Claude Code 2.1.268

企業為團隊架設 Claude 應用程式閘道器（gateway）時，現在能享有更一致的計費與安全管控。

只要在 gateway.yaml 設定 pricing:，已登入的 Claude Code 用戶端就會透過受管設定接收相同費率，使 /cost 指令與遙測數據精準對齊支出計量表。當 access_control.allow_cidrs 保持空白時，閘道器在啟動時會主動發出警告，且首次有請求從公開位址抵達時也會發出一次性警告。此外，管理員可配置 gatewayInternalNetworks 受管設定，允許使用者在組織自有的公開 IPv4 網段向閘道器執行 /login 登入。

工件（artifact，在獨立視窗展示的成果文件）迎來視覺與權限更新。發布時瀏覽器分頁會由 Claude 挑選符合頁面的專屬圖示。終端機權限提示改以提問開頭，資料編輯審核改用卡片顯示文件數與可開啟對象。在跳過核准的本機協同工作階段（session）中，安全防護進一步加固。工件工具遇到資料夾外或符號連結（指向另一個位置的捷徑檔）後的檔案會直接拒絕讀取。一般的 WebFetch 規則不再套用至工件讀寫，需改用專屬規則把關：

# 阻擋或把關工件讀寫
Artifact
WebFetch(domain:claude.ai)

自 2.1.265 版起第三方相容端點因結構中的正規表達式（用符號匹配字串的語法）導致回傳 HTTP 400 的問題也已修復。

想要透過腳本自動化管理工具，外掛指令現在全面補齊了結構化輸出，清單各列也補上錯誤與備註細節：

claude auth status --json             # 輸出包含 configDirectory
claude plugin install --json          # 安裝外掛
claude plugin uninstall --json        # 解除安裝
claude plugin update --json           # 更新外掛
claude plugin enable --json           # 啟用外掛
claude plugin disable --json          # 停用外掛
claude plugin list --json             # 每列含 errorDetails 與 noteDetails

在終端機操作 /plugin 選單進行安裝、啟用或停用時，只要關閉選單就會直接生效，不必手動輸入 /reload-plugins。此外，claude plugin validate 先前誤拒以兩個點開頭的外掛路徑，以及外掛檢查時靜默略過預設 monitors 檔案或根目錄 SKILL.md 的問題均已修復。

透過各大雲端平台租用 Claude 進行開發，推論環境與工具載入機制變得更加標準化。在 Bedrock、Vertex 與 Foundry 上的工作階段（session）中，環境、模型與設定改以附件形式送達，與原生對話一致。這些平台上的工具清單在對話中保持位元組層級穩定，較晚連線的工具改以延遲載入而非重寫清單。任務追蹤工具現在預設僅提供給指定模型族系：

# 僅支援 Claude 3.x、Opus 4.0 至 4.7、Sonnet 4.0 至 4.6 與 Haiku 4.5
# 其餘模型需在環境變數開啟：
CLAUDE_CODE_ENABLE_TODO_TOOLS=1

包含 TaskCreate、TaskGet、TaskUpdate、TaskList 與 TodoWrite 工具。此外，使用 excludeDynamicSections 的 SDK 工作階段首則訊息不再於每次請求重新渲染，修復了提示快取（把算過結果暫存起來直接用）與延伸思考中途斷裂的問題。

長時間對話後的歷史摘要與專案記憶，不再因為特殊符號而產生文字損毀。在終端機執行 /compact 指令或觸發自動壓縮時，產生的對話摘要不再損毀包含 $ 字元序列的文字。以 /compact 結尾的對話在恢復時，還原檔案的註記現在於每次恢復時都能維持相同載入順序。SDK 的提示建議、旁支提問與 /rename 指令，修復了在壓縮後仍送出壓縮前舊對話的問題。當專案內的 MEMORY.md 記憶檔案因篇幅超標遭到截斷時，警告訊息現在會清楚說明切除了多少行，以及裁切是從檔案的何處開始。

讓 AI 抓取線上網頁不再陷入無止境的乾等，網路擷取工具正式迎來明確的連線時限。WebFetch 工具遇到伺服器維持連線不結束傳輸時，現在會在 300 秒後逾時判定失敗；你可設定環境變數 CLAUDE_CODE_WEBFETCH_DEADLINE_MS 自訂時限，設為 0 則關閉上限。讀取 localhost 或其他無小數點的主機名稱時，WebFetch 報錯會主動說明遭拒原因並建議改用 curl 指令。在 Claude in Chrome 瀏覽器擴充套件中，導覽網址帶有協定但主機名稱無法解析時，不再詢問是否允許主機「https」。閱讀長篇網頁時內容會直接在對話呈現，不再儲存至本機檔案後再讀回。

自託管執行器的清理預設關閉；只有加上 claude self-hosted-runner --remove-session-state，才會在工作階段結束時刪除 <base-dir>/_sessions/ 下的專屬目錄。權限請求掛鉤（hook，特定事件發生時被觸發的小程式）修復了在 --print 模式下未被觸發的問題。政策協助工具警告也能在無頭模式（headless，沒有終端機視窗的執行模式，即 -p）下印出。/resume 會把 /fork 產生的背景工作階段標成專屬 ⑂ 分岔名稱，不再誤列為父層名稱。未單獨設定超時的 SessionEnd 掛鉤不再強制於 1.5 秒後取消，能依 CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS 延長。

VS Code 擴充套件為習慣鍵盤操作與螢幕報讀工具的使用者，帶來了更順手的新導覽鍵。

在權限詢問畫面中現在能使用方向鍵向左鍵與向右鍵 ← →，切換「一律允許」規則的儲存層級。同時新增了「Claude Code: Focus last message」指令，能隨時將鍵盤焦點快速移回對話中的最新訊息。在外掛管理對話框中進行安裝、啟用、停用或解除安裝時，變更會立即套用至已開啟的工作階段（session），完全無須重新啟動編輯器。此外，部分工件（artifact）權限提示比照命令列終端機規範，移除了「不再詢問」選項以確保每次操作安全。

在編輯器中開啟多個對話分頁或切換帳戶，選單與模型狀態不再莫名閃爍留白。登入、登出或切換帳號後，已開啟分頁的模型標籤、模型選取器與指令選單，修復了在數秒間呈現空白的問題。當專案或本機設定覆蓋全域設定中的模型時，新分頁或重新整理的對話中「Auto」選項不再無故消失。配置了 SessionStart 掛鉤（hook，特定事件發生時被觸發的小程式）時，視窗重載後的命名問題已獲修正。工作階段（session）名稱重載後不再退回最後一次提問。視窗內已有其他分頁運行時，底部狀態列的模型與遠端控制狀態燈不再空等新行程啟動。若分頁啟動在讀取設定後超過半秒抵達，第二個 Claude 行程也不再重複完整啟動流程。

自訂目錄與 Windows 診斷，現在能切合實際開發環境。

當透過設定檔或 environmentVariables 指定自訂目錄 CLAUDE_CONFIG_DIR 時，工作階段（session）清單、設定開關與分頁能正確讀取目錄，樣式工具不再把使用者樣式存入命令列讀不到的位置。從清單恢復對話時，系統不再忽視 claudeCode.preferredLocation: "sidebar" 而強制開啟面板，程式化開啟也不會重設為「panel」。未安裝 WSL 的 Windows 電腦不再誤跳安裝提示，安裝了 WSL 也能正確回傳本機檔案診斷。

網頁版 Claude Code 的檔案儲存機制獲得延長，雲端工作階段成果現在能完整保留一整天。執行時間超過約六小時的雲端工作階段（session），過去會悄悄遺失存放在持久化工作階段資料夾內的檔案，如今儲存成果可穩定保留長達一天。組織管理員若對模型的努力程度（effort level）設限，相關報錯已獲修復。例行工作（routine）在恢復工作階段或啟動時若未指定努力程度，不再拋出「Invalid effort level」無效等級錯誤。從既有對話建立新例行工作時，若尚未配置任何連接器，Claude 不再只是單純回覆確認，而是會主動說明並提供具體的添加指引。

企業在 Slack 中管理 Claude Tag 的後台介面，遇到短暫斷線也能迅速點擊重試。

管理設定頁面遇到短暫載入失敗時，不再卡在骨架或留白，失敗區塊現在會提供「重試」按鈕。Slack 頻道的設定頁面中，新增了返回組織 Claude in Slack 管理後台的連結。在 Slack Enterprise Grid 架構下，管理員將頻道移動至其他工作區時，Claude 原有的專案存放庫、環境與存取設定均能妥善保留。若你的 Claude 帳號所屬組織與 Slack 工作區不同，系統訊息會指引如何將工作區連結至組織。頻道設定卡片被多次點擊時，在變更套用後不再誤向提議的工作階段（session）回報拒絕，結果只會發送一次。

團隊頻道中的 Claude Tag 阻擋說明變得更具體，各公開頻道的專屬筆記也正式各自獨立。

當 Claude 停止執行某項動作時，現在會具體說明背後原因，清楚指出究竟是權限檢查未過、它自身決定先向你確認，還是單純缺少存取權限。在允許訪客使用 Claude 的頻道中，工作區訪客發起頂層 @ 提及時，不再誤收到「Slack 帳號未連接」的錯誤回覆。頻道中長時間運作的工作階段（session）進行排程更新時，不再於對話途中強行換新，而是會靜候頻道與所有討論串無人發言時才進行替換。公開頻道的記憶機制也有所革新，每個公開頻道現在各自保存專屬筆記，不再彼此調閱其他公開頻道的筆記內容，而整個工作區的共用筆記則維持互通。

自動程式碼審查（Code Review）在追蹤修復進度與處理多代理人任務時更加縝密。在後續追蹤審查中，未結案清單下方新增了說明，明確提醒必須將問題討論串標示為已解決（resolve）而非僅留言回覆，後續審查才不會持續計為未結案。當負責驗證問題的代理人（agent）在中途發生異常時，審查系統會主動替換該代理人並完成裁決，不再懸在未完成狀態。當排隊等待的推送觸發審查在等待期間，若拉取請求已被轉為草稿（draft）狀態，輪到時不會再於事後張貼審查結果。此外，拉取請求修改根目錄檔案（例如 README.md）時，審查系統不再因檔名碰巧與某目錄清單中的檔案相同，就誤忽略該目錄所定義的 CLAUDE.md 規範。

When enterprises deploy a Claude apps gateway for their teams, they now get more consistent billing and security controls.

By setting pricing: in gateway.yaml, signed-in Claude Code clients receive the same rates through managed settings, keeping the /cost command and telemetry accurately matched to the spend meter. When access_control.allow_cidrs is empty, the gateway now issues a warning at startup and a one-time warning the first time a request arrives from a public address. In addition, administrators can configure the gatewayInternalNetworks managed setting to allow users to /login to the gateway on their organization's own public IPv4 block.

Artifacts receive visual and permission updates. When published, browser tabs now display a dedicated icon chosen by Claude to match each page. The terminal permission prompt now leads with the ask's question, and data-edit permission prompts switch to a card showing the document count and who can open the artifact. In local Cowork sessions configured to skip all approvals, security protections are further hardened. The Artifact tool now directly refuses to read local files outside session folders or behind symlinks. Plain WebFetch rules no longer apply to Artifact tool reads and updates, requiring dedicated rules to gate them:

# Block or gate Artifact reads and updates
Artifact
WebFetch(domain:claude.ai)

An issue since version 2.1.265 where third-party compatible endpoints returned HTTP 400 due to a regex in the schema has also been fixed.

For automating tool management via scripts, plugin commands now provide full structured output support, with error and note details added to list rows:

claude auth status --json             # Output includes configDirectory
claude plugin install --json          # Install plugin
claude plugin uninstall --json        # Uninstall plugin
claude plugin update --json           # Update plugin
claude plugin enable --json           # Enable plugin
claude plugin disable --json          # Disable plugin
claude plugin list --json             # Each row includes errorDetails and noteDetails

When installing, enabling, or disabling plugins through the /plugin terminal menu, changes take effect immediately upon closing the menu without running /reload-plugins. Additionally, fixes resolve claude plugin validate mistakenly rejecting plugin paths beginning with two dots, as well as plugin checks silently skipping default monitors files or root SKILL.md files.

When developing with Claude through major cloud platforms, inference environments and tool-loading mechanisms are now more standardized. In Bedrock, Vertex, and Foundry sessions, environment, model, and settings details are delivered as attachments, matching first-party sessions. Tool lists on these platforms remain byte-stable across conversations, with late-connecting tools loading deferred instead of rewriting the list. Task-tracking tools are now offered by default only on designated model families:

# Only Claude 3.x, Opus 4.0 to 4.7, Sonnet 4.0 to 4.6, and Haiku 4.5
# Enable for other models via environment variable:
CLAUDE_CODE_ENABLE_TODO_TOOLS=1

This includes TaskCreate, TaskGet, TaskUpdate, TaskList, and TodoWrite tools. Furthermore, for SDK sessions using excludeDynamicSections, the first message is no longer re-rendered on each request, fixing prompt caching and extended thinking breaking mid-session.

History summaries and project memory after lengthy conversations no longer suffer text corruption from special characters. When executing the /compact command in the terminal or triggering auto-compact, generated summaries no longer mangle text containing $ character sequences. When resuming conversations that ended with /compact, restored-file notes now maintain the exact same loading order on every resume. For SDK prompt suggestions, side questions, and /rename, fixes ensure they no longer send the pre-compaction transcript. When a project's MEMORY.md file is truncated due to length limits, the warning message now explicitly states how many lines were cut and where the truncation begins.

Fetching web pages with AI no longer hangs indefinitely, as web scraping tools now enforce explicit connection deadlines. When a server holds a connection open without finishing transmission, the WebFetch tool now times out and fails after 300 seconds; you can set the CLAUDE_CODE_WEBFETCH_DEADLINE_MS environment variable to customize this deadline, or set it to 0 to disable the limit. When accessing localhost or other dotless hostnames, WebFetch errors now explain why the URL was refused and suggest using curl instead. In the Claude in Chrome extension, navigation URLs with a scheme but unparseable hostnames no longer prompt to allow the host "https". When reading lengthy web pages, content is now presented directly inline rather than being saved to a local file and read back.

Cleanup for self-hosted runners is disabled by default; only when passing claude self-hosted-runner --remove-session-state will dedicated directories under <base-dir>/_sessions/ be deleted upon session completion. PermissionRequest hooks fix an issue where they failed to fire in --print mode. Policy-helper warnings now also print during headless (-p) runs. The /resume command now labels background sessions created by /fork with their dedicated ⑂ fork name instead of incorrectly listing them under their parent's name. SessionEnd hooks without a per-hook timeout are no longer forcibly cancelled after 1.5 seconds, allowing extensions via CLAUDE_CODE_SESSIONEND_HOOKS_TIMEOUT_MS.

The VS Code extension introduces smoother navigation keys for keyboard and screen reader users.

In permission prompts, Left and Right arrow keys ← → can now switch where an always-allow rule is saved. A new "Claude Code: Focus last message" command quickly shifts keyboard focus back to the newest message in the conversation. When installing, enabling, disabling, or uninstalling in the Manage plugins dialog, changes apply immediately to open sessions without requiring an editor restart. Additionally, some artifact permission prompts omit the "don't ask again" choice to match terminal security standards.

When opening multiple chat tabs or switching accounts in the editor, menus and model states no longer flicker or go blank. After logging in, logging out, or switching accounts, model pills, model pickers, and command menus in open tabs resolve an issue where they went blank for several seconds. When project or local settings override the model named in global settings, the "Auto" option no longer disappears in new tabs or reloaded conversations. When a SessionStart hook is configured, naming issues after window reloads are resolved. Session names no longer revert to the last prompt after a reload. When another tab is already running in the window, footer model and Remote Control pills no longer wait needlessly for a new process to launch. If a session tab's launch arrives more than half a second after reading configuration, a second Claude process no longer runs through the full startup cycle.

Custom directories and Windows diagnostics now properly match real development setups.

When CLAUDE_CONFIG_DIR is set in settings files or via the environmentVariables setting, session lists, settings toggles, and chat tabs read the directory correctly, and style builders no longer save user styles into paths unreadable by the CLI. Resuming sessions from the session list no longer ignores claudeCode.preferredLocation: "sidebar" by forcing panel views, and programmatic opens no longer reset the preference to "panel". Windows machines without WSL no longer show false WSL installation prompts, and systems with WSL installed now return local file diagnostics accurately.

File persistence mechanisms in Claude Code on the web are extended, allowing cloud session artifacts to stay intact for a full day. Cloud sessions running longer than about six hours previously lost files saved to persisted session folders, but saved results now reliably persist for up to a day. Error handling is resolved when organization administrators cap a model's effort level. When a routine resumes a session or starts without a specified effort level, it no longer throws "Invalid effort level" errors. When creating a new routine from an existing conversation that lacks connectors, Claude no longer just confirms creation, but instead explains the situation and provides concrete guidance on how to add them.

Enterprise admin interfaces for Claude Tag in Slack can now quickly recover from transient disconnects with a retry button.

When the admin settings page hits a transient loading failure, it no longer gets stuck on loading skeletons or blank screens; failed sections now provide a Retry button. A link has been added from Slack channel configure pages back to the organization's Claude in Slack admin settings. Under Slack Enterprise Grid, when administrators move channels to another workspace, Claude settings including repositories, environments, and access permissions remain intact. If your Claude account belongs to a different organization than the Slack workspace, system notices explain how to connect the workspace to your organization. When channel-settings cards are clicked multiple times, they no longer falsely report refusals to the proposing session after changes have applied, sending the outcome only once.

Block explanations for Claude Tag in team channels are now more specific, and public channels keep separate dedicated notes.

When Claude stops an action, it now provides specific reasons, clearly indicating whether a permission check failed, it decided to confirm with you first, or access was simply missing. In channels allowing guests to use Claude, workspace guests making top-level @mentions no longer mistakenly receive "Slack account not connected" errors. When long-running channel sessions undergo scheduled refreshes, they are no longer forcefully swapped mid-conversation, waiting instead until the channel and its threads are completely quiet. Memory in public channels is also updated so that each public channel maintains its own notes rather than recalling notes saved in other public channels, while workspace-wide shared notes remain accessible across channels.

Automated code reviews are more thorough when tracking fix progress and handling multi-agent tasks. In follow-up reviews, a note is added beneath the open findings list reminding users that resolving a finding's thread, rather than merely replying to it, prevents later reviews from counting it as open. When an agent responsible for verifying a finding fails midway, the review system replaces that agent to reach a verdict instead of remaining incomplete. When a queued push-triggered review is waiting and the pull request is converted to a draft, it no longer posts review results when its turn arrives. Additionally, when a pull request edits a root file such as README.md, the review system no longer ignores CLAUDE.md conventions for a directory just because the filename matches a file listed in that directory.