發布日期：2026-07-08
現在你打開代理（agent，幫你跑任務的 AI 分身）清單，版面不但不會被切掉，每行還會用有顏色的字配上重點標題來告訴你它在幹嘛，不再只是印一堆看不懂的指令程式碼。被卡住的任務點開也能清楚看到是在等什麼回應。不管代理是在編輯、合併還是留言，現在都會把對應的 PR（pull request，請求將程式碼合併的流程）直接連起來方便你點；甚至你把它從「失敗」喚醒，或者它沒說話時，狀態都不會再跳錯。在手機或網頁上遙控看進度，也不會再一直卡在過期的「執行中」。就像去餐廳看候位看板，以前只亮著一堆亂碼和死當的畫面，現在會清楚寫著「第 3 桌客人點菜中」或「在等廚房回覆」，一目了然。
把權限交給系統自己跑的自動模式（auto mode），現在加強了防護網。以前如果它遇到一個讀不懂的變數，可能會直接執行危險的刪除指令（rm -rf）把東西清光，現在它會先停下來問你。同時，它也學會防止自己被對話紀錄裡的假核准訊息騙去執行指令，背景任務的通知會明白標示「這沒有經過人類同意」，並且會封鎖任何試圖竄改對話紀錄檔的行為。就像請打掃阿姨來家裡，以前她遇到不認識的箱子可能就直接丟掉，或是聽信留聲機裡的假錄音就去開金庫；現在她遇到不確定的東西會先打電話問你，而且絕對不聽非本人的指令。
在背景幫你更新軟體時，以前會把檔案全塞進記憶體導致瞬間飆高，現在改用直接寫進硬碟的方式（串流），一口氣省下大約 400MB 的記憶體空間。同時，系統也修好了幾種會當機或卡死的狀況：例如在你跑指令時把資料夾刪掉、監控檔案的程式還在跑卻被關掉，或是代理正好在重啟時你硬要去連線它，現在都不會讓整個程式直接崩潰。另外，在 Windows 上移除工作樹（worktree，讓同一個專案多開平行工作副本的 Git 功能）時，如果裡面有捷徑連結（symlink），再也不會誤刪外面的檔案了。就像手機在背景下載更新不再讓你玩遊戲卡死，而你不管是在換房間還是搬東西，系統都能穩穩處理，不會一有意外就整台死機。
內建的 /doctor 指令現在能替你診斷並修好設定。它不但能幫你找出設定上的疑難雜症，還能直接動手幫你修好；同時你也可以打比較好記的 /checkup 來呼叫它。這就像去診所，醫生會幫你詳細檢查還順便開藥把病治好。
這次順手掃除了幾個煩人的小怪。像是：對話剛好到達次數上限（--max-turns）時你打的字不再會默默被吃掉；當 Bash 指令印出超過 3 萬字時，這份 PR（pull request，請求合併程式碼）也不會因此在清單中失蹤；如果你在 Cowork 虛擬機（VM-mode）跑本地代理，也不會無緣無故叫你重新登入。針對外掛（plugin，擴充功能的附加元件）的語言伺服器，如果其中一個當機了，現在不會連累另一個正常的伺服器罷工。而當你匯入 MCP（讓 AI 呼叫外部工具的標準協定）的伺服器名單時，如果遇到不支援的特殊字元不會再整個卡死，它會跳過壞名字並把剩下的順利匯入；還有，這版先佔用了「Claude Browser」這個保留名字以防衝突。這就像一條水管，把過去會漏水、塞住或是被一顆小石頭卡死整條水管的接口全部修好了。
When you open the agent list (the list of AI assistants running tasks for you), the view will no longer clip off the screen, and each row now replaces raw technical tool code with color-coded status words and a clean headline explaining what it's doing. If a session is blocked, peeking at it shows exactly what it is waiting for. Any agent working on a PR (pull request, a proposal to merge code) — whether editing, merging, or commenting — will now link that PR right in the list. State bugs are squashed too: agents won't stay stuck on "failed" after you resume them, nor will they falsely flip statuses when staying quiet. Remote viewing on mobile or web also stays fresh instead of freezing on "Running". It's like checking a restaurant's waitlist board: it used to display raw data or freeze, but now it clearly reads "Table 3 ordering" or "Waiting on kitchen," keeping you perfectly informed.
Auto mode (the mode where the system runs tasks autonomously) has added stronger safety nets. Previously, if it encountered a variable it couldn't resolve, it might dangerously execute a broad delete command (rm -rf) and wipe things out; now it will stop and ask you first. It also learned to defend against fake approvals slipped into the transcript: background task notifications will explicitly state when no human has actually given input, and the system actively blocks any attempts to tamper with session transcript files. It's like hiring a cleaner who used to toss out boxes she didn't recognize or open the safe because of a fake audio recording; now she stops to call you when unsure, and completely ignores orders not directly from you.
When auto-updating in the background, the download now streams directly to the disk instead of buffering entirely in memory, slashing peak memory usage by about 400 MB. Along with this, several crash and hang scenarios are fixed: deleting, locking, or unmounting the current directory while a command is running; closing a file watcher while a directory scan is in flight; or trying to attach to a background agent right when it's restarting mid-upgrade — none of these will cause the program to catastrophically crash anymore. Additionally, when removing a worktree (a Git feature that creates parallel working copies) on Windows, the presence of a symlink (shortcut) inside it will no longer cause it to accidentally delete files entirely outside the worktree. It's like your phone downloading an update without stuttering your game, and no matter how you shuffle furniture around, the house won't collapse from a sudden unexpected move.
The built-in /doctor command can now diagnose and fix your setup. It can not only diagnose configuration issues but actively fix them for you; you can also use the handy alias /checkup to summon it. It's like visiting a clinic where the doctor gives you a thorough exam and hands you the medicine to cure it on the spot.
A handful of annoying gremlins have been exterminated. For instance, messages you type right as the turn hits the --max-turns limit will no longer silently vanish into the void; and a PR (pull request, a proposal to merge code) created inside a Bash call that spits out over 30,000 characters won't be excluded from linking; Cowork VM-mode users won't be falsely locked out with login errors either. For plugin (add-ons that extend capabilities) language servers, one failing to initialize will no longer completely block a perfectly valid server from handling the same file type. Furthermore, importing MCP (a standard protocol letting AI use external tools) servers won't freeze entirely if one name contains unsupported characters; it now flags the bad ones and continues importing the rest. Also, "Claude Browser" is now a reserved MCP server name. It's like plumbing maintenance: fixing all the little joints that used to leak, clog, or cause the entire pipe system to back up over a single pebble.
Fixed background agents staying shown as "failed" or "completed" in the agent list after being resumed with `SendMessage`
Fixed background jobs flipping from "needs input" back to "working" in the agent list when the agent's turn contained no readable text
Fixed the agent view rendering one line too high and clipping its header when the job list slightly overflowed the screen
Improved agent view: sessions that edit, merge, comment on, or push to an existing PR now link it in `claude agents`
Improved agent view: rows now show a colored state word and a classifier-written headline instead of raw tool call text, and the peek opens with full status including the exact ask for blocked sessions
Fixed background tasks in the web and mobile Remote Control panels showing stale "Running" status by forwarding full task state on every membership changeAdded an auto mode rule that blocks tampering with session transcript files
Improved auto mode to ask before running `rm -rf` on a variable it can't resolve from context
Background task notifications now explicitly state that no human input has occurred, preventing fabricated in-transcript approvals from being acted onAuto-update binary downloads now stream to disk instead of buffering in memory, cutting the updater's peak memory usage by roughly 400 MB
Fixed `claude attach` erroring when a background agent was mid-upgrade restart instead of waiting for it to come back
Fixed a Windows crash when the directory Claude was launched from is deleted, locked, or unmounted while a command is running
Fixed a crash when a file watcher was closed while a directory scan was still in flight
Fixed Windows worktree removal deleting files outside the worktree when an NTFS junction or directory symlink existed inside it`/doctor` is now a full setup checkup that can diagnose and fix issues; `/checkup` is its aliasFixed a message sent while Claude was working being silently lost when the turn ended at the `--max-turns` limit
Fixed session-to-PR linking missing a PR created in a Bash call whose output exceeded the 30K inline limit
Fixed `claude mcp add-from-claude-desktop` getting stuck when a server name contains unsupported characters; invalid names are now reported and remaining servers still import
Fixed `--json-schema` silently producing unstructured output when the schema was invalid, and schemas using the `format` keyword being rejected
Fixed a plugin LSP server that fails to initialize preventing a valid LSP server from another plugin handling the same file extension
Reserved the "Claude Browser" MCP server name (alongside "Claude Preview") ahead of the Claude Desktop pane rename; user-configured MCP servers can no longer register under either name
Fixed Cowork VM-mode local-agent sessions failing to start with "Not logged in · Please run /login" on CLI 2.1.203+
Fixed project verify skills being rewritten on every session instead of only when a documented command changed