發布日期：2026-06-25
這次讓自動模式（auto mode）跟背景任務變得更聰明。現在自動模式拒絕你執行的動作時，會把「為什麼拒絕」的理由清楚寫在紀錄與權限面板裡；你也可以開啟新設定，讓所有指令都經過安全分類，而不只是針對特定程式碼。對於背景代理人，它現在學會了平行處理，不會再笨笨地為了啟動代理人而要求主對話中斷。系統也新增了自動記憶體清理機制（memory-pressure reaping），會把閒置的背景指令清掉以節省資源。此外，以前你切換到背景時，常會被假警報恐嚇說「任務會被放棄」，或是莫名其妙生出一個重複執行對話的幽靈任務；在更新過後，這些假警報、幽靈任務，以及代理人面板會把其他兄弟代理人藏起來的毛病，全都修好了。就像請了一個精明的管家，以前他拒絕你做事只會搖頭，而且每次分派任務就要你閉嘴等他；現在他會清楚告訴你原因，並且能一邊聽你說話一邊指揮傭人，還會自動把沒在用的房間燈關掉省電。
跟外部系統打交道變得更順暢了。現在你在終端機使用 bash 模式（打 ! 開頭）時，系統會即時幫你自動補齊檔案路徑，不用再辛苦地一個字一個字打。如果你的 MCP（外部工具伺服器）需要登入認證，系統一啟動就會溫馨提示你去 /mcp 處理；而且當工具呼叫因為權限不足（401/403 錯誤）被擋下時，MCP 輔助程式現在會自動重新連線並重試，不會直接死給你看。此外，當外部套件（plugin）在市集上改名時，你的設定檔也會自動跟著更新名字。登入後查看模型選單時，也不會再看到過期或空白的假畫面了。就像你的專屬助理，現在打字會自動幫你接下去，去別的部門調資料遇到門禁會自己想辦法重新刷卡，別人改名了也懂得幫你更新通訊錄，而且名片本打開永遠是最新狀態。
企業管理員現在有了更透明的監控工具。系統在 OpenTelemetry（企業用來收集日誌的標準框架）中新增了專門記錄 Claude 回答內容的事件。只要你原本就有設定記錄使用者的提示詞，系統升級後預設就會連同 AI 的回答一起記錄下來，讓你一眼看穿 AI 到底產出了什麼。當然，如果你有資安考量，也可以手動設定變數（OTEL_LOG_ASSISTANT_RESPONSES=0）來關閉這個功能，讓日誌只保留使用者的指令。就像公司的監視器原本只錄得到員工對對講機講的話，現在連對講機另一頭的回應也能一併錄音存證，當然，你想把喇叭靜音也完全沒問題。
Auto mode and background tasks just got a lot smarter. When auto mode denies an action, it now explicitly lists the reason in the transcript and permissions panel; you can also opt to route all terminal commands through the safety classifier instead of just specific code-execution patterns. For background agents, launching one no longer instructs the main AI to "end your response"—it can keep working on other things while the agent runs in parallel. We also added an automatic memory-pressure reaping system that sweeps up idle background shells to save resources. Moreover, several annoying bugs are squashed: sending tasks to the background no longer spooks you with false "tasks would be abandoned" warnings, nor does it spawn a phantom subagent that re-runs your entire conversation, and the agent panel won't erroneously hide sibling agents anymore. It's like hiring a savvy butler: previously, he just shook his head without explaining why you couldn't do something, and told you to stop talking while he dispatched chores; now he explains his logic, listens to you while delegating tasks, and remembers to turn off the lights in empty rooms to save power.
Interacting with external systems is significantly smoother. When using bash mode (starting with !), you now get live auto-complete for file paths so you don't have to manually type out long directories. If your MCP (external tool server) requires authentication, the app now clearly notifies you at startup to check /mcp; and if a tool call gets rejected with an auth error (401/403), the helper will now automatically reconnect and retry instead of immediately failing. Additionally, if a marketplace plugin changes its name, your settings will automatically update to follow suit. The /model menu and other data-gated UI elements will also no longer show stale or empty states right after you log in. It's like having a smart assistant who finishes your sentences, automatically swiping their badge again if the security door doesn't open the first time, keeping your address book updated when contacts change their names, and ensuring your rolodex is always perfectly current.
Enterprise admins now have a more transparent monitoring tool. The system added a new OpenTelemetry (a standard logging framework) event that specifically captures Claude's response text. If your deployment was already configured to log user prompts, the upgrade will seamlessly start logging the AI's answers as well, giving you full visibility into what the AI actually produced. Naturally, if you have privacy or security concerns, you can explicitly set a variable (OTEL_LOG_ASSISTANT_RESPONSES=0) to disable this and keep the logs strictly prompts-only. It's like a company security camera that previously only recorded employees speaking into an intercom; now it records the voice coming back from the other side too, though you always have the switch to mute the speaker if needed.
Added `autoMode.classifyAllShell` setting to route all Bash/PowerShell commands through the auto-mode classifier instead of only arbitrary-code-execution patterns
Added auto-mode denial reasons to the transcript, the denial toast, and `/permissions` recent denials
Improved background agents: the launch result no longer instructs Claude to "end your response" — it keeps working on other tasks while the agent runs
Added automatic memory-pressure reaping for idle background shell commands (disable with `CLAUDE_CODE_DISABLE_BG_SHELL_PRESSURE_REAP=1`)
Fixed backgrounding (←←) spuriously cancelling with "N background tasks would be abandoned" when all running tasks carry over to the new session
Fixed pinned background agents being re-prompted to "Continue from where you left off" after every auto-update
Fixed backgrounding the main turn spawning a phantom "general-purpose (resumed)" subagent that re-ran the main conversation
Fixed agent panel hiding sibling agents when viewing a subagentAdded live file path autocomplete to bash mode (`!`)
Added a startup notice when MCP servers need authentication, pointing at `/mcp`
Improved MCP `headersHelper` auth: the helper now re-runs and reconnects automatically when a tool call returns 401/403
Improved plugin auto-rename: marketplace `renames` maps are now followed automatically, updating your settings to the new name
Improved `/add-dir` message when the directory is already a working directory
Fixed `/model` and other client-data-gated UI showing stale/empty state immediately after `/login`Added `claude_code.assistant_response` OpenTelemetry log event containing the model's response text. Redacted unless `OTEL_LOG_ASSISTANT_RESPONSES=1`; when that var is unset it follows `OTEL_LOG_USER_PROMPTS`, so deployments that already log prompt content will start receiving response content on upgrade — set `OTEL_LOG_ASSISTANT_RESPONSES=0` to keep prompts-only.