發布日期：2026-07-22
你請 Claude 檢查一串改動，還想繼續問別的事時，/code-review 現在會交給背景 subagent（在旁邊獨立做事的小幫手）處理，不再把整個對話塞滿；連著輸入的 slash commands 也仍會以這次審查為目標。/ultrareview 可接受「檢查我的登入改動」這種說法：它會審查目前分支，並把這句當成結果備註。非互動執行 /code-review ultra 也會確實啟動雲端審查，而非悄悄改做本機審查；參數不對時，Claude 會拿到可修正的錯誤訊息。像把文件送去後台校稿，你仍能在前台寫下一封信，校稿員也知道該看哪一份。
使用 --ax-screen-reader 模式改句子時，按 Option+Delete、Ctrl+W 或 Ctrl+U 刪掉一個詞或一整行，螢幕閱讀器現在會唸出剛刪掉的文字，不必猜自己刪了什麼。輸入列尾端打一個空白，VoiceOver 也會回報空白，而不是誤唸「換行」。在 plugin 與 settings 面板用方向鍵移動時，終端機游標會跟著停在焦點列，讓螢幕閱讀器和放大鏡都找得到目前位置。這像有人一邊替你讀稿，一邊用手指指出你剛擦掉的字與正在看的那一行。
把好幾行指令貼進某些終端機時，換行曾被編成 Ctrl+J，Claude 卻把它當成字母 j，整段內容擠成一行；現在會保留原本的換行。剛編輯完就按左方向鍵，也不再無預警丟掉整段對話：它會先請你確認。若你進入 agent view（把工作放到背景看的畫面），按 Esc 則能回到剛背景化的對話。像寫信時貼上的段落不再被壓成一行，手肘碰到返回鍵前也有人先問你要不要真的把信丟掉。
外部工具突然不能用時，claude mcp list 或 /mcp 現在會顯示 HTTP 狀態與錯誤文字，不再只留一句「連不上」。MCP（Model Context Protocol，讓 AI 連接外部工具的協定）設定值前後若藏了看不見的空白，也會先警告你；這種空白常讓看似相同的網址或金鑰失效。啟動時「N 個 MCP servers 需要驗證」的通知，則不會再把尚未在 claude.ai 連接的 connector 多算進去。好比客服不只說電話不通，還告訴你是號碼錯、對方忙線，並把名單上的幽靈聯絡人劃掉。
Windows 使用者若要讓工具讀取 C:\Users\unicorn 這類路徑，\u 開頭的片段以前可能被誤當成 Unicode 編碼，變成中日韓字，於是檔案像被改了地址般找不到；現在會保留原路徑。IDE 裡選了一大段文字，若內容剛好在 emoji 中間被截斷，也不再產生亂碼。另有工具執行器出錯卻被悄悄吞掉的情形也已修正。這像郵差終於照信封地址送件，不會把 \u 當成暗號改寫，信封破到一半時也會說清楚，而不是裝作沒收到。
整理過長對話後查看 /context，token（AI 處理與計費的文字單位）用量不再顯示整理前的舊數字。工具中斷後，也不會亂留「使用者中斷請求」或缺了另一半的 tool_use 紀錄；prompt history 寫入碰撞或失敗時，歷史紀錄不再容易漏掉或重複。遇到 context overflow（對話塞超過可容納的量）又有很大的 thinking budget 時，重試不會一直重送同一個註定失敗的請求。背景化與 headless／SDK 的 fork session，整理後也保得住來源關係；壞掉的 delta attachment 不再讓續接對話每回合失敗。像一本帳本不再重複記帳、掉頁，換了書封後仍看得出每張收據從哪裡來。
透過 Bedrock 使用模型時，gateway 花費計量現在會依你設定的模型費率計算 application-inference-profile ARN 與其他對應到上游模型的 ID，不會因識別碼不同而算錯價。剛建立或連結 pull request（PR，請求把一組程式改動併入專案的單據）就結束 session，也較不會遺失 PR 事件。Bedrock 設定精靈對分區 AWS 區域的 assume-role profile（暫時借用另一個 AWS 身分的設定），或只能走 proxy 的網路，能正確驗證。系統時鐘被調整後，回合耗時也不會罕見地變負數或失真。就像計程車改按可靠的里程表計價，收據、身分證明與行車時間都不會因儀表跳動而弄錯。
自動模式遇到危險的 rm、背景 & 或可疑 Windows 路徑時，不會一律跳出權限視窗；會交由 auto-mode classifier（判斷是否安全的自動分類器）決定。plan mode 裡，靜態分析器無法證明是唯讀的 Bash 指令，也同樣交給它判斷，少掉不必要的打斷。IDE 互動的 sandbox（把程式限制在安全範圍的隔間）指令限制也加強了。信任對話框會明說授權涵蓋哪個 repository root（專案最外層資料夾），agent frontmatter 的 hook（特定事件觸發的小程式）則必須放在已接受工作區信任的資料夾。伺服器管理設定中的無害功能與成本開關，也不再每次都要求批准。這像門禁把一般訪客、危險包裹和哪棟樓的通行證分清楚。
你想快速問一個小問題時，Claude 不會自行啟動 /deep-research 把流程拉長；只有你手動輸入這個指令才會開始。相反地，若你透過 /config model=<x> 或 Remote Control 換模型，fast mode（優先加快回應的模式）因此改變時，現在會有公告，避免你以為速度設定還和剛才一樣。這像咖啡店不會擅自把你的外帶咖啡升級成慢工手沖，但換了吧台與出杯速度，會把牌子翻出來讓你知道。
自己寫 agent 或 skill 的 Markdown 設定時，agent 名稱若含 :，現在會直接拒絕，因為這個符號保留給 plugin namespacing（用名稱前綴區分外掛來源的規則），可避免日後撞名。skill 的 frontmatter（檔案最上方的設定區）若填 context: fork，預設會在背景執行；想留在前台，可加 background: false。布林值也不必只能寫 true／false：不分大小寫的 yes、no、on、off、1、0 都接受。像替社區信箱訂規格：冒號留給地址分區，寄件方式有清楚預設，而「開／關」也聽得懂多種自然說法。
專案裡有一層又一層的被監看資料夾時，整棵目錄樹被刪除或搬走，Claude 以前可能因呼叫層數太深而崩潰，畫面要繪製同樣深的 UI 樹也會出問題；現在已修正。這對一般專案不會多出新按鈕，卻讓大型、巢狀很深的專案整理資料夾時少一次突然斷線。像搬走一整排層層疊起的收納箱，搬運員不再數到最底層就昏倒，展示架也不會在畫到最深一格時垮掉。
If you ask Claude to inspect a batch of changes but want to keep chatting, /code-review now hands the job to a background subagent, a separate helper working alongside the conversation. Chained slash commands still point at that review. /ultrareview now accepts wording such as “review my auth changes”: it reviews the current branch and attaches those words as a note. In non-interactive runs, /code-review ultra now truly starts a cloud review instead of silently falling back to a local one, and bad arguments return useful feedback. It is like sending a document to an editor while you keep writing at your desk; the editor still knows which document to read.
When editing in --ax-screen-reader mode, deleting a word or line with Option+Delete, Ctrl+W, or Ctrl+U now makes the screen reader announce what disappeared. At the end of the input, VoiceOver reports a typed space instead of wrongly saying “new line.” In the plugin and settings panels, terminal focus now follows the row selected with arrow keys, so screen readers and magnifiers can locate it. It is like having someone read a draft aloud while pointing to the words you just erased and the line you are currently reading.
Pasting several lines into certain terminals used to turn their newline encoding, Ctrl+J, into literal j characters and squash everything into one line. The pasted line breaks are now preserved. Pressing Left Arrow immediately after editing no longer silently discards the whole conversation; Claude asks for confirmation first. If you enter agent view, the screen for work running in the background, Esc returns you to the conversation it backgrounded. It is like pasted paragraphs staying as paragraphs, with someone asking before an accidental elbow bump throws your letter away.
When an external tool suddenly will not work, claude mcp list and /mcp now show the HTTP status and error text instead of merely saying it could not connect. MCP, the Model Context Protocol for connecting AI to outside tools, also warns when a configuration value has invisible whitespace at its beginning or end; those spaces can break an otherwise identical URL or key. The startup notice about MCP servers needing authentication no longer counts claude.ai connectors that are not connected there. It is like support telling you whether a phone number is wrong or busy, then removing ghost contacts from the list.
For Windows users, a path such as C:\Users\unicorn could have its \u segment mistaken for Unicode encoding and turned into CJK characters, making the file unreachable as if its address had changed. The original path is now kept intact. A long selection from an IDE also no longer becomes mojibake when it is cut in the middle of an emoji, and a tool-executor error is no longer silently lost. It is like a postal worker finally delivering to the address written on the envelope, rather than treating \u as a secret code and pretending a torn envelope never arrived.
After compacting a long conversation, /context no longer reports the old pre-compaction token count, the text units AI processes and bills for. Interrupted tools no longer leave a false “Request interrupted by user” message or half a tool_use record, and prompt history is less likely to lose or duplicate entries when writes collide or fail. A context-overflow retry with a large thinking budget stops resending the same doomed request. Backgrounded and headless or SDK fork sessions also retain their lineage after compaction, while malformed delta attachments no longer doom every resumed turn. It is like an account book that stops duplicating or losing entries and still shows where each receipt came from after rebinding.
When using models through Bedrock, gateway spend metering now charges application-inference-profile ARNs and other IDs mapped to an upstream model at the configured model’s rate, rather than pricing them incorrectly because the identifier differs. Pull-request events are less likely to vanish if a session ends right after creating or linking a PR, a request to merge a set of code changes. The Bedrock setup wizard can also verify assume-role profiles in partitioned AWS regions and proxy-only networks. Turn durations no longer rarely become negative or wrong after a system clock change. It is like a taxi switching to a dependable meter, so fares, receipts, identity checks, and trip time do not drift when the dashboard clock jumps.
Auto mode no longer opens a permission dialog for every dangerous rm, background &, or suspicious Windows path; its safety classifier decides instead. In plan mode, Bash commands the static analyzer cannot prove read-only go to that classifier too, reducing needless interruptions. Sandbox restrictions, the boundaries that confine a program, are stronger for IDE interactions. Trust dialogs now name the repository root they cover, while hooks, small programs triggered by events, in agent frontmatter require the agent file’s own folder to be trusted. Benign feature and cost toggles in server-managed settings no longer request approval every time. It is like an entry desk that distinguishes an ordinary visitor, a dangerous parcel, and which building a pass actually opens.
When you want a quick answer, Claude no longer starts /deep-research on its own and stretches the process out; it begins only when you invoke the command manually. Conversely, if switching models through /config model=<x> or Remote Control changes fast mode, the mode that prioritizes quicker responses, Claude now announces that change. You no longer have to assume the speed setting is what it was a moment ago. It is like a coffee shop not upgrading your takeaway order to a slow hand brew without asking, while clearly flipping a sign if the bar and serving speed change.
When writing Markdown settings for an agent or skill, an agent name containing : is now rejected because that character is reserved for plugin namespacing, the prefix system that distinguishes plugin sources. This prevents future name collisions. A skill frontmatter, the settings block at the top of the file, with context: fork now runs in the background by default; add background: false to keep it in front. Boolean values no longer need to be only true or false: case-insensitive yes, no, on, off, 1, and 0 work too. It is like setting rules for community mailboxes: colons reserve address zones, delivery has a clear default, and several natural ways of saying on or off are understood.
In a project with watched folders nested many layers deep, deleting or moving the whole directory tree could make Claude crash from excessive call depth. Rendering an equally deep UI tree could fail too. Both cases are fixed. This adds no new button for an ordinary project, but it makes reorganizing a large, deeply nested one less likely to end in a sudden failure. It is like moving a tower of stacked storage boxes without the mover fainting at the bottom layer or the display rack collapsing at its deepest shelf.
Changed `/code-review` to run as a background subagent, so review work no longer fills your conversation and keeps stacked slash commands as its review target
Fixed `/ultrareview` failing on descriptive arguments like "review my auth changes" — they now run a review of your current branch with the text applied as a note to the findings
Fixed `/code-review ultra` silently running a local review in non-interactive sessions — it now launches the cloud review
Improved `/ultrareview` error feedback so Claude can correct an invalid argument instead of retrying it unchangedAdded screen-reader announcements of deleted text for word and line deletions (`Option+Delete`, `Ctrl+W`, `Cmd+Backspace`, `Ctrl+U`, `Ctrl+K`) in `--ax-screen-reader` mode
Fixed VoiceOver reading "new line" instead of echoing the typed space at the end of the input in `--ax-screen-reader` mode
Fixed plugin and settings panels not moving the terminal cursor to the focused row, so screen readers and magnifiers can follow arrow-key navigationFixed the left arrow key discarding the conversation with no undo: presses right after editing now ask to confirm, and Esc in the agent view returns to the conversation it backgrounded
Fixed multi-line paste collapsing into one line with `j` in place of newlines in terminals that encode pasted newlines as Ctrl+JAdded HTTP status and error text to `claude mcp list` and `/mcp` when a server fails to connect, and a warning for MCP config values with hidden leading or trailing whitespace
Fixed the "N MCP servers need authentication" startup notice over-counting claude.ai connectors that aren't connected in claude.aiFixed Windows paths with `\u`-prefixed segments (like `C:\Users\unicorn`) being corrupted into CJK characters in tool inputs, which made those files inaccessible
Fixed mojibake when a long IDE selection was truncated mid-emoji, and a case where a tool executor error could be silently droppedFixed `/context` reporting stale pre-compact token usage after compacting from the message picker
Fixed an engine teardown race that could start and abandon a phantom turn, and made input pushed after close consistently rejected
Fixed spurious "[Request interrupted by user]" messages after interrupted tool calls, and an unpaired `tool_use` block left in the transcript when a tool aborted mid-response
Fixed prompt history entries being dropped or duplicated when history writes raced or failed
Fixed a retry loop that re-sent identical doomed requests after a context-overflow error with a large thinking budget; `Ctrl+B` backgrounding now applies the same background-shell caps as other paths
Fixed fork-session lineage being lost after compaction in headless and SDK sessions
Fixed a resumed session failing every turn, or crashing on resume, when its history held a malformed delta attachment
Fixed remote sessions continuing to send heartbeats after their worker was replaced, which left long-lived desktop and IDE processes retrying a rejected request every few seconds foreverFixed gateway spend metering to price Bedrock application-inference-profile ARNs and other config-mapped upstream model IDs at the configured model's rates
Fixed pull request events occasionally being lost when a session exited immediately after creating or linking a PR
Fixed the Bedrock setup wizard failing profile verification for assume-role profiles in partitioned AWS regions and on proxy-only networks
Fixed rare negative or incorrect turn duration measurements after a system clock adjustment by timing turns with a monotonic clockFixed agent frontmatter hooks running from untrusted folders: hooks now require the agent file's own folder to have accepted workspace trust
Improved auto mode: the dangerous-rm, background-`&`, and suspicious-Windows-path checks no longer open permission dialogs; the auto-mode classifier adjudicates them instead
Improved sandbox command restrictions for IDE interactions
Improved trust dialogs to name the repository root the grant covers
Changed plan mode with auto to no longer prompt for Bash commands the static analyzer can't prove read-only; the auto-mode classifier judges them instead
Changed server-managed settings so benign feature and cost toggles no longer trigger the settings-approval promptChanged `/deep-research` to start only when invoked manually; Claude no longer launches it on its own
Added an announcement when fast mode changes as a result of switching models via `/config model=<x>` or Remote ControlChanged agent markdown files to reject agent names containing `:`, which is reserved for plugin namespacing
Changed skills with `context: fork` to run in the background by default; opt out per skill with `background: false`
Added `yes`/`no`/`on`/`off`/`1`/`0` (case-insensitive) as accepted values for skill and plugin frontmatter booleans, alongside `true`/`false`Fixed crashes (maximum call stack exceeded) when a deeply nested watched directory tree was deleted or moved, and when rendering deeply nested UI trees