Claude Code 2.1.203：速度大解放：Mac 切換不再卡頓，打字捲動更滑順

繁體中文

新功能

速度大解放：Mac 切換不再卡頓，打字捲動更滑順

這次把效能調到最好。以前在 macOS 上切換背景任務時，會因為誤判「記憶體不足」而卡個 15 到 20 秒；在一般對話中，系統為了算剩餘字數（context-usage），每講完一句話就會把歷史紀錄全部重算一次，吃掉一堆資源；長答案吐出來時畫面還會整個重繪——現在這些蠢事都不會發生了，連帶把開機需要的記憶體也省下了大約 7MB。這就像以前你的電腦每次開新網頁都要全身健檢一次，現在它只看你點了什麼，所以反應快如閃電。

新功能

視覺介面更乾淨，畫面不亂跳、模式不看錯

操作介面變得更直覺也更穩定了。當你在手動模式（系統每次做事都要你按同意的模式）時，左下角會一直有個灰色的 ⏸ 標記提醒你，不怕忘記自己開了什麼模式。在看很長的對話往上捲時，內容不再會莫名其妙自己亂跳；打字時如果跳出指令建議，終端機也不會跟著狂閃。而且現在代理（agent）清單一打開就會乖乖分好「需輸入 / 執行中 / 已完成」的區域。此外，開機時一堆嚇人的「指令壞掉」警告也不會再洗版，通通收進 /doctor 裡了。這就像看一本排版精美的書，以前翻頁會卡住、字會跳動，現在每一頁都服服貼貼，連標籤也貼得清清楚楚。

新功能

背景任務更可靠，斷線能救回、升級不中斷

這次大幅提升了背景代理（agent）的生存能力。如果你的登入狀態快過期了，系統現在會提早警告你，讓你有時間續命，不會讓跑了一半的任務直接死掉。就算背後的連線金鑰真的過期導致連不上，現在它也能自動恢復，不會永遠裝死。以前如果你把代理清單關掉再重開，正在跑的子任務會默默停掉並重頭來過；或者背景程式自己偷偷升級時，會把所有執行中的任務全部砍光——這些災難現在都修復了。同時，如果任務的工作資料夾突然被刪除，它會報個清楚的錯誤給你，而不是卡在無限崩潰的迴圈裡。就像一個在外奔波的業務，以前手機沒電或公司搬家就會直接失聯，現在他不但會提早說手機快沒電，遇到狀況也會自動重連回報。

問題修復

工作環境與 Git 不再打架，指令順暢執行

對於常在複雜專案裡工作的人，指令與環境不會再搞失蹤了。以前在 Windows 上，背景任務可能會拿到舊的路徑清單（PATH），導致找不到該用的工具；或者會不小心弄丟你設定好的網址環境變數，害你連錯伺服器。針對 Git 使用者，如果你的專案裡有很多工作樹（worktree，讓你多開幾個平行工作副本的功能），以前跑 Bash 會被抱怨「參數太長」而失敗；被隔離在工作樹裡的代理，甚至會不小心跑到主資料夾去亂動指令；現在它們都會乖乖待在自己的沙盒裡做事。這就像工人去工地，以前會拿錯舊工具箱，或是跑到隔壁別人的地盤敲敲打打，現在他們每個人都能精準拿到對的工具，並且只在自己負責的區域施工。

新功能

子代理與外掛更聰明，細節體驗再升級

代理團隊的合作與工具整合變得更順暢了。現在代理們不會動不動就把整包任務直接推給另一個子代理（subagent），而是會更負責地自己動手。如果你用指令去尋找別人派出去的子代理，以前會找不到，現在系統會清楚列出所有執行中代理的 ID 給你看。在代理清單裡打字時，如果不小心打出不支援的斜線指令，字也不會莫名其妙被清空了。對於使用外掛（plugin）或 MCP（讓 AI 呼叫外部工具的標準協定）的人，現在會把額外的工作路徑也加進清單，且外掛如果正常給出除錯建議，就不會被系統誤判為「沒在用」而標記。這就像一支原本溝通不良的團隊，以前會互相推託工作或是找不到聯絡人，現在大家不但各司其職，還能準確找到對口。

English

FEATURE

Speed Unleashed: No more Mac stalls, smoother typing and scrolling

Performance is now tuned to the max. Previously on macOS, switching background tasks would stall for 15 to 20 seconds due to a false "low memory" detection; in interactive chat, the system would needlessly recalculate the entire transcript history after every single turn just to update the remaining context usage, devouring resources; and streaming long answers would cause the whole screen to redraw. None of these happen anymore, shaving off about 7MB of startup memory too. It's like a computer that used to run a full virus scan every time you opened a new tab, but now just loads what you clicked, reacting lightning fast.

FEATURE

Cleaner UI: No jumping screens, mode is always clear

The interface is now more intuitive and stable. When you are in manual mode (where the system waits for your approval before acting), a persistent grey ⏸ badge sits in the footer so you always know what mode is active. Scrolling up through long transcripts won't cause the text to jump around inexplicably, and the terminal won't violently flicker when showing shell command suggestions. Also, opening an empty agent list now cleanly displays structured sections like "Needs input / Working / Completed" right from the start. Start-up screens are cleaner too, as scary "command missing or broken" warnings no longer spam you; they are tucked neatly into /doctor. It's like reading a beautifully formatted book: pages used to stick and letters used to dance around, but now every page turns smoothly, and the sticky notes are perfectly legible.

FEATURE

Reliable Background Tasks: Auto-recovery and safe upgrades

The survival rate of background agents has seen a massive boost. If your login is about to expire, the system now warns you early so you can re-authenticate before any half-finished tasks are violently interrupted. If the background daemon's session token does go stale, agents will gracefully auto-recover instead of playing dead forever. Previously, returning to the agent list could silently kill running subagents and force them to restart from scratch, and silent background daemon upgrades used to mercilessly axe all your running sessions; these catastrophes are now fixed. Also, if a task's working directory gets unexpectedly deleted, it now gives you a clear error once rather than getting stuck in an infinite crash loop. Like a traveling salesperson who used to vanish entirely if their phone died or the office moved; now they warn you before their battery runs out and always know how to reconnect when things go wrong.

BUGFIX

Environment and Git play nice, commands run smoothly

If you work in complex projects, your commands and environment variables won't magically vanish anymore. Previously on Windows, background tasks might inherit an old PATH list, making them blind to required tools; they also had a habit of dropping custom URL environment variables, causing them to query the wrong servers and fail. For Git users, repositories with many worktrees (a feature granting parallel working copies) used to crash Bash with an "argument list too long" error, and agents isolated inside a worktree would occasionally wander into the parent directory to run commands; now they stay obediently in their own sandbox. It's like sending construction workers to a site: they used to grab an outdated toolbox or start hammering on the neighbor's property, but now they bring the exact right tools and strictly build only in their assigned plot.

FEATURE

Smarter subagents and plugins, finer details upgraded

Collaboration among agent teams and tool integration flows much better now. Agents are now far less likely to lazily dump their entire task onto a subagent, taking more direct responsibility instead. If you used commands to search for a subagent spawned by someone else, you used to hit a brick wall; the system now lists the IDs of all running agents for you. When typing in the agent list, accidently entering an unsupported slash command will no longer frustratingly erase your entire typed message. For power users of plugins or MCP (a protocol letting AI call external tools), additional working directories are now correctly broadcasted, and plugins that provide diagnostics won't be falsely marked for disuse. It's like a dysfunctional office that used to pass the buck and lose contact with each other; now everyone does their own share of the work and knows exactly who to call.

原始 CHANGELOG

Original changelog速度大解放：Mac 切換不再卡頓，打字捲動更滑順

Fixed opening or switching background agent sessions on macOS stalling for 15–20 seconds due to a false low-memory detection (regression in 2.1.196)
Fixed a memory and per-turn CPU regression in interactive sessions: the context-usage indicator no longer re-analyzes the entire transcript after every turn
Reduced binary size by ~7 MB and startup memory by ~7 MB by loading a large bundled dependency lazily instead of inlining it
Improved responsiveness while long responses stream: live-preview updates no longer re-render the whole screen

Original changelog視覺介面更乾淨，畫面不亂跳、模式不看錯

Added a grey ⏸ badge to the footer when in manual permission mode, making the active mode always visible
Fixed content jumping when scrolling up through long transcript history
Fixed the terminal flickering and jumping while typing in bash mode when a shell-history suggestion was shown
Changed the empty `claude agents` view to always show the organized sections (Needs input / Working / Completed) with descriptions
Changed left arrow to no longer close the background tasks, diff, and workflow detail views — press Esc instead
Removed the startup "claude command missing or broken" warnings — they now appear in `/doctor` and `/status` instead
Removed a redundant navigation hint from the `claude agents` footer
Fixed literal `^[[I` / `^[[O` escape codes being printed when reattaching to a background session

Original changelog背景任務更可靠，斷線能救回、升級不中斷

Added a warning when your login is about to expire, so you can re-authenticate before background sessions are interrupted
Fixed background sessions becoming permanently unresponsive to attach, replies, and stop when the daemon's session token went stale — the session now recovers automatically
Fixed returning to `claude agents` silently stopping running subagents and re-running the prompt from scratch — their work now carries over
Fixed background agents crash-looping when their working directory was deleted, replaced by a file, or became an invalid path — they now fail once with a clear error
Fixed a background daemon auto-upgrade failure silently killing all running background sessions
Fixed background agent startup failures showing only "exit_with_message" instead of the actual error
Fixed background sessions showing "Needs input" in the agent list after the question was already answered
Fixed `/exit` incorrectly warning about running background agents after all named agents had completed
Fixed the agent list crashing when opening a stopped session whose conversation was already open in another session

Original changelog工作環境與 Git 不再打架，指令順暢執行

Fixed background agents inheriting a stale `PATH` from the daemon instead of the dispatching shell, causing missing tools on Windows
Fixed background and agent-view sessions dropping a shell-exported `ANTHROPIC_BASE_URL`, which sent API keys to the default endpoint and failed with 401
Fixed Bash failing with "argument list too long" in repos with many git worktrees
Fixed worktree-isolated subagents sometimes running shell commands in the parent checkout instead of their own worktree
Fixed worktree creation rejecting nested repositories in multi-repo workspaces, leaving background sessions unable to isolate and edit
Fixed background task output on Windows being permanently replaced by an empty file after `/clear`
Fixed background sessions started from a non-git directory unable to edit files when a `WorktreeCreate` hook was configured

Original changelog子代理與外掛更聰明，細節體驗再升級

Added the session's additional working directories to MCP `roots/list`, with `notifications/roots/list_changed` sent when the set changes
Fixed `TaskStop` and `TaskOutput` failing to find background agents spawned by another agent — errors now list running agents by id and description
Fixed the `claude agents` composer discarding your typed message when a slash command isn't available there
Fixed background sessions ignoring `effortLevel` changes in settings.json when forked through the daemon
Fixed attached background sessions ignoring `CLAUDE_CODE_DISABLE_MOUSE` and `CLAUDE_CODE_DISABLE_MOUSE_CLICKS` opt-outs
Fixed the `@` directory picker in `claude agents` not showing registered git worktrees
Fixed LSP-only plugins being incorrectly flagged for disuse when their language servers deliver diagnostics or answer navigation requests
Improved subagent behavior: agents are now less likely to re-delegate their entire task to another subagent
[VSCode] Added a Settings toggle for "Enable Remote Control for all sessions"