發布日期：2026-10-15
當你放著 Claude 在背景通宵跑大量任務時，最怕半夜遇到額度限制而卡住空轉。以前開啟無人值守重試模式（CLAUDE_CODE_RETRY_WATCHDOG，遇到錯誤時自動重試的背景守護機制）後，只要撞上額度上限，即使重設時間遠在好幾個小時後，程式也會一直停在那裡乾等。現在新增了環境變數 CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS，只要重設等待時間超出你設定的毫秒上限，就會果斷結束該次請求。
習慣在終端機用快捷鍵飛速跳轉程式碼的 Vim 族群，這版操作手感更貼近原生編輯器。Vim 模式新增了 % 移動指令，能跳到配對括號，也能搭配 d、c、y 執行刪除、修改或複製。以前在負載沉重的機器上按 Esc 鍵，有時會卡在 INSERT（插入模式，打字輸入內文的狀態）而無法切換。回答提問輸入框時，自由文字欄位現在支援 NORMAL（一般模式，用來執行移動與指令的瀏覽狀態）。按 Esc 只會先退回一般模式，只有在一般模式下再按 Esc 才會取消，不會直接抹掉你辛苦打好的回答。
想在企業儀表板掌握 Claude 呼叫工具的原因，觀測日誌現在記錄得更詳盡。在 claude-api 技能中加入 OpenTelemetry（跨系統追蹤運作軌跡的開源標準）的設定指引，包含如何使用 ant 工具配置追蹤目的地。工具決策事件 tool_decision 新增 classifier_verdict 與 permission_mode 欄位，讓管理員分清自動模式（auto mode，讓 Claude 自主執行的模式）分類器的決策與既有權限規則。啟用 OTEL_LOG_TOOL_DETAILS=1 時，分類器阻擋操作的 tool_decision 事件還會以 classifier_category 記錄命中的內建規則名稱。
編寫 Mod（擴充 Claude 功能的模組）時，介面與流程控制更精準。Input 元件新增 rows 屬性，終端機能多行輸入，Enter 送出、Shift+Enter 或反斜線換行。修復在 claude -p 或雲端工作階段，首次回覆未完成就呼叫 $.model.fork 報錯 nothing-to-fork。提問樹狀清單在桌面端、行動端與 VS Code 支援 Markdown 且最多 48 行，終端機維持 12 行。掛鉤（hook，特定事件觸發的小程式）的 .catch 改為只在拋出例外或逾時時執行，內部呼叫產生的事件不再轉給自身。設定或取消設定環境變數的方法 $.env.set，完成操作時回傳 {}；未執行寫入時，回傳附帶原因的 { deny }。
聊到一半想切換到上下文視窗較小的模型，長對話轉移現在處理得更順暢。上下文視窗是模型一次能讀取的最大文字量；若對話太長，會先請原模型精簡摘要，以前未經詢問就扣額度摘要的問題已修復。透過 Anthropic API 使用 fable 或 best 別名時，若組織尚未開放最新版 Fable，會自動退回前一版重試一次。非 Bedrock 的自訂模型名稱即便包含 application-inference-profile，也不會誤跑 awsAuthRefresh 耽擱掛鉤。透過閘道器連線時，伺服器回傳不同模型名稱，也不會再反覆用不是你選的模型壓縮對話。
在專案中安裝大量自訂技能或外掛工具時，指令呼叫與權限確認現在更俐落透明。過去技能（擴充特定領域知識與指令的手冊）與命令若包含過多或過長的參數名稱，呼叫時會卡頓數秒。現在限制僅填入前 64 個名稱且每個最多 64 個字元。修復了技能、命令與掛鉤在填入大量佔位符時耗用大量記憶體的問題。修復 /rewind 回退對話後技能清單在模型中消失的問題，並更正了 /claude-api 技能誤稱 AWS 上 Claude 平台不支援 CMEK 外部金鑰端點的說明。外掛或技能掛鉤若觸發 ask 詢問決策，彈出的權限確認提示現在會明確標示該掛鉤來自哪一個外掛或技能。
登入更新失敗時，錯誤提示更清楚；快速執行短指令後，也不會再因更新途中結束而意外登出。過去當硬碟空間被塞滿時，系統無法完成登入憑證更新，卻彈出讓人一頭霧水的「Please run /login」錯誤訊息。現在會具體指出是哪顆磁碟空間已滿，並明確提醒你刪除檔案釋放空間。當你執行 claude auth status 檢查認證或 claude plugin list 查詢外掛等耗時短暫的指令時，過去若指令在 token（驗證身分的授權令牌）更新途中結束，會引發更新逾時甚至意外登出。這種登入更新逾時與意外登出的問題已經修好。
請 Claude 產出獨立網頁或文件成品（artifacts）時，更新內容與手機預覽現在都更穩定。在設定了攔截編輯或寫入的 PreToolUse 掛鉤環境下，過去更新成品時常會被直接拒絕。現在 Claude 會先正常編輯檔案再重新發布，而不是試圖在發布呼叫中直接修改檔案。在行動端應用程式中，Claude 產出的網頁以前捲動時文字會穿透到置頂標題上方。現在上方色帶會確實著色覆蓋，捲動內容不再穿幫破圖。
盤點專案原始碼或比對版本差異時，底層檢查工具的回報比以往更誠實安全。Grep 與 Glob 等搜尋工具過去在搜尋程式意外崩潰時，會回傳查無結果；現在會老實拋出錯誤告知搜尋失敗。Read 讀取工具不再將 .xlsm、.parquet、.epub 與 .heic 等二進位檔案誤當成純文字讀入。使用 WorktreeCreate 掛鉤建立的 Git 工作樹（worktree，讓同一專案平行開多份副本的功能），若裡面還有尚未推送的提交，退出 Claude Code 時不會再不問一聲就悄悄刪除。在採用 reftable 格式的 Git 存放庫中，若底層檢查逾時或失敗，/diff 指令不再漏掉發生衝突的揀選或復原操作。
安排多代理人或腳本排程自動處理專案時，背景步驟與權限控管更為穩定。工作流程代理人過去讀到任務說明未附帶使用者授權的框架備註後，有時會擅自略過自己的步驟卻仍回報執行成功，這個問題已經修復。在同一個工作階段中執行 Bash cd 切換目錄後，讀取、編輯或重跑工作流程腳本不再莫名彈出權限確認或直接被拒絕。工具呼叫若在權限檢查期間被回合中途傳入的訊息中斷，現在不會再向模型誤報為使用者拒絕。在 Windows 環境下，被終止的背景 PowerShell 命令輸出檔過去會漏掉結尾的 [killed] 標記，現在均已正常補上。
開啟終端機執行指令時，螢幕畫面上的字元邊界與色彩層次變得更加乾淨易讀。在 RHEL 8 的 MATE 終端機等使用舊版 VTE（終端機模擬器核心元件，版本低於 0.53）的環境中，修復了 >0q 遮擋資料夾信任提示以及退出時出現 >4m 雜訊的問題。修復特定版本 iTerm2 啟動時提示字元前方出現殘留 GOK 字樣的問題。只有標題列的 Markdown 表格若寬度超出終端機，過去會完全無法顯示（包含出現在計畫核准卡片中的表格），現在均能正常繪製。在暗色 ANSI 主題與 16 色終端機下，填滿方框內的文字維持淡化，但在選取與搜尋醒目提示下不再淡化，清單說明文字也正確比照次要文字淡化處理。
進行長時間對話或派發子任務時，畫面不再因資料繁重而短暫凍結。在全螢幕模式下執行 /compact 或自動壓縮時，過去因重建可見對話紀錄會造成介面短暫卡頓，現在已修復。在長時間運作的子代理人（subagent，負責獨立執行特定子任務的代理人）工作時，進度歷程改為增量更新，不再每次渲染都整條重算。重新接續先前的工作階段時，壓縮提示列不再提供無法載入歷史紀錄的 Ctrl+O 捷徑，而是清楚說明先前訊息儲存在哪個路徑，且此提示在全螢幕下也能正常顯示。當你中途打斷回覆時，過去會立刻冒出建議輸入 /clear 清除對話省 token（AI 處理與計費的文字單位）的提示，現在不會再立刻跳出來干擾。
串接團隊通訊軟體並使用自託管 Runner 時，工作階段的啟動與釋放更值得信賴。修復從 Slack 啟動的工作階段在受管設定限制 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）伺服器時，無法找到或掛載存放庫的問題。若需停用特定工具，可在設定中配置 deniedMcpServers。修復無關的環境設定或工作階段失敗訊息引用 Git 遺失參照字眼時，工作階段被誤判為來源分支已刪除、因而不再重試的問題。即將結束的工作階段若在釋放資源時失敗或遭拒，日誌不再承諾重試；重複失敗時會顯示失敗次數。
在 VS Code 面板與 Claude 討論題目時，公式呈現更漂亮且對話操作更滑順。Claude 的回覆支援 LaTeX 數學公式，包含獨立段落的 $$、多行 \[ … \] 與行內 $…$、\(…\) 語法。VS Code 與 JetBrains 面板中，Pro 與 Max 方案的用量上限橫幅新增「Get more usage」連結。修復對話造成面板當機或凍結後，每次重新載入視窗又強制重開該對話的惡性循環。修復長時間思考時面板變慢的問題，不再每次更新都重新解析整段思考文字。部分語言伺服器回報問題時不再顯示 [object Object]，而是正確顯示規則代號（例如 MD013）。非 claude.ai 登入或遠端視窗等未提供聽寫的環境，不再載入原生音訊模組檢查麥克風權限。
When leaving Claude running extensive tasks in the background overnight, hitting usage limits and idling mid-job is a constant worry. Previously, with unattended retry mode enabled (CLAUDE_CODE_RETRY_WATCHDOG, a background watchdog that automatically retries on errors), hitting a usage limit caused the process to wait indefinitely, even if the reset time was hours away. The new CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS environment variable decisively terminates the request whenever the reset wait time exceeds your configured millisecond threshold.
For Vim users accustomed to navigating code quickly with terminal shortcuts, handling now feels closer to the native editor. Vim mode introduces the % motion to jump between matching brackets, which also pairs with d, c, and y for delete, change, or yank operations. Previously, pressing Esc on heavily loaded machines could get stuck in INSERT (the state for typing text) without switching. Free-text fields in prompt inputs now support NORMAL (the navigation state for motions and commands). Pressing Esc now drops back to normal mode first, and only cancels when pressed again in normal mode, preventing your typed answer from being wiped out.
Observability logs now record more detail when tracking why Claude invokes tools in enterprise dashboards. Configuration guidance for OpenTelemetry (an open-source standard for tracking execution traces across systems) is now included in the claude-api skill, covering how to configure export destinations with the ant tool. The tool_decision event adds classifier_verdict and permission_mode fields, helping administrators distinguish classifier decisions in auto mode (the setting allowing Claude to execute autonomously) from existing permission rules. When OTEL_LOG_TOOL_DETAILS=1 is enabled, tool_decision events where the classifier blocks an action also record the triggered built-in rule name in classifier_category.
Interface and flow control become more precise when authoring a Mod (modules that extend Claude's capabilities). The Input component adds a rows prop for multi-line terminal input, submitting on Enter and inserting newlines on Shift+Enter or backslash. Calling $.model.fork before the initial response finishes in claude -p or cloud sessions no longer throws nothing-to-fork. Prompt tree lists now support Markdown with up to 48 lines across desktop, mobile, and VS Code, while staying at 12 lines in the terminal. The .catch handler on a hook (small programs triggered by specific events) now runs only on thrown exceptions or timeouts, and events generated by internal calls are no longer forwarded back to the hook itself. The $.env.set method for setting or unsetting environment variables now returns {} on successful operations, and { deny } with a reason when writes are skipped.
Switching mid-conversation to a model with a smaller context window is now much smoother when transferring long dialogs. A context window represents the maximum text volume a model can read at once. When conversations run too long, the original model is asked to summarize them first, resolving an issue where quota was deducted for summarization without asking. When using fable or best aliases through the Anthropic API, organizations without access to the latest Fable release will automatically fall back to the prior version and retry once. Non-Bedrock custom model names that include application-inference-profile no longer mistakenly run awsAuthRefresh and delay hooks. When connecting through a gateway, differing model names returned by the server will no longer cause repeated conversation compression using an unselected model.
Invoking commands and confirming permissions is now cleaner and more transparent when projects install numerous custom skills or plugins. Previously, skills (playbooks expanding domain knowledge and commands) and commands with too many or excessively long parameter names stuttered for seconds during invocation. Hydration is now limited to the first 64 names with a maximum of 64 characters each. An issue where skills, commands, and hooks consumed excessive memory while hydrating numerous placeholders has been resolved. The skill catalog no longer vanishes from the model after running /rewind to roll back conversations, and documentation in /claude-api mistakenly claiming the Claude platform on AWS lacks CMEK external key endpoint support has been corrected. When a plugin or skill hook triggers an ask prompt for decisions, the confirmation dialog now clearly labels which plugin or skill originated the hook.
Error messages are clearer when login refresh fails, and running quick commands will no longer cause accidental logouts if they terminate mid-refresh. Previously, when disk space filled up completely, the system failed to refresh login credentials and instead displayed a confusing "Please run /login" error message. It now identifies the exact disk that is out of space and explicitly prompts you to remove files to free up room. When running quick commands like claude auth status or claude plugin list, if the command exited while refreshing a token (an authorization ticket used to verify identity), it previously triggered refresh timeouts and unexpected logouts. These login refresh timeouts and accidental sign-outs are now fixed.
Updating content and previewing on phones is now more reliable when asking Claude to produce standalone web pages or document artifacts. In environments configured with PreToolUse hooks that intercept edits or writes, updating artifacts previously met with outright rejection. Claude now edits the file normally before republishing, instead of attempting to modify files directly inside the publishing call. In the mobile app, text on Claude-generated web pages previously bled above the sticky header during scrolling. The top header band now renders solid background fill, preventing scrolled content from showing through.
Underlying inspection tools report results more faithfully and safely when auditing source code or comparing version diffs. When search processes crashed unexpectedly, search tools like Grep and Glob previously returned empty results. They now properly throw errors to report the search failure. The Read tool no longer mistakenly ingests binary files such as .xlsm, .parquet, .epub, and .heic as plain text. When exiting Claude Code, Git worktrees (capabilities allowing parallel copies of the same project) created via WorktreeCreate hooks are no longer silently deleted if unpushed commits remain. In Git repositories using the reftable format, the /diff command no longer misses conflicted cherry-pick or revert operations if underlying inspections time out or fail.
Background steps and permission controls are significantly more dependable when orchestrating multi-agent or scripted project automations. When workflow agents encountered framework notes indicating missing user authorization in task descriptions, they previously skipped steps while falsely reporting success, an issue that is now fixed. After running Bash cd to change directories within the same session, reading, editing, or re-running workflow scripts no longer prompts unexpected permission requests or fails with rejections. Tool calls interrupted by incoming mid-turn messages during permission evaluation are no longer falsely reported to the model as user rejections. On Windows, output files from terminated background PowerShell commands that previously missed the trailing [killed] indicator now append it properly.
When opening a terminal to run commands, character boundaries and color gradations on screen are now cleaner and easier to read. In environments such as MATE Terminal on RHEL 8 that use older versions of VTE (terminal emulator core component, versions below 0.53), fixes resolve >0q obscuring folder trust prompts and >4m noise appearing upon exit. A bug where leftover GOK text appeared in front of the prompt upon launching certain iTerm2 versions has been fixed. Markdown tables containing only a header row previously failed to render completely when their width exceeded the terminal (including tables appearing in plan approval cards), but they now draw properly. Under dark ANSI themes and 16-color terminals, text inside filled boxes remains dimmed, but it is no longer dimmed under selections and search highlights, and list descriptions are also correctly dimmed like secondary text.
When engaging in lengthy conversations or dispatching subtasks, the interface no longer freezes briefly from heavy data loads. A brief UI stutter caused by rebuilding visible message history when running /compact or automatic compaction in fullscreen mode has now been fixed. During long-running subagent (an agent that independently performs a specific subtask) tasks, progress history now updates incrementally instead of recalculating the entire trail on every render. When resuming a previous session, the compaction prompt bar no longer offers a Ctrl+O shortcut that fails to load history, but instead clearly explains the path where earlier messages are stored, and this prompt now displays properly in fullscreen mode as well. When you interrupt a response midway, a prompt that previously appeared immediately suggesting /clear to clear conversations and save tokens no longer pops up right away to cause disruption. Tokens are units of text used for AI processing and billing.
When integrating team chat apps and using self-hosted runners, session launch and teardown become significantly more reliable. A fix resolves an issue where sessions launched from Slack failed to find or mount repositories when managed settings restricted MCP (Model Context Protocol, a protocol that connects AI to external tools) servers. If specific tools must be disabled, deniedMcpServers can be configured in settings. A bug has been fixed where unrelated environment settings or session failure messages citing missing Git reference wording caused sessions to be misidentified as having a deleted source branch, preventing retries. If a terminating session fails or is rejected while releasing resources, logs no longer promise retries. Repeated failures now display the failure count.
Formulas render more cleanly and conversation interactions feel smoother when discussing problems with Claude in the VS Code panel. Claude's responses now support LaTeX mathematical formulas, including standalone paragraph $$, multiline \[ … \], and inline $…$ or \(…\) syntax. In VS Code and JetBrains panels, a "Get more usage" link has been added to the usage limit banner for Pro and Max plans. A fix resolves a vicious cycle where a conversation that crashed or froze the panel was forcefully reopened every time the window reloaded. Panel slowdowns during extended thinking have been resolved by no longer reparsing the entire thinking text on every update. When certain language servers report issues, they no longer display [object Object], correctly showing rule identifiers instead (such as MD013). Environments without dictation, such as non-claude.ai logins or remote windows, no longer load the native audio module to check microphone permissions.
Added `CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS` so unattended retry mode (`CLAUDE_CODE_RETRY_WATCHDOG`) ends a request at a usage limit that resets too far away instead of waitingAdded the `%` motion to vim mode: jump to the matching bracket, alone or with d/c/y
Fixed vim mode sometimes staying in INSERT after Escape on a slow or busy machine
Fixed Esc discarding a typed AskUserQuestion answer in vim mode: the free-text field now supports vim NORMAL mode, and Esc cancels only from NORMALAdded OpenTelemetry tracing guidance to the `claude-api` skill, including setting up trace destinations with the `ant` CLI
Added `classifier_verdict` and `permission_mode` to OpenTelemetry `tool_decision` events, so auto mode classifier decisions can be told apart from permission rules
Added `classifier_category` to OpenTelemetry `tool_decision` events, naming the built-in rule a classifier block matched, when `OTEL_LOG_TOOL_DETAILS=1`Added `rows` to a mod's `Input`: a text field of several lines in the terminal, where Enter submits and Shift+Enter or a backslash before Enter starts a new line
Fixed a mod's `$.model.fork` returning `nothing-to-fork` when called before the first response of a new Claude Code process had finished, as in `claude -p` runs and cloud sessions
Improved mods: in VS Code and the desktop and mobile apps, a tree drawn above the question dialog can now hold Markdown and up to 48 rows; the terminal keeps 12 plain rows
Changed a mod hook's `.catch` handler to run only when the hook throws or times out; events raised beneath the hook's own calls now skip that hook without asking its `.catch`
Changed a mod's `$.env.set` to resolve with `{}` once the variable is set or unset, or with `{ deny }` saying why when the write is not madeFixed the `fable` (and `best`) model alias erroring on the Anthropic API when the newest Fable model is not available to your organization: Claude Code now retries once on the previous Fable model
Fixed switching to a model whose name contains `application-inference-profile` running your `awsAuthRefresh` command and delaying model-switch hooks when the provider is not Bedrock
Fixed a case where switching to a model with a smaller context window could spend usage credits on a summary by the previous model without asking first
Fixed conversations being compacted over and over, on a model you didn't pick, behind an LLM gateway whose responses name a different model than the one requested
Improved switching to a model with a smaller context window: a conversation too long for the new model is now summarized by the previous model firstFixed skills and commands with very many or very long argument names stalling for seconds when invoked; only the first 64 names, each up to 64 characters, are now filled in
Fixed skills, commands and hooks using very large amounts of memory when their arguments fill many placeholders
Fixed the /claude-api skill saying CMEK external-key endpoints aren't available on Claude Platform on AWS
Fixed skills no longer being listed to the model after using `/rewind`
Fixed the permission prompt for a plugin or skill hook's `ask` decision not showing which plugin or skill the hook comes fromFixed a misleading "Please run /login" error when a full disk keeps the login from being refreshed; the error now names the disk and says to delete files
Fixed sign-in refresh timeouts and unexpected sign-outs after a short command such as `claude auth status` or `claude plugin list` exited in the middle of a token refreshFixed artifact updates being refused where a PreToolUse hook covers Edit or Write; Claude now edits the file and then republishes instead of changing it in the publish call
Improved artifacts: pages Claude writes now paint the band above a sticky header, so scrolled text no longer shows there in the mobile appsFixed quitting removing a worktree made by a `WorktreeCreate` hook without asking when it may hold unpushed commits
Fixed Grep and Glob reporting no matches when the search program failed unexpectedly before returning any results; they now return an error saying the search failed
Fixed the Read tool reading binary files such as .xlsm, .parquet, .epub and .heic as text
Fixed `/diff` not noticing a conflicted cherry-pick or revert in reftable repositories when one of its git checks timed out or failedFixed workflow agents sometimes skipping their step, and the run still reporting success, after reading the harness note that their task text carries no user authority
Fixed a tool call cut off by a message sent mid-turn during its permission check being reported to the model as rejected by the user
Fixed workflow scripts asking for permission or being refused when read, edited or re-run after a Bash `cd` in the same session
Windows: Fixed stopped background PowerShell commands missing the closing `[killed]` line in their output fileFixed stray characters such as `>0q` covering the folder trust prompt, and `>4m` before the shell prompt at exit, in terminals built on VTE older than 0.53 such as MATE Terminal on RHEL 8
Fixed stray `GOK` text appearing in the prompt at startup in some versions of iTerm2
Fixed a Markdown table with only a header row drawing nothing when too wide for the terminal, including in plans on the plan approval card
Fixed dim text on the dark ANSI theme and 16-color terminals: faint inside filled boxes, not faint under selection and search highlights, and list descriptions dimmed like other secondary textFixed the compaction row in a resumed session offering ctrl+o for history that is not loaded; it now says where the earlier messages are saved, and also shows in fullscreen
Fixed the "new task? /clear to save tokens" hint coming back right after you interrupt a response
Fullscreen mode: Fixed `/compact` and auto-compact briefly freezing the UI while the visible transcript is rebuilt
Improved rendering performance while a long-running subagent works: its progress trail now updates incrementally instead of being reprocessed in full on every renderSelf-hosted runner: Fixed sessions started from Slack being unable to find or attach a repository when managed settings restrict MCP servers; use `deniedMcpServers` to keep those tools off
Self-hosted runner: Fixed sessions being reported as having a deleted source branch, and not retried, when an unrelated setup or session failure quoted git's missing-ref wording
Self-hosted runner: Fixed the log promising a retry of a failed or declined session release for a session that is ending; repeated failures now show a count[VSCode] Added LaTeX math rendering in Claude's responses, including `$$` and multi-line `\[ … \]` blocks and inline `$…$` and `\(…\)` math
[VSCode] Added a "Get more usage" link to the usage-limit banner for Pro and Max plans in the VS Code and JetBrains panels
[VSCode] Fixed a conversation that crashes or freezes the panel being reopened on every reload
[VSCode] Fixed the chat panel slowing down during long thinking: it no longer re-parses the whole thinking text on every update
[VSCode] Fixed diagnostics showing "[object Object]" instead of the rule code (e.g. MD013) for some language servers
[VSCode] Fixed the native dictation audio module being loaded, to check microphone permission, where dictation is not offered (non-claude.ai sign-ins, remote windows)