發布日期：2026-06-22
使用終端機指令的體驗大升級了！現在你在終端機使用 bash 模式（打 ! 開頭加上指令）時，Claude 不會只默默把結果記下來，而是會主動對執行結果做出回應（你如果不喜歡，也可以去設定關掉）。對於需要登入的外部工具（MCP），你現在可以直接在指令列輸入 claude mcp login 來登入或登出，不需要再切換到互動選單；如果你是在別台電腦透過 SSH 連線過來，還能用無瀏覽器模式完成認證。萬一你打錯了工具伺服器的名字，系統現在會聰明地猜測你是不是拼錯了，並把太長的伺服器清單縮短顯示。這就像你的助理原本只會幫你抄寫跑腿回來的發票，現在他跑完腿會順便跟你報告結果；而且現在你要進金庫，不用再跑回大廳填表，在門口直接刷卡就能進去，就連你名字寫錯一點點，門禁系統也會認得你。
子代理人與背景任務現在變得更好管教，畫面也不再亂閃了。以前在背景跑的子代理人遇到需要權限時，要嘛自己亂拒絕，要嘛卡在那邊；現在它會把權限請求丟回你的主對話框裡，並清楚標示「是哪個小助手在要權限」，你按 Esc 就能單獨拒絕那個工具。如果你派出去的小助手因為要求特定格式（JSON schema）連續五次失敗，系統現在會自動中止它，不會讓它陷入無限重試的死迴圈。我們也修好了一大堆畫面干擾：像是從代理人介面退出來時畫面不會再黏著舊紀錄、背景任務的預覽框不會再把原始的工具名稱閃爍出來、切換深淺色終端機時不再被突如其來的白光或黑畫面閃瞎，連同在主視窗按 Esc 或 Ctrl+C 時，系統現在也能正確煞車了。這就像是你派出去實習的菜鳥，以前遇到問題只會自己瞎忙或裝死，現在懂得打電話回來請示；而且他們傳回來的監視器畫面終於修好了，不會再看到殘影或莫名其妙的雪花雜訊。
面對電腦休眠或網路環境變化，系統的防護力更強了。以前如果你在 AI 正在產生文字（streaming）時把電腦蓋上休眠，醒來後它會跳出「找不到內容」或格式解析錯誤；現在它能完美熬過這段時間，醒來繼續運作。對於企業用戶，系統現在會正確顯示按照用量計費的對話花費；如果是透過 AWS 登入的企業版，/login 介面也新增了「重新整理憑證」的選項，不用再等到憑證徹底過期才被踢下線。如果你使用 Chrome 擴充功能並且關掉了權限警告，多開終端機時不會再發生分頁群組無法隔離的問題。而對於帶領團隊的主代理人，現在如果開出新隊友（透過 tmux 等背景視窗），新隊友也會乖乖繼承隊長所設定的努力度（--effort）。此外，系統的無限重試限制最高設為 15 次，避免失控浪費。這就像你把辦公室電腦關掉睡個午覺，以前醒來資料會全毀，現在它會自己保存在雲端；而且大老闆交代專案的認真程度，現在也會準確傳達給每一個下屬，不會有人偷懶。
這次也針對工作環境做了許多小翻新。在管理外部套件（/plugin）的「已安裝」分頁裡，現在多了一個專屬的「技能（Skills）」區塊，幫你把東西分得更清楚；而且就算你已經滑到最上面，也不會再被一個「上面還有更多」的錯誤提示騙了。查看工作流詳細資料時，你現在可以按 f 鍵來篩選任務狀態。對於技能的設定檔（SKILL.md），系統現在對寫法的包容度更大，你可以用烤肉串（kebab-case）或是駝峰（camelCase）寫法，萬一格式寫錯一點點，它也會默默載入功能而不是直接死機。此外，當你用 /review 來審查 Pull Request 時，它現在會直接使用 /code-review medium 這個效能更好、更精準的引擎。針對畫面細節，AI 回答裡如果在文字加刪除線，不會再印出兩個波浪號 ~~ 而是會正確畫一條線；代理人面板上用滑鼠框選的字也不會再刪除後還殘留反白；權限面板如果選項字太多，左邊的編號也終於對齊了。這就像把辦公桌重新收納，原本混在一起的文具現在有了專屬抽屜，找檔案還能按類別翻找；而且即使你的請假單格式不小心寫錯一行，人事部也會幫你處理，不會直接退件。
The terminal command experience just got a massive upgrade! When you use bash mode (starting with !), Claude will no longer just silently absorb the terminal output as context; it will now proactively respond and analyze the result right away (if you prefer the old silent behavior, you can toggle it off in your settings). For external MCP tools that require authentication, you can now seamlessly log in or out directly from your shell using claude mcp login without being forced into an interactive menu. If you're working over an SSH connection, there's even a headless, no-browser fallback. And if you happen to typo a tool server's name, the CLI is now smart enough to suggest the closest match and cleanly truncate unwieldy lists. It's like having an assistant who used to just silently file away the receipts after running an errand, but now actively hands you a summary report; plus, you no longer have to walk all the way back to the lobby to fill out a security form to access the vault—you just swipe your card at the door, and the guard even recognizes you if you slightly misspell your own name.
Subagents and background tasks are now far more obedient and visually stable. Previously, a background subagent hitting a permissions barrier would either quietly auto-deny itself or get stuck. Now, it neatly surfaces the permission prompt back to your main session, clearly identifying which assistant is asking, and pressing Esc will cleanly deny just that specific tool. We've also added a failsafe: if an agent fails a structured JSON schema validation 5 times in a row, it now automatically aborts instead of spiraling into an infinite loop. On the visual side, we've squashed a swarm of UI glitches: closing the agents view no longer leaves ghost text painted on your screen, background task previews won't flash ugly raw tool names before loading, switching into the agents view on light terminals won't blind you with a sudden dark-theme flash, and pressing Esc or Ctrl+C in the main view while agents run in the background is now perfectly responsive. It's like managing a team of interns: they used to panic silently or spin in circles when they hit a locked door, but now they text you for permission; and the broken security monitors tracking them have finally been fixed so you no longer see weird static flashes or lingering burn-in images.
The system is much more resilient when your computer goes to sleep or network conditions shift. Previously, if you closed your laptop while the AI was streaming an answer, you'd wake it up to an ugly 'Content block not found' or a JSON parsing error; the system can now flawlessly survive the hibernation cycle. For Enterprise and Team users, usage-based session costs are now correctly displayed again. If you use the Claude Platform on AWS, the /login menu now provides a 'refresh credentials' option so you can proactively renew your token before getting kicked out. We also fixed a bug with the Chrome extension where disabling the permissions gate broke tab-group isolation when running multiple terminal sessions. In agent teams, new teammates spawned into background window panes (like tmux) will now obediently inherit the leader's designated effort level (--effort). Finally, we've hard-capped API retries at 15 to prevent run-away loops from burning your wallet. It's like taking a nap while a massive document is printing: you used to wake up to a jammed machine and shredded paper, but now it pauses perfectly and finishes the job when you wake up. Plus, the boss's strict instructions on how thoroughly to do the job now successfully trickle all the way down to the newest interns without anyone slacking off.
We've sprinkled in a bunch of neat workflow and UI quality-of-life tweaks. The Installed tab in the plugin manager (/plugin) now features a dedicated "Skills" section to keep your add-ons meticulously organized, and it will no longer falsely claim there are "more above" when you're already at the absolute top of the list. When diving into workflow details, you can now hit f to instantly filter by task status. The parser for your custom skill configurations (SKILL.md) is significantly more forgiving now: it gladly accepts kebab-case, snake_case, and camelCase syntax, and if your YAML format is slightly mangled, it will load the skill with empty metadata instead of silently crashing the whole thing. Additionally, the /review <pr> command now leverages the much more robust /code-review medium engine under the hood. On the cosmetic side: markdown strikethroughs now actually render as strikethroughs instead of showing literal ~~ tildes, text you selected with your mouse in the agent panel won't awkwardly remain highlighted after being deleted, and permission prompt option numbers are finally properly aligned even if the text overflows. It's like reorganizing your desk: all your specialized tools now have their own dedicated drawer, you can instantly filter your in-tray, and HR will successfully process your paperwork even if you accidentally capitalized a word wrong.
`!` bash commands now trigger Claude to respond to the output automatically; set `"respondToBashCommands": false` in settings.json to keep the previous context-only behavior
Added `claude mcp login <name>` and `claude mcp logout <name>` to authenticate MCP servers from the CLI without opening the interactive `/mcp` menu, with `--no-browser` stdin redirect support for completing over SSH
Improved `claude mcp get` and `claude mcp remove` to suggest the closest configured server name on a typo and truncate long server listsChanged background subagents to surface permission prompts in the main session instead of auto-denying; the dialog shows which agent is asking, and Esc denies just that tool
Fixed background session recaps being duplicated; the agent's own end-of-turn summary now shows as the recap line
Fixed subagent transcript scroll position bleeding into the main transcript on exit
Fixed background task previews flashing raw tool names before the agent's plan loaded
Fixed opening a background session from `claude agents` leaving the previous screen painted behind it
Fixed Esc and Ctrl+C not responding while background agents are still running after the main turn ends
Fixed background job status in `claude agents` showing a stale "needs input" message after replying
Fixed `Agent(type)` deny rules and `Agent(x,y)` allowed-types restrictions not being enforced for named subagent spawns
Fixed Workflow `agent({schema})` subagents looping forever on repeated schema validation failures instead of aborting after 5 attempts
Fixed pressing `x` on a finished subagent in the agent panel not dismissing it
Fixed a dark-theme flash when opening a background session from `claude agents` on a light terminalFixed streaming requests failing with "Content block not found" or JSON parse errors after the machine wakes from sleep
Fixed session cost not showing for usage-based Enterprise and Team subscribers
Added "Claude Platform on AWS - refresh credentials" option to `/login` when `awsAuthRefresh` is configured
Fixed Chrome tab-group isolation not applying when the in-product permissions gate is off for concurrent CLI sessions
Fixed agent teams: teammates spawned via tmux/pane backends now inherit the leader's `--effort` level
Added `teammateMode: "iterm2"` setting with a warning when auto mode cannot find the `it2` CLI
Changed `CLAUDE_CODE_MAX_RETRIES` to cap at 15; for unattended sessions, use `CLAUDE_CODE_RETRY_WATCHDOG` insteadAdded a "Skills" section to the `/plugin` Installed tab
Added status filtering (press `f`) to the `/workflows` agent detail view
Improved memory: the agent is now reminded to compact its `MEMORY.md` index when nearing the size limit
Changed `/review <pr>` to use the same review engine as `/code-review medium`
Improved skill frontmatter: `display-name`, `default-enabled`, `fallback`, and `metadata.*` keys now accept kebab-case, snake_case, and camelCase
Improved malformed `SKILL.md` YAML frontmatter handling: loads the skill body with empty metadata instead of failing silently
Fixed `~~strikethrough~~` showing literal tildes in assistant messages instead of rendering as strikethrough
Fixed `--tools` allowing feature-gated tools to slip through before flags loaded on a cold first launch
Fixed mouse-selected text staying highlighted after deleting it in `claude agents`
Fixed a misleading "MCP server disconnected" notice for intentionally retired tools when resuming older sessions
Fixed `/plugin` Installed showing a "more above" indicator when already scrolled to the top
Fixed misaligned option numbers in permission prompts when the option text overflows