Claude Code 2.1.199

當 API 伺服器超載或你的網路稍微不穩時，現在系統變得更有彈性了。以前如果 Claude 話講到一半伺服器出錯，已經講出來的字會全部不見，現在它會幫你保留下來並標記「尚未講完」；如果是被暫時性的限流（rate limit，防止瞬間請求過多的機制）擋下，系統現在會自動幫你稍後重試，不會直接判你失敗，且重試次數上限也大幅放寬。如果子任務（subagent）被中斷，它不再會騙你「任務成功」，而是會老實回報錯誤，或者把做了一半的半成品交接給主管。遇到 SSL 憑證這種根本無法連線的錯誤，它也會直接教你怎麼修，不會傻傻重試到天荒地老。這就像打越洋電話，以前只要雜音一來就直接斷線，而且聊一半的內容都不算數，現在它不但會保留剛才聽懂的話，還會自動幫你重新撥號。

背景運作的任務現在變得更強韌了。以前在 Linux 上如果沒正常關機，留下來的壞紀錄會導致系統每 50 秒就自殺一次，並砍掉所有正在跑的任務；在 Mac 上用 SSH 遠端連線時也常發生無法開機的狀況；或是當你下令叫任務停下來時，系統卻又手賤自動把它重開——這些詭異的現象都修好了。如果你電腦記憶體不夠，任務不會再丟出一句看不懂的錯誤代碼，而是會明確叫你「釋放記憶體」。在不同資料夾開 Claude 時，Chrome 的外掛也不會再瘋狂跳出重新連線的網頁了。這就像一間公司的打卡系統，以前只要前一天不小心沒打卡下班，隔天整間公司就會每分鐘自動跳電一次，現在它終於修好，大家可以安穩上班了。

你可一次載入最多五個斜線技能，且空開對話不再把紀錄檔越撐越大。現在你可以一次呼叫多個斜線技能（像 /skill-a /skill-b 幫我做某事），它會乖乖把前五個技能都載入，不會只讀第一個。看代理清單時，如果子代理沒事做，它們會乖乖收成一行，不會直接消失讓你想找也找不到；當你在看子代理的畫面時，如果不小心打了切換模型的指令，它會提醒你這只對主代理有效，不會偷偷幫別人換模型。遇到設定檔壞掉要重設時，它也學會先幫你備份，不會直接把你過去的設定刪光光了。這就像你去點餐，以前一口氣點兩杯不同的飲料店員只會做第一杯，現在他能一口氣記住你的要求，而且結帳時不會因為你沒買新東西就多印一張空白發票給你。

The system is now much more resilient against overloaded API servers or minor network blips. Previously, if the server errored out mid-sentence, everything Claude had already streamed would vanish; now it saves that partial output with an "incomplete" notice. If you get temporarily blocked by a rate limit (a mechanism preventing too many rapid requests), the system will automatically retry with a delay instead of instantly failing your turn, and the retry cap has been significantly raised. If a subagent is cut off, it no longer lies about "success" but honestly reports the error or hands its partial work back to the parent agent. For hard connection errors like SSL certificate issues, it won't pointlessly burn through retries but will fail immediately and tell you exactly how to fix it. It's like a long-distance phone call: a burst of static used to drop the call and erase everything you just said; now it remembers what it heard and automatically redials for you.

Background tasks are now vastly more robust. Previously on Linux, if the system shut down uncleanly, corrupted leftover records would cause the background program to commit suicide every 50 seconds, taking all running tasks down with it. On Macs, background agents frequently failed to start over an SSH connection, and if you explicitly told a task to stop, the system would sometimes stubbornly auto-restart it right back up. These bizarre behaviors are now fixed. If your computer is starved for memory, tasks won't spit out a cryptic error anymore; they'll explicitly tell you to free up resources. Also, running Claude in different folders will no longer cause the Chrome extension to spam you with reconnect pages. It's like a company's timeclock system that used to trigger a building-wide blackout every minute if someone forgot to punch out yesterday; it's finally fixed, and everyone can work in peace.

You can now load up to five slash skills at once, and opening empty sessions no longer bloats your transcript file. You can now stack multiple slash skills in one breath (like /skill-a /skill-b do XYZ) and it will properly load up to the first five, instead of stubbornly ignoring everything after the first. In the agent list, idle subagents now neatly collapse into a summary row instead of vanishing completely out of sight. If you accidentally type a model-switching command while looking at a subagent, it now warns you that this only applies to the main lead agent, rather than silently making unwanted changes. If your configuration file gets corrupted and needs resetting, the system now safely backs it up first instead of mercilessly wiping your old settings. It's like ordering at a cafe: the barista used to only make the first drink if you ordered two at once, but now remembers your whole order, and the register won't print a long blank receipt if you didn't end up buying anything.