繁體中文

新功能

組織專屬模型設定、任務預設有名稱,團隊協作更順暢

如果你在企業團隊中使用 Claude,現在管理員可以統一設定大家預設要用哪個模型(例如 Opus 或 Sonnet)了,當你沒特別選的時候,選單裡會顯示「組織預設」。為了讓你在一堆背景任務中不迷路,現在你一開啟新任務,系統就會自動幫它取個好讀的名字,方便你隨時找它或傳訊息給它。此外,當 Claude 在對話裡提到某個檔案時,你現在可以直接按住 Cmd 鍵並點擊那個檔案名稱,你的電腦就會自動打開對應的資料夾把檔案指給你看。這就像去圖書館,以前櫃檯只給你一堆亂碼編號叫你自己找書,現在不只每本書都有清楚的書名,你還能點一下書單上的名字,管理員就直接把書送到你面前。

新功能

背景任務大升級:程式關閉或更新,任務依然活著

背景任務現在有了「不死之身」。以前如果你把 Claude 關掉、重新啟動,或是程式剛好在背景自動更新,你那些跑到一半的長時間指令和任務就會全部慘死;現在它們能在這些劇烈變動中存活下來,甚至在 Windows 上也不例外。如果不小心發生系統重啟導致背景任務死掉,下次你打開清單時,它們會自動從上次斷掉的地方無縫接軌繼續做。這也順便解決了以前你喚醒一個卡住的任務時,系統可能會誤刪掉它整份對話紀錄並強迫重頭開始的毛病。這就像你在煮湯時家裡突然停電,以前整鍋湯會壞掉只能倒掉重煮,現在復電後瓦斯爐會自己點火,繼續把你那鍋湯熬完。

效能優化

效能與防呆優化:不再誤報錯誤、節省 25% 資源

跑 /code-review 時現在約省 25% token,長時間無回應也會自動重試。如果你使用 /code-review 讓 AI 幫你做程式碼審查,系統現在把五個找問題的步驟整合成一個,幫你省下大約 25% 的 token 額度(使用量)。防呆方面,當你輸入速度太快時,語音系統不會再自作主張偷偷幫你啟動錄音並吃掉你的空白鍵;如果 AI 回應的訊號卡住超過 5 分鐘都沒反應,系統會自動把它切斷重試,不會讓你枯等。針對安全,如果有人透過設定檔偷偷塞了一個假的 MCP(外部工具)伺服器,現在系統會強制標示「等待批准」,不會直接給它過關。而在 Windows 的 PowerShell 用 git diff 這類指令時,也不會再因為一些正常的輸出被系統大驚小怪當作執行失敗了。這就像請人幫你看合約,以前他會分五次跟你收費,現在一次看完而且打七五折;而且遇到來路不明的夾帶文件,他也會立刻幫你攔下來檢查。

問題修復

遠端任務自動復活,指令與權限錯誤修復

現在對於遠端控制(Remote session)有了更強的防呆保護。如果遇到伺服器重新啟動中斷了連線,任務現在懂得在下一個工作節點上自動接續復活,不會直接暴斃。當你用 /cd 換資料夾並遇到程式非正常關閉時,如果舊路徑有特殊字元,也不會再發生舊路徑像幽靈一樣一直卡在復原清單裡的毛病。在權限控制方面,當你使用 --dangerously-skip-permissions 這個危險參數時,系統現在會乖乖顯示免責聲明並套用跳過模式,不會再一聲不響地退回一般自動模式騙你。此外,原本在沒事做的時候按兩下 Esc 打不開時光倒流(rewind)選單的退化問題修好了;而從一般視窗要打開代理人面板時,現在也只要按一次左鍵就好,跟背景任務的行為統一,不用再傻傻連按兩次。

問題修復

減少無用耗能與閃爍,外掛與登入驗證更順利

終端機畫面的運作變得更省力了。在生成長回答(streaming)時,系統現在懂得跳過沒變動的區塊,不再浪費資源重繪整個畫面;在背景任務執行結構化輸出失敗時,也不會再把失敗的那行跟重試的那行重複印出來閃瞎你的眼。同時,它也修復了幾個常讓你誤會的狀態顯示,像是在 Bedrock 平台上 /context 不會再全部顯示 0 個 token,深研(/deep-research)功能在驗證失敗時也會老實顯示「未驗證」,而不是誇張地跟你說「所有聲明都被反駁」。此外,如果你把 外部套件(plugin)放在自己電腦上或用 Git 管理,系統以前可能會跳過檢查或忽視你鎖定的版本號,甚至在 MCP 驗證時因為要求過多權限(scopes)而導致在 GitLab 登入失敗,現在這些驗證流程與權限要求都精準修復了。最後,如果你手動把底層服務器設定為非 Anthropic 的主機,系統也會自動把你不需要的遠端控制功能關掉,免得浪費連線。就像開車把儀表板上亂跳的油表修好,還把冷氣效率調高;去辦通行證時,警衛也不會因為你資料放錯夾子就拒絕你,一切通關都變得順暢無阻。

English

FEATURE

Organization default models, named sessions, easier collaboration

If you use Claude in an enterprise team, admins can now configure a default AI model (like Opus or Sonnet) for the whole company; if you haven't picked one, your menu will simply show "Org default". To help you avoid getting lost in a sea of background tasks, the system now automatically assigns readable names to new sessions as soon as they start, making them much easier to identify and message. Also, when Claude mentions a file in the chat, you can now simply Cmd-click (or Ctrl-click) it to instantly reveal that file in your Finder or Explorer. It's like going to a library where the clerk used to just hand you a list of barcode numbers; now, every book has a clear title, and simply pointing at a title on the list magically summons the book right to your hands.

FEATURE

Background tasks survive app closes and updates

Background tasks are practically immortal now. Previously, if you closed Claude, restarted it, or if the app updated itself in the background, your long-running commands and workflows would instantly die; they now survive these severe disruptions and carry on, even on Windows. If a system restart does manage to kill a background worker, the next time you open the agent list, it will automatically resume exactly from where it left off. This also fixes a nasty bug where waking a stalled task could accidentally delete its entire conversation history and force it to restart from scratch. It's like cooking a slow stew when the house loses power: previously you had to throw out the ruined pot and start over, but now when the power returns, the stove turns itself back on and flawlessly finishes simmering your meal.

PERF

Performance and foolproof: no false errors, saves 25% tokens on code review

Running /code-review now saves roughly 25% of tokens, and stalled responses will auto-retry. The /code-review command has been streamlined: five separate cleanup steps are now merged into one, cutting your token usage (API costs) by roughly 25%. On the foolproof side, typing very fast will no longer accidentally trigger the voice dictation system or swallow your spaces; and if an AI response completely stalls for 5 minutes with no output, a new watchdog will automatically abort and retry instead of leaving you staring at a frozen screen. Security got a bump too: malicious MCP (external tool) servers trying to sneak in via project settings files are now firmly blocked with a "Pending approval" badge. Also, using commands like git diff in Windows PowerShell will no longer falsely trigger catastrophic error alarms just because of standard text output. It's like hiring a lawyer to review a contract: they used to charge you for five separate meetings, but now they do it all at once for a 25% discount, and they instantly red-flag any sketchy clauses slipped in by the other party.

BUGFIX

Remote crash recovery and permission command fixes

Remote sessions just got a major resilience upgrade. If a mid-turn server restart interrupts your Remote session connection, it will now gracefully auto-resume on the next available worker instead of dying abruptly. If you moved directories with /cd and the app crashed, your old path won't haunt the resume list forever just because it had special characters in its name. On the permissions side, using the hazardous --dangerously-skip-permissions flag now actually displays the bypass disclaimer and applies the bypass to spawned agents as promised, instead of silently falling back to regular auto mode. Additionally, the regression where pressing Esc twice at an idle prompt failed to open the rewind menu is fixed, and opening the agents view from a foreground session now requires just a single left arrow press instead of two, matching the background session experience.

BUGFIX

Reduced UI stuttering and smooth sailing for plugins and logins

The terminal interface has been optimized to save energy and reduce visual clutter. When streaming long responses, the UI now skips redrawing unchanged areas to cut down on rendering work. If a background agent's structured output fails validation, it will no longer confusingly print the rejected attempt right next to its retry. We've also fixed several misleading statuses: /context will no longer falsely show 0 tokens across the board on Bedrock, and the /deep-research tool will accurately report "unverified" instead of dramatically claiming "all claims refuted" when a verification fails. For plugins and integrations, validating local plugins from the "." path now works properly, dependency version pins are respected even in git-backed local folders, and MCP OAuth logins will no longer ask for an entire encyclopedia of scopes causing enterprise IdPs like GitLab to outright reject you. Finally, if you configure a custom non-Anthropic base URL, the app now automatically disables the unneeded Remote Control feature to match other Bedrock/Vertex environments. It's like fixing a twitchy fuel gauge and upgrading the AC efficiency in your car; and when you go to get your security badge, the guard no longer rejects you just because you asked for building access in the wrong format.

原始 CHANGELOG

Original changelog組織專屬模型設定、任務預設有名稱,團隊協作更順暢
Added support for organization default models — admins set it in the org console; it shows as "Org default" (or "Role default") in `/model` when you haven't picked one yourself
Added readable default names for sessions at start, making them easier to identify and message
Added clickable file attachments in chat — Cmd/Ctrl-click reveals the file in Finder/Explorer
Original changelog背景任務大升級:程式關閉或更新,任務依然活著
Improved background session reliability: long-running commands and workflows now survive the session's process being stopped, restarted, or updated — including on Windows, where background shells are handed off instead of being killed
Improved background agents: workers killed by a daemon restart are now automatically resumed from where they left off the next time the agents view opens
Fixed waking a background job permanently deleting its conversation and re-running the original prompt when the transcript probe misread a real transcript; the file is now set aside, never deleted
Fixed multiple `claude agents` side panel issues: keyboard focus getting stuck when opening an agent, background jobs losing their subagent types on every open, and sessions showing incorrect status while actively running
Fixed `claude agents` session status: completed rows no longer flip between "Done" and "Needs your input", stalled agents are now labeled "Needs attention", and results that mention a PR show a clickable link
Original changelog效能與防呆優化:不再誤報錯誤、節省 25% 資源
Improved `/code-review` workflow: merged five cleanup finders into one, cutting token usage by roughly 25%
Security: `claude mcp list`/`get` no longer spawn `.mcp.json` servers that a repo self-approved via a committed `.claude/settings.json`; untrusted workspaces show `⏸ Pending approval`
Fixed the rate-limit warning flickering off and rate-limit telemetry being over-counted when multiple parallel requests were in flight at the moment a usage limit was hit
Fixed PowerShell `git diff`/`git grep`, `egrep`/`fgrep`, and quoted search patterns containing `|` being reported as failures when they exit 1, matching Bash behavior
Fixed voice dictation swallowing spaces and spuriously starting a recording during very fast typing when voice mode is enabled
The streaming idle watchdog is now on by default for all providers — it aborts and retries when a response stream produces no events for 5 minutes. Set `CLAUDE_ENABLE_STREAM_WATCHDOG=0` to disable.
Original changelog遠端任務自動復活,指令與權限錯誤修復
Fixed mid-turn crash recovery for Remote sessions — sessions interrupted by a server restart now auto-resume on the next worker
Fixed sessions moved with `/cd` reappearing in the old directory's resume list after a non-graceful exit when the old path contained special characters
Fixed `claude agents --dangerously-skip-permissions` silently falling back to auto mode instead of showing the bypass disclaimer and applying bypass mode to spawned agents
Fixed Esc Esc at an idle prompt not opening the rewind menu (regression); use Ctrl+C or Ctrl+X Ctrl+K to stop background agents
Changed opening the agents view from a foreground session to require a single `←` press instead of two, matching the behavior in background sessions
Original changelog減少無用耗能與閃爍,外掛與登入驗證更順利
Reduced per-frame rendering work in the terminal UI by skipping no-op subtree walks during streaming
Fixed duplicate recap lines after a background session's turn: a schema-rejected StructuredOutput attempt no longer renders alongside its retry
Fixed `/context` showing 0 tokens for all tool groups on Bedrock
Fixed `/deep-research` misreporting verifier failures as "all claims refuted" instead of `unverified`
Fixed `claude plugin validate` skipping local plugins whose source is "." and stopping after the first error class
Fixed plugin dependency version pins not being honored when the marketplace was added as a local folder path backed by a git repo
Fixed MCP OAuth requesting the authorization server's full `scopes_supported` catalog when no scope is specified, causing `invalid_scope` failures on GitLab self-hosted and other enterprise IdPs
Remote Control is now disabled when `ANTHROPIC_BASE_URL` points at a non-Anthropic host, matching the existing behavior under `CLAUDE_CODE_USE_BEDROCK`/`_VERTEX`/`_FOUNDRY`