預設改為手動模式,不會再趁你不注意偷偷繼續
現在 Claude 做事前預設會先問你,放著不回覆也不會自己繼續。不管是指令行、VS Code 還是 JetBrains,權限模式都改成了「手動(Manual)」。當它彈出對話框問你問題(AskUserQuestion)時,以前如果放著不管,它可能會自己當作你同意而繼續跑;現在預設會一直等你回答,除非你自己去設定檔裡開啟「閒置超時自動繼續」的功能。就像教練本來會自己決定要不要換球員,現在預設改成換人前必須請示總教練,而且沒拿到簽名絕對不會自己放行。
發布日期:2026-07-03
現在 Claude 做事前預設會先問你,放著不回覆也不會自己繼續。不管是指令行、VS Code 還是 JetBrains,權限模式都改成了「手動(Manual)」。當它彈出對話框問你問題(AskUserQuestion)時,以前如果放著不管,它可能會自己當作你同意而繼續跑;現在預設會一直等你回答,除非你自己去設定檔裡開啟「閒置超時自動繼續」的功能。就像教練本來會自己決定要不要換球員,現在預設改成換人前必須請示總教練,而且沒拿到簽名絕對不會自己放行。
這次把背景代理(agent)從休眠中喚醒的毛病治好了。以前你的電腦如果睡著又醒來,正在跑的任務可能會默默停住不幹活;如果它卡住被系統重新啟動,還會把之前你按 Esc 取消掉的步驟又拿出來跑一次,現在這些亂象都沒了。另外,如果程式崩潰留下了一個寫著進程代號的鎖定檔(daemon.lock),而作業系統剛好又把這個代號分給別的程式,以前代理就會因為誤認而永遠不肯啟動,現在修好這個檢查邏輯了;被限流(rate limit)打斷的子任務現在也會明確報錯,不會假裝自己跑完卻丟出空結果。這就像警衛打瞌睡醒來,以前會忘記自己巡邏到哪,或是看到地上掉著上一班的臂章就以為別人還在上班而拒絕接班,現在他醒來後會確實確認身分並正確接手工作。
對於依賴螢幕報讀軟體(screen-reader)和放大鏡的使用者,現在的體驗變得更友善了。系統會自動隱藏那些只是用來裝飾的符號,對話紀錄裡的特殊記號會被念成簡短的標籤,而表格也會貼心地用「標題:內容」的格式一句句念出來,MCP(讓 AI 呼叫外部工具的協定)的清單也能正確被游標追蹤到了。對於使用 tmux(終端機的畫面分割工具)的人,現在開啟了同步輸出功能,畫面不會再像閃光燈一樣閃爍;代理清單裡也不會再跑出亂七八糟的控制碼。這就像以前導覽員會把裝飾花紋也念出來,讓聽眾一頭霧水,現在他只會把重點整理成好懂的句子,而且投影幕也不會再瘋狂閃爍傷眼睛。
用語音輸入沒聲音或安裝因記憶體不足失敗時,現在你會看到正確的錯誤原因。如果你錄了半天其實沒錄到聲音,以前它會報錯說「語音連線失敗」,現在它知道只是沒錄到聲音而已;安裝被作業系統強制砍掉時,安裝腳本也會清楚告訴你原因。載入外掛(plugin)時,即使將 --plugin-dir 參數放在 agents 指令後方也能正常載入,也不會因為你用了工作樹(worktree,平行工作的 Git 功能)就找不到專案專屬的外掛。還有,如果你在設定檔裡把 MCP 伺服器名單寫錯格式,開機也不會直接崩潰了。這就像去買東西,以前卡片沒錢刷不過店員只會說「連線失敗」,現在他會直接告訴你「餘額不足」,不會讓你瞎猜。
Claude now asks you before taking action by default, and won't auto-continue if you don't reply. Across the CLI, VS Code, and JetBrains, the permission mode is now explicitly "Manual". When it pops up a dialog asking you a question (AskUserQuestion), it will no longer silently assume you agreed if you step away; by default, it will wait patiently for your answer, unless you specifically opt into the "idle timeout auto-continue" feature in your settings. It's like a sports coach who used to swap players autonomously but is now required by default to get the head coach's signature before making any substitutions, strictly waiting until that signature is secured.
We've cured background agents of their bad waking habits. Previously, if your computer went to sleep and woke back up, a running task might silently stall and refuse to work; and if a stalled task respawned, it might foolishly re-run a step you had explicitly cancelled with Esc. These confusing behaviors are gone. Additionally, if the program crashed and left behind a stale lockfile (daemon.lock) whose process ID got recycled by the OS, the agent used to get confused and refuse to ever start again; this detection logic is now fixed. Subagents cut off by a rate limit will also now fail properly instead of pretending they finished successfully and returning an empty void. It's like a security guard waking up from a nap: they used to forget their patrol route or see an old name badge on the floor and refuse to start their shift; now they properly verify their status and resume duty flawlessly.
The experience for users relying on screen readers and magnifiers is now vastly improved. The system automatically hides purely decorative symbols, reads complex transcript symbols as short, descriptive labels, and parses nested tables line-by-line as Header: value, making it much easier to digest. The MCP (a protocol letting AI use external tools) server list also now correctly tracks focus for accessibility tools. For those using tmux (a terminal multiplexer), we enabled synchronized output so the screen no longer flickers like a strobe light; and raw control codes will no longer bleed into the agent view. It's like an audio tour guide who used to confusingly describe every decorative carving on the wall; now they gracefully summarize the important exhibits, and the museum's flickering lights have finally been replaced.
You will now see accurate error reasons when voice dictation captures silence or an installation fails due to out-of-memory. If your recording ends up capturing pure silence, the system used to misleadingly throw a "Voice connection failed" error; now it understands you just didn't record any sound, and the installer script will explicitly explain what happened if the OS kills it for memory instead of failing mysteriously. For plugins, the --plugin-dir flag now properly loads agents even if placed after the agents command, and project-scoped plugins won't get lost just because you are using a Git worktree (parallel working copies). Additionally, misformatting the MCP server list in your settings file will no longer crash the entire program at startup. It's like going to the store: the cashier used to just say "connection failed" when your card was empty, but now they directly tell you "insufficient funds" so you aren't left guessing.
Changed `AskUserQuestion` dialogs to no longer auto-continue by default; opt into an idle timeout via `/config`
Changed the "default" permission mode to "Manual" across the CLI, `--help`, VS Code, and JetBrains; `--permission-mode manual` and `"defaultMode": "manual"` are accepted alongside `default`Fixed background sessions silently stopping mid-turn after sleep/wake or when reopening a stalled session
Fixed background sessions re-running a turn cancelled with Esc after a stall respawn
Fixed background agents never starting again after a crash left a stale `daemon.lock` whose PID the OS reused
Fixed background-agent daemon handover so a reinstalled older build can no longer take over the daemon; build recency is now judged by the version's embedded build timestamp
Fixed background-agent roster issues: transient corruption permanently disabling orphan cleanup, older binaries not preserving fields written by newer versions, and socket auth tokens being stripped during daemon restarts
Fixed subagents cut off by a rate limit before producing any text output returning an empty result instead of failing cleanlyImproved screen-reader output: decorative glyphs are now hidden, transcript symbols read as short labels, and nested tables read as `Header: value.` lines
Fixed `/mcp` server list not tracking focus for screen readers and magnifiers
Fixed rendering flicker under tmux 3.4+ by enabling synchronized terminal output
Fixed control bytes from background-agent output reaching the terminal in the agent viewImproved the install script to explain when installation is killed by the system running out of memory
Fixed voice dictation showing a misleading "Voice connection failed" message when a recording captures no audio
Fixed a crash at startup when `disabledMcpServers` or `enabledMcpServers` in `.claude.json` is set to a non-array value
Fixed `claude agents --plugin-dir <dir>` not showing the plugin's agents and skills in the agent view when the flag is placed after `agents`
Fixed project-scoped plugins not loading correctly from git worktrees of the same repository