發布日期：2026-06-19
Claude 現在不會在你未明講放棄工作內容時執行破壞性的 Git 指令。如果你沒有明確叫它放棄目前的工作進度，系統會主動攔截那些危險的 Git 指令（像是 git reset --hard、git clean -fd 等會刪掉檔案的操作）；對於 git commit --amend，如果不是它自己這次對話做的提交，它也不准修改。如果你用的是 Terraform 或 Pulumi 等基礎設施工具，除非你點名要刪除特定環境，否則系統會嚴格禁止執行 destroy 把伺服器砍掉。此外，以前排程任務或 Webhook 收到的訊號會被系統誤當成你的鍵盤輸入，甚至不小心幫 AI 按下「同意」按鈕，現在這個安全漏洞也被補上了。這就像給除草機加裝了安全鎖，除非你明確解鎖並指著那片花圃，否則它絕對不會把珍貴的玫瑰花當成雜草一起絞掉；而且外頭飛進來的鳥也不會再誤觸開關啟動機器了。
讓 AI 分工合作的代理人團隊現在更可靠了。以前如果你用 tmux（一種終端機分割工具）開出新隊友，遇到電腦載入太慢就會啟動失敗，甚至你在主畫面打的字會莫名其妙跑到新隊友的視窗去，現在這些互相干擾的問題都解決了。而且，以前隊友交代給子代理人去跑的背景任務，只要隊友一結束回合就會被無情砍掉，現在任務會順利跑完。針對 AI 的搜尋與思考，以前子代理人去查網頁（WebSearch）老是查出空白結果，或是 AI 花了一大堆時間思考（thinking block）卻沒吐出任何實質對話就結束回合，現在會再提示一次模型產生可見輸出。最後，因為設定衝突導致生出子代理人時狂跳 400 Extra inputs are not permitted 錯誤的毛病也修好了。這就像是工廠裡的生產線，以前新來的機器人開機太慢會讓產線停擺，交代給實習生的工作也會因為主管下班就被強制丟進垃圾桶；現在大家交接順暢，就算主管下班，實習生也會把手上的工作乖乖做完。
這次的介面變得更清爽順手了。設定面板（/config）現在操作更直覺，你可以按 Enter 或空白鍵直接切換開關，按 Esc 也會乖乖存檔關閉，而不是像以前一樣取消還原；想知道有什麼快捷指令，也可以直接打 /config --help 查詢。畫面上，我們移除了每次開機都會顯示的「設定問題」警告，現在想檢查問題請直接用 /doctor 指令。選到已淘汰或會自動升級的模型時，現在會看到警告。針對各種當機與干擾也做了大掃除：修好了在 Windows 終端機跑大量代理人時整個畫面崩潰混亂的問題、解決了在 vim 模式下游標會卡在輸入框上面的靈異事件、拔掉了專注模式下每句話底下吵死人的 Hooks 執行計時文字，也確保你自己寫的技能不會在清單上像影分身一樣出現好幾次。最後，如果你在無頭環境（headless，沒有螢幕）需要認證 MCP 伺服器，現在也不會再因為系統亂拿假工具給 AI 用而當機了。這就像把雜亂的工作桌重新整理，不該響的警報器拔掉、壞掉的按鈕修好，連會讓螢幕閃爍的舊燈管也一併換新，讓你專心工作不被打擾。
Claude will no longer execute destructive Git commands unless you explicitly ask it to discard your local work. The system will now actively block destructive Git commands (like git reset --hard and git clean -fd that irreversibly delete files). It also refuses to run git commit --amend unless the agent itself made that specific commit during the current session. For infrastructure-as-code tools like Terraform or Pulumi, running a destroy command is now strictly forbidden unless you specifically name the target stack. Additionally, scheduled tasks and Webhook triggers used to be incorrectly interpreted as your keyboard inputs, sometimes accidentally 'approving' an AI action on your behalf; this dangerous loophole has been firmly closed. It's like putting a safety guard on a lawnmower: unless you explicitly flip the switch and point at a specific patch of dirt, it absolutely refuses to shred your prized rose bushes, and a stray bird landing on the handle can no longer accidentally start the engine.
Team-based AI collaboration just got a lot more dependable. Previously, spawning a teammate in a tmux pane (a terminal multiplexer) could fail if your shell initialization was slow, and keystrokes you typed on the main screen would bizarrely leak into the new teammate's window; these chaotic overlaps are now fixed. Also, if a teammate dispatched a background task to a subagent, that task used to be ruthlessly killed the moment the teammate finished its turn—now it survives to completion. On the cognitive side, subagents using WebSearch no longer return totally blank results, and if the AI spins its wheels generating a "thinking block" but fails to produce actual visible output, it won't just silently give up anymore—it will re-prompt the model once to produce visible output. Finally, we squashed the annoying 400 Extra inputs are not permitted errors that plagued subagent spawns on certain configurations. It's like a busy factory floor: previously, a slow-booting new robot would jam the assembly line, and any chore handed to an intern was immediately thrown in the trash the moment their manager clocked out for the day. Now, handoffs are flawless, and the interns faithfully finish their tasks even after their boss goes home.
The interface is now noticeably cleaner and more intuitive. The settings menu (/config) makes sense now: you can hit Enter or Space to toggle switches, and pressing Esc actually saves and closes instead of reverting all your changes. You can also run /config --help to easily look up shorthand keys. We've scrubbed the constant 'setup issues' nag from the startup logo; if you want a system health check, just run /doctor. You'll now get clear warnings if you request a deprecated model or one that automatically updates to a newer version. We also squashed a ton of screen-corrupting bugs: the total UI meltdown (merged text and duplicated spinners) in Windows Terminal under heavy subagent loads is fixed, your cursor won't get stranded above the prompt in vim mode anymore, the noisy 'PostToolUse' timing lines cluttering focus mode are gone, and user-level skills won't infinitely duplicate themselves in your autocomplete list. Finally, headless environments (running without a screen) will no longer crash by exposing fake auth-stub tools to the AI when an MCP server demands authentication. It's like thoroughly decluttering your workstation: broken buttons now work predictably, phantom alarms have been silenced, and the flickering monitor that scrambled your spreadsheets has finally been replaced, leaving you with a distraction-free space.
Improved auto mode safety: destructive git commands (`git reset --hard`, `git checkout -- .`, `git clean -fd`, `git stash drop`) are now blocked when you didn't ask to discard local work, `git commit --amend` is blocked when the commit wasn't made by the agent this session, and `terraform destroy`/`pulumi destroy`/`cdk destroy` are blocked unless you asked for the specific stack
Fixed scheduled task and webhook trigger deliveries being treated as keyboard input; they now classify as task notifications and can no longer approve a pending action or set the session title in auto modeFixed tmux teammate panes failing to launch when the shell has slow rc-file initialization, and keystrokes typed during agent spawn leaking into the new tmux pane instead of the leader prompt
Fixed background tasks started by a teammate being killed when the teammate finishes a turn
Fixed WebSearch returning empty results in subagents
Fixed turns silently completing with no visible output when the model returned only a thinking block; Claude now re-prompts once
Fixed `thinking.disabled.display: Extra inputs are not permitted` 400 errors on subagent spawns and session-title generation for affected configurationsChanged `/config` toggle behavior: Enter and Space both change the selected setting, and Esc now saves and closes instead of reverting
Added `/config --help` to list all available shorthand keys for `/config key=value`
Added `attribution.sessionUrl` setting to omit the claude.ai session link from commits and PRs in web and Remote Control sessions
Removed the startup "setup issues" line under the logo — run `/doctor` to see configuration issues or use `--debug`
Added a warning when the requested model is deprecated or automatically updated to a newer model, shown on stderr in print mode (`-p`) and now also covering models set in agent frontmatter
Fixed fullscreen TUI corruption (statusline mid-screen, duplicated spinner rows, merged text) in Windows Terminal under heavy nested-subagent load
Fixed the terminal cursor being stranded above the prompt after navigating history in vim mode with the native cursor enabled
Fixed user-level skills appearing multiple times in slash-command autocomplete when multiple plugins are enabled
Fixed MCP servers requiring authentication exposing auth-stub tools to the model in headless/SDK mode
Fixed focus mode showing "Ran N PostToolUse hooks" timing lines under each response