辨識這種位元組順序記號並順利載入檔案。

開啟語音輸入功能後，終端機啟動畫面變得更清爽。若你已經在設定檔 ~/.claude/settings.json 中加入了以下設定啟用語音聽寫：

{"voice": {"enabled": true}}

系統啟動時會正確識別你已開啟功能，不會再每次重複跳出無謂的 /voice 功能提示與教學。

當你執行 /insights 指令深入分析對話歷史與專案洞察時，產生的報表內容更加乾淨俐落。以前在使用部分模型生成分析結果時，回覆內容中常會夾雜模型輸出的字面 <message> 標籤，把報表排版撐得亂七八糟；現在系統會自動把這些標籤濾掉。

自訂介面主題的色彩細節獲得了完整支援。以前在終端機主題設定中自訂 effort 或 ultracode 等狀態徽章的顏色時，系統會默默忽略你的設定並維持原廠預設顏色；現在自訂主題覆寫已能被系統精準讀取，並如實套用至各項狀態徽章。

對話請求管線的多項重複執行問題已全數修復。排隊輸入提示時按 Esc 不再引發競態條件導致工作階段（session）提早閒置而重複執行動作。被組織政策檢查拒絕的請求，不會在顯示拒絕訊息前被錯誤重複發送。對話壓縮後的提示文字也完成改良，避免技能的原始參數被誤當成新請求再次重跑。

在 IDE 內頻繁修改程式碼或透過網頁擷取大量資料時，偶爾會遇到畫面卡住或記憶體持續攀升。在 JetBrains IDE 終端機連線 Claude Code 外掛時，執行 Edit 與 Write 工具時會停頓約 5 秒的卡頓問題已獲得解決。此外，WebFetch 快取（cache，暫存查過的網頁內容）在 15 分鐘後會如預期過期並釋放記憶體。這讓它不再在整個工作階段（session）中持續佔用系統資源。

當你在系統中設定攔截掛鉤並使用 OpenTelemetry 監控執行歷程時，遙測數據變得更加完整連貫。以前由 PreToolUse 掛鉤（hook，特定事件觸發攔截的小程式）延遲執行的工具呼叫，會錯誤開啟一條全新的追蹤鏈，導致遙測軌跡破碎分散；現在這些延遲執行的工具會正確接續回原對話輪次的追蹤歷程中，維持單一連貫的調試軌跡。

When setting strict budgets, knowing exactly what you are spending and when limits reset is essential. Now when running /cost, checking the status line, or passing --max-budget-usd 10.0, workspaces with data residency requirements automatically factor in the 1.1× US-only inference surcharge. If your monthly spend limit is reached, the warning dialog explicitly tells you when your session or weekly limit resets. Furthermore, persistent retry mode with CLAUDE_CODE_RETRY_WATCHDOG=1 now exits immediately on organization spending caps or out-of-credit errors rather than waiting forever. The "View usage" button in VSCode has also been shifted inline with the alert text.

Connecting through corporate proxies to reach Claude is now much safer and smoother. Previously, when using Bedrock behind proxies that stripped response Content-Type headers, the client mistakenly assumed streaming failed and secretly re-ran turns non-streaming, silently doubling API costs; this double-billing glitch is now completely resolved. Furthermore, when using Bedrock with SSO profiles and awsAuthRefresh, startup credential checks properly honor export HTTPS_PROXY="http://proxy.corp.com:8080", eliminating startup freezes. In web sessions, requests from Bash or other tools to non-API anthropic.com domains also route through the session proxy to enforce organizational allowlists.

Multi-agent coordination and session messaging are now fully refined. Running ListAgents or typing /list-agents displays active teammates alongside subagents and other sessions, so reachable peers never look absent. A session now discovers its own peer-facing address via ListAgents, and sending a SendMessage to your own name explicitly informs you that it is yourself rather than claiming the agent does not exist. Titles starting with / are no longer rendered as "(untitled)" or broken in routing, and Windows users now enjoy cross-session messaging on equal footing with macOS and Linux.

Running background tasks no longer interrupts your workflow with non-stop pings. When using /goal for long-running jobs, check-in prompts now back off progressively (from 30 minutes, to 1 hour, to every 2 hours) instead of constantly buzzing you every 30 minutes. Resuming a session with claude --resume automatically restores its active goal. For remote and cloud sessions, executing lengthy SessionStart or Setup hooks (scripts triggered on events) continuously sends keep-alive heartbeats to keep the hosting container from being reaped mid-hook. Cloud workers restarting after being idle also cleanly preserve their active plan mode.

Naming your conversations keeps your workspace organized, but long sessions used to lose their custom titles. Renaming a session now keeps its title in /resume even after the conversation grows beyond ~64 KB. Path matching in claude -c and /resume properly distinguishes directories that differ only by _, -, or . separators instead of accidentally loading foreign chats. Merely touching or reopening a session file no longer falsely marks it as updated or disrupts its display order. In all-projects mode, if a session's previous directory (such as a removed worktree) no longer exists, Claude resumes safely in your current directory rather than asking you to cd into a ghost folder.

Opening tall forms or configuration panels in fullscreen mode no longer clips essential buttons or buries recent messages. When an MCP elicitation form exceeds terminal height, it neatly fits the viewport with scrollable fields while the Accept and Decline buttons remain pinned and visible at all times. Slash-command panels like /config and /model keep the active conversation pinned above them instead of covering your latest messages. Opening the /workflows dialog while Claude is actively streaming a response also properly constrains window dimensions, ensuring dialogue headers never scroll out of view.

Terminal power users will find cursor navigation and line editing much more natural. Configuring {"keybindingFlavor": "readline"} in ~/.claude/settings.json matches Bash word manipulation exactly: Alt+F and Ctrl/Option+→ jump to word boundaries, Alt+D deletes forward to word ends (restorable with Ctrl+Y), and punctuation acts as word dividers. In search boxes, Ctrl+Backspace deletes entire words instead of single characters. Text selections extended with Shift+Arrow keys are reliably grabbed by selection:copy. Deleting text with Ctrl+W, Ctrl+U, Option+Backspace, Option+D, or vim df/dt while inside [Pasted text #N] placeholders no longer leaves corrupted placeholder markers. In agent view with vim mode active, pressing Escape switches cleanly to NORMAL mode while preserving typed text.

Entering credentials or moving the mouse in web terminals used to risk leaking sensitive secrets or injecting stray text. Masked fields like login verification codes no longer leak into prompt history or let cleared secrets be pasted via Ctrl+Y after double-pressing Esc. Browser-based terminals no longer leak fragmented mouse tracking strings like "35;150;7M" into your input prompt when mouse reports arrive across multiple writes.

When switching to shell mode with an exclamation mark (!) and pressing Tab to autocomplete executable scripts, commands now execute seamlessly. Previously, Tab completion on a ./script path dropped the leading ./ prefix, leaving a bare command name that the underlying shell could not resolve or run; completion now faithfully retains ./.

Managing plugins across local environments and cloud accounts is now completely conflict-free. In cloud sessions, plugins synced from claude.ai are designated with a name@synced suffix, can be independently controlled with claude plugin enable <name>@synced or claude plugin disable <name>@synced, and will never overwrite local custom plugins sharing the same name. In plugin marketplaces, configuring {"metadata": {"pluginRoot": "packages"}} now properly resolves bare package source names under that subdirectory as documented.

Remote MCP (protocols connecting AI to external tools) servers recover gracefully from transient connection drops. When reconnecting mid-session in cloud environments or reconfiguring servers dynamically via the SDK's setMcpServers(), encountering a momentary 5xx server error no longer leaves the remote MCP server permanently broken; it automatically recovers and resumes tool availability.

Cleaning up temporary project directories or managing complex Git ignore patterns is now error-free. Launching Claude Code from a deleted working directory prints a helpful diagnostic message instead of throwing an unhandled stack trace. If the session working directory disappears mid-session, hooks (event-driven helper scripts) automatically fall back to executing from the project root or home directory rather than failing with ENOENT errors. In Git worktrees (parallel repository working copies), .worktreeinclude patterns starting with **/ successfully match paths inside gitignored folders, and in repositories with extensions.worktreeConfig set, Linux sandboxes no longer break git commands when .git/config.worktree is missing. Adding rule exclusions:

{"claudeMdExcludes": [".claude/rules"]}

cleanly excludes symlinked rule files when referencing the link name or parent directory.

In lightweight Alpine Linux environments built on the musl standard C library, Claude Code now loads dedicated musl-compiled native add-ons instead of glibc binaries rejected by the runtime. Developers running inside Alpine containers can now enjoy native clipboard access, native image pasting, and audio capture without compatibility hitches.

Creating configuration docs in certain text editors often introduces invisible header bytes that caused custom behaviors to fail silently. Markdown .md definitions for declarative agents, skills, and custom commands that begin with a UTF-8 BOM (byte order mark) were previously ignored without warning. Claude Code now parses and activates them reliably.

Enabling voice dictation now keeps your terminal startup interface clean and uncluttered. If you have already enabled voice dictation by adding the following snippet to ~/.claude/settings.json:

{"voice": {"enabled": true}}

Claude Code properly recognizes your configuration on startup, permanently silencing the redundant /voice onboarding tip.

Running /insights to analyze conversation history and generate codebase usage trends now yields cleaner reports. When querying certain underlying models, literal <message> tags previously leaked into the generated insights output; they are now stripped before the report is shown.

Custom theme styling for terminal badges is now fully supported. Configuration overrides for the effort and ultracode status badge colors were previously ignored by the interface renderer; Claude Code now honors your custom theme colors across all status indicators.

Pipeline execution now cleanly avoids unintended request duplicates. Pressing Esc with a prompt queued no longer triggers a race condition that prematurely idles the session and causes resubmissions to repeat prior actions. Requests rejected by organizational policy checks are no longer re-sent before the rejection reason is displayed. Context compaction reminders have also been updated to ensure skill arguments are not accidentally re-executed as fresh user requests.

Editing files within IDEs and fetching web resources no longer suffers from unexpected pauses or memory leaks. In JetBrains IDE terminals, a 5-second pause during Edit and Write tool calls when connected to the Claude Code plugin has been resolved. Furthermore, WebFetch cache entries properly expire and free up system memory after 15 minutes as intended, rather than hoarding memory for the remainder of the session.

Observing tool execution flows with OpenTelemetry while using custom interception hooks is now reliable and unbroken. Previously, tool executions deferred by a PreToolUse hook started a brand-new trace, fragmenting telemetry data into disconnected chunks; deferred tool calls now properly resume within the parent turn's original OpenTelemetry trace.