Claude Code 2.1.217

工具回傳很長的結果時，畫面雖只留摘要，記憶體卻曾偷偷抱著完整原文直到工作階段結束。這次修掉 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）工具輸出被截短後仍佔滿記憶體的問題。Claude Opus 4.8 在 Bedrock 上的自動壓縮也終於會觸發，超過上限時 /compact 不再失敗；而 CLAUDE.md 或 SKILL.md 的 paths 設定塞進太多大括號，也不會讓 CLI（命令列介面）啟動時耗盡記憶體或卡死。像收納員只留下需要看的摘要，不再把整車舊紙箱偷偷堆在後房。

Windows 更新若半路失敗，最糟曾會把 claude.exe 一起弄丟；現在失敗時會自動放回保留的執行檔，Claude 不會憑空消失。另一個讓人無力的情況是把 session（一次持續工作的對話）送到背景後，背景 shell（執行命令的程式）有時再也停不下來，尤其 Windows 或電腦很忙時更明顯。這版把它修好，無論用 /background、左箭頭送進背景，或對話正結束，都能正常停止。這就像工廠換新機器時把舊機留在旁邊備援；若要停工，也終於能關掉仍在背景運轉的機器。

把背景 session（一次持續工作的對話）關在指定工作區，本來就該像把人留在同一間辦公室；但工作資料夾若經過 symlink（符號連結，指向另一個位置的捷徑檔），先前路徑沒有被認成真正位置，對話可能因此逃出 workspace folder（工作區資料夾）。現在背景隔離會先辨認捷徑最後通往哪裡，再套用邊界，避免工作跑到不該碰的資料夾。這是安全修復，不是新增限制：你照常從捷徑開工作，但門禁終於看得懂它通往哪一扇門。

在公司網路裡能開網頁，不代表 Claude Desktop 就連得出去：mTLS、TLS 驗證、OAuth 範圍與 proxy（代理伺服器，代替你轉送網路連線的中繼站）設定曾被它忽略。現在這些企業設定會在 Desktop session（一次持續工作的對話）生效。受管理的 OTEL_EXPORTER_OTLP_ENDPOINT 也能真正管住所有 telemetry（遙測資料，程式送出的運作紀錄）訊號，較低層級的個別設定不能再把資料改送到別處。像總公司替每輛車規劃好路線與回報站，現在所有車都會照著走，不會有一輛私自繞路。

使用螢幕閱讀器（把畫面文字念出來的輔助工具）時，啟動公告以前常被第一個提示畫面截斷；現在它會好好念完。思考狀態列也不再每幾秒為了更新經過時間與 token（AI 處理文字時使用的計數單位）數而重畫一次，少了反覆干擾。若有人在 Remote Control（遠端控制，讓他人連進同一工作階段）已出現授權提示後才加入，先前看不到那個待處理視窗，現在也能看見。就像會議室的廣播不再講到一半被蓋掉，晚進門的人也能看到桌上正等簽的文件。

A tool can return a huge result while the screen shows only a shortened version; previously, the full result could still sit in memory until the session ended. This fixes that MCP (Model Context Protocol, the standard for connecting AI to external tools) memory leak. Auto-compact now fires for Claude Opus 4.8 on Bedrock, /compact works past the limit, and excessive brace groups in CLAUDE.md or SKILL.md paths no longer freeze or exhaust the CLI at startup. It is like a clerk keeping the summary you need instead of secretly piling the whole truckload of boxes in the back room.

A failed Windows update could once leave you without claude.exe; now it automatically restores the preserved executable instead of making Claude vanish. Another frustrating case was a background shell that sometimes could not be stopped after a session was sent to the background, especially on Windows or a heavily loaded computer. This is fixed whether you use /background, the left arrow, or the session is exiting. It is like replacing a factory machine while keeping the old one nearby as a backup, then being able to shut down the machine still running in the back room.

A background session should stay inside its assigned workspace, like a worker staying in the right office. But if its working directory was reached through a symlink (a shortcut file pointing elsewhere), the path was not converted to its real location and the session could escape its workspace folder. Background isolation now resolves that shortcut first, then applies the boundary. This is a security fix, not a new restriction: you can still open work through a shortcut, but the access gate now knows which door it actually reaches.

Being able to open a website on a corporate network did not always mean Claude Desktop could connect: it was ignoring mTLS, TLS verification, OAuth scopes, and proxy settings. Those corporate settings now apply to Desktop sessions. A managed OTEL_EXPORTER_OTLP_ENDPOINT also truly governs every telemetry signal, so lower-scope, signal-specific overrides cannot redirect data elsewhere. It is like headquarters setting the route and reporting depot for every vehicle; now every one follows it rather than one quietly taking a detour.

With a screen reader (an accessibility tool that reads on-screen text aloud), the startup announcement used to be cut off by the first prompt; now it can finish. The thinking status row also stops re-rendering every few seconds merely to refresh elapsed time and token counts. In Remote Control, people who join after a permission prompt appears can now see that pending dialog. It is like a meeting-room announcement that is no longer talked over halfway through, while late arrivals can still see the document waiting for a signature.