發布日期：2026-07-15
把耗時的任務發送至背景執行，卻常常因為進度不明或代理人擅自重啟而感到困擾嗎？現在你派出的子代理（subagents）變得更聽話了。以前系統在任務尚未完成時，有時會捏造已完成的報告；現在會確實顯示仍在執行。被手動砍掉的背景工作階段（session，工作階段為原定譯法）也不會再悄悄復活或重複執行舊指令。當你想在串流中調閱子代理的完整思路，這正是加開「--forward-subagent-text」參數或設定環境變數的時刻，它們會老實交出思考紀錄。就像請了誠實的助手，沒做完不會撒謊，開除後也不會回來重做白工。
放心地讓 AI 在專案中自動執行指令，但有時會擔心被惡意程式碼的障眼法騙過？這些安全漏洞已被徹底補齊。以前指令確認視窗可能會被零寬度字元（zero-width characters，隱形墨水般的特殊字元）或假引號矇騙，誤導你按下執行，現在它們都會在畫面現出原形。自動模式下，只要掛鉤判定需要詢問，系統便不能擅自跳過，必須停下來請示。此外，你按下的「永遠允許」特權現在會寫在專案的根目錄，即使換了工作樹（worktree，專案的平行副本）也同樣生效。就像保全換上能看穿偽裝的墨鏡，並把通行證鎖在大門口。
每次上傳檔案或召喚瀏覽器外掛，總是遇到莫名其妙的卡死嗎？現在上傳與網頁操作更加順暢。以前檔名以點結尾或包含保留字（如 DOS 系統專用的 .prn 後綴）的檔案會被拒收，現已全部放開，並嚴格杜絕具多重連結的危險檔案。使用 Chrome 瀏覽器外掛時，即使 Chrome 本身未啟動，主程式也不再死當。如果你要留下瀏覽器操作的網頁證據，這正是執行截圖並透過「save_to_disk」寫入磁碟的時刻，系統會將圖片存好並回傳檔案路徑。就像郵局優化了流程，即使寄送怪名字的包裹也能安全送達。
早上打開電腦重啟工作，卻發現開著的對話視窗集體被登出，或背景工具全部斷線？這些睡眠喚醒後的疑難雜症已全被修好。以往多個工作階段（session，工作階段為原定譯法）睡眠喚醒後會連鎖登出，現已隔離。同時多個工作階段共用同一個憑證庫（credential store，存登入證明的地方）也不會互相干擾。閒置對話喚醒後，外部工具伺服器（MCP servers）也會自動重新連線。而在 Windows 的無畫面工作階段下，也不會再因為讀不到鍵盤輸入就直接崩潰。就像員工睡醒回到崗位，櫃檯不會弄丟認證，外部合作商的電話也重新接通。
頻繁切換設定分頁或清除對話時，是否覺得畫面載入總是慢半拍？多個拖慢速度的介面問題已被一掃而空。非同步內容載入時原本存在 300 毫秒的延遲，現在已被完全抹平，畫面渲染效能也大幅提升。以前在輸入框輸入問號「?」時指令會被吃掉，現在這異常現象已修正。此外，當你以「/clear」清除當前的工作階段（session，工作階段為原定譯法）時，一切將重新計量。顯示在最底下的狀態列（statusLine，狀態列為原定譯法）費用計算也會確實歸零。就像開車的儀表板，以前切換功能會卡頓，現在指針反應靈敏，且按重設時里程表會乖乖歸零。
擔心因為快取失效而被重複收取重複提問的運算費嗎？這批修復讓計費更精準，避免花冤枉錢。在 Bedrock 與 Vertex 雲端服務上，提示詞快取（prompt-caching，把重複提問暫存以節省成本的功能）的退化問題已解決，不會再把背景系統脈絡誤算為新文字。現在若遇到模型備援（fallback，備援為原定譯法）警報，系統也會依照你的設定，在載入時自動修正，不跳多餘的警示。就像去市場買菜的過磅機，只精確秤量你買的菜，不把額外包裝袋也算進重量多收錢。
Have you ever dispatched a long-running task to the background only to find your agent's status unclear or restarting itself? Now subagents behave much better. Claude no longer fabricates final results while an agent is still working, and manually killed background sessions will not respawn or replay old tasks. If you want to include subagent thoughts and texts in stream-json output, that is when to append --forward-subagent-text or set the environment variable. It is like hiring a reliable helper—they report their actual status instead of lying, and stay stopped when you let them go.
Are you sure your environment is safe from sneaky command execution? A set of critical security gaps has been closed. Attackers previously tried using zero-width characters (invisible characters acting like fake quotes) to visually trick you into approving bad actions, but the review screen now strips them out. In auto mode, when a hook specifies an ask prompt, the system will respect that. Finally, "always allow" rules are now saved at your repository root, keeping permissions active across a worktree (a parallel folder copy) without asking again. It is like security guards putting on special glasses that expose invisible ink and locking passes at the front gate.
Have you ever had your uploads block or Chrome extension hang at startup? Files ending with a dot or DOS-reserved suffixes like .prn are now accepted, while double-linked files are correctly rejected. The CLI will no longer freeze if you have the Chrome extension enabled but Chrome is closed. When you need to keep a record of your web screen, that is when to execute save_to_disk during screenshot actions, which now correctly saves the image to your disk and returns the path. It is like a post office expanding its parcel policies to accept irregular names and ensuring pictures are printed.
Are you tired of finding your terminal windows logged out or plugins disconnected after your computer wakes from sleep? The sleep-wake recovery bugs are now solved. Sessions sharing one credential store (where logins are saved) will not trigger simultaneous logouts when waking up. In addition, idle web sessions will now auto-reconnect to MCP servers (tool integration servers) right away. Background jobs using LLM gateway tokens stay authorized when the daemon restarts, and Windows headless processes will not crash if stdin goes dry. It is like staff returning to their desks—IDs are intact and partner phones reconnect.
Does switching panels or running /clear feel a bit sluggish? This update sweeps away UI lags and annoying quirks. The 300ms delay before loading Settings and Stats views is gone, and terminal rendering is faster. Entering a single question mark in the command prompt will not open the shortcuts panel by mistake anymore. When you run /clear to reset the current session, the cost counter will start fresh. The statusline (the bar showing info at the bottom) also resets to $0. It is like a dashboard where shifting screens is instant, and resetting your trip meter actually resets it to zero.
Want to save on recurring tokens but worried that cache slips are costing you extra? We have patched credit calculations to prevent double billing. Prompt caching (saving past prompts to lower token use) on Bedrock or Vertex will no longer bill the trailing system block as fresh inputs on every request. If repository models are configured explicitly, Claude won't trigger default configurations or spurious fallback alerts at startup. It is like a produce scale that only weighs the vegetables you buy, without charging you for the heavy plastic wrapping.
Added `--forward-subagent-text` flag and `CLAUDE_CODE_FORWARD_SUBAGENT_TEXT` environment variable to include subagent text and thinking in stream-json output
Fixed subagents spawned with an explicit model override reverting to the parent's model when resumed or sent a follow-up message
Fixed reopening a just-stopped background session from the agents view starting a blank conversation under the same session id
Fixed `/loop` hiding the session from `/resume` after a single use
Fixed `claude agents` jobs becoming permanently undeletable when git no longer recognizes their worktree — the row now shows why the delete was refused instead of silently reappearing
Fixed background session titles in the agents view showing the naming model's refusal text when the prompt contains a link
Fixed background agents killed by the user auto-respawning, and revived agents re-running stale prompts from old sessions
Fixed routines with no schedule reporting a next run time in the year 1
Improved background agent result reporting — Claude now reports the status of still-running agents and waits for the real completion instead of fabricating resultsFixed permission previews relayed to chat channels not neutralizing bidirectional-override, zero-width, and look-alike quote characters, so tool inputs cannot visually alter the approval message
Fixed auto mode overriding a PreToolUse hook's `ask` decision for unsandboxed Bash — a hook `ask` now floors the decision at a prompt
Fixed nested `.claude/rules/*.md` files loading even when setting sources exclude project settings
Changed "always allow" permission rules to save at the repository root, so approvals granted in a git worktree persist across sessions and worktreesFixed file upload validation: filenames ending in a DOS device suffix (`.prn`) or trailing dot are now accepted, and files with multiple hard links are refused
Fixed file uploads to Claude in Chrome from remote and CLI sessions
Fixed a startup hang when the Claude in Chrome extension is enabled but Chrome is not running
Fixed Claude in Chrome setup pages failing to open in the browser on Windows
Claude in Chrome: hardened file-upload path validation
Claude in Chrome: `save_to_disk` on screenshot actions now writes the image to disk and returns the path; previously it did nothingFixed parallel Claude Code sessions all logging out simultaneously after wake-from-sleep when many sessions share one credential store
Fixed plugin MCP servers not reconnecting after an idle web session woke, leaving MCP calls failing until the next message
Fixed background jobs on LLM gateway auth (`ANTHROPIC_AUTH_TOKEN` + `ANTHROPIC_BASE_URL`) coming back "Not logged in" after the daemon respawns them
Fixed headless print-mode sessions on Windows crashing or silently exiting when stdin is unreadable
Hardened synced skill/plugin directory naming on Windows and kept CCR web fetch/search proxies working after `/clear`Fixed edits that leave the input as "?" being silently swallowed and toggling the shortcuts panel
Fixed a 300ms delay revealing async content (Settings tabs, Stats, diff views, and other loading states)
Fixed screen reader users losing the audible terminal bell after `/terminal-setup` or onboarding terminal setup
Fixed `/clear` not resetting the session cost counter — the statusline's cost now starts at $0 after `/clear`
Improved terminal layout and rendering performance
Updated integer environment variables (timeouts, token budgets, retry counts) to accept scientific notation and digit-separator spellings like `1e6` and `64_000`
Updated documentation links to the current docs sites
Changed Vim mode `s` and `S` (substitute char/line) to work in NORMAL mode, matching vim behavior
[VSCode] Updated the Remote Control banner to describe what it doesFixed Claude Code on Vertex and Bedrock attempting the default Opus model at startup and printing a spurious fallback notice when a model is explicitly configured
Improved the memory index over-limit warning to measure only loaded content, excluding frontmatter and HTML comments
Changed `/usage-credits` to ask for confirmation before sending a request to organization admins
Fixed a prompt-caching regression on Bedrock, Vertex, Mantle, and Foundry that billed the trailing system context block as fresh input tokens on every request.