ai_updates.json

{
  "generated_at": "2026-10-23",
  "tools": [
    {
      "id": "claude_code",
      "name": "Claude Code",
      "versions": [
        {
          "version": "2.1.309",
          "period": "2026-10-22",
          "items": [
            {
              "title": {
                "zh-TW": "ultrareview 嚴格檢查工作區信任，未獲授權立即中斷",
                "en": "ultrareview Enforces Workspace Trust Checks, Halting Immediately Without Authorization"
              },
              "body": {
                "zh-TW": "當你在終端機使用 `claude ultrareview` 進行進階程式碼審查時，工具現在會主動確認目前工作區是否值得信任。在一般終端機下它會跳出確認提問，若沒有互動終端或在自動化環境則會直接拒絕執行。在自動化腳本中，你必須加入不帶參數的 `--trust-workspace` 旗標（需放在 PR 編號或基準分支後面）。或者設定環境變數（執行時讀取的設定值）`CLAUDE_CODE_DISABLE_ULTRAREVIEW_TRUST_CHECK=1`。⟦好比大樓警衛室只讓登記過的訪客上樓，以前審查員可能直接走進私人辦公室，現在必須先刷卡確認訪客名單，未登記者一律在門外擋下。⟧當你在持續整合伺服器上排程跑深度審查，卻發現指令中斷時，可按上述位置補上信任旗標，通過工作區信任檢查。",
                "en": "When you run advanced code reviews in the terminal using `claude ultrareview`, the tool now actively verifies whether the current workspace is trusted. In a standard terminal, it prompts for confirmation, but it directly refuses execution if there is no interactive terminal or inside an automated environment. In automated scripts, you must include the `--trust-workspace` flag with no arguments (placed after the PR number or base branch). Alternatively, you can set the environment variable (configuration values read at runtime) `CLAUDE_CODE_DISABLE_ULTRAREVIEW_TRUST_CHECK=1`. ⟦Picture a building security desk that only allows registered visitors upstairs: where reviewers could previously walk straight into private offices, they must now badge in against the visitor log, and anyone unlisted is stopped at the door.⟧ When a scheduled in-depth review on a continuous integration server stops unexpectedly, add the trust flag in the position described above to pass the workspace trust check."
              },
              "original": "Changed `claude ultrareview` to require workspace trust: it asks on a terminal, else refuses; scripts pass `--trust-workspace`, which takes no value (put a PR number or base branch before it), or set `CLAUDE_CODE_DISABLE_ULTRAREVIEW_TRUST_CHECK=1`"
            },
            {
              "title": {
                "zh-TW": "/claude-api 技能換裝多代理人架構，預設啟用動態工作流程",
                "en": "/claude-api Skill Adopts Multi-Agent Architecture, Enabling Dynamic Workflows by Default"
              },
              "body": {
                "zh-TW": "當你透過 `/claude-api` 技能規劃雲端代理人團隊時，背後生成團隊的底層範本已經升級。以往工具會採用舊有的協調員架構（coordinator）。現在改以 `multiagent_20261001` 架構建立 Managed Agents（由平台託管的代理人）子團隊，並預設開啟動態工作流程。⟦宛如專案工作室調整團隊編制，以前沿用舊的組織方式，現在改用新的團隊組織方式。⟧當你需要為多步驟任務設計多個助理分工時，直接呼叫 `/claude-api`，產出的代理人團隊就會採用這套架構，動態工作流程也已預設開啟。",
                "en": "When you design cloud agent teams through the `/claude-api` skill, the underlying template generating those teams has been upgraded. Previously, the tool relied on the legacy coordinator architecture. It now builds Managed Agents (agents hosted by the platform) subteams using the `multiagent_20261001` architecture, with dynamic workflows enabled by default. ⟦A project studio reorganizes its staffing: where it previously relied on an older organizational setup, it now adopts a new team structure.⟧ When you need to delegate multi-step tasks across multiple assistants, call `/claude-api` directly, and the generated agent team will use this architecture with dynamic workflows turned on by default."
              },
              "original": "Changed the `/claude-api` skill to build Managed Agents subagent teams on `multiagent_20261001` with dynamic workflows on by default, instead of `coordinator`"
            },
            {
              "title": {
                "zh-TW": "sed -i 採用標準指令確認，修改內容不再交由 /rewind 復原",
                "en": "sed -i Adopts Standard Command Prompts, Leaving Edits Outside /rewind Restoration"
              },
              "body": {
                "zh-TW": "在對話中要求 Claude 直接批次替換檔案文字時，`sed -i` 指令的執行與確認流程有了調整。以前這類就地修改指令會跳出模擬的檔案編輯預覽，現在它改為顯示標準的 Bash 執行提示，並依照輸入的原文直接運行。與其他透過 Bash 進行的修改一樣，執行過後的檔案變更不會再被 `/rewind` 指令復原。⟦跟在便條紙上動筆塗改一樣，以前動筆前會先在旁邊墊一張描圖紙給你看預覽，現在直接按指令在紙上落筆。⟧如果你打算在大量程式碼中批次取代變數名稱，執行前請確認修改內容，因為 `/rewind` 不會復原這類修改。",
                "en": "When you ask Claude to batch-replace file text in a conversation, the execution and confirmation flow for the `sed -i` command has been adjusted. Previously, this in-place modification command displayed a simulated file edit preview. It now displays a standard Bash execution prompt and runs directly with the verbatim input. Like other modifications made through Bash, resulting file changes will no longer be restored by the `/rewind` command. ⟦Think of writing corrections on a notepad: where you once placed tracing paper over the page to preview changes, you now put pen directly to paper as instructed.⟧ If you plan to batch-replace variable names across extensive code, verify your changes before running, because `/rewind` will not undo these modifications."
              },
              "original": "Changed `sed -i` to get the standard Bash prompt and run as written, not a simulated file-edit preview; like other Bash edits, it is no longer restored by `/rewind`"
            },
            {
              "title": {
                "zh-TW": "補強沙盒隔離防線，嚴格限制設定檔與捷徑路徑寫入",
                "en": "Sandbox Isolation Hardened with Stricter Write Limits on Configs and Symlinks"
              },
              "body": {
                "zh-TW": "當你在 macOS 與 Linux 上啟用沙盒執行指令時，多項環境隔離與檔案寫入的防護邊界獲得了修正。沙盒內的 Bash 指令以前會在執行前於沙盒外載入 Shell 啟動檔，現在不會在隔離環境外偷跑。當目錄透過符號連結（指向真實位置的捷徑檔）連接時，macOS 的 `denyWrite` 規則現在會確實比對目標的真實位置，設定檔儲存庫的防寫限制也不會被捷徑繞過。沙盒指令與自動核准的修改不再能更動由 `--mcp-config` 指定的 MCP（讓 AI 連接外部工具的協定）設定檔。在 Linux 環境中，背景指令即使存活超過原本的工作階段，沙盒禁止寫入的防線也不再提早解除。若沙盒處於啟用但無法運作的狀態，排除於沙盒外的指令在寫入允許清單的路徑前會主動跳出確認，並直接拒絕未隔離的備援操作。",
                "en": "When you enable sandboxing to run commands on macOS and Linux, several protection boundaries for environment isolation and file writes have been fixed. Bash commands inside the sandbox previously loaded Shell startup files outside the sandbox before execution. They no longer run outside the isolated environment. When directories are connected via symlinks (shortcut files pointing to real locations), macOS `denyWrite` rules now accurately match the real target location, preventing shortcuts from bypassing write restrictions on configuration repositories. Sandboxed commands and auto-approved modifications can no longer alter MCP (Model Context Protocol, enabling AI to connect with external tools) configuration files specified by `--mcp-config`. In Linux environments, even if background commands outlive their original session, write-prohibition boundaries in the sandbox are no longer lifted prematurely. If the sandbox is enabled but non-functional, commands exempted from the sandbox will prompt for confirmation before writing to allowlisted paths, while un-isolated fallback operations are rejected outright."
              },
              "original": "Fixed sandboxed Bash commands on macOS and Linux running the shell's startup files outside the sandbox before each command\nFixed macOS sandbox `denyWrite` glob patterns not applying at a path's real location when the pattern's folder is reached through a symlink\nFixed sandboxed commands and auto-accepted edits being able to modify the MCP config file passed with `--mcp-config`\nFixed profile store write protection not applying when the config directory is reached through a symlink\nLinux: Fixed sandbox write denies lifting when a background command outlived the session\nChanged commands the settings exclude from the sandbox to ask before writing to its write-list paths while the sandbox is enabled but cannot run; the unsandboxed fallback is refused there"
            },
            {
              "title": {
                "zh-TW": "終端機強化指令審核，嚴防重導向與目錄切換繞過授權",
                "en": "Terminal Strengthens Command Auditing, Blocking Redirection and Directory Changes Bypasses"
              },
              "body": {
                "zh-TW": "當 Claude 在終端機執行複合指令或刪除檔案時，幾處曾會漏掉權限提示或錯誤信任狀態的盲點已全數修復。Bash 權限檢查以前會盲目信任由指令設定的變數，即使該指令的輸出重導向根本無法開啟；現在不會再誤判為成功。當指令在執行 git 前先變更目錄，若 `cd` 被包裝或帶有前綴，Bash 檢查現在會像一般 `cd` 一樣確實跳出確認提示。當複合指令包含子 Shell 時，`Edit(path)` 阻擋規則現在能確實攔截指向該路徑的重導向寫入。在 Windows 上，若 `Set-Location` 或 `cd` 目標為當前目錄但帶有額外參數，PowerShell 權限檢查不再自動放行，而是會跳出核准詢問。桌面應用程式與 VS Code 在略過權限模式下若遇到危險的 `rm` 與 `rmdir` 指令，現在會直接拒絕並附上改寫建議，避免無人值守的工作階段卡在確認畫面。",
                "en": "When Claude runs compound commands or deletes files in the terminal, several blind spots that previously missed permission prompts or falsely trusted state have been fixed. Bash permission checks previously blindly trusted variables set by a command even if output redirection failed to open. They no longer misjudge such executions as successful. When a command changes directories before running git, Bash checks now properly prompt for confirmation if `cd` is wrapped or prefixed, matching standard `cd` behavior. When compound commands contain subshells, `Edit(path)` blocking rules now reliably intercept redirection writes targeting that path. On Windows, if `Set-Location` or `cd` targets the current directory with extra arguments, PowerShell permission checks no longer auto-approve the command, prompting for confirmation instead. When dangerous `rm` and `rmdir` commands run in bypass permissions mode, the desktop app and VS Code now reject them directly with rewrite suggestions, preventing unattended sessions from stalling on confirmation prompts."
              },
              "original": "Fixed the Bash permission check trusting a variable set by a command whose redirect can fail to open\nFixed Bash permission checks not prompting for some compound commands that change directory before running git when the `cd` is wrapped or prefixed, as they do for a plain `cd`\nFixed an `Edit(path)` deny rule not blocking a Bash redirect to that path in some commands that also held a subshell\nWindows: Fixed PowerShell permission checks auto-approving some `Set-Location` or `cd` commands that target the current directory and carry extra arguments; these now prompt for approval\nChanged the dangerous `rm` and `rmdir` prompt in bypass permissions mode in the Desktop app and VS Code to deny the command at once with a rewrite hint, so unattended sessions keep going"
            },
            {
              "title": {
                "zh-TW": "收緊工作樹隔離防護，Git 操作提速並預先載入物件",
                "en": "Worktree Isolation Guard Tightened, Accelerating Git Operations and Pre-fetching Objects"
              },
              "body": {
                "zh-TW": "當你啟用隔離環境讓多個對話平行處理不同任務時，版本控制與工作目錄的邊界防護有了重要改善。工作樹（Git 讓同一專案平行開多份副本的功能）隔離的工作階段中，先前部分帶有語系前綴、原本應被隔離守衛拒絕的指令會被放行，現在已修正；zsh 下原本會漏網的算術運算指令也會被阻擋。由 Claude Code 建立的工作樹與分支簽出，現在會像其他 Git 操作一樣主動忽略替換參照（replace refs）。Windows 上多項內部 Git 指令變快；在所有平台上，這些指令的錯誤訊息都改以英文顯示，不受 Git 語言設定影響。在一般複製品或設定了 Bash 沙盒的部分複製品中，建立與重複使用工作樹時不再隨選抓取物件；這類部分複製品會先抓取物件，再建立或重複使用工作樹。",
                "en": "When you enable isolated environments so multiple conversations can handle different tasks in parallel, boundary protections for version control and working directories have improved significantly. Inside sessions isolated by worktrees (Git features that maintain parallel copies of the same project), commands with locale prefixes that should have been rejected by the isolation guard previously slipped through, but this is now fixed. Arithmetic commands under zsh that previously went unnoticed are also blocked. Worktrees and branch checkouts created by Claude Code now actively ignore replace references, matching other Git operations. Multiple internal Git commands are faster on Windows. Across all platforms, error messages from these commands are now displayed in English, regardless of Git locale settings. In standard clones or partial clones configured with Bash sandboxing, creating and reusing worktrees no longer fetches objects on demand. Such partial clones now pre-fetch objects before creating or reusing a worktree."
              },
              "original": "Fixed worktree-isolated sessions accepting some locale-prefixed commands the isolation guard should refuse\nFixed worktree-isolated sessions on zsh accepting some arithmetic commands that the isolation guard should refuse\nFixed worktrees and branch checkouts that Claude Code creates so they ignore git replace refs, as its other git operations already do\nImproved the speed of many git commands Claude Code runs for itself on Windows; on every platform, errors from those commands now show in English even when git is set to another language\nChanged worktree creation and reuse so git no longer fetches objects on demand, in ordinary clones and in partial clones with a Bash sandbox configured; such a partial clone fetches them first"
            },
            {
              "title": {
                "zh-TW": "/claude-api 擴充雲端追蹤支援，查核即時文件並開放授權令牌",
                "en": "/claude-api Expands Cloud Tracing Support, Verifying Live Docs and Enabling Auth Tokens"
              },
              "body": {
                "zh-TW": "當你請 `/claude-api` 技能協助配置雲端服務與鏈路追蹤時，多項說明準確度與追蹤設定已獲得改善。技能說明不再誤稱 `ant beta:worker poll --on-work` 會傳遞追蹤上下文。修正 C# 追蹤配置遺漏雲端平台客戶端跨度（spans）的瑕疵。修正原本聲稱 AWS 上的 Claude Platform 請求跨度無法加入呼叫端追蹤的錯誤說法。要求加入 OpenTelemetry 時，技能現在會在 SDK 追蹤之外同時主動提供 API 追蹤選項。技能在回覆功能不受支援前，會先連線查核最新即時官方文件。服務帳號金鑰與工作負載身分同盟的授權令牌（用來驗證身分的數位令牌），現在已可用於管理 OpenTelemetry 的輸出目的地。",
                "en": "When you ask the `/claude-api` skill to configure cloud services and distributed tracing, several documentation accuracies and tracing settings have been improved. The skill description no longer mistakenly claims that `ant beta:worker poll --on-work` propagates trace context. It fixes an issue in C# tracing setups where cloud platform client spans were omitted. It corrects the inaccurate claim that Claude Platform request spans on AWS cannot join caller traces. When asked to add OpenTelemetry, the skill now proactively provides API tracing options alongside SDK tracing. Before responding that a feature is unsupported, the skill now connects online to verify against current official documentation. Service account keys and authorization tokens (digital credentials used for identity verification) from workload identity federation can now be used to manage OpenTelemetry export destinations."
              },
              "original": "Fixed the /claude-api skill saying that `ant beta:worker poll --on-work` passes trace context to the command\nFixed the /claude-api skill's C# tracing setup missing spans from the cloud platform clients\nFixed the /claude-api skill saying that API spans for requests to Claude Platform on AWS can't join the caller's trace\nImproved the /claude-api skill to offer API tracing alongside SDK tracing when asked to add OpenTelemetry\nImproved the /claude-api skill to check the live docs before saying a feature is unsupported\nUpdated the /claude-api skill: service account keys and Workload Identity Federation tokens can now manage OpenTelemetry destinations"
            },
            {
              "title": {
                "zh-TW": "技能命名支援目錄前綴區隔，資料圖表導入新版對比標準",
                "en": "Skill Names Support Directory Prefixes, Introducing Modern Contrast Standards to Dataviz"
              },
              "body": {
                "zh-TW": "當你在專案中載入多組外掛或自訂技能時，技能重複衝突與檔案讀取例外處理已得到修復。透過 `--add-dir` 載入的技能若與既有技能同名，以前會無法使用，現在會自動加上前綴顯示為 `<folder>:<name>`。在終端機 Shell 變更檔案後，`/verify` 與 `/code-review` 不再發生找不到巢狀套件自身技能的問題。從 claude.ai 同步的外掛與技能，在遭遇磁碟讀取錯誤時不會再於工作階段中途莫名被整組移除。內建的 dataviz 資料視覺化技能正式支援 APCA 文字對比度演算標準（在需要符合 AA 規範處仍維持 WCAG 2 比例），調色盤驗證腳本新增 `apca(text, background)` 函式，並能指引圖表著色區塊內的標籤該搭配白色或接近黑色。",
                "en": "When you load multiple plugins or custom skills into a project, duplicate skill collisions and file-read exception handling have been resolved. If a skill loaded via `--add-dir` shares the same name as an existing skill, it was previously unusable. It is now automatically prefixed as `<folder>:<name>`. After modifying files in the terminal Shell, `/verify` and `/code-review` no longer encounter issues locating skills belonging to nested packages. Plugins and skills synced from claude.ai are no longer removed entirely mid-session when encountering disk read errors. The built-in dataviz skill officially supports the APCA text contrast algorithm standard (while maintaining WCAG 2 ratios where AA compliance is required). Its palette verification script introduces the `apca(text, background)` function and guides whether labels inside shaded chart regions should pair with white or near-black."
              },
              "original": "Fixed a skill in an `--add-dir` folder being unreachable when another skill already held its name; it is now offered as `<folder>:<name>`\nFixed `/verify` and `/code-review` not finding a nested package's own skill when its files were changed from the shell\nFixed plugins and skills synced from claude.ai being removed, sometimes mid-session, after a disk read error\nUpdated the bundled dataviz skill to accept APCA as a text-contrast measure (keeping WCAG 2 ratios where AA conformance is required), add an `apca(text, background)` function to its palette validator scripts, and say whether a label inside a colored chart fill should be white or near-black"
            },
            {
              "title": {
                "zh-TW": "修正對話送出與檢視快捷鍵，工具預覽文字妥善處理表情符號",
                "en": "Shortcut Keys for Sending and Viewing Fixed, with Better Emoji Handling in Tool Previews"
              },
              "body": {
                "zh-TW": "當你在終端機快速輸入文字與切換畫面時，快捷鍵的觸發反應與對話紀錄排版獲得了修正。以前在模型回覆開始前按下 ctrl+enter 立即送出，訊息只會進入排隊；若在提示詞掛鉤（特定事件發生時被觸發的小程式）執行期間按下，還會重啟掛鉤並將訊息留在等待佇列，現在已徹底修正。在 `!` 開頭的 Shell 模式下，綁定快捷鍵的斜線指令不再錯誤地被丟進系統 Shell 中執行。按下 ctrl+o 時，來自其他代理人與工作階段中等待投遞的訊息，現在會完整標記為等待狀態並顯示在紀錄下方。工具輸出的長預覽文字若剛好截斷在表情符號中間，以往會導致工作階段紀錄損毀而無法被 `jq` 解析，現在已修正截斷處理。",
                "en": "When you quickly type text and switch views in the terminal, shortcut responsiveness and transcript formatting have been fixed. Previously, pressing ctrl+enter to send immediately before the model began responding merely queued the message. If pressed while prompt hooks (small programs triggered by specific events) were executing, it restarted the hook and left the message stranded in the queue, which is now resolved. In Shell mode prefixed with `!`, slash commands bound to shortcuts are no longer erroneously forwarded to the system Shell for execution. When pressing ctrl+o, pending messages awaiting delivery from other agents and sessions are now fully flagged as pending and displayed at the bottom of the log. If long preview text from tool output was truncated in the middle of an emoji, it previously corrupted the session transcript and prevented `jq` from parsing it, which has now been fixed."
              },
              "original": "Fixed ctrl+enter (send now) only queuing a message when pressed before the response had started, and restarting a prompt's hooks while leaving the message queued when pressed while those hooks were running\nFixed key-bound slash commands running in the shell from `!` mode\nFixed ctrl+o not showing messages from other agents and sessions that are still waiting to be delivered; they now appear in full under the transcript, marked as waiting\nFixed long tool output previews being cut through an emoji, leaving session transcripts that `jq` could not parse"
            },
            {
              "title": {
                "zh-TW": "隊友代理人繼承工具限制，自架執行器接軌中斷操作",
                "en": "Teammate Agents Inherit Tool Restrictions, Bridging Interrupted Operations on Self-Hosted Runners"
              },
              "body": {
                "zh-TW": "當你啟動分割視窗的隊友代理人或在自架伺服器上執行任務時，跨代理人設定傳遞與狀態銜接更為一致。分割視窗中的隊友代理人現在會確實繼承隊長在指令列指定的工具禁用規則與 `--tools` 限制。透過 `/rename` 重新命名的工作階段，在 VS Code 視窗重新載入或非互動式 `--resume` 恢復後，不會再以隨機產生的名稱顯示給其他對話。若該回合透過 `/code-review --fix`、子代理人或 Shell 指令修改了已納入檢查點的檔案，`/rewind` 清單不再誤顯示「沒有程式碼變更」。在自架執行器環境中，若首個提示詞仍在子代理人中運作，`--startup-timeout-min` 不會誤殺工作階段，但凍結的行程仍會被停止。當工作階段轉移到新的自架執行器時，Claude 現在能清楚看見被中斷的工具呼叫，並將其標記為結果未知。",
                "en": "When you launch split-pane teammate agents or run tasks on self-hosted servers, cross-agent configuration propagation and state transitions are more consistent. Teammate agents in split panes now properly inherit tool disablement rules and `--tools` restrictions specified on the command line by the lead agent. Sessions renamed via `/rename` no longer revert to randomly generated names when viewed from other conversations after a VS Code window reload or a non-interactive `--resume`. If a turn modifies checkpointed files through `/code-review --fix`, subagents, or Shell commands, the `/rewind` list no longer erroneously displays 'no code changes'. In self-hosted runner environments, `--startup-timeout-min` will not terminate the session prematurely if the initial prompt is still active in a subagent, though frozen processes are still stopped. When a session transitions to a new self-hosted runner, Claude can now clearly see interrupted tool calls and mark them as having an unknown result."
              },
              "original": "Fixed split-pane agent teammates not inheriting the tool deny and `--tools` rules given on the leader's command line\nFixed sessions renamed with `/rename` appearing to other sessions under a generated name after a VS Code window reload or a non-interactive `--resume`\nFixed the `/rewind` list showing \"No code changes\" for a turn that changed checkpointed files through `/code-review --fix`, a subagent or a shell command\nSelf-hosted runner: Fixed `--startup-timeout-min` stopping a session whose first prompt, such as `/code-review`, was still working in a subagent; a frozen process is still stopped\nSelf-hosted runner: Improved sessions that move to a new runner: Claude now sees the tool call that was interrupted, marked as outcome unknown"
            },
            {
              "title": {
                "zh-TW": "修復 PowerShell 掛鉤阻擋機制，別名模型自動解析",
                "en": "PowerShell Hook Blocking Fixed, Aliased Models Automatically Resolve"
              },
              "body": {
                "zh-TW": "當你在工作流程中設定自訂掛鉤（特定事件發生時被觸發的小程式）以攔截工具或步驟時，Windows 與模組載入的相容性獲得了提升。在 PowerShell 中運行的掛鉤程式若以代碼 2 結束，以往無法成功阻擋操作，現在已能正確攔截。在未安裝 Git for Windows 的 Windows 環境中，指定使用 PowerShell 的工具前置與權限請求掛鉤也不再因缺少該工具而直接失守放行。自訂模組的 `turn.step` 掛鉤中若以別名（例如 `haiku`）指定模型，以前每次請求都會失敗，現在會正確解析出對應模型。當掛鉤模組因巢狀過深而無法載入時，錯誤訊息現在會明確點名出問題的檔案路徑，不再只印出簡略的堆疊溢位提示。",
                "en": "When configuring custom hooks (small programs triggered when specific events occur) in your workflow to intercept tools or steps, compatibility with Windows and module loading has improved. Hook programs running in PowerShell that exit with code 2 previously failed to block operations, but now intercept them correctly. On Windows environments where Git for Windows is not installed, pre-tool and permission-request hooks specifying PowerShell no longer fail open simply due to the missing tool. If a model is specified by an alias (such as `haiku`) in the `turn.step` hook of a custom module, requests that previously failed every time now correctly resolve the corresponding model. When a hook module fails to load due to excessive nesting, error messages now explicitly identify the problematic file path instead of printing a terse stack overflow notice."
              },
              "original": "Fixed hooks that run in PowerShell not blocking when the hook's program exits with code 2, and on Windows `PreToolUse` and `PermissionRequest` hooks with `\"shell\": \"powershell\"` failing open when Git for Windows isn't installed\nFixed a mod's `turn.step` hook that names a model by alias, such as `haiku`, failing every request with a model error: the alias now resolves to its model\nImproved the error when a hooks module is nested too deep to load: it now names the file instead of showing a bare stack-overflow message"
            },
            {
              "title": {
                "zh-TW": "大型說明檔避免重複壓縮，成品發布與權限不再頻繁卡關",
                "en": "Large Instruction Files Avoid Repeated Compaction, Artifact Publishing and Permissions No Longer Stall"
              },
              "body": {
                "zh-TW": "當對話中包含大量說明文件或正在發布網頁成品時，對話壓縮與成品更新流程不再反覆打擾。讀取受大型巢狀 `CLAUDE.md` 管轄的檔案後，工作階段以前會陷入反覆壓縮對話的循環；現在若巢狀說明檔超出容量上限，工具只會載入檔案開頭部分。發布或更新成品時，以往部分包含特定縮減程式碼（minified scripts）或未被預先讀取的常見頁面會遭到拒絕，現在已能正常發布。在同一個回合中，若對話曾觸發過內容壓縮，Claude 以前會在後續每次更新成品時不斷重複要求授權，這個問題現在已經修復。",
                "en": "When conversations involve extensive documentation or web artifacts are being published, conversation compaction and artifact update flows no longer cause repeated interruptions. After reading files governed by a large nested `CLAUDE.md`, sessions previously fell into loops of repeatedly compacting conversations. If a nested instructions file exceeds the size limit, the tool now loads only the initial portion of the file. When publishing or updating artifacts, certain common pages containing specific minified scripts or pages that were not preread previously faced rejection, but can now be published normally. If a conversation triggered content compaction within the same turn, Claude previously asked for permission repeatedly on every subsequent artifact update, an issue that is now fixed."
              },
              "original": "Fixed sessions compacting repeatedly after reading files under a very large nested CLAUDE.md; a nested CLAUDE.md over the size limit now loads only its start\nFixed artifact publishes and updates being refused for some ordinary pages, such as pages with certain minified scripts or artifacts Claude could not read first\nFixed Claude asking for permission on each artifact update for the rest of a turn after the conversation was compacted in that turn"
            },
            {
              "title": {
                "zh-TW": "桌面工作階段維持連線，遠端狀態列與停用工具不再錯亂",
                "en": "Desktop Sessions Stay Connected, Remote Status Line and Disabled Tools No Longer Misbehave"
              },
              "body": {
                "zh-TW": "當你透過桌面應用程式或遠端控制操作 Claude Code 時，帳號狀態與背景連線的穩定度得到了改善。在列印模式與 SDK 工作階段中，已經停用的 MCP（讓 AI 連接外部工具的協定）伺服器，以前會在後續跳出權限提示時被意外重新連線，現在已維持關閉。在由 Claude 桌面應用程式託管的工作階段中，網路連線重新整理時對話訊息偶爾遺失的問題已經排除。當同一台電腦上的終端機登入了不同帳號或組織時，桌面應用程式中的工作階段以前會遺失部分功能，現在不會再因此缺少功能。針對非 Git 儲存庫的普通資料夾，`claude remote-control` 的狀態列（顯示在終端機最底下的狀態列）不再將分支名稱錯誤顯示為 `HEAD`。",
                "en": "When operating Claude Code through the desktop app or remote control, the stability of account status and background connections has improved. In print mode and SDK sessions, disabled MCP (a protocol that connects AI to external tools) servers previously reconnected unexpectedly when permission prompts appeared later, but now remain closed. In sessions hosted by the Claude desktop app, an issue where conversation messages were occasionally lost during network refreshes has been eliminated. When terminals on the same computer are logged into different accounts or organizations, sessions inside the desktop app previously lost certain features, but will no longer miss capabilities as a result. For standard folders that are not Git repositories, the status line (the bar displayed at the bottom of the terminal) of `claude remote-control` no longer mistakenly shows the branch name as `HEAD`."
              },
              "original": "Fixed a disabled MCP server being reconnected by a later permission prompt in print and SDK sessions\nFixed messages from a Claude Desktop-hosted session sometimes being lost when its connection refreshed\nFixed features missing from Claude Code sessions in the Claude desktop app when the terminal on the same computer was signed in to a different account or organization\nFixed the `claude remote-control` status line showing `HEAD` as the branch for a folder that is not in a git repository"
            },
            {
              "title": {
                "zh-TW": "隱私設定即時同步，意見回饋與 GitHub 標記精確呈現",
                "en": "Privacy Settings Sync in Real Time, Feedback and GitHub Mentions Render Accurately"
              },
              "body": {
                "zh-TW": "當你在終端機檢查隱私設定或向官方回報問題時，回傳的訊息說明與對外發布內容更加清晰。輸入 `/privacy-settings` 時，以前可能顯示過期的「協助改善我們的 AI 模型」開關狀態；現在會即時呈現最新設定，無法取得設定時也會明確告知。當你所屬的組織未開放意見回饋時，送出 `/feedback` 不再只顯示含糊的「請稍後再試」，而是會顯示伺服器提供的具體原因並指引回報錯誤的管道。含有大量紀錄的超大 `/feedback` 報告在傳送失敗時，現在會自動改為排除對話紀錄後重新送出，不再直接報錯中斷。在發布 Pull Request、議題或留言時，Claude 不會再將角色名稱或佔位符寫成 `@` 標籤，避免意外通知到無關的 GitHub 使用者。",
                "en": "When checking privacy settings in the terminal or reporting issues to the team, returned explanations and public content are much clearer. Running `/privacy-settings` previously might have displayed an outdated toggle state for helping improve AI models, but now displays current settings in real time and clearly states when settings cannot be retrieved. When feedback is not enabled for your organization, submitting `/feedback` no longer merely displays a vague request to try again later, instead presenting the specific reason provided by the server and directing you to bug-reporting channels. When an oversized `/feedback` report containing extensive logs fails to transmit, it now automatically retries after excluding conversation transcripts, rather than aborting directly with an error. When publishing pull requests, issues, or comments, Claude no longer formats role names or placeholders as `@` mentions, preventing unintended notifications to unrelated GitHub users."
              },
              "original": "Fixed `/privacy-settings` showing an out-of-date \"Help improve our AI models\" value; it now also says when the setting can't be fetched\nFixed feedback saying \"try again later\" when your organization does not accept feedback; it now shows the server's reason and where to report a bug\nFixed oversized /feedback reports failing instead of being resent without the transcript\nFixed Claude notifying unrelated GitHub users by writing role names and placeholders as mentions in PRs, issues and comments"
            },
            {
              "title": {
                "zh-TW": "磁碟空間滿載提供清理指引，修復串流逾時與工具安裝彈窗",
                "en": "Full Disk Space Provides Cleanup Guidance, Streaming Timeouts and Tool Installation Popups Fixed"
              },
              "body": {
                "zh-TW": "當底層系統空間不足或身處特定雲端環境時，錯誤提示與連線等待機制變得更為實用。在共用 `/proc` 的 Linux 沙盒中，讀取工具讀取特定 PDF 頁面範圍不再誤報檔案不存在。在 Amazon Bedrock 將 `ANTHROPIC_BEDROCK_BASE_URL` 設為預設端點時，細粒度串流不會再被錯誤關閉。透過僅發送保持連線訊號（keepalive pings）的閘道連線時，等待 API 回覆提示不再閃爍，非 Anthropic 後端會等待 45 秒。在未安裝 Xcode，也未安裝 Command Line Tools（命令列開發工具）的 Mac 上，啟動 Claude Code 不再每次跳出安裝視窗。磁碟滿載導致 Bash 指令中斷時，會說明原因並點名應釋放空間的資料夾；暫存空間已滿時啟動也會顯示清楚訊息而非晦澀代碼。",
                "en": "When the underlying system is low on space or running in specific cloud environments, error prompts and connection waiting mechanisms are more practical. Inside Linux sandboxes sharing `/proc`, the read tool no longer falsely reports that files do not exist when reading specific PDF page ranges. When Amazon Bedrock configures `ANTHROPIC_BEDROCK_BASE_URL` as the default endpoint, fine-grained streaming is no longer mistakenly disabled. When connecting through gateways that only send keepalive pings, the prompt waiting for API responses no longer flickers, and non-Anthropic backends wait for 45 seconds. On a Mac with neither Xcode nor Command Line Tools (developer tools for the command line) installed, launching Claude Code no longer triggers an installation popup every time. When Bash commands abort because the disk is full, the tool explains the reason and names the folder where space should be freed. Starting up when temporary storage is full also displays a clear message rather than cryptic codes."
              },
              "original": "Fixed Read reporting \"File does not exist\" for PDF page ranges in Linux sandboxes that share the parent's `/proc`\nFixed fine-grained tool streaming staying off on Amazon Bedrock when `ANTHROPIC_BEDROCK_BASE_URL` is set to the default `bedrock-runtime.<region>.amazonaws.com` endpoint\nFixed the \"Waiting for API response\" hint flickering behind gateways that send only keepalive pings: routes off Anthropic's backend now wait 45 seconds\nmacOS: Fixed the \"Install Command Line Developer Tools\" dialog opening each time Claude Code starts on a Mac without Xcode or the Command Line Tools\nImproved disk-full handling: a Bash command cut off by a full disk now says so and names the folder to free up, and starting on a full temp volume prints a clear message, not a raw ENOSPC error"
            },
            {
              "title": {
                "zh-TW": "VS Code 聊天面板長訊息支援摺疊，快取損壞自動修復",
                "en": "VS Code Chat Panel Supports Collapsing Long Messages, Corrupted Cache Automatically Recovers"
              },
              "body": {
                "zh-TW": "當你在 VS Code 側邊面板中使用 Claude Code 時，幾處版面繪製與快取異常已經獲得排除。當編輯器因快取（把算過或查過的結果暫存起來直接用）了不完整的檔案而導致 Claude 面板一片空白時，擴充套件現在會自行恢復，或明確提示問題原因與手動解決方法。像 `/goal` 這類過長的斜線指令提示詞，現在會像其他長訊息一樣自動摺疊在展開按鈕後方，即使把聊天面板調窄，長訊息也不會撐破版面。輸入框上方標示來自特定外掛的提示標籤，不再錯誤遮擋斜線指令與 `@` 標註的選單。貼上複製的 LaTeX 數學公式時，不再產生多餘的空白清單項目、空白段落或程式碼區塊，編號清單也不會再被強制從 1 重新起算。",
                "en": "When using Claude Code in the VS Code side panel, several layout rendering and caching anomalies have been resolved. When the editor caches an incomplete file and causes the Claude panel to go completely blank, the extension now recovers on its own or clearly explains the cause and manual resolution. Lengthy slash command prompts such as `/goal` now automatically collapse behind an expand button just like other long messages, and narrowing the chat panel will no longer cause long messages to break the layout. Hint tags above the input box indicating specific plugins no longer mistakenly obscure menus for slash commands or `@` mentions. Pasting copied LaTeX mathematical formulas no longer creates superfluous empty list items, blank paragraphs, or code blocks, and numbered lists are no longer forced to restart from 1."
              },
              "original": "[VSCode] Fixed the Claude panel staying blank after the editor cached an incomplete copy of its files; it now recovers by itself, or says what went wrong and how to fix it\n[VSCode] Fixed long slash-command messages such as `/goal` filling the chat panel; they now fold behind Show more like other prompts (anthropics/claude-code#97469)\n[VSCode] Fixed long messages not folding behind Show more after the chat panel is made narrower\n[VSCode] Fixed a mod's \"From the … plugin\" label above the prompt being drawn over the slash command and @-mention menus (anthropics/claude-code#102981)\n[VSCode] Fixed an empty list item, paragraph or code block, and list items renumbered from 1, when pasting copied LaTeX math"
            },
            {
              "title": {
                "zh-TW": "VS Code 編輯等待時間縮短，憑證警告提示與空紀錄修復",
                "en": "VS Code Edit Wait Time Shortened, Certificate Warning Guidance and Empty Logs Fixed"
              },
              "body": {
                "zh-TW": "當你在 VS Code 中編修程式碼或發出語音指令時，操作延遲與背景紀錄邏輯獲得了調整。聽寫或套件自身請求的憑證（確認連線對象身分的證明）驗證失敗時，例如經過會檢查加密流量的代理伺服器，日誌現在會提供處理指引。每次開啟新對話卻未發送訊息時，歷史紀錄不再被空工作階段塞滿。若在 `~/.claude/keybindings.json` 的 Chat 區段將 `\"ctrl+j\"` 設為 `null`，在聊天輸入框按下 Ctrl+J 不再強制換行，改交由 VS Code 與作業系統處理。當組織不接受意見回饋時，錯誤提示不再只回傳 HTTP 狀態碼，會顯示具體原因。聊天面板中的編輯與寫入速度加快，每次編輯後等待偵測新問題的時間從原本的 1 至 1.5 秒縮短為多數情況下的 0.75 秒。",
                "en": "When editing code or issuing voice commands in VS Code, operational latency and background logging logic have been adjusted. When certificate verification (checking digital proof of the connection target's identity) fails for dictation or extension requests, such as when passing through a proxy that inspects encrypted traffic, logs now provide troubleshooting guidance. Opening a new conversation without sending a message no longer clutters history with empty sessions. If `\"ctrl+j\"` is set to `null` in the Chat section of `~/.claude/keybindings.json`, pressing Ctrl+J in the chat input box no longer forces a newline, deferring handling to VS Code and the operating system. When an organization does not accept feedback, error prompts no longer return only an HTTP status code, displaying the specific reason instead. Editing and writing in the chat panel are faster, reducing the wait time to detect new problems after each edit from 1 to 1.5 seconds down to 0.75 seconds in most cases."
              },
              "original": "[VSCode] Added a log warning that says what to do when dictation or the extension's own requests fail certificate verification, such as behind a TLS-inspecting proxy\n[VSCode] Fixed Session history filling up with empty \"Untitled\" sessions each time you start a new session without sending a message\n[VSCode] Fixed Ctrl+J in the chat input always inserting a newline: it is now left to VS Code and the OS when \"ctrl+j\" is set to null in the Chat context of ~/.claude/keybindings.json\n[VSCode] Fixed the feedback error showing only an HTTP status when your organization does not accept feedback\n[VSCode] Improved Edit and Write speed in the chat panel: the wait for new problems after each edit is now 0.75 seconds in most cases, down from 1 to 1.5 seconds"
            },
            {
              "title": {
                "zh-TW": "雲端工作階段修正環境變數引號，管理介面與變更清單明朗化",
                "en": "Cloud Sessions Fix Environment Variable Quotes, Admin Interface and Changes List Clarified"
              },
              "body": {
                "zh-TW": "當你在瀏覽器管理雲端工作階段與連線設定時，表單填寫與文件顯示體驗有了改善。在例行表單中新增連接器時，清單在載入期間不再誤報沒有可用的連接器。雲端環境編輯器不再打亂變數順序，當數值同時包含單引號與雙引號時，也不會再多餘地自動補上反斜線。管理設定中的自架環境對話框獲得調整，標題、分頁標籤與按鈕在滾動時會固定在可視範圍，只有中間內容會捲動。雲端工作階段中的閱讀文件工具列不再顯示跳脫後的 JSON 原始碼，而是會呈現易讀的排版內容。變更面板現在會主動標明目前是僅列出已推送的提交，還是僅顯示 Claude 的檔案修改。",
                "en": "When managing cloud sessions and connection settings in the browser, the experience for filling out forms and viewing documents has improved. When adding connectors in routine forms, the list no longer falsely claims that no connectors are available while loading. The cloud environment editor no longer scrambles variable order, nor does it insert redundant backslashes when values contain both single and double quotes. The self-hosted environment dialog in administrative settings has been adjusted so that headers, tabs, and buttons stay pinned in view while scrolling, leaving only the middle content scrollable. The document-reading toolbar in cloud sessions no longer displays escaped raw JSON, rendering cleanly formatted content instead. The changes panel now explicitly clarifies whether it is listing only pushed commits or displaying only file modifications made by Claude."
              },
              "original": "[Cloud sessions] Fixed the Add connector list in routine forms saying no connectors were available while they were still loading\n[Cloud sessions] Fixed the cloud environment editor shuffling environment variables and adding backslashes to values that contain both kinds of quote\n[Cloud sessions] Fixed the self-hosted environment details dialog in admin settings so its title, tabs and buttons stay in view while only the content scrolls\n[Cloud sessions] Fixed the \"read documentation\" tool row in cloud sessions, which showed escaped JSON instead of a readable page\n[Cloud sessions] Improved a cloud session's Changes pane, which now says when it lists only pushed commits or only Claude's edits"
            },
            {
              "title": {
                "zh-TW": "Slack 標註草稿正確轉發頻道，審核按鈕支援重啟對話",
                "en": "Slack Channel Drafts Dispatch Correctly, Review Buttons Support Restarting Sessions"
              },
              "body": {
                "zh-TW": "當你的團隊在 Slack 頻道中使用 Claude 進行協作時，訊息派送與互動卡片的穩定性顯著提升。你在頻道中核准的草稿，以前可能被 Claude 誤貼到私訊中；現在沒有指明頻道的貼文會被直接拒絕，確保內容確實發布至指定的頻道。在審核卡片上按下核准或拒絕時，若背後的工作階段已經停止，你的操作現在會自動重啟該工作階段繼續執行。在工作區轉移至不同 Claude 組織後（包括 Enterprise Grid 上的未共用頻道），Claude 不會再漏看頻道中未被標註的討論訊息。在對討論串發布回覆之前，Claude 現在會先讀取其他人剛發送的最新留言，避免發言內容脫節。管理後台的 Slack 設定頁面若載入失敗，不再誤報尚未連接工作區，而是會顯示錯誤訊息並提供重試按鈕。",
                "en": "When your team collaborates with Claude inside Slack channels, the reliability of message dispatching and interactive cards has significantly improved. Drafts approved in a channel were previously at risk of being mistakenly posted by Claude into direct messages. Posts without a specified channel are now rejected outright, ensuring content is delivered to the designated channel. When clicking approve or reject on a review card, if the underlying session has stopped, your action now automatically restarts that session to proceed. After workspaces are migrated to a different Claude organization (including unshared channels on Enterprise Grid), Claude no longer misses unmentioned discussion messages in channels. Before posting replies to a thread, Claude now reads the latest comments just posted by others, preventing disjointed responses. If the admin Slack settings page fails to load, it no longer falsely reports that no workspace is connected, instead displaying an error message alongside a retry button."
              },
              "original": "[Claude Tag] Fixed Claude in Slack posting a draft you approved for a channel into your DM by mistake; a post with no channel is now refused so it can go to the approved channel\n[Claude Tag] Fixed Allow and Deny on Claude's approval cards in Slack doing nothing when the session had stopped; your answer now restarts the session\n[Claude Tag] Fixed Claude ignoring untagged messages in a Slack channel after its workspace moved to a different Claude organization, including unshared channels on Enterprise Grid\n[Claude Tag] Fixed Claude in a Slack channel posting a threaded reply before reading a newer message someone had just added to that thread\n[Claude Tag] Fixed the Slack page in Claude Tag admin settings saying no workspaces were connected when the list failed to load; it now shows an error with a Retry button"
            },
            {
              "title": {
                "zh-TW": "PR 審查檢查標記自動補全，手動停止審查及時關閉檢查",
                "en": "Pull Request Review Check Marks Automatically Backfilled, Manual Stops Promptly Close Checks"
              },
              "body": {
                "zh-TW": "當你在 GitHub Pull Request 上等待自動程式碼審查時，檢查狀態更新結果現在更為準確。當審查已經順利完成，但先前建立檢查狀態的嘗試曾失敗過，工具在最終重試時現在會確實補上已完成的檢查標記，不再讓 PR 懸空呈現沒有檢查結果的狀態。若在審查執行期間推播了新的程式碼並在後方排隊，即使輸入了 `@claude review stop` 手動停止，以往排隊的審查仍會被啟動；現在排隊的檢查會立即關閉，並具體註明被停止的原因。⟦如同公文簽核系統在審批完畢後自動蓋上最終核准章，中途撤回的案件也會立刻標記撤回原因，避免審核進度懸而未決。⟧",
                "en": "When waiting for automated code reviews on GitHub pull requests, check status updates are now more accurate. When a review finishes successfully but earlier attempts to create the check status failed, the tool on its final retry now reliably backfills the completed check mark, preventing the pull request from hanging in a state with no check results. If new code is pushed during an ongoing review and queued up behind it, typing `@claude review stop` to halt the review manually previously still allowed queued reviews to start. Queued checks are now closed immediately with the specific cancellation reason noted. ⟦Just as an official document approval system automatically applies a final approval stamp once sign-off is complete, any case recalled midway is immediately marked with the withdrawal reason to prevent review progress from remaining in limbo.⟧"
              },
              "original": "[Code Review] Fixed a finished review showing no check on the pull request when earlier tries to create it failed; a final try now adds a completed check\n[Code Review] Fixed a push queued behind a running review still starting a review after `@claude review stop`; its check now closes and says why"
            }
          ]
        },
        {
          "version": "2.1.307",
          "period": "2026-10-21",
          "items": [
            {
              "title": {
                "zh-TW": "Code Review 支援停止指令，快速解除 PR 自動審查",
                "en": "Code Review adds stop commands to end automatic PR reviews"
              },
              "body": {
                "zh-TW": "當你在 GitHub 的拉取請求（Pull Request）啟用自動審查後，若是功能已經測試完畢準備合併，往往不需要機器人再次介入。⟦跟退訂每日鮮乳配送一樣，以前聯絡客服終止可能還會多送一天的牛奶，現在辦理完成立即停送，不會在退訂當天又往門口放一瓶多餘的飲品。⟧現在 Code Review 新增了 `@claude review stop` 指令，同時也支援 `@claude review off` 與 `@claude review pause`。輸入後能直接結束該拉取請求上的 `@claude review always` 訂閱，而且不會觸發新一輪審查。在合併程式碼的最後關頭，你可以俐落收尾，不必再被多餘的審查留言打擾。",
                "en": "After you enable automatic reviews on a GitHub pull request, you often do not need the bot to step in again once the feature has been tested and is ready to merge. ⟦Canceling a daily milk delivery used to mean contacting customer service and possibly receiving milk for another day. Now deliveries stop as soon as the cancellation is processed, without an extra bottle arriving at your door that day.⟧ Code Review now supports `@claude review stop`, along with `@claude review off` and `@claude review pause`. These commands directly end the `@claude review always` subscription for that pull request without triggering another review. As you prepare to merge the code, you can wrap things up without extra review comments getting in the way."
              },
              "original": "[Code Review] Added `@claude review stop` (also `off` and `pause`) to Code Review, ending an `@claude review always` subscription on a pull request without starting a new review"
            },
            {
              "title": {
                "zh-TW": "內建指令選單顯示精簡描述，新增 API 版本切換",
                "en": "Built-in command menus show shorter descriptions and add API version switching"
              },
              "body": {
                "zh-TW": "平常在終端機按下 `/` 或是輸入 `/help` 查詢功能時，畫面呈現的整潔度直接影響尋找指令的效率。以前針對 `/claude-api` 等內建技能，清單會直接倒出寫給模型閱讀的長篇說明，把選單畫面撐得相當冗長。⟦讀者翻閱工具書的索引目錄時，以前條目旁印滿給排版校對看的長篇備註，現在全部換回簡短摘要，版面呈現乾淨明瞭。⟧現在選單已改為顯示給使用者看的精簡說明，閱讀起來不再雜亂。若你需要將專案內的 Claude API 程式碼推進至 `2026-10-21` 版本，直接輸入 `/claude-api api-version` 即可快速遷移。",
                "en": "When you press `/` in the terminal or enter `/help` to look up features, a tidy display directly affects how quickly you can find a command. For built-in skills such as `/claude-api`, the list previously displayed lengthy instructions written for the model, making the menu cumbersome. ⟦When readers browsed a reference book's index, entries used to be crowded with long notes intended for typesetters and proofreaders. Those notes have now been replaced with short summaries, leaving the layout clear and easy to read.⟧ The menu now shows concise descriptions written for users, making it easier to read. If you need to migrate your project's Claude API code to version `2026-10-21`, enter `/claude-api api-version` to make the switch quickly."
              },
              "original": "Added `/claude-api api-version` to move Claude API code to API version `2026-10-21`\nFixed `/help` and the `/` command menu showing long model-facing descriptions for bundled skills such as `/claude-api` instead of their short descriptions"
            },
            {
              "title": {
                "zh-TW": "安全事件支援遙測追蹤，權限清單清楚列出被拒指令",
                "en": "Safety events gain telemetry tracking, and permission lists identify denied commands"
              },
              "body": {
                "zh-TW": "當你透過自動模式或安全規則控管各項工具時，操作被擋下的細節與放行流程現在更加清楚。只要工具呼叫被權限規則、自動模式、安全檢查、掛鉤（特定事件發生時被觸發的小程式）或沙盒阻止、提示或退回，OpenTelemetry 都會記錄 `claude_code.safety_event` 遙測事件。以前在 `/permissions` 的 Recently denied 畫面中不會顯示被拒絕的 shell 指令，且自動模式的本機分類器在審查重試指令時還會忽略此處的核准；現在核准記錄會直接標明具體指令名稱而非抽象描述。當網頁搜尋預算耗盡時，通知也會同步告誡 Claude 不得透過瀏覽器或電腦操作（computer use）轉向外部搜尋引擎查詢。以前指令被擋下後難以追查原貌、核准重試又可能再次受阻，現在每一道安全關卡都有跡可循，放行指令也能正確生效。",
                "en": "When you control tools through auto mode or safety rules, the details of blocked actions and the approval process are now clearer. Whenever permission rules, auto mode, safety checks, hooks (small programs triggered by specific events), or the sandbox block a tool call, prompt for approval, or send it back, OpenTelemetry records a `claude_code.safety_event` telemetry event. Previously, the Recently denied view in `/permissions` did not show denied shell commands, and auto mode's local classifier ignored approvals made there when reviewing retried commands. Approval records now identify the specific command rather than giving an abstract description. When the web search budget runs out, the notification also tells Claude not to turn to external search engines through the browser or computer use (operating a computer's interface). Previously, it was difficult to trace the original command after it was blocked, and an approved retry could be blocked again. Now each safety check leaves a record, and command approvals take effect correctly."
              },
              "original": "Added the `claude_code.safety_event` OpenTelemetry event for each tool call a permission rule, auto mode, a safety check, a hook or the sandbox stopped, prompted or sent back\nFixed `/permissions` > Recently denied not showing the denied shell command, and auto mode's local classifier ignoring an approval from there when it checks a retried command; approving a row now names the command, not its description\nChanged the notice Claude gets when the WebSearch budget is used up to also tell it not to query search engines in a browser or through computer use"
            },
            {
              "title": {
                "zh-TW": "修復 MCP 工具連線與休眠逾時，避免重複發送寫入請求",
                "en": "MCP connection and sleep timeout fixes prevent duplicate write requests"
              },
              "body": {
                "zh-TW": "當你依賴 MCP（讓 AI 連接外部工具的協定）擴充功能時，以下連線問題已修好。在 macOS 上，若登入鑰匙圈鎖定，`claude -p` 會說明驗證受阻的原因，解鎖後下次執行便能連線，無須重新登入。過去 macOS 或 Linux 電腦在 MCP 伺服器連線途中進入睡眠，伺服器會被回報為逾時，並在無頭模式（沒有視覺介面的執行模式）下一直顯示失敗，這已修好。HTTP 呼叫中途斷線時，現在僅限唯讀或具備冪等特性（重做也不改變結果）的工具才會自動重試，避免寫入操作觸發兩次。背景任務執行時，新一輪對話有時未取得 SDK MCP 伺服器的最新工具，這已修好。桌面版 Code 標籤頁中，直接設定的 GitHub MCP 伺服器寫入工具，表單關閉後未建立內容的狀況也已修正。",
                "en": "If you rely on MCP (a protocol that connects AI to external tools) for added functionality, the following connection problems have been fixed. On macOS, if the login keychain is locked, `claude -p` explains why authentication is blocked. After unlocking it, the next run can connect without requiring you to log in again. Previously, if a macOS or Linux computer went to sleep while connecting to an MCP server, the server was reported as timed out and continued to appear failed in headless mode (running without a visual interface). This has been fixed. If an HTTP call loses its connection midway, automatic retries are now limited to tools that are read-only or idempotent (repeating the operation does not change the result), preventing writes from being triggered twice. A new conversation turn sometimes failed to get the latest tools from SDK MCP servers while background tasks were running; this has been fixed. In the desktop app's Code tab, write tools from directly configured GitHub MCP servers showed a form that closed without creating anything. This has also been fixed."
              },
              "original": "Fixed MCP servers showing \"needs authentication\" in `claude -p` on macOS while the login keychain was locked; they now say why, and the next run after an unlock connects without a new sign-in\nFixed MCP servers being reported as timed out, and staying failed in headless runs, when a macOS or Linux computer went to sleep while they were connecting\nFixed HTTP MCP tool calls being sent twice when the connection closed mid-call; only read-only or idempotent tools are now retried\nFixed a turn sometimes starting without an SDK MCP server's latest tools when a background task was running\nFixed write tools on a directly configured GitHub MCP server showing a form that closed without creating anything in the desktop app's Code tab"
            },
            {
              "title": {
                "zh-TW": "子代理避免重複接工，工具呼叫格式錯誤時支援自動重試",
                "en": "Subagents avoid duplicate assignments, and malformed tool calls can retry automatically"
              },
              "body": {
                "zh-TW": "當你借助子代理或 Agent SDK 執行複雜的多工流程時，多項非同步協作的容錯機制已完成補強。針對同時啟動的多個 fork 子代理，現在各自分發時會明確指派身分，不再容易互相搶接對方的工作。⟦工廠領班指派多人協同組裝時，以前組員容易爭搶同一份工單，現在每個人都清楚領取專屬工作卡，遇到辨識不清的指示還會主動重新確認，最多兩次。⟧工具呼叫無法解析時，現在最多重試兩次；仍無法解析才會報錯並結束這一輪。在文字稿記錄方面，串流回應不再填入佔位的 token（AI 處理與計費的文字單位）數值與空白停止原因；透過 SDK 的 `setModel()` 切換模型時，確認等待時間也由 5 秒放寬至 15 秒。這些調整減少了自動化流程中的非預期中斷，讓複雜排程執行得更平穩。",
                "en": "If you use subagents or the Agent SDK for complex multitasking workflows, several safeguards for asynchronous collaboration have been strengthened. Multiple fork subagents launched at the same time now receive explicit identities when dispatched, making them less likely to take each other's assignments. ⟦When a factory foreman assigned several workers to assemble something together, workers used to compete for the same work order. Now each worker clearly receives a dedicated job card and actively asks for clarification when instructions are unclear, up to two times.⟧ Tool calls that cannot be parsed now retry up to two times. An error is reported and the turn ends only if parsing still fails. In transcripts, streaming responses no longer include placeholder token (a unit of text AI processes and bills for) counts or empty stop reasons. When switching models through the SDK's `setModel()`, the wait for confirmation has also been extended from 5 seconds to 15 seconds. These changes reduce unexpected interruptions in automated workflows and help complex schedules run more smoothly."
              },
              "original": "Fixed subagent transcripts recording placeholder token usage and an empty stop reason for streamed responses\nFixed model switches through the Agent SDK's `setModel()`, and apps built on it, being refused when the API took more than 5 seconds to confirm the model; Claude Code now waits up to 15 seconds\nImproved recovery from unparseable tool calls: Claude Code now retries the call up to twice before ending the turn with an error\nImproved fork subagents launched together: each is now told which launch is its own, so forks are much less likely to try to start each other's work"
            },
            {
              "title": {
                "zh-TW": "修正自訂閘道標頭與遙測錯誤，完善模型設定判定",
                "en": "Custom gateway header and telemetry fixes improve model setting handling"
              },
              "body": {
                "zh-TW": "若你使用自訂 ANTHROPIC_BASE_URL 閘道（代為轉送連線請求的中間服務）或 Foundry，無效的 prompt-caching-scope 標頭已不再送出。Claude apps 閘道工作階段不再忽略 CLAUDE_STREAM_FIRST_BYTE_TIMEOUT_MS；其 Claude Desktop 工作階段設定 OTEL_EXPORTER_OTLP_HEADERS 後遙測遭拒的問題已修復。CLAUDE_CODE_DISABLE_1M_CONTEXT 搭配帶 [1m] 的模型別名時不再報錯找不到模型。Artifact 讀取重用連線遭伺服器關閉時不再出現連線重設錯誤。自 2.1.305 起，axScreenReader 與 env 區塊中的 CLAUDE_AX_SCREEN_READER，只從使用者、受管理或 --settings 生效。",
                "en": "If you use a custom ANTHROPIC_BASE_URL gateway (an intermediary service that forwards connection requests) or Foundry, the invalid prompt-caching-scope header is no longer sent. Claude apps gateway sessions no longer ignore CLAUDE_STREAM_FIRST_BYTE_TIMEOUT_MS. Telemetry rejection after setting OTEL_EXPORTER_OTLP_HEADERS in their Claude Desktop sessions has been fixed. Using CLAUDE_CODE_DISABLE_1M_CONTEXT with a model alias containing [1m] no longer produces a model-not-found error. Artifact reads no longer produce connection reset errors when the server has closed a reused connection. Starting with 2.1.305, axScreenReader and CLAUDE_AX_SCREEN_READER in the env block take effect only from user settings, managed settings, or --settings."
              },
              "original": "Fixed requests failing with a model-not-found error when `CLAUDE_CODE_DISABLE_1M_CONTEXT` is set and the model is an alias with the `[1m]` suffix, such as `opus[1m]`\nFixed the `prompt-caching-scope` beta header being sent behind custom `ANTHROPIC_BASE_URL` gateways and on Foundry, where it has no effect\nFixed `CLAUDE_STREAM_FIRST_BYTE_TIMEOUT_MS` being ignored in Claude apps gateway sessions\nFixed telemetry from Claude Desktop sessions behind a Claude apps gateway being rejected when `OTEL_EXPORTER_OTLP_HEADERS` is set\nFixed Artifact tool reads failing with a connection reset when the server had closed a reused connection\nChanged the `axScreenReader` setting and `CLAUDE_AX_SCREEN_READER` in a settings `env` block to apply only from user, managed or `--settings` settings (since 2.1.305)"
            },
            {
              "title": {
                "zh-TW": "瀏覽器登入即刻顯示成功頁面，加快遠端與雲端連線速度",
                "en": "Browser sign-in shows success immediately, and remote and cloud connections are faster"
              },
              "body": {
                "zh-TW": "當你開啟瀏覽器登入帳號或使用遠端連線功能時，整體等待體驗與狀態回報已獲得多項改進。現在瀏覽器登入只要一完成，瀏覽器便會立刻跳出成功畫面，不需要乾等帳號個人資料慢慢載入。桌面應用有某個 MCP 外部工具伺服器連不上時，雲端工作階段等待首次回覆的時間也縮短了。如果對話文字稿中塞滿了貼上的大型圖片或 PDF 讀取內容，遠端控制（Remote Control）的連線建立過程也變得更快。透過桌面應用程式封存閒置工作階段而中止的背景指令，通知會清楚標明輸出檔案名稱，不再向 Claude 誤報成是使用者手動停止。",
                "en": "When you sign in through a browser or use remote connections, several changes improve waiting times and status reporting. As soon as browser sign-in finishes, the browser now immediately shows a success page without waiting for account profile data to load. In cloud sessions, the wait for the first response is shorter when one of the desktop app's MCP servers can't connect. Remote Control connections also start faster when the conversation transcript contains large pasted images or PDF content that has been read. When archiving an idle session through the desktop app stops a background command, the notification clearly identifies the output filename and no longer incorrectly tells Claude that the user stopped it manually."
              },
              "original": "Fixed a background command stopped through a connected app (such as the desktop app archiving an idle session) being reported to Claude as stopped by the user; the notice now names its output file\nImproved browser sign-in: the success page now appears as soon as sign-in completes, without waiting for your account profile to load\nImproved the time to the first reply in cloud sessions when one of the desktop app's MCP servers can't connect\nImproved Remote Control connect time for sessions whose transcript contains large pasted images or PDF reads"
            },
            {
              "title": {
                "zh-TW": "修正對話自動精簡震盪，避免重啟後舊訊息復活",
                "en": "Fixes reduce autocompaction thrashing and keep cleared messages from returning after restarts"
              },
              "body": {
                "zh-TW": "長時間對話並夾帶大量圖表資料時，歷史訊息的壓縮與清理機制決定了對話能否順利延續。⟦辦公室碎紙機處理厚重的設計圖冊時，以前一次塞入太多大張圖紙容易反覆卡紙倒退，清空後的廢紙匣在重開電源後竟然還會把舊草稿吐出來；現在卡紙少了一些，清空的廢紙匣也不會再吐出舊草稿。⟧過去在單一輪次中讀取大量 PDF 頁面、圖片或龐大工具輸出時，可能出現「Autocompact is thrashing」錯誤，代表自動精簡反覆震盪；這次修好了其中一部分情況。在雲端工作階段中，自訂的 `/autocompact` 視窗範圍在工作階段重新啟動後不會再跳回預設值。透過 `/clear` 清除的對話紀錄，在雲端或自託管執行器重啟後，也不會再莫名其妙復活。",
                "en": "In long conversations with lots of charts and other data, the way message history is compressed and cleared determines whether the conversation can continue smoothly. ⟦When an office shredder handled a thick book of design drawings, feeding in too many large sheets used to make it repeatedly jam and reverse. Even an emptied waste bin would spit out old drafts after the power was turned back on. Now there are fewer jams, and an emptied bin no longer spits out old drafts.⟧ Previously, reading many PDF pages, images, or large tool outputs in a single turn could cause an “Autocompact is thrashing” error, meaning automatic compaction was repeatedly oscillating. Some of these cases have been fixed. In cloud sessions, a custom `/autocompact` window range no longer resets to its default after the session restarts. Conversation history cleared with `/clear` also no longer unexpectedly returns after a cloud or self-hosted runner restarts."
              },
              "original": "Fixed some \"Autocompact is thrashing\" errors in turns that read many PDF pages, images or other large tool results\nFixed `/autocompact` in cloud sessions going back to the default window after the session restarted\nFixed a conversation cleared with `/clear` sometimes coming back when a cloud or self-hosted runner session restarted"
            },
            {
              "title": {
                "zh-TW": "雲端搜尋提供即時進度回饋，網頁讀取錯誤回歸正常結果",
                "en": "Cloud search shows live progress, and web fetch failures return normal results"
              },
              "body": {
                "zh-TW": "當你在雲端環境挑選儲存庫或讓 Claude 抓取網頁資料時，多處介面狀態與錯誤回饋現在更為貼近真實情況。在儲存庫選擇器中，搜尋進行中會明確顯示「Searching…」，不會再於結果抵達前誤跳「找不到相符儲存庫」的假警報。面對自託管環境，「Waiting for a runner」的等待提示不再被截斷，若是隨選執行器，還會從你最後一則訊息起算等待秒數，並顯示失敗次數。在網路擷取方面，Cowork 雲端工作階段的 WebFetch 遇到網站回傳 4xx 狀態碼或 robots.txt 阻擋時，會如同本機環境般作為一般結果呈現，不再拋出工具錯誤。在雲端或自託管執行器工作階段，搜尋遭 403 拒絕且回應附有原因時，現在會顯示該原因，不再只顯示驗證錯誤（Authentication error）。",
                "en": "When you select a repository in the cloud or have Claude fetch web content, several interface states and error messages now more accurately reflect what is happening. The repository picker explicitly shows “Searching…” while a search is running, instead of prematurely showing a false “No repositories match” message before results arrive. In self-hosted environments, the “Waiting for a runner” message is no longer cut off. For on-demand runners, it also shows the seconds spent waiting since your last message and the number of failures. For web fetching, when a website returns a 4xx status code or robots.txt blocks access, WebFetch in Cowork cloud sessions now presents this as a normal result, as it does locally, instead of raising a tool error. In cloud or self-hosted runner sessions, when a search is denied with a 403 response that includes a reason, that reason is now shown instead of just “Authentication error”."
              },
              "original": "Fixed WebFetch in Cowork cloud sessions reporting a tool error when a site answered with a 4xx status or a robots.txt refusal; it is now a normal result, as in local sessions\nImproved web search errors in cloud and self-hosted runner sessions: a search refused with a 403 that gives a reason now shows that reason instead of a generic \"Authentication error\"\n[Cloud sessions] Fixed the repository picker saying \"No repositories match\" while a search was still running; it now shows \"Searching…\" until the results arrive\n[Cloud sessions] Fixed the \"Waiting for a runner\" status on self-hosted environments being cut off; with on-demand runners it now counts up from your last message and shows failed attempts"
            },
            {
              "title": {
                "zh-TW": "自託管執行器預設掛鉤超時上限，防止殘留程序佔用槽位",
                "en": "Self-hosted runners add a default hook timeout and prevent leftover processes from occupying slots"
              },
              "body": {
                "zh-TW": "自行架設執行器（Self-hosted runner）跑自動化流程的人，這版要留意四項改動。專案準備階段的 checkout 掛鉤（特定事件發生時被觸發的小程式）現在預設以 30 分鐘為限，超時會直接判定設定失敗；若任務需要更久，可自訂 `--checkout-hook-timeout-min` 延長時間或設為 `0` 解除上限。過去若在掛鉤執行期間封存或重新排入工作階段，或停止執行器，掛鉤的行程會留著繼續執行，執行器的槽位也一直被佔住，這已修好。啟用 `--use-anthropic-git-proxy` 時，也不會再誤將前一次工作階段留下的 `.claude/settings.local.json` 提交進版本控制中。替 Anthropic 管理的 Git 建立憑證檔案時，現在比照 CLI 過濾受信任憑證，並以 `[runner:warn]` 記錄排除數量。",
                "en": "If you run automation on a self-hosted runner, there are four changes to note in this release. Checkout hooks (small programs triggered by specific events) during project preparation now have a default limit of 30 minutes, after which setup is marked as failed. If a task needs longer, use `--checkout-hook-timeout-min` to extend the limit or set it to `0` to remove the limit. Previously, archiving or requeuing a session while a hook was running, or stopping the runner, left the hook's process running and kept the runner's slot occupied. This has been fixed. With `--use-anthropic-git-proxy` enabled, `.claude/settings.local.json` left over from a previous session is also no longer mistakenly committed to version control. When creating certificate files for Anthropic-managed Git, the runner now filters trusted certificates the same way as the CLI and logs the number excluded with `[runner:warn]`."
              },
              "original": "Self-hosted runner: Fixed sessions with `--use-anthropic-git-proxy` committing a `.claude/settings.local.json` left in the clone by an earlier session\nSelf-hosted runner: Fixed archiving or requeueing a session, or stopping the runner, during its checkout hook leaving the hook's processes running and the runner's slot held\nSelf-hosted runner: Changed checkout hooks to be stopped after 30 minutes by default, failing the session's setup; if yours needs longer, raise `--checkout-hook-timeout-min` or set it to `0`\nSelf-hosted runner: Changed the certificate file built for Anthropic-managed git to filter its trusted certificates as the CLI does; a `[runner:warn]` line gives the number left out"
            },
            {
              "title": {
                "zh-TW": "VS Code 限制展開提示詞高度，修正 LaTeX 數學式貼上格式",
                "en": "VS Code limits expanded prompt height and fixes pasted LaTeX formatting"
              },
              "body": {
                "zh-TW": "在 VS Code 中使用 Claude Code 撰寫複雜指示或整理筆記時，對話框排版與文字貼上體驗更為順暢。過去只要點擊 Show more 展開較長的提示詞，整張輸入卡片會直接霸佔整個對話面板；現在釘選的提示詞高度被限制在面板的 40% 以內，超出範圍時會自動產生捲軸方便上下滾動。當你複製含有 LaTeX 數學式的清單項目或表格儲存格時，貼進富文字編輯器中不會再被錯誤切碎成全新的清單或表格。遇到擴充套件重啟後殘留的 Claude 背景行程時，「仍在其他地方開啟」的提示視窗現在也能清楚識別，並說明該行程通常在約 10 分鐘內便會自行退出。以前長文字一展開就遮蔽對話歷史、貼上公式格式又容易碎裂，現在排版邊界更穩定，編輯與查閱操作不再互相打架。",
                "en": "Writing complex instructions or organizing notes with Claude Code in VS Code now feels smoother, with better conversation layouts and text pasting. Previously, clicking Show more to expand a long prompt made the entire input card take over the conversation panel. Pinned prompts are now limited to 40% of the panel’s height, with a scrollbar appearing automatically when the content exceeds that limit. When you copy list items or table cells containing LaTeX equations into a rich text editor, they no longer get incorrectly split into new lists or tables. The “still open somewhere else” notice now clearly identifies Claude background processes left over after an extension restart and explains that they usually exit on their own within about 10 minutes. Previously, expanding long text obscured the conversation history, and pasting equations could break their formatting. With more stable layout boundaries, editing and reviewing no longer get in each other’s way."
              },
              "original": "[VSCode] Fixed a long prompt expanded with Show more covering the whole chat panel: a pinned prompt is now capped at 40% of the panel and scrolls\n[VSCode] Fixed part of a list item or table cell with LaTeX math pasting as a new list or table in rich text editors\n[VSCode] Changed the \"still open somewhere else\" notice to also cover a Claude process left over from an extension restart, which usually ends on its own within about 10 minutes"
            },
            {
              "title": {
                "zh-TW": "VS Code 語音聽寫新增中斷提示，完善 TLS 憑證支援",
                "en": "VS Code voice dictation adds stop explanations and improves TLS certificate support"
              },
              "body": {
                "zh-TW": "在 VS Code 裡對著麥克風念出程式碼需求時，自動停止的機制與公司內網憑證相容性已經獲得完善。當語音聽寫因連續 15 秒沒有辨識出語音或達到 2 分鐘上限而自動關閉時，麥克風按鈕會跳出說明告知原因，並確認剛才說過的文字皆已妥善保留。針對經 TLS 封包檢查的代理網路，若僅在 `~/.claude/settings.json` 設定 `NODE_EXTRA_CA_CERTS` 曾導致聽寫連線失敗，現在已修好，聽寫也會遵循 `CLAUDE_CODE_CERT_STORE`。語音聽寫與擴充套件本身的請求，現在都比照 CLI 過濾受信任憑證。擴充套件本身的請求也會信任環境變數或使用者設定指定的 `NODE_EXTRA_CA_CERTS` 憑證檔案。",
                "en": "When you dictate coding requirements into a microphone in VS Code, automatic stopping and compatibility with corporate network certificates now work more reliably. If voice dictation stops automatically after 15 consecutive seconds without recognized speech or upon reaching the 2-minute limit, the microphone button displays an explanation and confirms that everything you dictated has been preserved. On proxy networks with TLS inspection, setting `NODE_EXTRA_CA_CERTS` only in `~/.claude/settings.json` could cause dictation connections to fail. This is now fixed, and dictation also follows `CLAUDE_CODE_CERT_STORE`. Voice dictation and the extension’s own requests now filter trusted certificates the same way the CLI does. The extension’s own requests also trust the `NODE_EXTRA_CA_CERTS` certificate file specified through environment variables or user settings."
              },
              "original": "[VSCode] Added a notice on the microphone button when voice dictation stops by itself (15 seconds without recognized speech, or the 2-minute limit), saying why and that the text so far is kept\n[VSCode] Fixed dictation failing behind a TLS-inspecting proxy when `NODE_EXTRA_CA_CERTS` is set only in `~/.claude/settings.json`; it now also follows `CLAUDE_CODE_CERT_STORE`\n[VSCode] Changed voice dictation to filter its trusted certificates as the CLI does\n[VSCode] Changed the extension's own requests to filter their trusted certificates as the CLI does, and to trust the `NODE_EXTRA_CA_CERTS` file named in the environment or in user settings"
            }
          ]
        },
        {
          "version": "2.1.306",
          "period": "2026-10-21",
          "items": [
            {
              "title": {
                "zh-TW": "雲端工作階段裡的 `claude -p` 不再誤報沒登入",
                "en": "`claude -p` in cloud sessions no longer falsely reports \"Not logged in\""
              },
              "body": {
                "zh-TW": "明明登入好了，腳本卻回你「Not logged in」，很容易讓人摸不著頭緒。這發生在雲端工作階段裡另外執行 `claude -p`（不開互動畫面、直接回一次答案）時：只要沒帶上該工作階段的環境變數，就會這樣誤報。現在不會了。",
                "en": "Running `claude -p` (which answers once without opening the interactive screen) inside a cloud session without that session's environment variables used to report \"Not logged in\", as if you had never signed in. It no longer does."
              },
              "original": "Fixed `claude -p` inside cloud sessions reporting \"Not logged in\" when started without the session's environment variables"
            },
            {
              "title": {
                "zh-TW": "接很多 MCP 工具時，API 錯誤不再被誤報成對話太長",
                "en": "With many MCP tools, API errors are no longer mislabeled as \"prompt is too long\""
              },
              "body": {
                "zh-TW": "明明才聊幾句，畫面卻說「prompt is too long」，你可能以為得開新對話了。這發生在接了很多 MCP 工具（讓 Claude 連外部服務的擴充）的工作階段：碰到 API 錯誤時，錯誤訊息會被換成這句。它還會順手做一次不必要的 compaction（把舊對話濃縮成摘要）。現在你看到的是原本的 API 錯誤，對話也不會白白被壓縮。",
                "en": "In sessions with many MCP tools (extensions that connect Claude to outside services), an API error used to show up as \"prompt is too long\", making you think the conversation had grown too long. It also triggered an unneeded compaction (condensing older conversation into a summary). Now you see the original API error, and your conversation is not compacted for nothing."
              },
              "original": "Fixed an API error being replaced by \"prompt is too long\", and an unneeded compaction, in sessions with many MCP tools"
            }
          ]
        }
      ]
    },
    {
      "id": "codex",
      "name": "Codex",
      "versions": [
        {
          "version": "0.164.0",
          "period": "2026-10-23",
          "items": [
            {
              "title": {
                "zh-TW": "/presets 新增十組快捷欄位，常用提示詞與指令隨選即用",
                "en": "/presets adds ten shortcut slots for instant access to frequent prompts and commands"
              },
              "body": {
                "zh-TW": "當你反覆在對話中輸入相同的檢查指令或固定提問時，重新打字往往耗費時間。⟦宛若汽車音響上的電台記憶功能，以前想聽固定頻道每次都要手動調頻，現在可以選取已儲存的頻道。⟧現在你可以打 `/presets` 將常用的提示詞與斜線指令儲存到十個編號欄位中。呼叫時還會跳出專屬的鍵盤選單，讓你透過鍵盤快速選取所需的預設項目。例如每次開始整理報告時，就能透過鍵盤選單選取儲存好的審查指令，省去反覆翻找與複製貼上的麻煩。",
                "en": "When you repeatedly enter the same inspection commands or standard questions during a conversation, retyping them is often time-consuming. ⟦Picture a car radio with station presets, where listening to a favorite channel once required manual tuning each time, but now only takes selecting a saved station.⟧ You can now type `/presets` to store frequently used prompts and slash commands into ten numbered slots. Invoking it brings up a dedicated keyboard menu so you can quickly choose the preset you need. For example, whenever you start compiling a report, you can select a saved review command through the keyboard menu, eliminating the hassle of searching around and copy-pasting."
              },
              "original": "Save reusable prompts and slash commands in ten numbered `/presets` slots, with a keyboard menu for quick access. (#54380, #54381)"
            },
            {
              "title": {
                "zh-TW": "/model 與 /permissions 支援行內參數，單行直接套用設定",
                "en": "/model and /permissions support inline arguments to apply settings in one command"
              },
              "body": {
                "zh-TW": "臨時需要切換模型或調整執行權限時，你可以把要用的設定直接接在指令後面。⟦如同在義式咖啡機直接轉動濃淡旋鈕，以前分開調整的出杯速度與風味濃度，現在點選飲品時順手一扭就能一併指定。⟧現在打 `/model` 或 `/permissions` 時，可以直接在同一行指令接上參數。你能直接在指令後方指定模型名稱、運算速度與推理解題力度（reasoning effort）。如果需要變更授權層級，也可以透過行內參數指定權限設定檔（permission profile）。這讓你輸入一行指令就能立刻切換到所需的執行環境。",
                "en": "When you need to switch models or adjust execution permissions on the fly, you can append your desired settings directly to the command. ⟦Think of turning a brew strength knob on an espresso machine, where cup speed and flavor intensity once had to be adjusted separately, but now can be set with a quick twist when selecting a beverage.⟧ When typing `/model` or `/permissions`, you can now pass arguments directly on the same line. You can specify the model name, compute speed, and reasoning effort right after the command. If you need to change authorization levels, you can also specify a permission profile through inline arguments. This lets you switch immediately to the required runtime environment with a single command."
              },
              "original": "Set model, reasoning effort, speed, and permission profiles directly with inline `/model` and `/permissions` arguments. (#54378, #54383)"
            },
            {
              "title": {
                "zh-TW": "支援 Kitty 與 Sixel 協定，全螢幕紀錄直接預覽圖片",
                "en": "Kitty and Sixel protocols supported to preview images directly in fullscreen logs"
              },
              "body": {
                "zh-TW": "在純文字終端機中請 AI 產生或查看圖片時，常常必須切換到外部檢視器才能確認成果。⟦彷彿按下數位相機快門後直接在電子觀景窗回放相片，以前非得把記憶卡插進電腦螢幕才能檢視構圖，現在眼睛不用離開目鏡就能看清影像細節。⟧如果你的終端機支援 Kitty 或 Sixel 圖形協定，現在全螢幕對話紀錄中可以直接預覽圖片。無論是讓工具讀取既有的本機圖檔，或是由模型新產生的圖片，畫面都會直接在紀錄中渲染顯示，不必再特地跳出視窗開啟檔案。",
                "en": "When asking AI to generate or inspect images in a plain-text terminal, you often have to switch to an external viewer to verify results. ⟦Much like checking a digital camera's electronic viewfinder right after pressing the shutter, where reviewing composition once required plugging a memory card into a computer monitor, you can now inspect image details without taking your eye off the eyepiece.⟧ If your terminal supports the Kitty or Sixel graphics protocol, you can now preview images directly inside fullscreen conversation transcripts. Whether the tool reads an existing local image file or displays a new image generated by the model, the graphic renders directly within the log without opening an external window."
              },
              "original": "Preview viewed and generated images in the fullscreen transcript on terminals supporting Kitty or Sixel graphics. (#54216)"
            },
            {
              "title": {
                "zh-TW": "企業級 MCP 支援專用登入與帳號授權，納入金鑰用戶端",
                "en": "Enterprise MCP supports dedicated login and account authorization with secret client access"
              },
              "body": {
                "zh-TW": "企業團隊在將內部系統串接至外部工具時，往往受限於嚴格的身分核驗與安全規範。⟦跟持員工證進入特約商務中心一樣，以前訪客通道只能辨識一般識別貼紙，現在閘門不僅支援企業總部的數位權限卡，也能核對附帶安全密鑰的專屬通行條。⟧現在企業級 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）連線支援更多元的驗證方式。你可以使用符合資格的 Codex 帳號授權直接通行，或是使用專屬登入通道完成驗證。對於需要用戶端密鑰（client secret）的 OAuth 用戶端，現在也能順利通過身分確認，安全連接各項企業服務。",
                "en": "When enterprise teams connect internal systems to external tools, they are often constrained by strict identity verification and security policies. ⟦Just as entering a partner business center with an employee badge once only supported basic visitor decals, the gate now accepts corporate digital access cards as well as dedicated passes with security keys.⟧ Enterprise MCP (a protocol connecting AI to external tools) connections now support more diverse authentication methods. You can authenticate directly using eligible Codex account authorization or complete verification through dedicated login channels. OAuth clients requiring client secrets can now also pass identity verification smoothly and securely connect to various enterprise services."
              },
              "original": "Authenticate enterprise MCP connections using eligible Codex account grants or dedicated sign-in, including OAuth clients requiring secrets. (#55334, #55859, #56001)"
            },
            {
              "title": {
                "zh-TW": "Amazon Bedrock 設定精靈預設啟用，自動整合 AWS 登入方式",
                "en": "Amazon Bedrock setup wizard enabled by default to integrate AWS login methods"
              },
              "body": {
                "zh-TW": "初次在工具中設定雲端模型提供者時，複雜的認證流程常讓人無所適從。⟦好比智慧電視初次開機的引導流程，以前要自己翻出網路線與遙控器代碼手動輸入，現在一開機就自動跳出連線精靈，還會把同品牌的登入選項整合在同一個畫面。⟧現在 Amazon Bedrock 的設定引導精靈改為預設開啟。當系統偵測到環境中的憑證時，API 金鑰輸入選項會自動與其他 AWS 登入方式歸類在同一個區塊，讓你清楚檢視並選取合適的登入途徑，不必再四處尋找設定入口。",
                "en": "When configuring a cloud model provider in the tool for the first time, complex authentication flows can feel overwhelming. ⟦Consider the initial setup guide of a smart TV, where you once had to find Ethernet cables and enter remote codes manually, but now a setup wizard appears automatically and brings brand login options together on one screen.⟧ The configuration setup wizard for Amazon Bedrock is now enabled by default. When credentials are detected in the environment, the API key input option is automatically grouped into the same section alongside other AWS login methods. This allows you to clearly view and select the appropriate sign-in path without searching around for configuration menus."
              },
              "original": "The Amazon Bedrock setup wizard is now enabled by default, with API key entry grouped alongside other AWS sign-in methods when credentials are detected. (#56033, #56023)"
            },
            {
              "title": {
                "zh-TW": "對話壓縮失敗自動備援，恢復工作階段完整保留討論串設定",
                "en": "Compaction failures fall back smoothly while resumed sessions preserve thread settings"
              },
              "body": {
                "zh-TW": "長時間與 AI 討論累積了大量對話紀錄時，背後的摘要壓縮機制是維持流暢對答的關鍵。⟦就像編輯長篇文件時的自動存檔機制，以前存檔失敗會在草稿夾留下一半的殘缺摘要，重開檔案時個人偏好也被重設成原廠設定；現在即使背景壓縮失敗也會平穩備援，重新開啟時還能精準還原原本的排版。⟧現在對話歷史壓縮（compaction）若不幸失敗，不會再於對話紀錄中留下半截殘缺的摘要內容，背景壓縮失敗時也會安全備援而不引發程式異常崩潰。恢復工作階段（session）時，工具現在能確實找到已壓縮的歷史紀錄，並完整復原原本討論串儲存的設定，而不是擅自將它們覆蓋成全域預設值。",
                "en": "When extensive discussions with AI accumulate substantial conversation history, the underlying summary compaction mechanism is essential for maintaining smooth exchanges. ⟦Like an autosave feature while editing a long document, where a save failure once left a broken half-summary in drafts and reset personal preferences to factory defaults upon reopening, background compaction failures now fall back smoothly and restore original layouts upon relaunch.⟧ If conversation history compaction fails, it no longer leaves partial summary fragments in transcripts, and background compaction failures now fall back safely without triggering crashes. When resuming a session, the tool reliably finds compacted history and fully restores the thread's saved settings instead of overwriting them with global defaults."
              },
              "original": "Failed compaction no longer leaves partial summaries in conversation history, and background compaction failures fall back without panicking. (#53932, #55581)\nResuming sessions now finds compressed histories and restores saved thread settings instead of replacing them with global defaults. (#54008, #55923)"
            },
            {
              "title": {
                "zh-TW": "自動清理未使用的安裝與市集暫存，設定日誌上限釋放磁碟空間",
                "en": "Prune unused installs and marketplace caches while capping daemon logs to free disk space"
              },
              "body": {
                "zh-TW": "工具在背景持續運作或頻繁下載外掛後，磁碟空間常常在不知不覺間被暫存檔案佔滿。⟦類似手機定期執行儲存空間大掃除，以前沒在用的舊版本與未再使用的檔案包會一直堆在角落，背景服務的紀錄簿也無限膨脹；現在清理工具會主動清空廢棄包裝，並替紀錄簿設下頁數上限。⟧現在工具加入了多項磁碟維護機制來降低空間佔用。系統會主動修剪不再使用的安裝版本，清理市集中已棄置的 checkout（取出的檔案副本）。同時替常駐程式（daemon）的日誌檔案設定大小上限，防止日誌無限制增長消耗硬碟空間。",
                "en": "When tools run continuously in the background or download plugins frequently, disk space often fills up unnoticed with temporary files. ⟦Similar to periodic storage cleanups on a smartphone, where unused old versions and abandoned package archives once piled up in corners while background logs grew without bound, cleanup tools now purge obsolete packages and cap log pages.⟧ The tool now includes several disk maintenance mechanisms to reduce disk usage. The system actively prunes unused installation versions and cleans up abandoned marketplace checkouts (extracted working copies of files). It also sets a size limit on daemon log files to prevent logs from growing indefinitely and consuming disk storage."
              },
              "original": "Reduced disk usage by pruning unused installations, cleaning abandoned marketplace checkouts, and bounding daemon logs. (#53827, #55149, #55689, #55690, #55856)"
            },
            {
              "title": {
                "zh-TW": "強化執行環境安全防護，遮蔽終端機憑證並嚴格檢驗路徑",
                "en": "Harden runtime security by masking terminal credentials and strictly validating paths"
              },
              "body": {
                "zh-TW": "在執行腳本或調用系統環境時，防止敏感資訊外洩與防堵潛在執行漏洞是保障系統安全的核心關鍵。\n\n終端機操作方面，支援的遠端 Bash 指令與 Zsh 快照工作階段（session）現在會自動遮蔽各類憑證，包含來自自訂提供者的敏感認證金鑰，避免機密在紀錄或螢幕輸出中曝光。\n\n沙盒防護也有所強化，針對 macOS 系統上大小寫變形的中繼資料路徑，以及用來尋找執行檔的符號連結（symlink，指向另一個位置的捷徑檔），都加入了嚴格的檢驗機制，防止路徑被惡意替換。\n\n內嵌的 JavaScript 執行環境也完成更新，為最佳化後的 `Array.sort` 排序操作加入安全檢查，防範異常。以前在遠端執行或排查腳本時容易擔心金鑰外洩或路徑被掉包，現在各個執行環節都有專門的安全防線把關。",
                "en": "When executing scripts or invoking system environments, preventing sensitive information leaks and closing potential execution vulnerabilities are essential to securing the system.\n\nFor terminal operations, supported remote Bash commands and Zsh snapshot sessions now automatically mask various credentials, including sensitive authentication keys from custom providers, preventing secrets from being exposed in logs or screen output.\n\nSandbox defenses have also been strengthened by adding strict validation mechanisms against case-variant metadata paths on macOS and symlinks (shortcut files pointing to another location) used to locate executables, preventing paths from being maliciously substituted.\n\nThe embedded JavaScript runtime has also been updated to add safety checks to optimized `Array.sort` operations, guarding against anomalies. While running or debugging scripts remotely once raised concerns about leaked keys or swapped paths, dedicated security safeguards now protect each execution stage."
              },
              "original": "Mask credentials during supported remote Bash commands and Zsh snapshot sessions, including credentials from custom providers. (#54309, #54432, #56049)\nHardened sandbox checks for case-variant metadata paths on macOS and replaceable symlinks used to locate executables. (#54273, #55384)\nUpdated the embedded JavaScript runtime with safety checks for optimized `Array.sort` operations. (#54519)"
            },
            {
              "title": {
                "zh-TW": "Windows 支援代理伺服器驗證，指令列設定覆寫與排隊流程修復",
                "en": "Windows supports proxy authentication alongside fixes for CLI config overrides and prompt queuing"
              },
              "body": {
                "zh-TW": "在企業網路環境或複雜的指令列流程中，連線驗證與參數傳遞的正確性直接影響工作進展。\n\n在 Windows 系統中，HTTPS 與 WebSocket 連線現在能正確處理代理伺服器（proxy）的驗證要求，自動使用已登入使用者的 Negotiate 或 NTLM 憑證完成身分查核，避免連線失敗中斷。\n\n在終端機執行多層子指令時，若在不同層級重複帶入 `-c` 或 `--config` 參數，現在各層級的設定覆寫值都會完整保留，不再被意外覆蓋或遺失。\n\n排隊等待執行的提示詞也調整了執行邏輯，現在會確實等候前方擱置中的權限確認選單完成選擇，才會繼續往下執行。以前遇到內網代理或連續提問時容易卡在連線中斷或權限錯亂，現在整套指令與排程流程都能順暢運行。",
                "en": "In enterprise network environments or complex command-line workflows, the reliability of connection authentication and parameter passing directly impacts work progress.\n\nOn Windows, HTTPS and WebSocket connections now properly handle proxy authentication challenges by automatically verifying identities with the logged-in user's Negotiate or NTLM credentials, preventing connection failures.\n\nWhen running nested subcommands in the terminal with `-c` or `--config` arguments repeated across different levels, configuration overrides at each tier are now fully preserved instead of being accidentally overwritten or lost.\n\nQueued prompts have also received updated execution logic, now reliably waiting for preceding pending permission menus to finish their selection before continuing execution. While corporate proxies or consecutive prompts once led to dropped connections or permission mix-ups, the entire command and scheduling pipeline now runs smoothly."
              },
              "original": "Windows HTTPS and WebSocket connections now handle proxy authentication challenges using the signed-in user’s Negotiate or NTLM credentials. (#55994, #56000)\nCLI configuration overrides are preserved when `-c` or `--config` appears at multiple subcommand levels. (#56022)\nQueued prompts now wait for pending permission selections to resolve before continuing. (#55682)"
            }
          ]
        },
        {
          "version": "0.163.0",
          "period": "2026-10-15",
          "items": [
            {
              "title": {
                "zh-TW": "/config 支援跨類別搜尋設定，即時修改並提示重啟",
                "en": "/config Supports Cross-Category Setting Search, Instant Edits, and Restart Prompts"
              },
              "body": {
                "zh-TW": "當你想調整 Codex 的各項參數，卻記不清具體設定名稱藏在哪個分類時，現在直接輸入 `/config` 就能解決。⟦彷彿走進大型量販店使用智慧導覽機，以前要一層層貨架慢慢找商品，現在直接搜尋關鍵字就能立刻取貨，機器還會貼心提醒哪些生鮮結帳後必須馬上冷藏。⟧輸入 `/config` 後，你可以跨越不同分類快速搜尋偏好設定，支援的選項能當場修改；哪些變更需要重啟，介面也會清楚標示，不必再自己猜測設定是否需要重啟。",
                "en": "When you want to adjust parameters in Codex but cannot recall which category a specific setting is tucked under, typing `/config` now solves the problem immediately. ⟦Picture using a smart kiosk in a large wholesale store: previously you had to browse through aisle after aisle of shelves to find an item, but now searching by keyword lets you pick it up immediately, and the machine even helpfully reminds you which perishables must be refrigerated right after checkout.⟧ After entering `/config`, you can quickly search for preferences across different categories, and supported options can be modified on the spot. The interface also clearly indicates which changes require a restart, sparing you from guessing whether a setting demands one."
              },
              "original": "Open `/config` to search preferences across categories, change supported settings immediately, and see which changes require a restart. (#52990, #53429)"
            },
            {
              "title": {
                "zh-TW": "生成中途可即時追問轉向，中斷時同步停止程式碼與工具呼叫",
                "en": "Steer Generations Mid-Stream with Follow-Up Prompts, Halting Code and Tool Calls on Interrupt"
              },
              "body": {
                "zh-TW": "看著 AI 劈頭往錯誤的方向洋洋灑灑寫下去，你不用再乾等它講完廢話才能打字糾正。⟦好比教練在場邊即時對球員大喊變更戰術，以前只能眼睜睜看著整套戰術打完才能檢討，現在哨音一響場上球員立刻改變跑位，正在熱身準備上場的替補球員也會全部停下動作。⟧Codex 正在回答的當下，你可以隨時送出後續追問來即時引導思考方向。如果你決定中斷這次回應，正在執行的 Code Mode（逐格執行程式碼的模式）儲存格會立刻停止，連同其巢狀工具呼叫也一併結束，避免背景程序白白空轉。",
                "en": "Watching the AI dash off in the wrong direction with paragraph after paragraph, you no longer have to wait idly until it finishes before typing a correction. ⟦Think of a coach shouting tactical changes from the sidelines to players in real time: previously you had to watch the entire play run its course before reviewing it, but now players shift positions the moment the whistle blows, while bench substitutes warming up to enter the game halt their movements as well.⟧ While Codex is in the middle of responding, you can send follow-up prompts at any time to steer its train of thought on the fly. If you decide to interrupt the response, currently executing Code Mode (a mode that executes code cell by cell) cells stop immediately, and their nested tool calls are terminated alongside them, preventing background processes from spinning in vain."
              },
              "original": "Send a follow-up while Codex is responding to steer it immediately. Interrupting a turn also stops active Code Mode cells and their nested tool calls. (#51812, #51835)"
            },
            {
              "title": {
                "zh-TW": "/keymap 支援自訂前導鍵，快捷鍵提示常駐且語音靜音改為 F9",
                "en": "/keymap Supports Custom Leader Keys, Sticky Shortcut Prompts, and Moves Voice Mute to F9"
              },
              "body": {
                "zh-TW": "習慣用鍵盤快速叫出各種操作的讀者，現在調整自訂按鍵更加得心應手。⟦如同樂團指揮設定了預備拍，以前比劃動作稍縱即逝容易漏看，現在提示手勢會穩穩停在空中等你接招，也把靜音操作換到另一個指定琴鍵。⟧在 `/keymap` 設定中，你可以自由配置前導鍵（按下一組按鍵後等待後續指令的引導鍵），系統預設組合為 `Ctrl+X`。按下後跳出的快捷鍵提示會持續停留在畫面上，直到你做出選擇或取消為止。語音靜音功能則移到了 `F9` 按鍵。",
                "en": "Readers accustomed to invoking operations swiftly via the keyboard can now customize key bindings with far greater ease. ⟦Imagine an orchestra conductor establishing a preparatory beat: previously the fleeting gesture was easily missed, but now the cueing hand rests steadily in the air waiting for your response, while the mute control is reassigned to another designated piano key.⟧ Within `/keymap` settings, you can freely configure a leader key (a guiding key that waits for follow-up inputs once pressed), with the system defaulting to `Ctrl+X`. The shortcut prompt that appears after pressing it remains on screen until you make a choice or cancel. Voice muting has been moved to the `F9` key."
              },
              "original": "Configure a leader key in `/keymap`. The default is `Ctrl+X`, shortcut hints stay open until you choose or cancel, and voice mute moves to `F9`. (#52273)"
            },
            {
              "title": {
                "zh-TW": "/daemon 支援重啟背景伺服器，中斷前主動警告並自動重連",
                "en": "/daemon Supports Restarting Background Servers with Disruption Warnings and Auto-Reconnect"
              },
              "body": {
                "zh-TW": "當背景執行的伺服器連線卡住或需要重載時，你不必再大費周章手動終止處理程序。⟦就像大樓總電源維護前廣播通知，以前電閘一拉整棟直接斷電，現在管理員會先警告哪些電梯與設備即將暫停，重啟完畢後還會自動幫大家把照明重新接通。⟧現在從 `/daemon` 指令就能直接重啟受管的背景伺服器（常駐在系統底層默默運作的背景程序）。重啟前，Codex 會先跳出警告，提醒這次重啟可能會中斷哪些進行中的工作。確認執行後，系統會在伺服器重啟完成時自動重新連線，省去手動重連的繁瑣步驟。",
                "en": "When a background server connection hangs or requires a reload, you no longer need to go through the hassle of manually killing the process. ⟦Consider a building announcement prior to main electrical maintenance: previously throwing the main switch cut power to the entire building without ceremony, but now the superintendent first warns which elevators and facilities will temporarily halt, and automatically reconnects lighting once the restart completes.⟧ You can now restart managed background servers (daemons running quietly at the system level) directly from the `/daemon` command. Before restarting, Codex displays a warning outlining which active tasks might be disrupted. Once confirmed, the system automatically reconnects as soon as the server finishes rebooting, eliminating tedious manual reconnection steps."
              },
              "original": "Restart a managed background server from `/daemon`. Codex warns about work the restart could interrupt, then reconnects automatically. (#53173)"
            },
            {
              "title": {
                "zh-TW": "唯讀工具支援平行執行，技能、記憶與對話紀錄查詢不再排隊",
                "en": "Read-Only Tools Support Parallel Execution: Skills, Memory, and Session History Queries No Longer Queue Up"
              },
              "body": {
                "zh-TW": "提問時需要 AI 同時調閱多項過往資訊的讀者，現在能感受到明顯更快的查詢反應。⟦跟圖書館調派多位館員同時分頭找書一樣，以前一個人要先找完歷史檔案才能去翻閱典藏技能，現在三位館員各自朝不同書架出發，找齊資料後一口氣交到你手上。⟧Codex 現在能夠平行（同時多工處理而不依序等待）執行彼此互不干擾的獨立唯讀工具。這項機制涵蓋了擴充技能檢索、記憶調閱，以及對話歷史紀錄查詢。這些互不衝突的讀取工作現在能同步展開，縮短等待回答的時間。",
                "en": "Readers whose questions require the AI to retrieve multiple pieces of past information at once will now notice noticeably faster query responses. ⟦A library dispatches multiple librarians to fetch books simultaneously: where one person previously had to search historical archives before checking archived skills, three librarians now head toward separate shelves at once and deliver everything into your hands the moment it is gathered.⟧ Codex can now execute mutually independent read-only tools in parallel (processing tasks concurrently without waiting in line). This mechanism covers skill retrieval, memory lookups, and session history queries. These non-conflicting read tasks now run concurrently, cutting down the wait time for answers."
              },
              "original": "Codex can run independent read-only tools in parallel, including skill, memory, and conversation-history lookups. (#52245)"
            },
            {
              "title": {
                "zh-TW": "Daybreak 模型對合格帳號預設啟用，codex exec 支援金鑰指定選用",
                "en": "Daybreak Model Enabled by Default for Eligible Accounts, with Key-Based Selection in codex exec"
              },
              "body": {
                "zh-TW": "想體驗最新 Daybreak 模型的讀者，只要帳號符合資格，打開程式就能直接用上。⟦宛若高鐵會員自動升級商務車廂，以前必須特地到櫃檯出示證件申請，現在符合條件的旅客刷票進站就能直接入座，在自動售票機輸入個人憑證也能順利劃位。⟧針對符合資格的帳號，Codex 現已預設啟用 Daybreak 模型。如果你習慣使用終端機指令 `codex exec` 執行任務，在伺服器提供存取權限的前提下，只要附上 OpenAI API 金鑰（存取雲端服務的通關密鑰），也能透過該指令自由選用 Daybreak 模型。",
                "en": "Readers eager to experience the latest Daybreak model can now use it right out of the box, provided their account is eligible. ⟦A high-speed rail membership automatically upgrades passengers to business class: where travelers once had to present identification at the counter to apply, eligible passengers can now swipe their tickets at the gate to take their seats directly, while entering personal credentials at automated ticketing machines also secures a seat smoothly.⟧ Codex now enables the Daybreak model by default for eligible accounts. If you prefer running tasks with the terminal command `codex exec`, you can also freely select the Daybreak model through that command by supplying an OpenAI API key (the secret pass for accessing cloud services), provided the server grants access."
              },
              "original": "Daybreak is enabled by default for eligible accounts. `codex exec` can also select it with an OpenAI API key when the server offers access. (#53701, #53717)"
            },
            {
              "title": {
                "zh-TW": "對話串切換支援 /side 回到分支，重啟時完整恢復授權與網路許可",
                "en": "Thread Switching Supports /side to Return to Branches, Fully Restoring Authorizations and Network Permissions on Reopen"
              },
              "body": {
                "zh-TW": "頻繁在多個任務之間切換對話串（在同一專案中分頭進行的獨立對話紀錄）的讀者，現在不用擔心進度與設定跑位。⟦有如在辦公室借調多份專案公文，以前換個案子審核印章就得重新蓋過，現在換回舊案能直接調閱已批准的用印紀錄，派出去的公文回函也會精準送回當初發文的專案夾。⟧切換對話串後，你可以隨時輸入 `/side` 指令回到先前的旁支對話。如果切換前有尚未回覆完畢的提示詞，產生的答案也會準確送回當初提問的對話串。當你重新打開已保存的對話串時，先前擁有的權限與網路存取許可都會完整還原，同時依然遵守手動覆寫與現行的管理政策。",
                "en": "Readers who frequently switch between threads (independent conversation logs running concurrently in the same project) no longer need to worry about losing progress or configuration state. ⟦Handling multiple project dossiers in an office setting once required restamping approvals whenever you switched cases, but returning to an earlier case now pulls up previously approved seal records directly, and replies to dispatched memos return precisely to the originating project folder.⟧ After switching threads, you can type the `/side` command at any point to return to a previous side discussion. If a prompt was still generating when you switched, the resulting response is routed accurately back to the thread where the question originated. When reopening a saved thread, your previously granted authorizations and network access permissions are fully restored, while continuing to respect manual overrides and active administrative policies."
              },
              "original": "Resuming a saved thread restores its permissions and network approvals, while respecting explicit overrides and current managed policy. (#53150, #53419, #53424)\nReturn to an existing side conversation with `/side` after switching threads. Answers to pending prompts also go back to the thread that asked. (#52184, #53436)"
            },
            {
              "title": {
                "zh-TW": "網路代理加強防護：DNS 查詢前先審核、阻擋重複標頭與憑證別名繞過",
                "en": "Network Proxy Security Tightens: Pre-DNS Query Audits, Blocking Duplicate Headers and Credential Alias Bypasses"
              },
              "body": {
                "zh-TW": "在需要透過代理伺服器連接外部網路的環境中，這項更新替連線安全築起了更嚴格的防線。⟦猶如機場海關加強查驗流程，以前旅客報出目的地前不會先核驗護照，現在查驗官在查詢航線前就先確認出境許可，嚴查重複填寫的申報表，也不再讓換了別名的違禁證件繞過檢查站。⟧網路代理現在會在發起網域名稱查詢（將網址轉譯為伺服器位址的查詢）前，先行檢查該連線是否具備相應權限。代理伺服器也會主動拒絕帶有重複 `Host` 標頭的請求，並防止憑證別名繞過檢查掛鉤（特定事件發生時被觸發的小程式），杜絕未經授權的潛在連線漏洞。",
                "en": "In environments that route external network traffic through a proxy server, this update erects a stricter line of defense for connection security. ⟦Customs inspectors at an airport checkpoint tighten screening procedures: where officers previously did not verify passports before travelers stated their destinations, inspectors now confirm exit permits before looking up flight routes, strictly scrutinize duplicate declaration forms, and block contraband credentials disguised under aliases from slipping past checkpoints.⟧ The network proxy now verifies whether a connection holds appropriate permissions before initiating a domain name query (resolving a web address into a server destination). The proxy server also actively rejects requests bearing duplicate `Host` headers and prevents credential aliases from bypassing check hooks (small routines triggered when specific events occur), plugging potential unauthorized connection loopholes."
              },
              "original": "The network proxy now checks permission before a DNS lookup, rejects duplicate `Host` headers, and prevents credential aliases from bypassing inspection hooks. (#51650, #52661, #53506)"
            },
            {
              "title": {
                "zh-TW": "Windows 支援快捷鍵貼圖，WSL 剪貼簿重複讀取並防止複製卡死",
                "en": "Windows Gains Image Pasting Shortcuts, While WSL Reuses Clipboard Readers and Prevents Copy Freezes"
              },
              "body": {
                "zh-TW": "在 Windows 或 WSL（Windows 系統內建的 Linux 子系統）環境下操作終端機的讀者，剪貼內容時將順暢許多。⟦恰似換了一組靈敏的櫃檯讀卡機，兩組指定快捷鍵都能把文字與圖片直接貼進視窗；頻繁讀取時能共用讀取模組，複製操作也會逾時退出，不再一直卡住。⟧在 Windows 中，`Ctrl+V` 與 `Shift+Insert` 現在都能貼上文字與圖片。而在 WSL 環境下，連續多次貼上會重複使用剪貼簿讀取器；複製操作在遇到阻礙時則會逾時退出，不再無休止地卡死畫面。",
                "en": "Readers running terminals under Windows or WSL (Windows Subsystem for Linux) will experience far smoother clipboard workflows. ⟦Upgrading to a responsive card reader at a checkout counter allows two designated shortcuts to feed both text and images straight into the window, frequent scans share the reader module, and stalled copy operations time out instead of hanging indefinitely.⟧ On Windows, both `Ctrl+V` and `Shift+Insert` can now paste text and images. Under WSL, successive pastes reuse the clipboard reader, while copy operations encountering obstacles time out gracefully rather than freezing the screen indefinitely."
              },
              "original": "On Windows, `Ctrl+V` and `Shift+Insert` paste text as well as images. On WSL, repeated pastes reuse the clipboard reader, and copy attempts time out instead of hanging. (#53224, #52967, #53453)"
            },
            {
              "title": {
                "zh-TW": "指令輸出修復：本地與遠端傳輸更穩，codex exec 不再重複印出差異",
                "en": "Command Output Fixes: Steadier Local and Remote Transfers, codex exec Stops Reprinting Diffs"
              },
              "body": {
                "zh-TW": "經常在終端機執行本機指令或遠端遙控工作的讀者，這次更新修復了多項令指令卡住或資訊混亂的輸出問題。⟦宛如貨運物流中心整頓了出貨流程，以前只要下游包裝箱還沒封口，送貨司機就算卸完貨也只能在原地枯等，現在確認卸貨完成就能迅速出發，晚到的包裹也能妥善入庫而不被遺漏。⟧以前執行本機指令時，只要底層的子行程（由主程式衍生出來獨立處理任務的分支程式）沒有關閉輸出管道，整個指令就會乾等而不肯結束。現在本地指令在工作完成後能迅速結束。執行遠端指令時，晚到的輸出內容現在會妥善保留，確保最後一次輪詢時依然能完整收集。使用 `codex exec` 執行無關指令後，終端機也不會再重複印出先前的檔案差異清單。",
                "en": "For readers who regularly run local commands or remote tasks in the terminal, this update resolves several output glitches that previously hung commands or cluttered information. ⟦A freight logistics hub overhauls its shipping workflow: previously, as long as downstream packing crates remained unsealed, delivery drivers had to wait idly on site even after finishing unloading, but now drivers depart promptly once unloading is confirmed, and late-arriving parcels are safely warehoused rather than missed.⟧ Previously when running local commands, as long as underlying subprocesses (child processes spawned by the main program to handle tasks independently) kept output pipes open, the overall command would stall indefinitely rather than terminate. Local commands now exit promptly once their work is done. When executing remote commands, late-arriving output is now safely retained, ensuring it is fully collected during the final poll. Running unrelated commands via `codex exec` also no longer causes the terminal to reprint previous file diffs."
              },
              "original": "Local commands finish promptly when a child keeps an output pipe open. Remote commands preserve late-arriving output so the final poll can still collect it. (#52652, #53331, #53622)\n`codex exec` no longer prints the same file diff again after an unrelated command. (#53444)"
            },
            {
              "title": {
                "zh-TW": "隨附技術文件更新：收錄 GPT-6 遷移指引與 Linux 沙箱架構說明",
                "en": "Bundled Documentation Update: GPT-6 Migration Guide and Linux Sandbox Architecture Details"
              },
              "body": {
                "zh-TW": "正在規劃將專案升級至新世代模型，或關注系統底層安全防護的讀者，現在能從隨附文件中查閱更詳盡的官方指南。⟦正如工廠設備手冊換上新版擴建指南，不僅詳細列出升級新型機具時的運轉特性與資料所在地考量，也公開了安全防護罩如何隔絕危險區域並過濾高壓氣流。⟧隨附的 OpenAI 文件已更新針對 GPT-6 系列模型的遷移指引，詳細說明了各模型特定的推理特性、外部工具使用方式，以及資料駐留（資料儲存與處理的地理位置）相關考量。隨附說明也進一步釐清了 Linux 沙箱機制的運作方式，詳細介紹系統如何透過 bubblewrap 達成檔案系統隔離，並藉由 seccomp 進行系統呼叫過濾。",
                "en": "Readers planning to upgrade projects to next-generation models or keeping an eye on low-level system security can now consult more detailed official guidance in the bundled documentation. ⟦A factory equipment manual swaps in a fresh expansion guide, not only detailing the operational characteristics and data location considerations when upgrading to new machinery, but also revealing how safety shields isolate hazardous areas and filter high-pressure airflow.⟧ The bundled OpenAI documentation updates migration guidance for the GPT-6 model family, detailing model-specific reasoning characteristics, external tool usage, and data residency (the geographic location where data is stored and processed) considerations. The bundled documentation also clarifies how the Linux sandbox mechanism operates, detailing how the system achieves filesystem isolation through bubblewrap and filters system calls via seccomp."
              },
              "original": "Updated the bundled OpenAI Docs guidance for migrations to the GPT-6 family, including model-specific reasoning, tool use, and data residency considerations. (#51768)\nClarified how the Linux sandbox uses bubblewrap for filesystem isolation and seccomp for syscall filtering. (#52610)"
            },
            {
              "title": {
                "zh-TW": "安裝包體積縮減，改用 Codex 的標準 shell 執行器並提示設定遷移",
                "en": "Package Size Reduced with Codex's Standard Shell Runner and Config Migration Notice"
              },
              "body": {
                "zh-TW": "在 Linux 與 macOS 環境下安裝或更新 Codex 的讀者，這次會發現下載更輕盈且底層執行器更單純。⟦與出差輕裝上陣的概念相同，以前行李箱裡硬塞了一整套自備工具組，現在改用原有的標準工具，並把非必要的裝飾吊牌全數卸下，讓行李箱體積與重量減輕不少。⟧官方透過從安裝二進位檔中剝除符號（除錯用的內部名稱標籤），縮減了 Linux 與 macOS 版本的安裝套件體積，原始建置產物則依然保留完整符號。在執行架構上，Codex 現在改用 Codex 的標準 shell 執行器（負責執行命令的工具），不再隨附修補版的 `zsh`。如果既有設定仍包含已淘汰的 shell 標記，系統會跳出遷移通知協助你平順過渡。",
                "en": "Readers installing or updating Codex on Linux and macOS will find lighter downloads and a simpler underlying runner. ⟦A business traveler packs light under the same principle: rather than stuffing a custom toolkit into a suitcase, they rely on existing standard tools and strip away nonessential decorative tags, significantly reducing suitcase size and weight.⟧ The team reduced package sizes for Linux and macOS releases by stripping symbols (internal name labels used for debugging) from installation binaries, while raw build artifacts still retain full symbols. In terms of execution architecture, Codex now switches to Codex's standard shell runner (a tool responsible for executing commands), no longer bundling a patched `zsh`. If existing configurations still contain deprecated shell flags, the system displays a migration notice to help you transition smoothly."
              },
              "original": "Codex now uses its standard shell executor and no longer bundles a patched `zsh`. Configurations using the retired shell flags receive a migration notice. (#52160, #52213)\nReduced Linux and macOS release package size by stripping symbols from packaged binaries while preserving them in the original build outputs. (#52178)"
            }
          ]
        },
        {
          "version": "0.162.1",
          "period": "2026-10-09",
          "items": [
            {
              "title": {
                "zh-TW": "終端機顯示多行提問不再當機",
                "en": "Terminal Displaying Multiline Prompts No Longer Crashes"
              },
              "body": {
                "zh-TW": "當你在文字視窗等待背景任務，途中跳出落落長的確認提問時，畫面不會再突然卡死關閉。以前遇到非同步提問（程式在背後運作時插進來詢問意見的對話）包含多行文字，終端機文字介面常會直接當機。這個問題現在已經修好，訊息裡的換行排版會原樣保留，附帶的超連結完整網址也不會遺失。以前常因為介面當機而看不到完整問題，現在長內容與跳轉連結都能順利讀取。",
                "en": "When you wait for background tasks in a text window and a lengthy confirmation prompt pops up along the way, the screen will no longer suddenly freeze and close. Previously, when asynchronous prompts (dialogues that interrupt to ask for input while programs run in the background) contained multiline text, the terminal text interface often crashed directly. This issue is now fixed, with line breaks in messages preserved as formatted and full URLs in attached hyperlinks no longer lost. While users previously often missed full questions due to interface crashes, long content and redirect links can now be read smoothly."
              },
              "original": "Fixed a TUI crash when asynchronous questions contain multiple lines, preserving line breaks and complete hyperlink destinations. (#51866)"
            },
            {
              "title": {
                "zh-TW": "命令列啟動不再因背景伺服器設定不同而失敗",
                "en": "Command-Line Startup No Longer Fails Over Differing Background Server Settings"
              },
              "body": {
                "zh-TW": "如果你習慣讓工具在背後持續運作，開新視窗輸入指令時，不必再擔心程式卡在開頭進不去。以前只要正在執行的背景伺服器與命令列介面（CLI，用文字指令操作的工具）預設設定不同，啟動就會報錯失敗。現在相容性檢查只會在命令列明確覆寫功能時才進行，啟動失敗的問題已經修好。以前常因預設值不同而被擋在門外，現在就算背景設定有別也能順利開啟。",
                "en": "If you are used to letting tools run continuously in the background, you no longer need to worry about programs getting stuck at startup when entering commands in a new window. Previously, whenever the running background server and the command-line interface (tools operated with text commands) had different default settings, startup would fail with an error. Now, compatibility checks only run when the command line explicitly overrides features, and the startup failure issue has been fixed. While users were previously locked out due to differing default values, the tool can now launch smoothly even when background configurations differ."
              },
              "original": "Fixed startup failures caused by differences between a running background server's feature settings and CLI defaults. Compatibility checks now apply only to explicit command-line feature overrides. (#52648)"
            }
          ]
        }
      ]
    },
    {
      "id": "agy",
      "name": "Antigravity",
      "versions": [
        {
          "version": "1.4.6",
          "period": "2026-10-23",
          "items": [
            {
              "title": {
                "zh-TW": "外掛清單新增 MCP 伺服器設定列，不必停用再啟用即可改變數",
                "en": "Plugin List Adds MCP Server Configuration Row, Updating Variables Without Disabling and Re-enabling"
              },
              "body": {
                "zh-TW": "平時替終端機安裝各種外掛後，若想微調連線金鑰或伺服器參數，現在不必再先停用、再重新啟用外掛。輸入 `/plugin` 檢視已安裝的外掛清單，現在每一筆外掛底下都會出現 `Configure MCP servers...` 設定選項。⟦彷彿在舞台音響架旁微調效果器的獨立旋鈕，以前更換參數就得拔掉總電源重新開機；現在面板清楚標出哪些線路仍待調整，轉完旋鈕後新設定立刻接手生效。⟧若外掛還有未填寫的變數，清單會標示 `⚠ needs setup` 提醒補齊；每個伺服器也會標出狀態，已備妥的會顯示 `✓ configured`，尚未完成的則列出具體缺少的變數名稱。儲存後，已啟用的外掛會自動重啟對應的 MCP（讓 AI 連接外部工具的協定）伺服器，讓新數值立即生效。",
                "en": "When fine-tuning connection keys or server parameters after installing plugins in your terminal, there is no longer a need to disable and re-enable them. Entering `/plugin` to view the list of installed plugins now displays a `Configure MCP servers...` option beneath every entry. ⟦Picture tweaking an individual knob on a stage audio rack, where changing parameters once required pulling the main power plug to restart the entire system; now the panel clearly marks which lines still need adjustment, and after turning the knob, new settings take effect immediately.⟧ If a plugin still has empty variables, the list marks them with `⚠ needs setup` as a reminder to fill them in; each server shows `✓ configured` for ready ones and lists specific missing variable names for incomplete ones. After saving, enabled plugins automatically restart their corresponding MCP (a protocol connecting AI to external tools) servers to apply new values immediately."
              },
              "original": "Added a `Configure MCP servers...` row for installed plugins in `/plugin`, so you can set or change a plugin's MCP server variables at any time without disabling and re-enabling it; plugins with missing values show `⚠ needs setup`, each server shows `✓ configured` or the variables it still needs, and saving restarts an enabled plugin's servers so the new values take effect."
            },
            {
              "title": {
                "zh-TW": "終端機沙盒預設保護更多憑證，存取 Docker 與 Kubernetes 需明確授權",
                "en": "Terminal Sandbox Protects More Credentials by Default, Requiring Explicit Authorization for Docker and Kubernetes Access"
              },
              "body": {
                "zh-TW": "在沙盒環境執行各類指令時，本機存放的敏感憑證現在有了更嚴格的預設隔離防護。沙盒指令預設不再允許讀取或修改 `~/.docker`、`~/.kube` 與 `~/_netrc` 這三處路徑，除非你明確授權放行。⟦跟入住出租套房的生活規範一樣，以前抽屜裡的私人證件預設敞開任人翻動；現在重要暗格直接上鎖，除非屋主親自授權放行，否則誰也不能隨意翻看。⟧過去沙盒內部預設開放寫入 `~/.docker` 與 `~/.kube`，因此現在任何需要使用 Docker registry（存放容器映像檔的倉庫）或 Kubernetes 叢集憑證的沙盒指令，都必須先取得你的明確授權才能繼續執行。",
                "en": "When running commands inside sandbox environments, locally stored sensitive credentials now receive stricter default isolation safeguards. Sandboxed commands no longer allow reading or modifying `~/.docker`, `~/.kube`, or `~/_netrc` by default unless you grant explicit authorization. ⟦Much like living rules in a rented studio, private documents inside drawers were previously left open for anyone to rummage through; now important compartments are locked directly, and no one may browse them without the owner's explicit permission.⟧ In the past, sandbox environments permitted writes to `~/.docker` and `~/.kube` by default. Consequently, any sandboxed command requiring Docker registry (a repository storing container images) credentials or Kubernetes cluster credentials must now obtain your explicit authorization before proceeding."
              },
              "original": "Improved terminal sandbox defaults to protect more credentials: sandboxed commands can no longer read or modify `~/.docker`, `~/.kube`, or `~/_netrc` unless you grant access explicitly. Previously `~/.docker` and `~/.kube` were writable inside the sandbox, so sandboxed commands that use Docker registry or Kubernetes cluster credentials now need an explicit grant."
            },
            {
              "title": {
                "zh-TW": "對話壓縮保留上傳圖片與 PDF，精簡歷史不漏看附件",
                "en": "Conversation Compaction Preserves Uploaded Images and PDFs, Keeping Attachments Accessible After History Pruning"
              },
              "body": {
                "zh-TW": "面對累積了大量訊息的長對話，手動或自動執行對話壓縮時，不用再擔心先前上傳的檔案就此消失。現在無論是系統自動觸發歷史壓縮，或是你在終端機手動輸入 `/compact`，對話摘要都會完整保留你先前上傳的圖片、PDF 與各類附件檔案。⟦翻閱裝訂成冊的案件卷宗，以前把過往紀錄濃縮成大綱時，夾在中間的現場照片與附件常被一併收進封存箱；現在所有參考文件都會直接釘在摘要首頁，回顧時隨時都能對照細節。⟧這讓 AI 即使在舊對話被精簡整合之後，依然能清楚讀取這些附加檔案，繼續針對檔案內的細節回答問題。",
                "en": "When facing long conversations with accumulated messages, you no longer need to worry about previously uploaded files disappearing during manual or automatic conversation compaction. Whether the system automatically triggers history compaction or you manually enter `/compact` in the terminal, the conversation summary completely preserves your previously uploaded images, PDFs, and various attachment files. ⟦Leafing through a bound case file where past records were condensed into outlines, crime scene photos and attachments clipped inside once got tucked away into storage boxes; now all reference documents remain pinned directly to the front page of the summary, ready for checking details at any review.⟧ This ensures the AI can still clearly inspect these attached files and continue answering questions regarding details within them, even after older exchanges are condensed and consolidated."
              },
              "original": "Improved context compaction, including `/compact`, to keep images, PDFs, and other files you attached in the conversation summary, so the agent can still see them after older history is summarized."
            },
            {
              "title": {
                "zh-TW": "授權範圍涵蓋主對話與子代理，同項操作不必重複確認",
                "en": "Authorization Scope Extends Across Main Conversation and Subagents, Eliminating Repeated Confirmations for the Same Operation"
              },
              "body": {
                "zh-TW": "把複雜任務交給子代理（在背景協助分工的獨立 AI 助理）分頭處理時，頻繁跳出的授權確認常打斷工作節奏。現在選擇針對整個對話允許這項操作後，無論你是在主對話還是子代理的授權提示中做出選擇，這項權限都會直接套用到整個主對話與旗下所有子代理。⟦持同一張專案通行證進出實驗大樓，以前只要換了不同助理人員進場操作，門禁警衛就得把主管叫出來重新簽核；現在只要主專案獲准通行，旗下所有支援團隊都能共用這項授權直接開工。⟧這省去了子代理遇到相同操作時再次請你確認的麻煩，分工作業不再頻繁中斷。",
                "en": "When delegating complex tasks to subagents (independent AI assistants helping divide work in the background), frequent authorization prompts often interrupt your workflow. Now, after choosing to allow an operation for the entire conversation, that permission applies directly across the entire main conversation and all of its subagents, regardless of whether you approved it from the main thread or a subagent's prompt. ⟦Holding a single project pass to access a research facility, bringing in a new assistant previously required security guards to call a manager for fresh sign-offs; now once the main project clears access, all supporting teams share that authorization to begin work immediately.⟧ This saves you from repeated approval prompts when subagents perform the same operation, keeping distributed work uninterrupted."
              },
              "original": "Improved permission prompts so that allowing an action for the conversation, in the main agent or in a subagent, applies across the main conversation and all of its subagents instead of asking again for the same action."
            },
            {
              "title": {
                "zh-TW": "狀態列顯示具體等待進度，模型重試附帶動態倒數",
                "en": "Status Line Displays Specific Waiting Progress with Dynamic Countdown for Model Retries"
              },
              "body": {
                "zh-TW": "送出複雜任務等待回應時，終端機不再只丟給你一個模糊的處理中標籤。終端機底部的 status line（顯示在終端機最底下的狀態列）現在會具體列出 AI 當下正在等待的環節，取代過去籠統的運作標籤。例如它正在初始化 MCP（讓 AI 連接外部工具的協定）伺服器，或是遇到模型暫時無法使用而自動重試時，都能即時掌握當前狀態。遇到重試狀況時，狀態列會直接出現類似 `Model unavailable (attempt 2/10), retrying in 12s` 的動態倒數秒數與嘗試次數。⟦看著候機室的電子告示牌，以前遇到班機延誤只顯示含糊的處理中，讓人只能盯著螢幕乾等；現在看板會即時標明正在調度登機門，並附上精確的倒數時間。⟧",
                "en": "When submitting complex tasks and waiting for responses, the terminal no longer leaves you with a vague in-progress label. The status line (the status bar displayed at the very bottom of the terminal) now specifically lists the exact step the AI is waiting on, replacing previous generic activity indicators. Whether it is initializing an MCP (a protocol connecting AI to external tools) server or automatically retrying when a model is temporarily unavailable, you can monitor current status in real time. During retry scenarios, the status line directly shows a dynamic countdown in seconds alongside the attempt count, such as `Model unavailable (attempt 2/10), retrying in 12s`. ⟦Flight display boards in an airport departure lounge once showed only a vague processing notice during flight delays, leaving passengers to stare helplessly at screens; now the board clearly states that a gate is being scheduled alongside an exact countdown.⟧"
              },
              "original": "Improved the in-progress status line to show what the agent is waiting on, such as initializing MCP servers or retrying an unavailable model (for example `Model unavailable (attempt 2/10), retrying in 12s` with a live countdown), instead of a generic working label."
            },
            {
              "title": {
                "zh-TW": "`/diff` 逐字突顯行內修改差異，各配色主題皆能清楚辨識",
                "en": "`/diff` Highlights Word-Level Inline Changes with Clear Visibility Across All Color Themes"
              },
              "body": {
                "zh-TW": "想看清程式碼或文字修改前後究竟動了哪幾個字，比對畫面現在能精準指出變更核心。只要在終端機輸入 `/diff` 查看修改差異，在所有色彩配置主題下，畫面都會直接在變動的那一行當中，標出真正發生變更的字詞。⟦拿螢光筆校對合約草稿時，圈出實際替換的字詞，就能把目光集中到修改的位置；現在的比對畫面也會替你標出這些重點，讓你查看每個修改行內究竟換了哪些字詞。⟧",
                "en": "When you want to see exactly which words were altered before and after code or text modifications, the diff view now pinpoints the core of every change with precision. Simply entering `/diff` in the terminal to inspect modifications highlights the exact words that changed right within the modified line across all color themes. ⟦Using a highlighter to proofread contract drafts, circling replaced words once helped focus attention on actual modifications; now the diff view flags those focal points for you so you can inspect exactly which words were swapped within each modified line.⟧"
              },
              "original": "Improved `/diff` to highlight the exact words that changed within each modified line, in every color scheme."
            },
            {
              "title": {
                "zh-TW": "外掛驗證支援說明選單，全面檢驗設定結構與技能檔",
                "en": "Plugin Validation Adds Help Flag, Thoroughly Verifying Config Structures and Skill Files"
              },
              "body": {
                "zh-TW": "開發自訂外掛時若想排查設定檔結構是否合法，現在有了更詳盡的驗證工具。在終端機執行外掛驗證指令，現在支援 `--help` 查看說明參數，並針對各項設定進行完整檢驗：\n\n```bash\nagy plugin validate --help\n```\n\n它會檢查情境提供者（context providers）的設定細節，抓出缺少 `label` 或 `command`、觸發字詞無效或重複、未知欄位，以及執行指令檔遺失等問題。針對技能設定，若 `plugin.json` 裡宣告的 `skills` 項目不存在或缺少 `SKILL.md`，驗證工具會主動發出警告，並詳細回報載入器跳過的所有 `SKILL.md`，包含透過 `plugin.json` 或 `skills.json` 引入的技能檔案。⟦檢查即將出廠的組裝模型盒，以前只粗略核對外盒印製的零件總數；現在檢驗員逐一對照說明書，抓出缺漏的組裝步驟與對不上的零件，連跳過未裝的模組都詳細列出。⟧",
                "en": "When troubleshooting configuration file structures while developing custom plugins, more detailed validation tooling is now available. Running the plugin validation command in the terminal now supports `--help` to inspect parameter documentation and performs comprehensive verification across settings:\n\n```bash\nagy plugin validate --help\n```\n\nIt inspects configuration details for context providers, catching missing `label` or `command` entries, invalid or duplicate trigger words, unknown fields, and missing executable script files. For skill configuration, if a `skills` entry declared in `plugin.json` does not exist or lacks `SKILL.md`, the validation tool issues a warning and reports in detail all `SKILL.md` files skipped by the loader, including skill files brought in via `plugin.json` or `skills.json`. ⟦A factory inspector examining a boxed model kit once only checked the total part count printed on the packaging; now the inspector cross-checks the instruction manual step by step to catch missing assembly steps and mismatched components, listing even the unmounted modules that were skipped.⟧"
              },
              "original": "Improved `agy plugin validate`: it now has `--help`, checks a plugin's context providers (missing `label` or `command`, invalid or duplicate triggers, unknown fields, missing command files), warns about `skills` entries in `plugin.json` that don't exist or contain no `SKILL.md`, and reports every `SKILL.md` the plugin loader skips, including skills listed through `plugin.json` or `skills.json`."
            },
            {
              "title": {
                "zh-TW": "終端機游標修正全形字與換行偏誤，上下方向鍵精準對位",
                "en": "Terminal Cursor Resolves Offsets for Full-Width Characters and Line Wraps, Aligning Up and Down Arrow Keys Accurately"
              },
              "body": {
                "zh-TW": "打字輸入中文、日文、韓文或表情符號時，終端機游標現在能精準停在插入點。以前在指令提示列，以及設定、權限修改、外掛設定、意見回饋、掛鉤與成品搜尋框中，輸入中日韓全形字（CJK）、表情符號、定位字元、特殊空白或文字遇到自動換行時，游標常會偏離插入點，按上下方向鍵也會跳到錯誤的行數。這個問題已經修好，游標與方向鍵現在都能正確停在對應文字位置。",
                "en": "When typing Chinese, Japanese, Korean, or emoji, the terminal cursor now stays accurately positioned at the insertion point. Previously, entering full-width CJK (Chinese, Japanese, and Korean) characters, emoji, tab characters, special whitespace, or wrapped text in the command prompt, as well as in the settings, permission amendment, plugin setup, feedback, hooks, or artifact search boxes, often caused the cursor to drift from the insertion point and made up and down arrow keys jump to the wrong line. This issue has been resolved, and both the cursor and arrow keys now land on the correct text positions."
              },
              "original": "Fixed text cursors drifting away from the insertion point after wide characters (CJK, emoji), tabs, other Unicode spaces, and on soft-wrapped lines, including `Up`/`Down` landing on the wrong row, in the prompt and in the settings, permission amend, plugin setup, feedback, hooks, and artifact search boxes."
            },
            {
              "title": {
                "zh-TW": "指令啟動、檔案編輯提示與企業登入的三項異常修復",
                "en": "Three Fixes for Command Execution, File Edit Prompts, and Enterprise Login Issues"
              },
              "body": {
                "zh-TW": "平時呼叫工具跑指令、想退回工作區外的變更，或是重新啟動後準備上工時，你可能曾碰過莫名報錯或被迫重新登入。\n\n工具呼叫附帶很大的 reasoning signature（推論簽章，模型隨工具呼叫回傳的資料）時，指令以前偶爾無法啟動並回報 `argument list too long`；這個問題已修好。\n\n提示畫面多出一個選項時，例如在接受編輯模式（accept-edits mode）下修改工作區外的檔案，警告標語 `Esc disabled during file edits` 先前會寫錯拒絕鍵的名稱，現在會寫出正確的拒絕鍵。\n\nmacOS 與 Windows 上的企業帳號先前若因登入憑證超出系統鑰匙圈的大小上限，重新啟動後登入狀態便會失效；這個問題已經修好，登入狀態現在能在重新啟動後正常延續。",
                "en": "When calling tools to execute commands, reverting changes outside your workspace, or preparing to work after a restart, you might have occasionally encountered unexpected errors or forced re-logins.\n\nWhen tool calls included a large reasoning signature (data returned by models alongside tool calls), commands occasionally failed to launch with `argument list too long`; this issue has been resolved.\n\nWhen the prompt shows an extra choice, such as for edits outside the workspace in accept-edits mode, the warning banner `Esc disabled during file edits` previously named the wrong reject key; it now names the correct one.\n\nWhen enterprise accounts on macOS and Windows had login credentials exceeding the system keychain size limit, their login state was previously lost upon restarting; this issue has been resolved, and login status now persists properly across restarts."
              },
              "original": "Fixed terminal commands occasionally failing to start with `argument list too long` when the model returned a very large reasoning signature with the tool call.\nFixed the `Esc disabled during file edits` warning naming the wrong reject key when the prompt shows an extra choice, such as for edits outside the workspace in accept-edits mode.\nFixed sign-in not persisting across restarts on macOS and Windows for enterprise accounts, because the saved credential exceeded the system keychain's size limit."
            }
          ]
        },
        {
          "version": "1.4.5",
          "period": "2026-10-22",
          "items": [
            {
              "title": {
                "zh-TW": "回覆正文與收合思考區塊空行區隔，頂端僅保留標題",
                "en": "Blank line separates response text from collapsed thought block, leaving only the header at the top"
              },
              "body": {
                "zh-TW": "閱讀代理人剛生成的回覆時，若思考過程與正文擠成一團，往往會打斷閱讀節奏。⟦像在會議中翻閱專案報告，以前草稿筆記直接黏在結論上方容易分散焦點；現在頂端只留一張寫著構思時間的簡潔書籤，中間隔出一行空白，正文一目了然，需要時再翻開整份草稿。⟧現在代理人開始輸出回覆後，收合的思考區塊只會顯示 `Thought for …` 標題，不再重複列出思考過程的文字片段。收合的思考區塊與正文之間也多了一行空行隔開，文字段落不會再擠在一起。如果你想確認剛才完整的推導邏輯，只要手動展開該區塊，依然能查閱完整的思考內容。",
                "en": "When reading a newly generated response from an agent, having the thought process clustered against the main text often breaks your reading flow. ⟦Picture reviewing a project report during a meeting: previously, draft notes stuck directly above the conclusion would distract focus, whereas now only a neat bookmark noting thinking time remains at the top with a blank line in between, keeping the main text clear at a glance while letting you unfold the full draft whenever needed.⟧ Now, once the agent begins outputting a response, the collapsed thought block only displays the `Thought for …` header instead of repeating text snippets from the thought process. A blank line also separates the collapsed thought block from the main text, so text paragraphs no longer bunch together. If you want to verify the complete reasoning logic behind it, you can still view the full thoughts simply by expanding the block manually."
              },
              "original": "Improved how agent responses read: a blank line now separates the collapsed thought block from the response, and once the response starts, the collapsed thought shows only its `Thought for …` header instead of repeating a snippet of the thinking; expand it to see the full thinking."
            },
            {
              "title": {
                "zh-TW": "比對檢視支援折疊搜尋計數，游標樣式與停駐位置對齊",
                "en": "Diff view supports folded search counts, aligning cursor style and resting position"
              },
              "body": {
                "zh-TW": "檢閱程式碼變更時，折疊起來的未更動行數常讓搜尋結果變得難以掌握。⟦好比翻閱厚重的帳冊索引，以前夾在折頁裡的項目算不進筆數，各分冊的標記色塊也不一致；現在標籤直接註記折疊處藏有幾筆帳目，選取欄位也統一換成箭頭符號，翻到當前回合就直接對準第一筆明細。⟧在 `/diff` 檔案檢視中以 `/` 搜尋時，工具現在會計算折疊行內的相符項目，以 `match 2/5 · +3 hidden` 顯示隱藏數量，`f` 鍵展開提示也改為 `show N hidden matches`。介面上，`Commit Tree` 改用與 `All Changes` 及 `Per Turn` 相同的 `>` 游標與選取色，不再用高亮色塊；切換到 `Per Turn` 時游標也會直接停在第一個檔案，不必再跳過回合標題。",
                "en": "When reviewing code changes, collapsed unchanged lines often make search results hard to track. ⟦Much like browsing a thick ledger index, items tucked inside folded pages were previously excluded from match counts and marker color blocks across volumes were inconsistent, whereas tabs now directly annotate how many entries are hidden in folds, selection columns uniformly switch to arrow symbols, and turning to the current turn lands straight on the first detail.⟧ When searching with `/` inside the `/diff` file view, the tool now counts matches within folded lines, displaying the hidden count as `match 2/5 · +3 hidden`, and the `f` key expand hint is now updated to `show N hidden matches`. On the interface, `Commit Tree` now uses the same `>` cursor and selection color as `All Changes` and `Per Turn` instead of highlight blocks; when switching to `Per Turn`, the cursor also rests directly on the first file without needing to skip past turn headers."
              },
              "original": "Improved `/` search in the `/diff` file view to count matches inside folded unchanged lines, such as `match 2/5 · +3 hidden` or `no matches · +1 hidden`, and the `f` hint now reads `show N hidden matches`.\nImproved `/diff`'s `Commit Tree` to mark the selected commit with the same `>` cursor and selection color as `All Changes` and `Per Turn` instead of a highlighted bar, and made `Per Turn` start with the cursor on the first file instead of its turn header."
            },
            {
              "title": {
                "zh-TW": "循環排程引入偏移時間差，避免整點邊界流量暴增",
                "en": "Recurring schedules introduce time offsets to prevent traffic spikes on hour boundaries"
              },
              "body": {
                "zh-TW": "若有多個背景任務排在整點同時啟動，瞬間湧入的連線很容易讓伺服器塞車。⟦跟社區大樓安排住戶分流一樣，以前所有單位全擠在整點出門容易在閘門大塞車；現在為每個單位分配固定的前後時間差，整點前後平順出發，需要分秒不差的專車依然能申請在準點出發。⟧由代理人透過 `schedule` 工具建立的循環 cron 排程，現在觸發時會依每個排程套用固定的偏移時間，不再剛好卡在整點邊界，有效分散尖峰流量。在一天內多次重複的任務（例如每 30 分鐘或每小時一次），會在各自週期內的特定時間點觸發。每天早上 9:00 等固定時間排程，依然會在預定的該分鐘內執行。若排程需要嚴格對準時刻，代理人也能要求精確時間，而 `@every` 間隔排程則維持原樣不變。",
                "en": "When multiple background tasks are scheduled to launch at the same time on the hour, an instant surge of connections easily congests the server. ⟦Just as a residential complex staggers departures for its tenants, having every household head out right on the hour previously caused massive gate congestion, but assigning a fixed offset window to each unit now smooths out traffic around the hour while reserved shuttles that require exact timing can still request sharp departure times.⟧ Recurring cron schedules created by an agent via the `schedule` tool now apply a fixed offset time to each schedule when triggered, no longer landing exactly on hour boundaries and effectively spreading out peak traffic. Tasks that repeat multiple times throughout the day (such as every 30 minutes or once an hour) trigger at specific points within their respective intervals. Fixed-time schedules like 9:00 AM every morning still execute within their scheduled minute. If a schedule requires strict time alignment, the agent can also request exact timing, while `@every` interval schedules remain unchanged."
              },
              "original": "Improved recurring cron schedules created with the agent's `schedule` tool to fire at a consistent per-schedule offset instead of exactly on the cron boundary, to avoid traffic spikes: schedules that repeat within a day, such as every 30 minutes or hourly, now fire somewhere within their period, fixed-time schedules such as daily at 9:00 still fire within their scheduled minute, and the agent can request exact timing when a schedule needs it. `@every` schedules are unchanged."
            },
            {
              "title": {
                "zh-TW": "降低閒置與執行資源消耗，長對話與啟動載入更流暢",
                "en": "Reduced idle and runtime resource consumption with smoother long conversations and startup loading"
              },
              "body": {
                "zh-TW": "長時間在終端機開啟工具或處理大型專案時，背景消耗的硬體資源現在明顯下降。\n\n當畫面沒有任何文字變動時，終端機會暫停螢幕重繪，讓閒置時的 CPU 使用率約減少一半。而在模型輸出串流回覆的過程中，工具會略過重複的畫面更新，並降低額度餘額的刷新頻率以節省運算效能。\n\n長篇對話的渲染效能有所改善，重新恢復長對話工作階段時的載入速度也提升。命令列介面啟動並刷新對話摘要時，CPU 開銷也一併降低。\n\n在記憶體管理方面，輸入提示詞與各類文字時改為打字時才動態配置儲存空間，不再預先保留整塊記憶體。閒置時的 CPU 消耗降低，長對話顯示與恢復載入也更快。",
                "en": "When keeping the tool open in the terminal for long periods or working on large projects, background hardware resource consumption has now noticeably decreased.\n\nWhen there are no text changes on screen, the terminal pauses screen repaints, cutting idle CPU usage roughly in half. Meanwhile, while the model streams output, the tool skips redundant screen redraws and lowers the refresh frequency of quota balances to save computing power.\n\nRendering performance for long conversations has improved, and loading speed when resuming long conversation sessions has also increased. When the command-line interface launches and refreshes conversation summaries, CPU overhead is similarly reduced.\n\nRegarding memory management, storage space for prompt inputs and various texts is now dynamically allocated during typing instead of pre-allocating entire memory blocks. Idle CPU consumption drops, and long conversation display as well as resumption loading become faster."
              },
              "original": "Improved idle CPU usage by pausing screen refreshes while nothing on screen changes, roughly halving the CLI's CPU use when it is idle.\nImproved CPU usage while a response streams by skipping redundant screen updates and refreshing the credit balance less often.\nImproved rendering speed for long conversations, and made long conversations load faster when you resume them.\nImproved startup CPU usage when the CLI refreshes its conversation summaries.\nImproved memory usage by allocating storage for the prompt and other text inputs as you type instead of reserving it up front."
            },
            {
              "title": {
                "zh-TW": "模型遭安全機制阻擋時停止無止盡重試",
                "en": "Stop infinite retries when the model is blocked by safety mechanisms"
              },
              "body": {
                "zh-TW": "遇上模型回覆被內容安全過濾機制阻擋的突發狀況，代理人過去容易陷入停不下來的重試迴圈。以前遇到回應被安全過濾器阻擋時，代理人會在背景靜默重試，甚至可能陷入無止盡的重試迴圈；現在這個問題已經修好，遇到阻擋時不再默默持續重試。",
                "en": "When encountering unexpected situations where a model response was blocked by content safety filters, agents previously tended to get trapped in an endless retry loop. In the past, when a response was blocked by safety filters, the agent would silently retry in the background and could even enter an infinite retry loop; this issue has now been resolved, and it no longer keeps silently retrying when blocked."
              },
              "original": "Fixed the agent silently retrying, potentially without end, when a model response was blocked by content safety filters."
            },
            {
              "title": {
                "zh-TW": "外掛設定支援以字串指定技能路徑，避免載入失敗",
                "en": "Plugin configuration supports specifying skill paths as strings to prevent loading failures"
              },
              "body": {
                "zh-TW": "想載入為其他工具設計的外掛程式時，設定檔格式只要稍有不同就會直接報錯。以前如果外掛的 `plugin.json` 將 `skills` 欄位寫成單一路徑字串（例如 `\"skills\": \"./skills/\"`）而非陣列清單，外掛就會載入失敗。現在這個問題已經修好，工具能正確辨識單一路徑字串並完成載入。",
                "en": "When trying to load plugins designed for other tools, any slight difference in configuration format would directly trigger an error. Previously, if a plugin's `plugin.json` specified the `skills` field as a single path string (such as `\"skills\": \"./skills/\"`) instead of an array list, the plugin failed to load. This issue has now been resolved, and the tool correctly recognizes a single path string and completes loading."
              },
              "original": "Fixed plugins failing to load when `plugin.json` gives `skills` as a single path string, such as `\"skills\": \"./skills/\"`, instead of an array, as some plugins written for other tools do."
            },
            {
              "title": {
                "zh-TW": "市集外掛安裝與更新遇網路中斷自動重試一次",
                "en": "Marketplace plugin installation and updates automatically retry once on network interruption"
              },
              "body": {
                "zh-TW": "從市集安裝或自動更新外掛程式時，偶發的網路震盪容易造成程序中斷。以前下載內容短暫過時或下載中斷時，市集外掛安裝與自動更新可能因校驗碼不符或網路錯誤而失敗；現在命令列介面會自動重試下載一次。",
                "en": "When installing or automatically updating plugins from the marketplace, occasional network disruptions easily interrupted the process. Previously, when downloaded content was temporarily stale or downloads were interrupted, marketplace plugin installations and automatic updates could fail due to checksum mismatches or network errors; the command-line interface now automatically retries the download once."
              },
              "original": "Fixed marketplace plugin installs and automatic updates failing with a checksum mismatch or network error when a download was briefly stale or interrupted; the CLI now retries the download once."
            }
          ]
        },
        {
          "version": "1.4.4",
          "period": "2026-10-21",
          "items": [
            {
              "title": {
                "zh-TW": "Auto 模式登陸 Windows 11，加強沙盒外高風險指令審查",
                "en": "Auto Mode Arrives on Windows 11 with Stricter Out-of-Sandbox Command Review"
              },
              "body": {
                "zh-TW": "在 Windows 11 24H2 或更新版本上使用個人 Google 帳戶時，現在你可以體驗能自行決定工具呼叫的 Auto 模式。⟦如同具備主動防護的車輛巡航系統，以前只照既定速限行駛而難以應對突發變換，現在遇到危險路況會主動放慢並要求駕駛接手確認。⟧你可以透過 `--mode auto` 參數、在 `/config` 設定選單的 Agent Mode 選取、按下快捷鍵 `Shift+Tab`，或是開啟 Remote Control 設定進行切換，預設選項仍維持為 Agent 模式。當 Auto 模式審查即將脫離終端機沙盒（sandbox，隔離指令執行的安全環境）的指令時，防護更為嚴謹。面對管線指令會逐一檢查其中的每個子指令；若是 `yes` 等簡短回覆，會連同助理的前一個提問一併納入評估，並在執行具破壞性的 Git 或檔案操作前主動跳出提示確認。",
                "en": "When running on Windows 11 24H2 or later with a personal Google account, you can now experience Auto mode, which decides tool calls on its own. ⟦Picture a car cruise control system with active safety: where it once only held a preset speed and struggled with sudden shifts, it now slows down on hazardous roads and prompts the driver to take over.⟧ You can switch modes via the `--mode auto` flag, by selecting Agent Mode in the `/config` settings menu, by pressing `Shift+Tab`, or by enabling Remote Control settings, while Agent mode remains the default. When Auto mode reviews commands about to leave the terminal sandbox (an isolated security environment for command execution), safeguards are significantly stricter. For pipeline commands, it inspects every sub-command individually; for short replies like `yes`, it evaluates them alongside the assistant's preceding question, prompting for confirmation before running destructive Git or file operations."
              },
              "original": "Added Auto mode on Windows 11 version 24H2 and later for personal Google Accounts, selectable via `--mode auto`, `Agent Mode` in `/config`, `Shift+Tab`, or Remote Control Settings (`Agent` remains the default).\nImproved Auto mode's review of commands that leave the terminal sandbox to include the assistant's preceding question on short replies like `yes`, check every subcommand in pipelines, and prompt before destructive Git or file commands."
            },
            {
              "title": {
                "zh-TW": "/diff 支援區塊折疊與整檔切換，修正清單換行與標題截斷",
                "en": "/diff Adds Hunk Folding and Full File Toggle, Fixing List Wrapping and Title Truncation"
              },
              "body": {
                "zh-TW": "在終端機輸入 `/diff` 檢視程式碼變更時，Commit Tree 與 Per Turn 兩種檢視模式新增了多項閱讀控制。⟦跟查閱裝訂成冊的工程藍圖一樣，以前只能整頁攤開翻找細節，現在可以隨手折疊次要章節，還能把同一次施工的多處修改彙整在同一份清單。⟧現在按下快捷鍵 `f` 可以切換顯示整份檔案，按下 `Ctrl+O` 則能折疊個別變更區塊（hunk，版本控制中連續的一段程式碼修改）。在 Per Turn 檢視中，同一輪對話內對同一檔案進行的多次編輯，會直接合併為單一比對畫面，不再被拆得零散。清單顯示的排版問題也已修復，清單中超出終端機寬度的長行會正常截斷而不再強行折行，Per Turn 頂部標題也不會再把提示詞截斷在 50 個字元。",
                "en": "When typing `/diff` in the terminal to inspect code changes, Commit Tree and Per Turn views add several reading controls. ⟦Like consulting a bound set of engineering blueprints where you previously had to spread out entire pages to search for details, you can now collapse secondary sections at will and consolidate edits from the same job into a single list.⟧ Pressing `f` now toggles the full file view, while `Ctrl+O` folds an individual hunk (a contiguous section of code changes in version control). In Per Turn view, multiple edits made to the same file within a single turn merge directly into a unified diff screen instead of being scattered. List formatting issues are also fixed: lines exceeding terminal width now truncate cleanly rather than wrapping forcefully, and Per Turn headers no longer truncate prompts at 50 characters."
              },
              "original": "Added whole-file toggle (`f`) and hunk folding (`Ctrl+O`) to `Commit Tree` and `Per Turn` views in `/diff`, and consolidated `Per Turn` diffs so multiple edits to a file in one turn appear as a single combined diff.\nFixed long rows wrapping instead of truncating at the terminal width in `/diff`'s `Commit Tree` and `Per Turn` lists, and `Per Turn` headers truncating prompts at 50 characters."
            },
            {
              "title": {
                "zh-TW": "MCP 工具呼叫支援追溯最上層對話，清理伺服器連線洩漏",
                "en": "MCP Tool Calls Support Root Conversation Tracing, Fixing Server Connection Leaks"
              },
              "body": {
                "zh-TW": "當你使用 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）串接自訂伺服器時，子代理呼叫工具的追蹤能力有了重大改善。⟦就像委託跨國快遞寄件，以前轉交給各區投遞員後就查不到源頭，現在每筆派工紀錄都烙印總單號，無論轉發幾手都能追溯最初的寄件委託。⟧現在即使工具是由深層的巢狀子代理所觸發，發出的 `tools/call` 請求都會自動帶上 `_meta[\"antigravity.google/root_conversation_id\"]` 欄位，讓外部伺服器能辨識最頂層的對話識別碼。在伺服器連線與資源釋放方面，先前若在伺服器啟動期間執行 `tools/list` 失敗，或是在連線尚未完成時直接關閉命令列工具，背景連線常式與用戶端工作階段容易殘留。這兩種情況造成的用戶端工作階段與背景連線常式洩漏，現在已修好。",
                "en": "When connecting custom servers with MCP (Model Context Protocol, a protocol that connects AI to external tools), tracking tool calls made by subagents is substantially improved. ⟦Think of shipping a package through an international courier: previously, handoffs to regional delivery agents obscured the origin, but now every dispatch record stamps a master tracking number so the initial request remains traceable across any number of transfers.⟧ Even when tools are invoked by deeply nested subagents, outgoing `tools/call` requests now automatically include the `_meta[\"antigravity.google/root_conversation_id\"]` field so external servers can identify the top-level conversation ID. In terms of server connections and resource cleanup, background connection routines and client sessions previously lingered if `tools/list` failed during startup or if the CLI closed before the connection finished. Leaks of client sessions and background connection routines caused by both scenarios are now resolved."
              },
              "original": "Added `_meta[\"antigravity.google/root_conversation_id\"]` to MCP `tools/call` requests so MCP servers receive the top-level conversation ID even when a tool is called from a nested subagent.\nFixed MCP client sessions and background connection routines leaking when `tools/list` failed during server startup or when the CLI shut down while a server was connecting."
            },
            {
              "title": {
                "zh-TW": "外掛程式驗證加強嚴格檢查，修復子代理繼承與技能重複載入",
                "en": "Plugin Validation Enforces Stricter Checks, Fixing Subagent Inheritance and Skill Duplication"
              },
              "body": {
                "zh-TW": "在開發或安裝自訂擴充套件時，外掛程式的驗證機制與執行環境獲得了多項修正。\n\n指令 `agy plugin validate` 導入更嚴格的檢查，全面檢驗 `plugin.json`、MCP 伺服器設定、`hooks.json` 與 sidecar 清單，防範無效 JSON 語法、未知欄位、錯誤變數佔位符、無法辨識的掛鉤（hook，事件觸發時執行的自動腳本）工具比對器及遺失腳本。\n\n子代理方面，修復了子代理遺失父工作階段繼承外掛的問題；若打包外掛的子代理設定 `plugins.inherit_user: false` 且限制 MCP 工具，先前啟動拋出 `mcp spec not found for provider` 的異常也已修復。\n\n當自訂規則或 `SKILL.md` 技能同時存在於獨立客製化資料夾與已啟用外掛程式時，系統不再重複載入兩次。",
                "en": "When developing or installing custom extensions, plugin validation mechanisms and execution environments received multiple fixes.\n\nThe `agy plugin validate` command introduces stricter checks that thoroughly inspect `plugin.json`, MCP server configurations, `hooks.json`, and sidecar manifests, guarding against invalid JSON syntax, unknown fields, malformed variable placeholders, unrecognized tool matchers for hooks (automated scripts triggered by specific events), and missing scripts.\n\nFor subagents, an issue where subagents lost plugins inherited from parent sessions has been resolved. If a packaged plugin subagent sets `plugins.inherit_user: false` and restricts MCP tools, the startup error throwing `mcp spec not found for provider` is also fixed.\n\nWhen custom rules or `SKILL.md` skills exist simultaneously in standalone customization folders and enabled plugins, the system no longer loads them twice."
              },
              "original": "Improved `agy plugin validate` with stricter checks across `plugin.json`, MCP server configs, `hooks.json`, and sidecar manifests for invalid JSON syntax, unknown fields, bad variable placeholders, unrecognized hook tool matchers, and missing hook scripts.\nFixed subagents losing plugins inherited from the parent session, and fixed plugin-bundled subagents with `plugins.inherit_user: false` and restricted MCP tools failing to start with `mcp spec not found for provider`.\nFixed rules and `SKILL.md` skills loading twice when discovered from both a standalone customization folder and an enabled plugin."
            },
            {
              "title": {
                "zh-TW": "串流回應暫存降低資源消耗，壓縮請求保留工具宣告重用快取",
                "en": "Stream Response Caching Cuts Resource Overhead, Compaction Retains Tool Declarations"
              },
              "body": {
                "zh-TW": "當你請模型生成長篇回應或修改大量程式碼時，回應串流與對話壓縮的快取（cache，把處理過的結果暫存起來重複使用）機制減輕了運算負擔。⟦好比在料理檯旁備妥常備菜，已切洗好的食材直接下鍋而不重複動刀，收拾廚房時也把標準菜單留在黑板上，省去重新謄寫核對的時間。⟧在文字串流輸出期間，終端機現在會跨影格暫存程式碼區塊的語法標記（token，程式碼結構的分析記號）、算繪後的 Markdown 區塊，以及行內檔案編輯的確認比對畫面，降低 CPU 與記憶體使用量。而在執行 `/compact` 指令或系統觸發自動壓縮時，對話摘要請求會完整保留工具宣告，讓模型 API 能順暢重複使用先前已快取的提示詞前綴。",
                "en": "When asking models to generate lengthy responses or modify extensive code, caching mechanisms for response streaming and conversation compaction reduce computing overhead. ⟦Imagine keeping prepped ingredients ready by the kitchen counter, where washed and chopped items drop straight into the pot without re-cutting, while standard menus stay on the chalkboard during cleanup to save time re-copying.⟧ During streamed text output, the terminal now caches code-block syntax tokens (structural analysis units for code) across frames, along with rendered Markdown blocks and inline file edit confirmation diffs, lowering CPU and memory usage. When running `/compact` or when the system triggers automatic compaction, conversation summary requests preserve tool declarations intact, allowing model APIs to smoothly reuse previously cached prompt prefixes."
              },
              "original": "Improved CPU and memory usage while streaming responses by caching streamed code-block syntax tokens, rendered Markdown blocks, and inline file-edit confirmation diffs across frames.\nImproved `/compact` and automatic compaction to retain tool declarations on summarization requests so the model API reuses its cached prompt prefix."
            },
            {
              "title": {
                "zh-TW": "Remote Control 支援平板多欄排版，修正行動端介面與算繪錯誤",
                "en": "Remote Control Adds Multi-Column Layout on Tablets, Fixing Mobile UI and Rendering Bugs"
              },
              "body": {
                "zh-TW": "在平板與行動裝置上使用 Remote Control 遠端控制時，大螢幕排版與多項操作體驗獲得整體升級。\n\n平板與摺疊螢幕裝置不再強制縮成單欄手機畫面，而是啟用具備觸控尺寸控制項的多欄排版，拖曳捲軸時彈出的控制面板（popover）也會持續保持開啟。\n\n行動端的介面細節也獲得修正。手機上標題與對話框遭截斷、iOS Safari 點選輸入框時畫面被自動放大、設定載入中 Enable Telemetry 與 Marketing Emails 開關自行跳回，以及拖曳網址時遺失查詢參數或井字標記（#）的問題已全數排除。\n\n輸出算繪與檔案匯出同樣更為穩定。先前巢狀程式碼區塊排版破裂、帶錢字號的 Shell 變數（$VAR）被誤判為 LaTeX 公式、媒體預覽無法載入、複製圖片或匯出 PDF 時圖表模糊或空白，以及背景分頁計時器偏差等現象，現在都已修正完畢。",
                "en": "When using Remote Control on tablets and mobile devices, large-screen layouts and various operational experiences received an overall upgrade.\n\nTablets and foldable devices are no longer forced into a single-column phone view; instead, they enable a multi-column layout with touch-sized controls, and popover panels remain open while dragging scrollbars.\n\nMobile interface details have also been refined. Issues where titles and dialogs truncated on phones, tapping input fields in iOS Safari triggered automatic zoom, the Enable Telemetry and Marketing Emails toggles flipped back while settings loaded, and dragging URLs lost query parameters or hashes (#) are all resolved.\n\nOutput rendering and file exports are likewise more stable. Prior issues with broken nested code block layouts, Shell variables ($VAR) being misidentified as LaTeX formulas, media previews failing to load, blurry or blank charts when copying images or exporting PDFs, and background tab timer drift have all been fixed."
              },
              "original": "Improved Remote Control on tablets and foldables to use the multi-pane layout with touch-sized controls instead of collapsing to the phone view, and kept popovers open while dragging scrollbars.\nFixed several UI issues in Remote Control, including clipped headers and dialogs on phones, iOS Safari zoom on input focus, `Enable Telemetry` and `Marketing Emails` reverting while Settings loads, and dragged URLs losing query or `#` parameters.\nFixed rendering and export bugs in Remote Control, including broken nested code blocks, `$VAR` shell variables parsed as LaTeX, broken media previews, blurry or blank diagrams in `Copy Image` and `Export as PDF`, and background-tab timer drift."
            },
            {
              "title": {
                "zh-TW": "終端機輸入框回歸底部定位，修復 Ctrl+Z 復原與表情符號截斷",
                "en": "Terminal Input Box Re-Anchors to Bottom, Fixing Ctrl+Z Undo and Emoji Truncation"
              },
              "body": {
                "zh-TW": "在終端機中與模型對話時，輸入框的排版位置與文字復原行為迎來了重要修正。⟦正如在紙上書寫與擦拭草稿，以前橡皮擦常常擦掉多餘的字跡甚至弄縐紙張，現在每一次復原動作都能精準撤回上一筆紀錄。⟧先前當你關閉面板、下拉選單或行內提示時，輸入框有時會卡在螢幕中間不回彈，在 Medium 詳細度模式下輸入框上方還會出現空白縫隙；現在這些排版異常都已修復，輸入列會穩定貼齊底部。文字編輯方面，快捷鍵 Ctrl+Z 的復原機制已重新調整。過去貼上文字時復原容易錯亂，新輸入的內容甚至會與歷史紀錄錯誤合併。現在無作用的退格鍵與 Delete 不再被記錄為多餘步驟，在表情符號（emoji）旁執行刪除或復原時，也不會再發生字元遭到截斷的狀況。",
                "en": "When chatting with models in the terminal, the input box layout position and text undo behavior received major fixes. ⟦A writer sketching and erasing drafts on paper once risked smudging extra marks or wrinkling the sheet, but now each undo step accurately retracts the previous entry.⟧ Previously, closing panels, dropdown menus, or inline prompts could leave the input box stuck midway up the screen, and a blank gap appeared above the input box in Medium verbosity mode; these layout anomalies are now resolved, keeping the input bar securely anchored to the bottom. For text editing, the Ctrl+Z undo mechanism has been retuned. Pasting text previously scrambled undo history, sometimes incorrectly merging newly typed content with older history. Ineffective Backspace and Delete keystrokes are no longer recorded as extra steps, and deleting or undoing next to emojis (pictographic symbols) no longer causes characters to be truncated."
              },
              "original": "Fixed the prompt box staying mid-screen after closing panels, dropdowns, or inline prompts, and blank gaps appearing above the prompt in `Medium` verbosity mode.\nFixed `Ctrl+Z` undo in the prompt mishandling pasted text, merging typing with recalled history entries, recording no-op `Backspace` or `Delete` steps, and truncating characters near emoji."
            },
            {
              "title": {
                "zh-TW": "修正規則優先順序措辭，明確涵蓋前方內建指示",
                "en": "Clarify Rule Precedence Phrasing to Explicitly Cover Earlier Built-In Instructions"
              },
              "body": {
                "zh-TW": "當你在專案根目錄撰寫自訂規範時，可以留意系統提示詞對規則優先順序的說明。以前系統提示詞（system prompt，引導 AI 行為的核心指示）的措辭，只寫出 GEMINI.md 與 AGENTS.md 的使用者規則可覆寫排在它們後方的指示。現在系統提示詞明確寫出 GEMINI.md 與 AGENTS.md 規則可覆寫較早出現的內建指示。",
                "en": "When writing custom rules in project roots, note the updated system prompt phrasing for rule precedence. Phrasing in the system prompt (core instructions guiding AI behavior) previously only stated that user rules in GEMINI.md and AGENTS.md could override instructions placed after them. The system prompt now explicitly clarifies that GEMINI.md and AGENTS.md rules can override earlier built-in instructions."
              },
              "original": "Fixed the system prompt's user rules wording so `GEMINI.md` and `AGENTS.md` rules explicitly override earlier built-in instructions instead of only instructions that follow them."
            },
            {
              "title": {
                "zh-TW": "支援帶編碼參數的媒體附件，避免音訊檔案遭錯誤拒絕",
                "en": "Support Media Attachments with Codec Parameters, Preventing Erroneous Audio Rejections"
              },
              "body": {
                "zh-TW": "當你在終端機對話中上傳音訊或多媒體檔案時，附件格式的相容性獲得了修正。以前若媒體附件的 MIME 類型（標示附件內容格式的類型名稱）帶有額外參數，例如常見的音訊格式 audio/webm; codecs=opus，會被誤判為不支援而直接遭到拒絕。這類原本受支援的附件，現在不會再因 MIME 類型帶有參數而被誤拒。",
                "en": "When uploading audio or multimedia files in terminal conversations, attachment format compatibility has been fixed. Previously, if a media attachment's MIME type (a label identifying attachment content formats) included extra parameters, such as the common audio format audio/webm; codecs=opus, it was incorrectly flagged as unsupported and rejected. Such natively supported attachments are no longer falsely rejected due to parameterized MIME types."
              },
              "original": "Fixed supported media attachments with parameterized MIME types such as `audio/webm; codecs=opus` being rejected as unsupported."
            }
          ]
        }
      ]
    },
    {
      "id": "grok_build",
      "name": "Grok Build",
      "versions": [
        {
          "version": "1.0.62",
          "period": "2026-10-16",
          "items": [
            {
              "title": {
                "zh-TW": "無頭模式等待背景工作接續對話，逾時中斷改回傳錯誤",
                "en": "Headless mode waits for background follow-up turns and exits with error on timeout"
              },
              "body": {
                "zh-TW": "當你寫腳本透過指令列自動處理任務時，最怕背景工作還沒做完程式就提早離場。使用 `grok -p` 啟動無頭模式（沒有終端機畫面的執行模式）時，若已完成的背景工作觸發了接續對話，現在會耐心等待該輪對話結束後才退出。⟦如同自助洗衣機的運轉流程，以前主行程一結束就立刻切斷電源，不管追加的防皺運轉跑完了沒有；現在機器會等接續動作徹底收尾才關機，要是設定的時間上限強行中斷了未完工項，面板更會亮起錯誤紅燈而不是回報順利完成。⟧如果加上 `--background-wait-timeout` 參數且因逾時停止了待處理工作，程式結束時會回傳錯誤代碼，不再誤報為成功執行。",
                "en": "When scripting automated tasks through the command line, the biggest risk is a process exiting prematurely before background work finishes. When launching headless mode (an execution mode with no terminal screen) with `grok -p`, if completed background tasks trigger follow-up conversation turns, the runner now waits patiently for that turn to conclude before exiting. ⟦Like the cycle of a laundromat washing machine, the power previously cut off the moment the main cycle finished, regardless of whether the extra anti-wrinkle cycle had completed; now the machine shuts down only after follow-up actions wrap up completely, and if a configured time limit forcibly stops unfinished work, the panel flashes a red error light instead of reporting a successful run.⟧ If you provide the `--background-wait-timeout` option and pending work is stopped due to a timeout, the program now exits with an error code instead of falsely reporting success."
              },
              "original": "Headless mode (`grok -p`) waits for the follow-up turn that a finished background task starts before exiting, and a `--background-wait-timeout` that stops pending work exits with an error instead of success"
            },
            {
              "title": {
                "zh-TW": "儀表板冷啟動恢復工作階段，保持浮層開啟並支援按鍵返回",
                "en": "Dashboard cold boot resumes sessions while keeping popover open with keyboard back navigation"
              },
              "body": {
                "zh-TW": "你在儀表板翻找過去的歷史紀錄時，可能會想先點進某個舊存檔看看進度。在儀表板浮層中針對未預載的存檔執行冷啟動 `/resume` 時，介面現在會讓浮動面板保持附著開啟。⟦彷彿在電子導覽機前點選歷史路徑，以前只要觸控未預載的深層檔案，選單面板就會冷不防關閉脫離，讓人想反悔返回都無從按起；現在清單畫面會持續固定在視野中，隨時都能按上一頁安全退回。⟧如此一來，載入舊工作階段時不會丟失操作畫面，你可以隨時按下向左方向鍵（Left）或 `Ctrl+\\` 順利退出，返回原本的清單繼續瀏覽。",
                "en": "When browsing past history in the dashboard, you might want to open an older saved state to inspect its progress. Triggering a cold `/resume` on an unloaded session from the dashboard popover now keeps the floating panel open and attached. ⟦At a digital directory kiosk, tapping an unloaded deep record previously caused the menu panel to snap shut and detach unexpectedly, leaving no way to step back; now the list display stays anchored in view so you can safely return at any moment.⟧ This way, you never lose your context when loading an older session, and you can press the Left arrow key (Left) or `Ctrl+\\` at any time to exit cleanly and return to the list."
              },
              "original": "Cold `/resume` from the dashboard overlay keeps the overlay attached, so you can still back out with Left or `Ctrl+\\`"
            },
            {
              "title": {
                "zh-TW": "選取帶參數斜線指令按 Enter 直接執行，不再殘留空格",
                "en": "Selecting slash commands with arguments and pressing Enter executes immediately without trailing spaces"
              },
              "body": {
                "zh-TW": "在終端機輸入框使用斜線指令時，大家常會透過選單挑選想跑的動作。當你在提示詞撰寫區反白選取帶有可選參數的斜線指令並按下 Enter 鍵時，現在會立刻執行該指令。以前輸入框（composer，撰寫提示詞的文字區）只會在指令後面留下一格多餘的空格，停在原地等待你再次確認。⟦好比在車載導航螢幕上選取附帶途經點的路線，以前按下確認鈕，畫面只會在輸入欄末端留下一格空白等候；現在一按便直接開始指引，不再卡在輸入框等待多餘的操作。⟧這樣不必在選好指令後還得手動多敲一次鍵，命令一選定就能順暢送出。",
                "en": "When using slash commands in the terminal input box, users often pick actions from the menu. Selecting a slash command with optional arguments in the prompt area and pressing Enter now executes that command immediately. Previously, the composer (the text area for writing prompts) only appended an extra space after the command, sitting idle while waiting for your confirmation. ⟦On an in-car navigation screen, selecting a route with waypoints and hitting confirm previously only left a blank space at the end of the input field; now a single press starts guidance immediately instead of stalling for redundant input.⟧ You no longer need to press an extra key manually after choosing a command, allowing selected actions to submit smoothly."
              },
              "original": "Enter on a highlighted slash command with optional arguments runs the command instead of leaving a trailing space in the composer"
            },
            {
              "title": {
                "zh-TW": "多項 CLI 指令支援管道提早關閉與串流限制，避免輸出異常崩潰",
                "en": "Multiple CLI commands handle early pipe closures and stream limits without crashing"
              },
              "body": {
                "zh-TW": "平時在終端機串接多道指令過濾資料時，下游工具提早結束往往會引發非預期的中斷。包含 `grok models`、列出工作階段的 `grok sessions` 與 `grok export` 等指令，當輸出管道提早關閉（例如串接 `head` 只取前幾行）或輸出裝置已滿時，現在都不再發生當機。⟦就像工廠輸送帶上的分裝包裝線，以前只要下游集貨箱裝滿或接收閘門提早關閉，傳送機械便會當場卡死並引發整條產線警報；現在機台一旦察覺出口無法收件，就會平順停止輸出並安穩關機。⟧`grok update` 與 `grok login` 兩項指令在標準輸出（stdout）或標準錯誤關閉且無法寫入時，現在同樣能安全結束，確保批次腳本不再因串流截斷而崩潰。",
                "en": "When piping multiple commands in the terminal to filter data, a downstream tool exiting early often triggers unexpected disruptions. Commands including `grok models`, `grok sessions` for listing sessions, and `grok export` no longer crash when their output pipe closes early (such as when piped into `head` for the first few lines) or when the output device fills up. ⟦In a factory conveyor packaging line, if a downstream collection bin filled up or a receiving gate closed early, the transport machinery previously jammed on the spot and triggered plant-wide alarms; now the equipment halts output smoothly and shuts down safely as soon as it detects that the exit cannot accept packages.⟧ `grok update` and `grok login` now also exit safely when standard output (stdout) or standard error is closed and unwritable, ensuring batch scripts no longer crash due to truncated streams."
              },
              "original": "CLI commands such as `grok models`, `grok sessions` and `grok export` no longer crash when their output pipe closes early, such as when piped to `head`, or when the output device is full\n`grok update` and `grok login` no longer crash when stdout or stderr is closed or unwritable"
            }
          ]
        },
        {
          "version": "1.0.61",
          "period": "2026-10-16",
          "items": [
            {
              "title": {
                "zh-TW": "/resume 精確比對單一工作階段直接開啟，選單支援任意詞序搜尋",
                "en": "/resume Directly Resumes an Exact Single-Session Match and Supports Any-Word-Order Search in the Picker"
              },
              "body": {
                "zh-TW": "當你累積了大量對話紀錄想重新接續進度時，不必再忍受死板的關鍵字順序或多餘的確認步驟。⟦翻找辦公室抽屜裡的專案夾時，以前封面標題上的單字順序只要顛倒就翻不出來，就算報出唯一編號也非得拉開整個清單；現在只要寫齊關鍵字無論順序都能對應，代碼剛好只吻合一份時更能直接抽取出夾。⟧現在在終端機輸入 `/resume <id or name>` 時，只要指定的 ID 或標題剛好只吻合單一工作階段，Grok Build 就會直接進入該工作階段，不再每次都強制彈出挑選器讓你多點一次。如果呼叫挑選器手動尋找，只要輸入的內容包含標題中的每一個單字，無論先後順序為何，挑選器都能精確列出符合的工作階段。",
                "en": "When you have accumulated a large conversation history and want to resume progress, you no longer have to endure rigid keyword ordering or redundant confirmation steps. ⟦Picture someone searching for a project folder in an office drawer. Previously, if the word order on the cover was reversed, it could not be found, and reciting a unique number still required opening the entire list. Now, entering all keywords matches regardless of sequence, and matching exactly one code lets you pull the folder straight out.⟧ Now when you enter `/resume <id or name>` in the terminal, as long as the specified ID or title matches exactly one session, Grok Build directly enters that session instead of forcing a picker popup every time for an extra click. If you bring up the picker to search manually, the picker accurately lists matching sessions as long as your input includes every word in the title, regardless of their order."
              },
              "original": "Resume picker finds sessions whose titles contain every word you type, in any order\n`/resume <id or name>` opens the session directly when the ID or title matches exactly one session, instead of always opening the picker"
            },
            {
              "title": {
                "zh-TW": "儀表板預覽文字支援滑鼠拖曳，放開即自動複製到剪貼簿",
                "en": "Dashboard Preview Text Supports Mouse Dragging, Automatically Copying to the Clipboard on Release"
              },
              "body": {
                "zh-TW": "想要從儀表板快速擷取一段預覽內容轉貼給同事時，現在不需要再按任何鍵盤快捷鍵。⟦拿著螢光筆在文件上劃記重點時，以前畫完標記還得手動按相機快門留底；現在筆尖劃過文字並在放開筆桿時，內容就自動傳送至隨身筆記。⟧在儀表板檢視預覽文字時，你可以直接按住滑鼠游標拖曳選取想要的字串。只要在選取完成後放開滑鼠按鍵，被標記的文字片段就會自動複製到剪貼簿。無論是需要摘錄報錯內容還是比對執行結果，拖曳選取後就能直接切換到其他對話視窗貼上，省下反白後還要按鍵複製的瑣碎手續。",
                "en": "When you want to quickly grab preview content from the dashboard to paste for a colleague, you no longer need to press any keyboard shortcuts. ⟦Imagine marking highlights on a document with a highlighter. Previously, after finishing the mark, you still had to press a camera shutter manually to keep a record; now, as the pen tip glides across the text and releases, the content is transmitted automatically to a pocket notebook.⟧ While viewing preview text on the dashboard, you can click and drag the mouse cursor to select the desired string. As soon as you release the mouse button after completing the selection, the marked text snippet is automatically copied to the clipboard. Whether you need to extract error details or compare execution results, you can paste directly into another chat window right after dragging to select, saving the tedious step of pressing keys to copy after highlighting."
              },
              "original": "Dashboard peek text can be selected by dragging over it with the mouse, and the selection is copied to the clipboard when you release"
            },
            {
              "title": {
                "zh-TW": "工作流程檢視向下滾動名冊時，畫面持續朝較新代理人平順推進",
                "en": "Workflow View Smoothly Advances Toward Newer Agents When Scrolling Down the Roster"
              },
              "body": {
                "zh-TW": "追蹤多個代理人協同運作的工作流程時，檢視進度最怕視窗畫面不聽使喚。⟦順著排隊走道向前推動檢視鏡頭時，以前只要滾輪往下滾動，視野就冷不防被彈簧拉回起點；現在畫面會依循操作，順暢朝著後方新加入的成員推進。⟧在工作流程的執行檢視畫面中，當你在代理人名冊上向下滾動滑鼠滾輪時，畫面現在會順著操作方向持續朝向較新的代理人捲動，而不是突然彈回原處。這讓你在依序盤點每個代理人的最新執行進度時，視野能夠一路平穩跟隨，不必反覆重新尋找焦點。",
                "en": "When tracking a workflow where multiple agents collaborate, the biggest frustration when checking progress is an uncooperative viewport. ⟦Think of tracking an inspection camera forward along a queue corridor. Previously, whenever the scroll wheel moved downward, the view was abruptly pulled back to the starting point like a spring; now, the frame follows the action, smoothly advancing toward newly joined members at the rear.⟧ In the workflow execution view, when you scroll down on the agent roster with the mouse wheel, the view now continuously scrolls toward newer agents in the direction of your movement, rather than snapping back to the origin. This allows your line of sight to smoothly follow along as you inspect each agent's latest execution progress in sequence, without having to repeatedly hunt for focus."
              },
              "original": "Workflow run views scroll toward newer agents when you wheel down the agent roster, instead of snapping back"
            },
            {
              "title": {
                "zh-TW": "Windows 執行掛鉤失敗時點名未啟動的程式，不再誤指指令碼出錯",
                "en": "Windows Identifies Unstarted Programs When Hook Execution Fails, Rather Than Blaming the Script"
              },
              "body": {
                "zh-TW": "在 Windows 環境下使用掛鉤（在特定事件發生時被觸發的小程式）除錯自動化流程時，最怕看到語意模糊的錯誤訊息。⟦呼叫外勤車隊出門執行任務時，以前只要外包貨車在門口無法發動，控制台卻只會回報派工單寫錯；現在通報會明確點名是哪一輛車發動失敗，不再冤枉原本寫好的任務清單。⟧在 Windows 系統中若掛鉤產生行程失敗，終端機現在會直接點名究竟是哪個外部程式無法啟動，例如具體標示出 PowerShell 或 Git Bash 發生錯誤，而不是怪罪掛鉤指令稿本身。面對這類問題時，你能立刻看出是環境缺少執行工具還是路徑設定遺漏，省下逐行檢查指令碼內容的時間。",
                "en": "When debugging automated workflows on Windows using hooks (small programs triggered by specific events), the worst experience is seeing ambiguous error messages. ⟦Dispatching a field fleet for a mission previously meant that whenever an outsourced truck failed to start at the gate, the control center only reported an error in the work order; now, the alert clearly names which vehicle failed to start without falsely blaming the original task list.⟧ If a hook fails to spawn a process on Windows, the terminal now directly names which external program failed to start, such as specifically indicating an error in PowerShell or Git Bash, rather than blaming the hook script itself. When confronting such issues, you can immediately tell whether the environment lacks an execution tool or is missing a path configuration, saving time spent inspecting the script line by line."
              },
              "original": "Hook spawn failures on Windows name the program that failed to start, such as PowerShell or Git Bash, instead of blaming the hook script"
            }
          ]
        },
        {
          "version": "1.0.60",
          "period": "2026-10-15",
          "items": [
            {
              "title": {
                "zh-TW": "grok --resume 依更新時間排序工作階段，找不到本機紀錄時即刻回報",
                "en": "grok --resume sorts sessions by update time and immediately reports when local records are missing"
              },
              "body": {
                "zh-TW": "當你累積了大量對話想重新接續先前的工作進度，只憑模糊印象輸入名稱往往難以精準定位。⟦如同在圖書館櫃檯查詢借閱紀錄，以前一旦館內找不到藏書，館方電腦便主動發起外館調閱並要求重新刷卡驗證；現在螢幕直接按最後更新時間由近到遠條列項目，若沒有相符紀錄還會提示相近名稱，找不到時立即告知結果。⟧\n\n執行 `grok --resume <title>` 恢復對話時，列出的相符工作階段會依據最後更新時間由新到舊排序，並清楚附上各自的最近更新時間。要是輸入的名稱完全沒有符合的紀錄，命令列會主動推薦相似的標題供你挑選。\n\n當指定的標題在本地電腦中完全不存在時，命令列現在會立即回報未找到，而不是嘗試發起遠端還原程序並要求你登入帳號。",
                "en": "When you have accumulated many conversations and want to resume earlier progress, typing a name from vague memory often makes precise targeting difficult. ⟦Think of checking borrowing records at a library counter. Previously, whenever a book was missing from the collection, the computer initiated an interlibrary loan and required card reverification; now, the screen lists items from newest to oldest by last update time, suggests similar names if no record matches, and reports the result immediately when nothing is found.⟧\n\nWhen running `grok --resume <title>` to resume a conversation, matching sessions are listed in order from newest to oldest based on their last update time, clearly displaying their respective recent update times. If the entered name has no matching records at all, the command line recommends similar titles for you to choose from.\n\nWhen the specified title does not exist at all on the local machine, the command line now immediately reports that it was not found, instead of attempting to initiate a remote restore process and asking you to log in."
              },
              "original": "`grok --resume <title>` lists matching sessions newest first with when each was last updated, and suggests similar titles when nothing matches\n`grok --resume <title>` with a title that matches no local session reports the miss right away instead of trying a remote restore and asking you to sign in"
            },
            {
              "title": {
                "zh-TW": "全螢幕 Markdown 檢視器按 Shift+Y 複製游標所在的程式碼區塊",
                "en": "Full-screen Markdown viewer copies the code block under the cursor with Shift+Y"
              },
              "body": {
                "zh-TW": "閱讀長篇技術文件或檢視終端機輸出時，切換到全螢幕檢視器能讓排版更為乾淨，但想取用其中的範例段落往往需要額外步驟。⟦彷彿拿著手持掃描器在貨架間盤點，以前想記錄規格得拿出紙筆沿著邊界逐字謄寫，現在光點只要對準商品標籤扣下按鍵，整組資料便立刻收進隨身清單。⟧\n\n在全螢幕 Markdown 檢視器中，現在只要將游標移動到目標程式碼上方，按下快捷鍵 `Shift+Y`，就會複製游標所在的整個程式碼區塊。如此一來，閱讀長篇文件時便能直接擷取所需的指令或程式碼片段，不必再特地動用滑鼠手動框選。",
                "en": "When reading long technical documents or inspecting terminal output, switching to a full-screen viewer keeps the layout cleaner, but grabbing example passages often takes extra steps. ⟦Picture taking inventory among store shelves with a handheld scanner. Previously, recording specifications meant pulling out pen and paper to transcribe character by character along the margins; now, aiming the light beam at a product label and pressing a button instantly pulls the whole dataset into a portable list.⟧\n\nIn the full-screen Markdown viewer, moving the cursor over target code and pressing `Shift+Y` now copies the entire code block beneath the cursor. Reading long documents this way lets you capture needed commands or snippets directly without reaching for a mouse to highlight them manually."
              },
              "original": "`Shift+Y` in the fullscreen markdown viewer copies the code block under the cursor"
            },
            {
              "title": {
                "zh-TW": "從儀表板開啟工作階段時，按 Esc 聚焦回捲歷程而非退回儀表板",
                "en": "Pressing Esc focuses scrollback instead of returning to dashboard when opening a session"
              },
              "body": {
                "zh-TW": "從儀表板開啟特定工作階段準備輸入指令時，要是想先檢閱先前的上下文，直覺按鍵往往可能帶來非預期的畫面切換。⟦跟參觀特展展示台一樣，以前在互動導覽的第一道題目按下返回鍵，畫面會冷不防退回接待大廳；現在只會把檢視游標移到上方的過往解說板，與獨立參觀時的行為完全相同。⟧\n\n以前從儀表板開啟工作階段後，若在第一個提問處按下 `Esc`，畫面會直接關閉並退回儀表板。現在按下 `Esc` 時，焦點會切換到上方儲存的回捲歷程（scrollback，可往上捲動查看歷史輸出的區域）。這樣的一致行為與直接在終端機啟動的獨立工作階段相同，避免你為了查看先前的對話而中斷手頭上的流程。",
                "en": "When opening a session from the dashboard to enter commands, intuitive keystrokes can often trigger unexpected view switches if you want to inspect earlier context first. ⟦Like visiting a special exhibition booth, pressing back at the first question of an interactive guide used to kick you back to the reception hall without warning; now, it only shifts the viewing cursor up to the previous display board, behaving exactly like a self-guided tour.⟧\n\nPreviously, after opening a session from the dashboard, pressing `Esc` at the first prompt would close the view and return directly to the dashboard. Now, pressing `Esc` shifts focus to the saved scrollback (the buffer area for scrolling up to inspect past output). This consistent behavior matches standalone sessions started directly in the terminal, preventing interruptions to your workflow when you only want to review prior dialogue."
              },
              "original": "`Esc` on the first question in a session opened from the dashboard moves focus to the scrollback, as it does in a standalone session, instead of returning to the dashboard"
            }
          ]
        }
      ]
    },
    {
      "id": "muse_code",
      "name": "Muse Code",
      "versions": [
        {
          "version": "1.4.11",
          "period": "2026-10-21",
          "items": [
            {
              "title": {
                "zh-TW": "內建 `web3d` skill，幫你做單檔 3D 網頁和瀏覽器小遊戲",
                "en": "Built-in `web3d` skill builds single-file 3D web pages and browser games"
              },
              "body": {
                "zh-TW": "想做一個能在瀏覽器裡轉動的 3D 產品展示頁，或一款小遊戲，現在可以直接交給 Muse。內建的 `web3d` skill（教 AI 照固定步驟做某類工作的說明書）會引導它做出單檔的 3D 網頁或瀏覽器遊戲，交件前先開瀏覽器檢查。動工前，它會先規劃每個部分的形狀、周遭環境和材質；第一版能跑之後，再派一個全新的測試子代理實際玩一遍、寫下批評，批評沒處理完不准交件，審查輪數也有時間上限。⟦餐廳出菜前先請另一位廚師試吃挑毛病，挑出的問題改完才端上桌，但試吃不會無限輪下去。⟧你說「停」，或只要它打開頁面時，它會立刻停止檢查。",
                "en": "Want a 3D product page you can spin in the browser, or a small game? You can now hand it straight to Muse. The built-in `web3d` skill (a playbook that teaches the AI fixed steps for one kind of job) guides it in building single-file 3D web pages and browser games and checking them in a browser before handoff. Before building, it plans each part's shape, surroundings and materials; after the first working build, a fresh tester subagent plays the page and writes a critique that must be closed before delivery, with a time guard on review rounds. ⟦A kitchen has a second chef taste each dish and point out flaws before it goes out, but the tasting does not go on forever.⟧ When you say stop, or just ask it to open the page, it stops checking right away."
              },
              "original": "Added a built-in `web3d` skill that guides building single-file 3D web pages and browser games and checking them in a browser before handoff\nThe built-in `web3d` skill now plans each part's shape, surroundings and materials before building, and after the first working build a fresh tester subagent plays the page and writes a critique that must be closed before delivery, with a time guard on review rounds\nThe built-in `web3d` skill now stops checking a page as soon as you say stop or ask it to just open the page"
            },
            {
              "title": {
                "zh-TW": "新增 `/game-3d-visuals`，`/blender` 也能做一般 3D 工作",
                "en": "New `/game-3d-visuals`, and `/blender` takes on general 3D work"
              },
              "body": {
                "zh-TW": "手上有幾張參考圖，想做出長得一樣的 3D 模型或遊戲畫面？新的 `/game-3d-visuals` 指令會依參考圖和一份視覺規格驅動 3D 模型與遊戲製作，並反覆比對渲染圖（電腦算出來的成品畫面）和參考圖，直到合格。`/blender` 則擴大到 Blender 裡的一般 3D 工作：建模、照一張圖重建場景、動畫和大型環境都能做，附帶的輔助腳本在 Blender 4 和 5 都能用。它也能拿渲染圖跟參考圖對量，逐格檢查動畫有沒有突然跳動、物體下沉或跑出鏡頭。折射的反彈次數也保留得更多，疊在一起的玻璃不會再渲染得那麼暗。",
                "en": "Have a few reference images and want a 3D model or game scene that matches them? The new `/game-3d-visuals` command drives 3D models and games from reference images and a visual spec, comparing each render (the finished image the computer calculates) against the references until it passes. `/blender` now covers general 3D work in Blender: object modeling, recreating a scene from an image, animation and large environments, with helper scripts that work on Blender 4 and 5. It can also measure a render against its reference image and check animations frame by frame for pops, sinking and objects leaving the shot. It keeps more refraction bounces too, so stacked glass renders less dark."
              },
              "original": "Added `/game-3d-visuals` to drive 3D models and games from reference images, a visual spec, and a render-vs-reference acceptance loop\n`/blender` now handles general 3D work in Blender, including object modeling, scene recreation from an image, animation and large environments, and ships helper scripts that work on Blender 4 and 5\n`/blender` can now measure a render against its reference image and check animations frame by frame for pops, sinking and objects leaving the shot, and keeps more refraction bounces so stacked glass renders less dark"
            },
            {
              "title": {
                "zh-TW": "工作階段協定新增 `session/leave`，可把對話交給另一台主機",
                "en": "The session protocol adds `session/leave` to hand a session to another host"
              },
              "body": {
                "zh-TW": "自己寫用戶端、透過工作階段協定（讓其他程式操控 Muse 對話的溝通規則）接 Muse 的開發者，現在多了 `session/leave`。用戶端可以把一個已載入的工作階段交給另一台主機，原本這台主機和上面其他工作階段照常運作。⟦客服把一通電話轉給另一個部門，自己這條線和其他來電都不受影響，跟這個動作一模一樣。⟧另一個修正也跟搬動工作階段有關：主機把閒置的分支卸載後，或主機重新啟動後，用同一個指令重試 `session/fork`，以前會回內部錯誤，現在會拿回那個分支和它的歷史紀錄。",
                "en": "Developers who write their own clients on top of the session protocol (the rules other programs use to drive Muse conversations) now get `session/leave`. A client can hand one loaded session to another host while this host and its other sessions keep running. ⟦A support agent transfers one call to another department while their own line and every other caller carry on untouched.⟧ A related fix: retrying `session/fork` with the same command after the host unloaded the idle fork, or after a host restart, used to return an internal error; it now returns that fork and its history."
              },
              "original": "Added `session/leave` to the session protocol, so a client can hand one loaded session to another host while this host and its other sessions keep running\nRetrying `session/fork` with the same command after the host unloaded the idle fork, or after a host restart, now returns that fork and its history instead of an internal error"
            },
            {
              "title": {
                "zh-TW": "SDK 能回答 agent 的提問，續接時也收得到待核准的請求",
                "en": "The SDKs can answer the agent's questions and catch pending approvals on resume"
              },
              "body": {
                "zh-TW": "用 TypeScript 或 Python SDK 把 Muse 嵌進自己程式的人，可能遇過 agent 問了問題，程式卻沒辦法回答。現在 TypeScript 用 `Session.answerUserInput`、Python 用 `Session.answer_user_input` 就能回答，續接工作階段時還沒答的問題也行。續接時也能傳入核准處理器和核准失敗觀察器，續接當下還在等核准的請求會送到處理器，失敗則通知觀察器。SDK 以前會對不接受意見的核准選項送出意見，讓核准一直卡著；現在改回報新的 `feedbackNotAccepted` 失敗類型，列出哪些選項收意見，並改選提供的拒絕或中止選項讓這一輪繼續。程式若逐一處理每種 `ApprovalFailure`，要補上這個新類型。",
                "en": "If you embed Muse in your own program with the TypeScript or Python SDK, you may have had the agent ask a question your code could not answer. Now TypeScript can answer with `Session.answerUserInput` and Python with `Session.answer_user_input`, including a question still open when a session is resumed. When resuming, you can also pass an approval handler and an approval-failure observer, so an approval still pending at resume reaches the handler and any failure reaches the observer. The SDKs used to send feedback on an approval choice that does not accept it, which left the approval open; they now report the new `feedbackNotAccepted` failure kind, listing the choices that accept feedback, and answer with the offered deny or abort choice so the turn goes on. Code that handles every `ApprovalFailure` kind needs a case for it."
              },
              "original": "The TypeScript and Python SDKs can answer the agent's questions with `Session.answerUserInput` and `Session.answer_user_input`, including a question still open when a session is resumed\nThe TypeScript and Python SDKs accept an approval handler and an approval-failure observer when resuming a session, so an approval still pending at resume reaches the handler and any failure reaches the observer\nThe TypeScript and Python SDKs no longer send feedback on an approval choice that does not accept it, which left the approval open; they report the new `feedbackNotAccepted` failure kind (listing the choices that accept feedback) and answer with the offered deny or abort choice so the turn goes on, and code that handles every `ApprovalFailure` kind needs a case for it"
            },
            {
              "title": {
                "zh-TW": "SDK 補收漏掉的事件時，不再提早收工",
                "en": "The SDKs no longer stop filling event gaps too early"
              },
              "body": {
                "zh-TW": "SDK 和 Muse 之間的事件偶爾會漏送，這時 SDK 要把缺口補齊。以前較早的缺口通知如果在較晚那則之後又送到一次，Python 和 TypeScript SDK 會提早停止補缺口，甚至回報工作階段已是最新，其實還少了事件。⟦郵差把上週的補寄通知晚於本週的送來，以前你就以為信都到齊、不再等了，現在你認得出那是舊通知，直接丟掉。⟧現在重複的舊通知，Python 會回 `IgnoredStaleFrame`，TypeScript 會回 `ignoredStaleFrame`。TypeScript 的 `StaleDroppableMethod` 也多了 `view/gap`，程式若逐一比對它的每個值，要補一個新分支。",
                "en": "Events between the SDK and Muse sometimes go missing, and the SDK then has to fill the gap. When an earlier gap notice was redelivered after a later one, the Python and TypeScript SDKs used to stop filling the gap early, or even report the session current while events were still missing. ⟦When last week's redelivery notice arrived after this week's, you used to assume all the mail was in and stop waiting; now you recognize it as an old notice and toss it.⟧ The repeated notice now returns `IgnoredStaleFrame` in Python and `ignoredStaleFrame` in TypeScript. TypeScript's `StaleDroppableMethod` also gains `view/gap`, so code that matches it exhaustively needs a new case."
              },
              "original": "The Python SDK no longer stops filling a delivery gap early, or reports the session current with events missing, when an earlier gap notice is redelivered after a later one; the repeat now returns `IgnoredStaleFrame`\nThe TypeScript SDK no longer stops filling a delivery gap early, or reports the session current with events missing, when an earlier gap notice is redelivered after a later one; the repeat now returns `ignoredStaleFrame`, and `StaleDroppableMethod` gains `view/gap`, so code that matches it exhaustively needs a new case"
            },
            {
              "title": {
                "zh-TW": "agent 外掛能附 MCP 伺服器，用 `muse mcp login` 登入",
                "en": "Agent plugins can bring MCP servers with `muse mcp login` sign-in"
              },
              "body": {
                "zh-TW": "想裝一個外掛，順便把它要用的外部服務也接好？agent 外掛現在可以附帶 MCP 伺服器（讓 AI 連接外部工具的協定），每個伺服器各自審核，並用 `muse mcp login` 登入。Windows 上，外掛的 hook（特定時機自動觸發的小程式）若用 `/bin/bash` 這類根目錄路徑啟動程式，以前安裝會失敗並顯示「plugin capability path must stay inside plugin root」，現在裝得起來。審核過的 skill 在登入後重新啟用時，授權也不會掉：審核過的 `allowed-tools` 保留預先核准，不會在未登入啟動之後退回跳出核准視窗。",
                "en": "Want a plugin to set up the outside services it needs at install time? Agent plugins can now supply MCP servers (the protocol that connects AI to external tools), each with its own review and `muse mcp login` sign-in. On Windows, plugins whose hooks (small programs triggered automatically at certain moments) start a program by a rooted path such as `/bin/bash` used to fail to install with \"plugin capability path must stay inside plugin root\"; they now install. A reviewed skill also keeps its grant when it is re-armed after login: reviewed `allowed-tools` keep their preapproval instead of falling back to an approval prompt after a logged-out start."
              },
              "original": "Agent plugins can supply MCP servers with per-server review and `muse mcp login` sign-in\nOn Windows, plugins whose hooks start a program by a rooted path such as `/bin/bash` now install instead of failing with \"plugin capability path must stay inside plugin root\"\nA reviewed skill's invocation-grant resolver now survives the post-login re-arm, so reviewed `allowed-tools` keep their preapproval instead of falling back to an approval prompt after a logged-out start"
            },
            {
              "title": {
                "zh-TW": "`muse exec --json` 列出每次執行的 token 用量，`/usage` 也算進背景工作",
                "en": "`muse exec --json` reports token usage per run, and `/usage` counts background work"
              },
              "body": {
                "zh-TW": "用腳本跑 Muse、想知道每次花了多少，現在不用自己估。`muse exec --json` 會在每次執行結束前印一行 `run.usage`，列出這次各模型用了多少 token（AI 計算用量的單位），以及你訂閱方案的額度。`/usage` 的總數也把背景工作算進去了，並在另一列 Observers 單獨列出。⟦水費帳單以前只算客廳的水龍頭，現在連後院自動灑水也計入，還分開列一行讓你看清楚。⟧",
                "en": "If you run Muse from scripts and want to know what each run cost, you no longer have to guess. `muse exec --json` prints a `run.usage` line before each run ends, with the run's token usage (the unit AI usage is measured in) per model and your subscription allowance. `/usage` totals now include background work too, broken out in a separate Observers row. ⟦A water bill that used to count only the kitchen tap now includes the backyard sprinklers, listed on their own line so you can see them.⟧"
              },
              "original": "`muse exec --json` prints a `run.usage` line before each run ends, with the run's token usage per model and your subscription allowance\n`/usage` totals now include background work and break it out in a separate Observers row"
            },
            {
              "title": {
                "zh-TW": "`muse serve` 可用環境設定調容量，當機後重開的對話不再卡住",
                "en": "`muse serve` takes capacity settings, and chats reopened after a crash no longer look stuck"
              },
              "body": {
                "zh-TW": "把 Muse 架成服務給多個程式共用的話，`muse serve` 現在讀三個環境設定，分別控制工作執行緒、指令受理容量和最多載入幾個工作階段；值寫錯會報出來，並改用正式環境的值。服務執行中你才登入、帳號卻還沒完成初始設定時，回合以前只叫你再登入一次，現在會說明該怎麼做。當機把工作階段紀錄的結尾截掉後，重開的對話以前看起來像卡住，現在歷史、新回合和核准請求都會正常顯示。讀長訊息、工作流程子任務的提示或原始網頁搜尋結果的下一頁時，如果沒有新存檔，也不再把整段對話重讀一遍。",
                "en": "If you run Muse as a service shared by several programs, `muse serve` now reads three environment settings for its worker threads, command admission capacity and loaded-session ceiling; an invalid value is reported and the production value is used. When you log in while it is running and your account still needs onboarding, turns used to just ask you to log in again; they now say what to do. A chat reopened after a crash cut the end of its session log used to look stuck; its history, new turns and approval requests now show again. Reading the next page of a long agent message, a workflow child's prompt or a raw web-search result no longer rereads the whole chat while nothing new has been saved."
              },
              "original": "`muse serve` reads three environment settings for its worker threads, command admission capacity, and loaded-session ceiling; an invalid value is reported and the production value is used\nWhen you log in while `muse serve` is running and your account still needs onboarding, turns now say what to do instead of asking you to log in again\nA `muse serve` chat reopened after a crash cut the end of its session log now shows its history, new turns and approval requests again, instead of looking stuck\nIn `muse serve`, reading the next page of a long agent message, a workflow child's prompt or a raw web-search result no longer rereads the whole chat while nothing new has been saved"
            },
            {
              "title": {
                "zh-TW": "支援 OSC 7501 的終端機，看得到 Muse 每一項背景工作",
                "en": "Terminals with OSC 7501 show every piece of background work Muse is running"
              },
              "body": {
                "zh-TW": "Muse 同時在背景跑好幾件事時，你很難知道它到底在忙什麼，又是哪一件在等你。支援程式狀態回報（OSC 7501）的終端機，現在會顯示工作階段正在跑的每個監控和背景終端機。工作流程、子代理和背景的後續回合也會一一列出，還會標出哪一個正在等你回應。⟦機場看板以前只掛一班飛機，現在每班的狀態都上板，哪一班在等你登機一眼就看到。⟧",
                "en": "When Muse runs several things in the background, it is hard to tell what it is busy with and which one is waiting on you. Terminals that support program status reports (OSC 7501) now show each monitor and background terminal a session is running. Each workflow, subagent and background follow-up turn is listed too, along with which one is waiting for you. ⟦An airport board that used to show just one flight now lists every flight's status, so you spot the one waiting for you to board at a glance.⟧"
              },
              "original": "Terminals that support program status reports (OSC 7501) now also show each monitor and background terminal a session is running\nTerminals that support program status reports (OSC 7501) now also show each workflow, subagent and background follow-up turn a session is running, and which one is waiting for you"
            },
            {
              "title": {
                "zh-TW": "網路核准預設選「這次不再問」，自動審查不再把本機請求全丟給你",
                "en": "Network approvals default to \"don't ask again\", and auto-review stops handing you every local request"
              },
              "body": {
                "zh-TW": "Muse 要連某個網址時會先問你，以前同一個目的地可能一問再問。現在網路核准提示預設停在「Yes, don't ask again this session」，只針對那個主機和連接埠，直接按 Enter，同一個目的地就不會再問。自動審查（Auto-review）以前看到連往 `localhost` 或 `127.0.0.1` 的請求，光憑位址或連接埠就整個交給你決定，現在會像其他動作一樣自己審。續接一個明確關掉審查員的 Auto-review 工作階段時，以前會被拒絕，現在照當時記下的人工設定繼續。",
                "en": "Muse asks before connecting to an address, and the same destination could ask again and again. Network approval prompts now start on \"Yes, don't ask again this session\" for that exact host and port, so pressing Enter stops the same destination from asking again. Auto-review used to hand you every request to a local server on `localhost` or `127.0.0.1` just because of its address or port; it now reviews them like any other action. Resuming an Auto-review session with the reviewer explicitly turned off used to be refused; it now proceeds under the recorded human override."
              },
              "original": "Network approval prompts now start on \"Yes, don't ask again this session\" for that exact host and port, so pressing Enter stops the same destination from asking again\nAuto-review no longer hands you every request to a local server on `localhost` or `127.0.0.1` just because of its address or port; it reviews them like any other action\nResuming an Auto-review session with the reviewer explicitly off now proceeds under the recorded human override instead of refusing, matching gate-off"
            },
            {
              "title": {
                "zh-TW": "工作流程當機後續跑，不再重跑已完成的子代理",
                "en": "Resuming a Workflow after a crash no longer reruns finished subagents"
              },
              "body": {
                "zh-TW": "長時間的工作流程（Workflow，串聯多個子任務自動執行的機制）跑到一半當機，最怕續跑時從頭再來。現在續跑時，在不同時間點已完成的平行子代理不會再重跑，也不會警告存下的復原狀態不可信，除非當機時還有更早啟動的子代理在跑。續跑被拒或啟動後失敗，現在會記下最終結果，用戶端不會一直顯示工作流程還開著。工作流程啟動後的部分歷史遺失、讀不到工作階段的歷史檔，或缺的部分包含工作流程怎麼啟動的紀錄時，續跑都會停下來說明原因，請你重跑那段工作，不再硬接著跑、從頭重跑或只丟一個沒頭沒尾的錯誤。",
                "en": "When a long Workflow (a mechanism that chains subtasks to run automatically) crashes halfway, the worst outcome is starting over on resume. Resuming now no longer reruns parallel subagents that had already finished at different times, or warns that the saved recovery state was not trustworthy, unless a subagent started before them was still running at the crash. A Workflow resume that is refused or fails after launch now records its final outcome, so clients no longer show the Workflow as open. When part of the saved history written after the Workflow started is missing, when Muse cannot read the session's history file, or when the missing part includes how the Workflow started, resuming now stops with the reason and asks you to run that work again, instead of pressing on, rerunning from the start or failing with a bare error."
              },
              "original": "Resuming a Workflow after a crash no longer reruns parallel subagents that had already finished at different times, or warns that the saved recovery state was not trustworthy, unless a subagent started before them was still running at the crash\nA Workflow resume that is refused or fails after launch now records its terminal reconciliation, so clients no longer show the Workflow as open until the host exits\nWhen part of a session's saved history written after a Workflow started is missing, resuming that Workflow now stops with the reason and asks you to run that work again, instead of continuing it\nWhen Muse can't read a session's saved history file, resuming a Workflow from that session now stops with the reason and asks you to run that work again, instead of re-running the Workflow from the start\nWhen the missing part of a session's saved history includes the record of how a Workflow started, resuming that Workflow now also stops with the reason and asks you to run that work again, instead of failing with a bare error"
            },
            {
              "title": {
                "zh-TW": "工作流程的子代理指定模型時不再被誤拒",
                "en": "Workflow subagents that name a model are no longer wrongly refused"
              },
              "body": {
                "zh-TW": "明明指名的就是正在用的模型，子代理卻不肯開工，讓人摸不著頭緒。現在工作流程裡的子代理指名的若正是工作階段在用的模型，就算它在模型清單裡被隱藏或不見，也照常執行，跟不指定模型一樣。在舊版開始的工作階段裡，只要 `/model` 對你的供應商列出該模型一次，指名它的子代理也能執行。工作流程裡的代理讀取存下的結果時，會拿到來源和欄位標籤；摘要清單省略較早的結果時，就算完整結果還在，也會收到通知。ParallelGroup 拒絕執行時，錯誤訊息會點名 members 參數和 .defer 改法，不再只給籠統的資料錯誤或內部代理編號。",
                "en": "Naming a model for a subagent in a Workflow could get refused for no clear reason. A Workflow subagent that names the model your session is running now runs even when that model is hidden or missing from the model list, the same as leaving the model out. In sessions first started on an older version, a subagent that names a model now runs when `/model` lists that model once for your provider. Workflow agents receive source and field labels when reading saved results, and get a notice when the summary list omits earlier results, even when full results are available. ParallelGroup refusals now name the members argument and the .defer correction instead of a generic plain-data error or an internal agent identifier."
              },
              "original": "Workflow agents receive source and field labels when reading saved results\nParallelGroup refusals now name the members argument and the .defer correction instead of a generic plain-data error or an internal agent identifier\nA Workflow subagent that names the model your session is running now runs even when that model is hidden or missing from the model list, the same as leaving the model out\nWorkflow agents now get a notice when the summary list omits earlier results, even when full results are available\nA Workflow subagent that names a model now runs in sessions first started on an older version, when `/model` lists that model once for your provider, instead of being refused"
            },
            {
              "title": {
                "zh-TW": "Muse 只說下一步卻沒動手時，會被催著繼續做",
                "en": "When Muse only announces its next step, it is nudged to keep going"
              },
              "body": {
                "zh-TW": "請 Muse 做事，它回一句「接下來我會……」就停了，你只好再打一次「繼續」。現在模型只交代下一步、沒有呼叫任何工具時，agent 會請它接著做，每次執行最多催兩次。子代理因設定的模型不能用而改用工作階段模型時，啟動它的 agent 仍會知道實際跑的是哪個模型。Muse 當機重啟後，agent 直接啟動、還沒開始跑的子代理不會自己再啟動一次，而是給上層一個「結果不明」，讓它檢查後決定要不要重開。你回答問題或傳新訊息前備好的 skill 提醒，現在會被丟掉，Muse 不會照著已經過時的建議行動；背景任務完成時，也不再重問還沒答的問題。",
                "en": "You ask Muse to do something, it replies \"next I will…\" and stops, and you have to type \"continue\" again. Now, when the model only says what it will do next and calls no tool, the agent asks it to go on, up to twice per run. When a subagent runs on the session model because its configured model is unavailable, the agent that started it is still told which model ran. After Muse restarts from a crash, a subagent the agent started directly that had not started running yet is no longer started again by itself; its parent gets one result saying the outcome is unknown, so it can check and start it again if still needed. A skill reminder prepared before you answered Muse's question or sent another message is now dropped, so Muse no longer acts on guidance your answer already settled, and the agent no longer repeats a pending question when a background task finishes."
              },
              "original": "After Muse restarts from a crash, a subagent the agent started directly (not from a workflow) that had not started running yet is no longer started again by itself; its parent gets one result saying the outcome is unknown, so it can check and start the subagent again if still needed\nWhen a subagent runs on the session model because its configured model is unavailable, the agent that started it is still told which model ran\nWhen the model only says what it will do next and calls no tool, the agent asks it to go on (up to twice per run) instead of ending the turn there\nA skill reminder prepared before you answered Muse's question or sent it another message is now dropped instead of reaching Muse as current advice, so Muse no longer acts on guidance your answer already settled\nAgent no longer repeats a pending question when a background task finishes"
            },
            {
              "title": {
                "zh-TW": "壓縮時檢查點沒存成，下一回合會補存，長對話重開又快了",
                "en": "Missed checkpoints during compaction are saved next turn, so long sessions reopen fast again"
              },
              "body": {
                "zh-TW": "很長的工作階段要重新打開，靠的是檢查點（記錄對話進度的存檔）才能快速接上。壓縮對話時如果檢查點沒存成，下一回合現在會補存，而且不再多做一次摘要，重開長對話又快了。側邊對話在壓縮時正在跑工具，或取消的工作流程工具存不下來時，那次壓縮現在也會存好續接用的檢查點。磁碟滿或寫入失敗排除後，下一個檢查點會正常儲存，不再又跳過一次。壓縮檢查點失敗又無法回復時，工作階段以前每一步都重新摘要同一段歷史，現在會保留每份摘要繼續往下。長工作階段也會在檢查點修剪已涵蓋的紀錄，不再無限長大。",
                "en": "Reopening a long session quickly depends on checkpoints (saves that record the conversation's progress). When a checkpoint could not be saved during a compaction, the next turn now saves it without another summary, so reopening a long session is fast again. Compactions that happen while a side chat is mid-tool, or after a cancelled workflow tool could not be saved, now save their resume checkpoint too. After a full disk or a failed write clears, the next checkpoint is saved instead of being skipped once more. A session whose compaction checkpoint failed and could not be rolled back used to re-summarize the same history on every step; it now keeps each summary and continues. Long sessions also trim covered records at checkpoints instead of growing without bound."
              },
              "original": "When a context checkpoint could not be saved during a compaction, the next turn now saves it without another summary, so reopening a long session is fast again\nResuming a session stays fast after a side chat was mid-tool during a compaction, or after a cancelled workflow tool could not be saved; those compactions now save their resume checkpoint\nLong sessions trim covered records at checkpoints instead of growing without bound\nAfter a full disk or a failed write clears, the next context checkpoint is saved instead of being skipped once more\nA session whose compaction checkpoint failed and could not be rolled back no longer re-summarizes the same history on every step; it keeps each compacted summary and continues"
            },
            {
              "title": {
                "zh-TW": "有背景任務或分支對話時，檢查點也照常儲存",
                "en": "Checkpoints keep saving alongside background tasks and branched chats"
              },
              "body": {
                "zh-TW": "Muse 多線並行時，記錄對話進度的檢查點（重開對話時用來快速接上的存檔）以前常被默默漏存。工具執行中有監控或背景任務回報時，這一回合剩下的檢查點現在不再被跳過。在來源對話同一次開啟裡打開的分支、倒回或側邊對話，現在會存自己的檢查點，不必等到 app 重開。重開對話後才完成的背景任務，就算啟動它的回合無法還原，也不再擋住壓縮存檢查點。背景任務還在跑時重開已存的對話，任務完成後，檢查點和任務結果都會照常記錄下來。",
                "en": "With background tasks, forks or side chats open, checkpoints (saves that record the conversation's progress) used to get skipped quietly. Checkpoints are no longer skipped for the rest of a turn when a monitor or background task reports in while tools are running. A forked, rewound or side chat opened in the same app run as its source now saves its checkpoints instead of skipping them until the app restarts. A background task that finishes after you reopen a chat no longer stops compaction from saving checkpoints when the turn that started it could not be restored. Reopening a saved chat while a task it started is still running no longer stops the checkpoint, or the task's result, from being recorded when that task finishes."
              },
              "original": "Context checkpoints are no longer skipped for the rest of a turn when a monitor or background task reports in while tools are running\nA forked, rewound or side chat opened in the same app run as its source now saves its context checkpoints, instead of skipping them until the app restarts\nA background task that finishes after you reopen a chat no longer stops compaction from saving checkpoints when the turn that started it could not be restored\nReopening a saved chat while a background task it started is still running no longer stops compaction from saving its checkpoint, or the task's result from being recorded, when that task finishes"
            },
            {
              "title": {
                "zh-TW": "`/compact` 在續接和分支的對話裡都能正常用",
                "en": "`/compact` works properly in resumed and forked chats"
              },
              "body": {
                "zh-TW": "手動打 `/compact` 壓縮對話，卻看不出到底成功沒有？續接工作階段後，`/compact` 現在會在對話紀錄裡顯示「◆ Context compacted」結果列，不再默默毫無變化。續接的分支對話裡，`/compact` 可以重複使用。正在看背景結果時打 `/compact`，現在壓縮的是最新的主要執行，不是那份背景結果。重開的對話裡，如果子代理的結果剛好在先前某次壓縮前才存進去，以前之後每次壓縮都會失敗，現在不會。agent 工作時你先打好的排隊輸入，會在回合結束後的壓縮做完就執行，不再停在閒置的提示前空等。",
                "en": "Type `/compact` and cannot tell whether it worked? After resuming a session, `/compact` now shows its `◆ Context compacted` outcome row in the transcript instead of silently leaving it unchanged. In resumed forked conversations, `/compact` works repeatedly. Issued while viewing a background delivery, `/compact` now compacts the newest top-level run instead of the delivery run. A reopened chat no longer fails every compaction when a subagent delivery was saved late, just before an earlier compaction. Input you queued while the agent was working now runs after a post-turn compaction instead of waiting at an idle prompt."
              },
              "original": "Queued input typed while the agent was working now runs after a post-turn context compaction instead of waiting at an idle prompt\nAfter resuming a session, `/compact` now shows its `◆ Context compacted` outcome row in the transcript instead of silently leaving the transcript unchanged\nA reopened chat no longer fails every compaction when a subagent delivery was saved late, just before an earlier compaction\n`/compact` works repeatedly in resumed forked conversations\n/compact issued while viewing a background delivery now compacts the newest top-level run instead of the delivery run"
            },
            {
              "title": {
                "zh-TW": "磁碟滿造成的寫入失敗，不再讓對話打不開",
                "en": "Write failures from a full disk no longer lock you out of a chat"
              },
              "body": {
                "zh-TW": "磁碟滿了是常見的意外，但不該讓整段對話報銷。工具結果因磁碟滿等原因寫不進去的對話，現在不必重開 app 就能再打開。工具呼叫寫不進去、結果卻寫進去的對話，重開後也打得開，只把那一回合擱置，Muse 會說明一次。記錄工作流程啟動結果時遇到磁碟錯誤，對話以前會一直停止存檔到重開為止，現在不會。用圖片倒回所建立的對話，在自動保留歷史的檢查點之後，倒回更早的回合也不再失敗。",
                "en": "A full disk is a common mishap, but it should not wreck a whole conversation. A saved chat whose tool result could not be written to disk, for example on a full disk, now opens again without a restart. A chat whose tool call could not be written while its result was written now opens after a restart; only that turn is set aside, and Muse says so once. A disk error while recording a workflow launch result no longer stops the chat from saving until restart. In a conversation created by rewinding with an image, rewinding to an earlier turn no longer fails after an automatic history-preserving checkpoint."
              },
              "original": "A chat no longer stops saving until restart after a disk error while recording a workflow launch result\nA saved chat whose tool result could not be written to disk (for example on a full disk) now opens again without a restart\nRewinding to an earlier turn no longer fails after an automatic history-preserving checkpoint in a conversation created by rewinding with an image\nA chat whose tool call could not be written to disk (for example on a full disk) while its result was written now opens again after a restart; only that turn is set aside, and Muse says so once"
            },
            {
              "title": {
                "zh-TW": "續接工作階段時，歷史、推理和背景狀態都會保留",
                "en": "Resumed sessions keep their history, reasoning and background status"
              },
              "body": {
                "zh-TW": "隔天回來想接著昨天的工作做，最怕打開是空的或少了一截。自動更新時還開著的工作階段，現在續接會帶回歷史，不再空空如也。有待處理背景結果的對話，續接時不再冒出不必要的上下文上限錯誤。檢查點修剪後續接，那些在新一代沒提供工具的 MCP 伺服器（連接外部工具的服務），工具身分綁定不會再弄丟。透過 `muse serve` 續接最早在 `muse` 供應商標籤下記錄、又用了記錄事件外掛的對話，先前的推理會保留，不再被當成別家的丟掉。重開側邊對話或分支後，背景工作狀態仍看得到；續接時也會自動取好名字。",
                "en": "Coming back the next day to pick up where you left off, the worst surprise is an empty or truncated session. A session left running through an auto-update now resumes with its history instead of coming back empty. Conversations with pending background results resume without unnecessary context limit errors. Resuming after a checkpoint prune no longer loses MCP (servers that connect external tools) tool identity bindings for servers that contribute no tools in the new generation. Resuming over `muse serve` a chat first recorded under the `muse` provider label that also used a plugin recording session events now keeps its earlier reasoning instead of dropping it as another provider's. Background work status stays available after reopening a side conversation or fork, and a resumed session gets its automatic name."
              },
              "original": "Resuming a session after a checkpoint prune no longer loses MCP tool identity bindings for servers that contribute no tools in the new generation\nResume conversations with pending background results without unnecessary context limit errors\nResuming over `muse serve` a chat first recorded under the `muse` provider label that also used a plugin recording session events now keeps its earlier reasoning instead of dropping it as another provider's\nA session left running through an auto-update now resumes with its history instead of coming back empty\nGate-on resume settles the session's automatic name like gate-off instead of leaving the reopen unnamed\nKeep background work status available after reopening a side conversation or fork"
            },
            {
              "title": {
                "zh-TW": "續接清單載入變快，分支裡的 `/fork` 不再失敗",
                "en": "The resume picker loads faster, and `/fork` works inside forks"
              },
              "body": {
                "zh-TW": "工作階段一多，打開續接選單往下捲就越來越慢。現在每多載一頁，不再把已載入的每一列重掃一遍，清單再長也維持同樣速度；打開分支工作階段時的重複歷史掃描也減少了。選單會註明沒有任何回合的工作階段不會保留。續接撞到衝突時，訊息會點名是哪個工作階段伺服器佔著，不再建議你把它砍掉。從已壓縮對話分出來的分支裡再用 `/fork`，以前會失敗並顯示「source Goal state changed or failed validation」，現在可以正常用。",
                "en": "The more sessions you have, the slower the resume picker got as you scrolled. Paging in more sessions no longer rescans every loaded row for each new one, so long lists keep loading at the same speed, and opening forked sessions repeats fewer history scans. The picker notes that sessions with no turns are not kept. A resume conflict now names the session server holding it instead of suggesting you kill it. `/fork` in a forked chat branched from a compacted chat used to fail with \"source Goal state changed or failed validation\"; it now works."
              },
              "original": "`/fork` now works in a forked chat that was branched from a compacted chat, instead of failing with \"source Goal state changed or failed validation\"\nPaging more sessions into the resume picker now stays linear instead of rescanning every loaded row for each new one, so long session lists keep loading pages at the same speed\nResume conflict names the holding session server instead of suggesting killing it\nResume picker notes that sessions with no turns are not kept\nReduce repeated history scans when opening forked sessions"
            },
            {
              "title": {
                "zh-TW": "按量計費時，存好的貢獻者模型不再被換成標準價模型",
                "en": "Pay-as-you-go sessions keep your saved contributor model"
              },
              "body": {
                "zh-TW": "你特地存好要用的貢獻者模型，工作階段一開頭跑的卻不是它。這發生在按量計費（用多少付多少）的工作階段：以前在訂閱檢查回覆之前，會先換成標準價的基礎模型。現在從第一個請求起就保留你存的模型；啟動時被重新對應過的，訂閱檢查一回覆也會換回來。⟦在餐廳指定了主廚，以前櫃台還沒查完會員資格，就先派一般廚師上菜，現在第一道菜就是你點的那位主廚做的。⟧",
                "en": "In pay-as-you-go sessions (you pay for what you use) with a saved contributor model, the model used to be mapped to the standard-price base until the subscription check answered. Now the saved model is kept from the first request, and one remapped at startup is restored as soon as the subscription check answers. ⟦You requested a particular chef, but the front desk used to send out the regular cook until it finished checking your membership; now your chef makes the very first dish.⟧"
              },
              "original": "Pay-as-you-go sessions restore a saved contributor model remapped at startup once the subscription check answers, instead of staying on the standard-price base\nPay-as-you-go sessions keep a saved contributor model from the first request, instead of mapping it to the standard-price base until the subscription check answers"
            },
            {
              "title": {
                "zh-TW": "服務重啟前會先核對身分，不再誤停長得像的工作階段",
                "en": "Restarts check ownership first and leave look-alike sessions running"
              },
              "body": {
                "zh-TW": "同時開好幾個 Muse 工作階段時，重啟一個服務不該順手把別的也關掉。服務和迴圈重啟前，現在會拿擁有權紀錄逐一核對每個程序，長得像、但不屬於它的工作階段會留著繼續跑。側邊對話的嘗試現在記在來源工作階段上，不再直接丟掉；搶不到資源的側邊對話照樣開得起來，重試時會補上或接上先前的結果。工作階段保留機制以前只要背景讀一次紀錄失敗，就永遠放棄那個工作階段，現在會在之後再讀，過了保留期限一樣會正常到期。",
                "en": "With several Muse sessions open, restarting one service should not take the others down with it. Service and loop restarts now verify every process against an ownership record before stopping it, so look-alike sessions are left running. Side chats now record their attempts on the source session instead of dropping them; a contended side chat still opens, and a retry backfills or joins the earlier outcome. Session retention used to give up on a session forever when one background read of its log failed; the read is now retried later, so the session still expires once it is past the retention period."
              },
              "original": "Side chats now record their attempts on the source session instead of dropping them, matching gate-off; a contended side still opens and a retry backfills or joins the prior outcome\nService and loop restarts verify every process against an ownership record before stopping it, so look-alike sessions are left running\nSession retention no longer gives up on a session forever when one background read of its log fails; the read is retried on a later pass, so the session still expires once it is past the retention period"
            },
            {
              "title": {
                "zh-TW": "表格每列都有分隔線，`/new` 和 `/clear` 立刻清畫面",
                "en": "Tables get row lines, and `/new` and `/clear` clear the screen at once"
              },
              "body": {
                "zh-TW": "在終端機裡看 Muse 回覆，有幾個小地方變順眼了。回覆裡的表格現在每一列之間都畫分隔線。打 `/new` 或 `/clear` 會立刻清空畫面，不再讓舊對話停留好幾秒；新工作階段啟動時你同時打字，第一個字母也不會被吃掉。TUI（在終端機裡畫出來的操作介面）收到工作階段列時，不再清掉終端機原本的捲動紀錄，對話紀錄裡的工作階段列也不會畫兩次。",
                "en": "Reading Muse's replies in the terminal got a little easier on the eyes. Tables in replies now draw a line between every row. `/new` and `/clear` clear the screen immediately instead of leaving the old chat visible for seconds, and text you type while the fresh session starts no longer loses its first letter. The TUI (the interface drawn inside the terminal) no longer erases the terminal's existing scrollback when session rows arrive, and session rows no longer render twice in the transcript."
              },
              "original": "Tables in assistant replies draw a line between every row\nThe TUI no longer erases the terminal's pre-existing scrollback when session rows arrive, matching gate-off\nText you type while `/new` or `/clear` starts a fresh session no longer loses its first letter\nSession rows no longer render twice in the TUI transcript\n/new and /clear now clear the screen immediately instead of leaving the old chat visible for seconds"
            },
            {
              "title": {
                "zh-TW": "讀影片改在本機擷取畫格，錯誤訊息不再出現兩次",
                "en": "Video reads extract frames locally, and errors show only once"
              },
              "body": {
                "zh-TW": "交給 Muse 一支影片、要它看內容時，它現在會引導在本機擷取畫格，不必把影片上傳。回合失敗前已經收到部分輸出時，錯誤現在只顯示一次，不再出現兩次。排隊輸入面板在訊息更新失敗時會說明，也會告訴你輸入仍然有效。`/stop all` 期間剛好完成的任務，會回報為已完成，不再讓停止動作失敗。同一段對話裡，也不再一再載入表格格式的 skill。",
                "en": "Hand Muse a video to look at, and it now guides local frame extraction without uploading the video. When partial output arrived before a turn failed, the error now shows once instead of twice. The queued input panel says when message updates fail and that your input still works. Tasks that finish during `/stop all` are reported as already finished instead of failing the stop. The table-formatting skill is no longer loaded again and again within a conversation."
              },
              "original": "Video file reads guide local frame extraction without uploading the video\nThe queued input panel says when message updates fail and that your input still works\nAvoid repeated table-formatting skill loads within a conversation\nReport tasks that finish during `/stop all` as already finished instead of failing the stop\nA failed turn now shows its error once instead of twice when partial output arrived before the failure finished"
            },
            {
              "title": {
                "zh-TW": "工具執行的指令併入同一條追蹤，hook 的 `model_provider` 恢復正確",
                "en": "Tool-run commands join the turn's trace, and the hook `model_provider` is correct again"
              },
              "body": {
                "zh-TW": "想追查 Muse 一回合裡到底跑了哪些指令，以前會看到一堆找不到上層的零散追蹤紀錄。現在工具執行的指令，包括 shell、MCP（讓 AI 連接外部工具的協定）呼叫、hook、背景終端機和 `!` shell，都會併進該回合的追蹤，像 `meta` 這類工具產生的下游紀錄，會掛在執行它的工具呼叫底下。⟦家譜上原本有一群人找不到父母，現在每個人都接回自己那一支。⟧在每條 Meta 閘道連線上，包括公司 Mac 啟動器，Muse 模型的 hook（特定時機自動觸發的小程式）輸入 `model_provider` 又會是 `meta`；透過 Meta 閘道提供的 Claude 和 GPT 模型，則照舊不帶這個欄位。",
                "en": "Trying to trace which commands ran in a Muse turn used to turn up a pile of orphaned traces with no parent. Commands run by tools, including shell, MCP (the protocol that connects AI to external tools) calls, hooks, background terminals and `!` shell, now join the turn's trace, so downstream spans from tools like `meta` nest under the tool call that ran them. ⟦A family tree that had a group of people with no parents now has everyone reattached to their own branch.⟧ On every Meta gateway connection, including the corp Mac launcher, the hook (a small program triggered automatically at certain moments) input `model_provider` is `meta` again for Muse models, and it is still omitted for Claude and GPT models served through the Meta gateway."
              },
              "original": "Hook input `model_provider` is `meta` again for Muse models on every Meta gateway connection, including the corp Mac launcher, and is still omitted for Claude and GPT models served through the Meta gateway\nCommands run by tools (shell, MCP calls, hooks, background terminals, `!` shell) now join the turn's trace instead of starting orphan traces, so downstream spans from tools like `meta` nest under the tool call that ran them"
            }
          ]
        },
        {
          "version": "1.4.10",
          "period": "2026-10-21",
          "items": [
            {
              "title": {
                "zh-TW": "/blender 指令登場，參考圖直接重建為 3D 可編輯場景",
                "en": "/blender Command Debuts to Reconstruct Reference Images Directly into Editable 3D Scenes"
              },
              "body": {
                "zh-TW": "當你手邊有一張概念設計圖或參考照片，想快速建立立體模型時，往往需要在建模程式裡從零搭建物件與燈光。新推出的 `/blender` 指令能依據參考圖片，自動重建出可自由編輯的 Blender（一套開源三維繪圖工具）場景。⟦就像拿著一張平面設計草圖走進工作室，模型師當場組裝出能拆卸零件、隨意更換視角的立體模型。⟧這讓你省去繁瑣的手動對位，產出的檔案也能立刻在編輯器內接續調整視角、材質與燈光。",
                "en": "When you have a concept design sketch or reference photo and want to quickly create a 3D model, you often need to build objects and lighting from scratch in modeling software. The newly introduced `/blender` command automatically reconstructs a freely editable Blender (an open-source 3D modeling tool) scene based on reference images.\n\n⟦Picture walking into a studio with a 2D design sketch, where a modeler immediately assembles a 3D model with detachable parts and freely adjustable viewing angles on the spot.⟧\n\nThis saves you from tedious manual alignment, and the generated files can immediately be used to adjust perspectives, materials, and lighting directly in the editor."
              },
              "original": "Added `/blender` to reconstruct reference images as editable Blender scenes"
            },
            {
              "title": {
                "zh-TW": "電腦操作精簡工具呼叫，跳過重複查閱與多餘清單直奔目標",
                "en": "Computer Use Streamlines Tool Calls, Skipping Redundant Checks and Lists to Head Straight for Goals"
              },
              "body": {
                "zh-TW": "讓 AI 代理操作桌面螢幕時，每多呼叫一次工具，都會拉長等待時間並消耗額度。電腦操作（Computer Use，讓 AI 透過視覺畫面模擬人類點擊與輸入的機制）現在減少了許多冗餘的工具呼叫。當任務明確指定某個應用程式時，它會直接啟動該程式，不再先列出所有應用程式清單；遇到檔案或系統查詢，會直接以終端機指令回答；它還會跳過重複閱讀自身說明的動作。⟦好比熟練的助理接獲指令便立刻動手，不再反覆翻閱員工手冊或把工具箱裡的物件逐一清點一遍。⟧就算畫面當下看起來像是已經做完，它仍會依照指示確實執行任務。",
                "en": "When having an AI agent operate desktop screens, every extra tool call prolongs waiting times and consumes quota. Computer Use (a mechanism that lets AI simulate human clicks and keystrokes through visual screens) now cuts out many redundant tool calls. When a task explicitly specifies an application, it launches that program directly instead of first listing all applications. When encountering file or system queries, it answers directly with terminal commands. It also skips repeatedly re-reading its own instructions.\n\n⟦Think of a skilled assistant getting straight to work upon receiving an order, instead of repeatedly leafing through the employee handbook or counting every item in the toolbox.⟧\n\nEven if the screen currently looks like the task is already finished, it still carries out the assignment faithfully according to instructions."
              },
              "original": "Computer Use makes fewer tool calls: it skips re-reading its own instructions, opens the app a task clearly points to without listing apps first, answers file and system questions with the shell, and does the task even when the screen already looks done"
            },
            {
              "title": {
                "zh-TW": "背景任務提示頻率降至單次，監聽狀態即刻顯示且長程目標不中斷",
                "en": "Background Task Prompts Reduced to Once, Displaying Listening States Instantly Without Halting Goal Runs"
              },
              "body": {
                "zh-TW": "在終端機啟動耗時較長的工作時，對話視窗的提示頻率與狀態顯示有了更清晰的安排。執行超過 30 分鐘的背景指令，現在只會主動提示代理一次，不再每隔半小時就提醒，需要時由代理自行排定後續檢查時機。當代理收到長時間背景指令的提示時，對話紀錄會顯示「Waiting for background task」，讓你明白它為何開始動作；針對仍在執行的背景指令，重複發出的等待請求也不再卡住整個等待時間。即使電腦負載偏高或硬碟緩慢，新啟動的監聽器也會即刻出現在頁尾與 `/tasks` 清單中。在長程目標模式（goal run，讓 AI 自主推進多步驟目標的執行流程）下，若工具執行後產生了新的下一步指引，它會替換既有指引並繼續前進，不再提早停下；僅在指引完全相同時才會如常停止。",
                "en": "When launching long-running jobs in the terminal, prompt frequencies and status displays in conversation windows are now arranged much more clearly. Background commands running longer than 30 minutes will now proactively prompt the agent only once instead of reminding it every half hour, leaving the agent to schedule follow-up checks when needed. When an agent receives a prompt for a long-running background command, the conversation log displays \"Waiting for background task\" so you understand why it started taking action. Repeated wait requests for background commands that are still executing no longer stall the entire wait duration. Even under heavy machine load or slow disk speeds, newly started listeners appear immediately in the footer and the `/tasks` list. In goal run mode (an execution flow where AI autonomously advances multi-step goals), if a tool run produces new next-step guidance, it replaces existing guidance and continues moving forward rather than stopping early. It only stops as usual when the guidance is completely identical."
              },
              "original": "A background command that runs past 30 minutes now prompts the agent once instead of every 30 minutes; the agent schedules its own later check if it needs one\nThe transcript shows \"Waiting for background task\" when the agent is prompted about a long-running background command, so you can see why it started working\nA duplicate wait for a background command that is still running no longer holds the chat for its full requested wait\nA monitor you start now appears in the footer and in `/tasks` right away, even when the machine is busy or the disk is slow\nA long goal run no longer stops early when fresh next-step guidance replaces still-active guidance after new tool work; identical guidance still stops as before"
            },
            {
              "title": {
                "zh-TW": "沙盒環境嚴格隔離設定目錄，Git 技能禁止擅自隱藏未追蹤檔案",
                "en": "Sandbox Environment Strictly Isolates Config Directories, Forbidding Git Skill from Hiding Untracked Files"
              },
              "body": {
                "zh-TW": "把終端機指令交給 AI 時，機密資料與檔案庫的安全界線至關重要。沙盒（sandbox，隔離執行的受限環境）指令，無法再讀取或修改 `~/.config/muse`、`~/.config/tbh` 或 `XDG_CONFIG_HOME` 下的登入資訊、設定、核准規則與被封鎖程式。任務只要求乾淨工作樹時，Git 技能拒絕用忽略規則或 stash 隱藏未追蹤檔案。掛鉤（hook，特定事件觸發的小程式）參數 `model_provider` 改標註模型供應商；僅 Meta 閘道列為 Meta 擁有的模型回報 `meta`，其餘省略，其上的 Claude 或 GPT 工作階段不再回報為 `meta`。⟦像訪客證開不了保險櫃，清潔工也不准把雜物塞進抽屜掩蓋混亂。⟧",
                "en": "When delegating terminal commands to AI, the security boundaries of sensitive credentials and repositories are critical. Sandboxed commands (a restricted environment for isolated execution) can no longer read or modify login credentials, settings, approval rules, or blocked applications under `~/.config/muse`, `~/.config/tbh`, or `XDG_CONFIG_HOME`. When a task only requires a clean worktree, the Git skill refuses to hide untracked files using ignore rules or stash. The hook (a small program triggered by a specific event) parameter `model_provider` now indicates the model provider. Only models listed as Meta-owned by the Meta gateway report `meta`, while others omit it, so Claude or GPT sessions running on it no longer report `meta`.\n\n⟦Just as a visitor badge cannot open a safe, a cleaner is also forbidden to shove clutter into drawers to hide a mess.⟧"
              },
              "original": "Security: Sandboxed commands can no longer read or change sign-in details, settings, approval rules or blocked apps stored in `~/.config/muse`, `~/.config/tbh` or the same folders under `XDG_CONFIG_HOME`\nSecurity: The git skill now refuses to hide your untracked files with ignore rules or a stash when a task only asks for a clean tree\nSecurity: Hook input `model_provider` now names the vendor of the selected model instead of the connection it was routed through: it is `meta` only for a model the built-in Meta gateway catalog lists as Meta-owned, and is omitted otherwise, so a Claude or GPT session on the Meta gateway no longer reports `meta`"
            },
            {
              "title": {
                "zh-TW": "工作流程子任務獨立重試，崩潰重啟精準標記中斷狀態",
                "en": "Workflow Subtasks Retry Independently, Accurately Tagging Interrupted States upon Crash Restarts"
              },
              "body": {
                "zh-TW": "執行包含多個子任務的工作流程時，單一環節發生中斷不再導致整個流程無故卡死。如果工作流程（Workflow，串聯多個子任務自動化執行的機制）的子任務因啟動紀錄無法儲存，或該任務已存有另一筆不同的啟動紀錄而停止，現在會以白話說明未執行任何新操作、其餘子任務仍持續運作，並指引下一步做法。排定重試後立刻遭中斷的任務，恢復時會依照原記錄的模型與思考力度接續運作。若恢復的任務初次儲存在任何內容寫入前就失敗，僅有該任務會重試而非中斷整個流程；重試的儲存再失敗，也僅標記該任務失敗，並告訴你下一步該怎麼做。遇崩潰重新啟動時，若已終止的代理重啟遭拒，其紀錄列會直接標為失敗並讓整體流程完工，不再誤顯為執行中。",
                "en": "When executing workflows containing multiple subtasks, an interruption in a single step no longer stalls the entire flow without explanation. If a subtask in a workflow (a mechanism that chains multiple subtasks for automated execution) stops because its launch record cannot be saved, or because a different launch record already exists for it, it now explains in plain language that no new operations were run, confirms that other subtasks are still running, and guides next steps. Tasks interrupted immediately after scheduling a retry will resume operation using the originally recorded model and reasoning effort. If an initial save of a resumed task fails before any content is written, only that task retries instead of interrupting the entire workflow. If the retried save fails again, it only marks that task as failed and tells you what to do next. Upon restarting after a crash, if a terminated agent is rejected for restart, its record row is directly marked as failed and allows the overall workflow to finish, rather than mistakenly displaying as running."
              },
              "original": "A Workflow child that stops because its start record can't be saved, or because a different start record was already saved for it, now says in plain words that nothing new ran, that the other Workflow children keep running, and what to do next\nA Workflow task that was interrupted right after it scheduled a retry now resumes with its recorded model and effort instead of being held\nWhen a resumed Workflow task's first save fails before anything is written, only that task retries instead of the whole Workflow stopping; if the retry's save fails too, only that task fails and tells you what to do next\nA Workflow resumed after a crash no longer shows a killed agent as still running when its relaunch is refused before it starts; the killed attempt's row now ends Failed and the Workflow completes"
            },
            {
              "title": {
                "zh-TW": "工作流程恢復後保留自訂名稱，控制台與即時進度完整顯示",
                "en": "Workflows Retain Custom Names After Resuming, Displaying Fully in Console and Live Progress"
              },
              "body": {
                "zh-TW": "為工作流程取了清楚好記的專屬名稱，卻常在重新啟動或查閱紀錄時被打回預設代號。現在工作流程中途暫停並重新恢復後，各項流程訊息與結束提示都會完整保留先前記錄的自訂名稱。控制室面板以及點開的執行結果頁面，也會持續顯示該流程原本的名稱。新儲存的流程歷史紀錄與即時動態更新都會採用所記錄的名稱；如果遇上過去未儲存名稱的舊版執行紀錄，介面則會明確標記為「name unavailable」，方便你清楚辨識。⟦跟在專案檔案夾貼上手寫標籤一樣，中途交接給其他班次時，標籤依然牢牢貼在封面，不再因為換人接手就退回空白編號。⟧",
                "en": "Giving a workflow a clear and memorable custom name often used to result in it reverting to a default identifier upon restarting or checking records. Now, after a workflow is paused mid-run and resumed, workflow messages and finish prompts fully retain the previously recorded custom name. The Mission Control panel and opened run result pages also consistently display the workflow's original name. Newly saved workflow history records and real-time live updates will use the recorded name. If encountering legacy execution records where names were not previously saved, the interface clearly marks them as \"name unavailable\" for easy identification.\n\n⟦Much like a handwritten label taped to a project folder stays firmly on the cover during a handover to another shift, it no longer reverts to a blank number just because someone else took over.⟧"
              },
              "original": "Workflow messages and exit prompts keep the recorded name after resume\nWorkflow control rooms and opened results keep the recorded workflow name\nNewly saved workflow history and live updates use recorded names, with \"name unavailable\" for older runs that have none"
            },
            {
              "title": {
                "zh-TW": "子代理忠實遵從自訂思考力度，支援父任務等待時並行處理",
                "en": "Subagents Faithfully Honor Custom Reasoning Effort, Supporting Concurrent Processing While Parent Waits"
              },
              "body": {
                "zh-TW": "調派多個子代理分頭協同作業時，每位代理的推理深度與模型指派現在都能精準維持。透過工作階段協定啟動的工作階段，現在會遵循所儲存的思考力度（reasoning effort，決定模型在回答前投入多少深度推理的參數）。子代理會嚴格遵循其代理定義中設定的思考力度與專屬模型；當定義的模型無法使用時，會發出通知並改為備援使用當前工作階段的模型。若工作流程中的子代理指名了你工作階段的模型，即使該模型在重啟後被隱藏或自模型清單中移除，子代理依然會維持運行。在父代理處於等待狀態時，子代理現在獲准同步處理待處理的結果，不再平白枯等。⟦彷彿專案主管在等待外部回覆時，被分派任務的助理無須跟著停工，能繼續整理其他已送達的數據資料。⟧",
                "en": "When dispatching multiple subagents to collaborate on tasks, the reasoning depth and model assignment of each agent can now be maintained precisely. Sessions launched via the session protocol now follow the saved reasoning effort (a parameter determining how much deep reasoning the model invests before answering). Subagents strictly adhere to the reasoning effort and dedicated model specified in their agent definitions. When a defined model is unavailable, a notification is sent, and it falls back to using the current session's model instead. If a subagent in a workflow specifies your session's model, it keeps running even if that model is hidden or removed from the model catalog after a restart. While the parent agent is in a waiting state, subagents are now permitted to process pending results concurrently instead of waiting idly.\n\n⟦Imagine a project manager waiting for an external reply, where assigned assistants need not sit idle and can keep organizing other data that has already arrived.⟧"
              },
              "original": "Sessions started through the session protocol honor saved reasoning effort\nSubagents now honor the reasoning effort set in their agent definition\nA Workflow subagent that named your session's model keeps running after a restart even when that model is now hidden or missing from the model list, the same as one that left the model out\nSubagents use their definition's model, falling back to the session model with a notice when unavailable\nAllow subagents to process pending results while the parent waits"
            },
            {
              "title": {
                "zh-TW": "muse serve 載入對話不再卡死其他請求，健全歷史還原與帳號指引",
                "en": "muse serve No Longer Blocks Other Requests While Loading Chats, Solidifying History Restoration and Account Guidance"
              },
              "body": {
                "zh-TW": "將 Muse 當作本機背景伺服器運行時，多個工作階段並行存取的排程機制有了顯著改善。當你透過 `muse serve` 處理龐大對話時，伺服器在首次載入歷史紀錄、恢復或分支工作階段、開啟側邊對話、列出待審批項目或讀取工具輸出時，都能持續回應其他請求，不再拖住其他對話；嘗試恢復載入中的工作階段也會依序等待，而非誤報已被佔用。它現在能直接顯示 Muse 1.4.8 仍運行中的對話紀錄，即便儲存檢視檔案遭刪除也能正常還原。若啟動時缺乏模型清單且帳號需補填付款方式、新手導覽或年齡確認，對話輪次會明確說明該做什麼，不再無端要求重新登入。",
                "en": "When running Muse as a local background server, the scheduling mechanism for concurrent session access has improved significantly. When handling large conversations via `muse serve`, the server continues responding to other requests while initially loading history, resuming or branching sessions, opening side chats, listing pending approval items, or reading tool outputs, without stalling other conversations. Attempts to resume a session that is currently loading will now wait in turn rather than falsely reporting that it is already occupied. It can now directly display ongoing conversation records from Muse 1.4.8 and cleanly restore them even if saved view files were deleted. If the model list is missing at startup and the account requires adding a payment method, completing onboarding, or confirming age, the turn explicitly explains what needs to be done instead of prompting for a re-login without explanation."
              },
              "original": "`muse serve` now shows the history of a session that Muse 1.4.8 is still running, instead of failing to read it until that session closes\nWhen a `muse serve` client resumes a chat whose saved history view file was deleted, the chat shows its history again instead of none\n`muse serve` keeps answering other requests while a long session's history loads for the first time, and resuming a session that is still loading waits for it instead of reporting the session as in use\n`muse serve` keeps answering other requests while it resumes or forks a long session, opens a side chat on one, lists its pending approvals or reads one of its tool outputs, instead of stalling every other thread until that session's history is read\nWhen `muse serve` starts with no saved model list and your account still needs a payment method, onboarding or age confirmation, turns now say what to do instead of asking you to log in again"
            },
            {
              "title": {
                "zh-TW": "手動壓縮支援外接工具，修正檢查點儲存並加速大型工作階段分支",
                "en": "Manual Compaction Supports External Tools, Fixing Checkpoint Saves and Accelerating Large Session Branching"
              },
              "body": {
                "zh-TW": "對話內容過長而需要精簡上下文時，手動輸入 `/compact` 的觸發條件與儲存機制現在更為可靠。設定了外接工具 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）的工作階段，現在可以直接執行手動 `/compact`，不再回報無可壓縮內容。在儲存檢查點時若恰好遇到其他寫入操作，不會再遺漏儲存；重新開啟已儲存的對話後，壓縮也能持續正常運作，不再重複摘要並最終報錯。如果在送出模型請求前就取消訊息，執行壓縮會如實回報沒有內容可摘要，避免留下錯誤檢查點而導致後續跳過開發者指令。當工具結果被清除或提醒過期時，`/compact` 的 token（AI 處理與計費的文字單位）計算已獲修正；在經歷壓縮的超大型工作階段中，執行 `/btw`、`/side` 與 `/fork` 的開啟速度也明顯加快。",
                "en": "When conversations grow too long and context needs trimming, the trigger conditions and saving mechanisms for manually running `/compact` are now much more reliable. Sessions configured with external MCP (Model Context Protocol, a protocol enabling AI to connect to external tools) tools can now directly run manual `/compact` without reporting that there is nothing to compact. Saving a checkpoint while another write operation is occurring no longer drops the save. After reopening a saved conversation, compaction continues working normally rather than repeatedly summarizing and ultimately erroring out. If a message is canceled before sending the model request, running compaction accurately reports that there is no content to summarize, preventing invalid checkpoints that could cause subsequent developer instructions to be skipped. The token (a text unit for AI processing and billing) calculation for `/compact` has been corrected when tool results are cleared or reminders expire. In very large compacted sessions, opening speeds for running `/btw`, `/side`, and `/fork` are also noticeably faster."
              },
              "original": "Manual `/compact` now compacts sessions with MCP servers configured, instead of answering \"no compactable run is available\"\nA context compaction no longer skips its checkpoint when another session write lands while that checkpoint is being saved\nRunning /compact in a chat whose turns all ended before any model request (for example, a message cancelled before Muse began preparing its model request) now reports that there is nothing to summarize instead of writing a summary checkpoint that could later let the chat skip setting up its developer instructions\nCompaction keeps working after you reopen a saved chat, instead of summarizing again at every boundary and eventually failing with \"could not prepare the conversation\"\nCorrect `/compact` token counts when tool results are cleared or reminders expire\n`/btw`, `/side`, and `/fork` open faster in very large compacted sessions"
            },
            {
              "title": {
                "zh-TW": "工作階段恢復嚴格檢驗啟動紀錄，掛鉤只在紀錄完整時觸發執行",
                "en": "Session Resumption Strictly Verifies Startup Logs, Firing Hooks Only When Logs Are Fully Intact"
              },
              "body": {
                "zh-TW": "重新開啟先前儲存的對話時，若背後的執行紀錄曾發生缺損，盲目載入往往會引發非預期的錯誤。如果工作階段的日誌無法證實其最後一次啟動紀錄，現在會直接拒絕恢復該對話，並在訊息中點名損毀的紀錄，不再默默開啟。在終端機介面恢復工作階段時，只有日誌能證實啟動完整的對話才會觸發 SessionStart 掛鉤（hook，特定事件發生時被觸發的小程式），其來源會標註為 `resume`。若對象是從未啟動的分支副本，來源則標註為 `fork`。若是無法證實啟動的副本，則完全不執行掛鉤並在啟動通知中說明。已儲存的工作階段格式也新增了標記欄位，可記錄儲存當下是否仍在啟動中；雖然目前尚未正式寫入，但未來一旦開始記錄，舊版 Muse 將拒絕開啟這類工作階段以防出錯。",
                "en": "When reopening a previously saved conversation, blindly loading it often causes unexpected errors if the underlying execution logs were corrupted. If a session log cannot verify its last startup record, it now explicitly refuses to resume that conversation and names the corrupted record in the message, rather than opening silently. When resuming a session in the terminal interface, only conversations whose startup can be verified by the log will trigger the SessionStart hook (a small program triggered by a specific event), with its source marked as `resume`. For branched copies that have never been started, the source is marked as `fork`. For copies whose startup cannot be verified, the hook is not executed at all, and an explanation is provided in the startup notification. The saved session format also adds a flag field to record whether it was still starting up at the time of saving. Although not actively written yet, once recorded in the future, legacy versions of Muse will refuse to open such sessions to prevent errors."
              },
              "original": "Saved sessions can record whether a run was still starting when they were saved; nothing records it yet, so no session opens differently today, and once a later Muse records it, Muse versions released before that one will not open such a session\nResuming a session whose log cannot prove its last start is refused with a message naming the broken record instead of opening silently\nSessionStart hooks run on every resumed TUI session whose log proves its last start, reporting `resume` (or `fork` for a never-activated clone) as their source; a replayed copy that cannot prove it runs none and says so in its startup notice"
            },
            {
              "title": {
                "zh-TW": "跨桌面與終端切換即時同步對話，分支重試具備冪等性不再撞號",
                "en": "Real-time conversation sync across desktop and terminal switches, with idempotent branch retries that prevent collisions"
              },
              "body": {
                "zh-TW": "在桌面應用程式與終端機之間切換工作環境時，不同版本之間的對話紀錄接軌現在省去了不必要的等待。若桌面端與終端機運行的 Muse Code 版本不同，切換時會直接同步另一端剛進行的簡短對話，不再從頭重新建構歷史紀錄（自 1.4.9 以前版本升級後的首次開啟仍會重建一次）。恢復進度落後的對話時，會在同一次恢復中直接補齊最新紀錄，不再出現歷史空白或遺漏即時回覆。使用相同指令識別碼重試分支時，會符合冪等（idempotent，多次執行結果與單次相同）原則直接回傳原始結果，不再報錯；指定不同切分時則維持衝突。即使對話含有中途送達的背景任務通知，分支在應用程式與 SDK 也能順利載入。工作階段中模型重複使用工具識別碼不再破壞後續壓縮，載入重複核准紀錄的工作階段時記憶體消耗也明顯降低。",
                "en": "When switching work environments between the desktop application and the terminal, bridging conversation history across different versions now eliminates unnecessary waiting. If the desktop app and terminal run different versions of Muse Code, switching will directly sync recent short conversations from the other side rather than rebuilding history from scratch (the first launch after upgrading from versions prior to 1.4.9 still rebuilds once). When restoring a conversation that has fallen behind, the latest records are backfilled in the same restoration pass, preventing history gaps or missing real-time replies. When retrying a branch using the same command identifier, it adheres to idempotent principles (yielding identical results when run multiple times as when run once) and directly returns the original result instead of throwing an error. Specifying a different split continues to raise a conflict. Even if a conversation contains background task notifications delivered midway, branches load smoothly across both the application and the SDK. A model reusing tool identifiers within a session no longer breaks subsequent compaction, and memory consumption drops significantly when loading sessions with duplicate approval records."
              },
              "original": "A forked chat whose history has a background-task notice delivered mid-turn now opens in apps and SDKs in more cases, instead of failing to load\nSwitching a conversation between the desktop app and the terminal when they run different versions of Muse Code now catches up a short exchange made on the other side instead of rebuilding its saved history from scratch; the first open after updating from a release before 1.4.9 still rebuilds it once\nResuming a thread whose saved history had fallen far behind now brings it up to date in that same resume, instead of showing no history and missing live replies until the next resume\nA session whose model reused a tool call id for two same-length results no longer becomes impossible to resume or compact after old tool results are cleared\nRetrying a fork with the same command id now returns the original outcome instead of answering command_id_conflict, matching gate-off idempotency; reusing a command id with a different cut still conflicts\nReduce memory use when loading sessions with repeated approval records"
            },
            {
              "title": {
                "zh-TW": "長篇分析回答改用短句分段，隱藏無回應標記並自動補問",
                "en": "Long analytical responses switch to short-sentence paragraphs, hiding no-reply markers with automatic follow-up questions"
              },
              "body": {
                "zh-TW": "閱讀 AI 產生的長篇分析報告時，密密麻麻的一整片文字往往讓人難以快速抓住重點。現在遇到長篇分析類型的回答，模型會主動採用短句與精簡段落進行排版，取代過去整面難以閱讀的文字牆。若先前對話中途意外中斷，使得某一輪的最終結果只剩下一行「前一輪未產生助手回覆」的內部標記，Muse 會自動發起一次有限度的重新提問，為你取得真正的實質回覆，並在執行摘要中主動剔除該標記。對話紀錄面板也不再顯示帶有菱形符號的「◆ [Previous turn ended without an assistant reply.]」提示列；如果回答只包含該標記，畫面不會渲染任何內容，若標記後方附有正常回答，則只乾淨呈現正式的回答內容。⟦恰似專業編輯重新編排文稿，不僅將密集論述切成清晰段落，也在印製時主動抽除排版留下的空白除錯標記。⟧",
                "en": "When reading long analytical reports generated by AI, dense walls of text often make it hard to quickly grasp key points. When encountering long analytical responses, the model now proactively formats output using short sentences and concise paragraphs, replacing the previously unreadable walls of text. If a previous conversation was interrupted unexpectedly so that a turn's final output was left with only an internal marker indicating that the prior turn produced no assistant reply, Muse automatically issues a bounded re-prompt to retrieve a genuine response, actively stripping that marker from the run summaries. The conversation history panel also no longer displays the prompt bar `◆ [Previous turn ended without an assistant reply.]` with its diamond symbol. If a response contains only that marker, the screen renders nothing, and if normal answers follow the marker, it cleanly presents only the formal response content. ⟦A professional editor reformats a manuscript by breaking dense arguments into clear paragraphs and stripping out leftover blank debugging marks before printing.⟧"
              },
              "original": "Long analysis answers use short sentences and paragraphs instead of dense walls of text\nA turn whose final answer consists only of the \"[Previous turn ended without an assistant reply.]\" marker now gets one bounded re-prompt for a real reply, and the marker is stripped from run summaries\nThe transcript no longer shows a \"◆ [Previous turn ended without an assistant reply.]\" row — a reply containing only that resume marker renders nothing, and a real answer prefixed by it shows just the answer"
            },
            {
              "title": {
                "zh-TW": "開發者指令阻擋改以白話引導，SDK 撤回狀態與技能範例修復",
                "en": "Developer command blocks switch to plain-language guidance, and SDK retraction states and skill examples receive fixes"
              },
              "body": {
                "zh-TW": "串接工具時，晦澀報錯常令人卡關。EdenFS 查全庫 Sapling（`hg files`、`hg locate`、`hg manifest`）卡死前阻擋，改指引 `search` 或 `hg biggrep`。衍生子行程遭拒改以白話說明，不再顯示內部列舉代碼。連接外部工具的 MCP 伺服器啟動存活未回，改報 `the server did not respond during startup` 而非泛用逾時。TypeScript 與 Python SDK 撤回遭拒不再誤標撤回或丟棄；前者的 `StaleDroppableMethod` 納入 `turn/retracted`，窮舉比對需補分支。內建 Three.js 技能改用可運作的 WebGPU 後製，不再引用已移除舊檔。",
                "en": "When integrating tools, cryptic error messages often cause roadblocks. Sapling repository-wide queries on EdenFS (`hg files`, `hg locate`, `hg manifest`) are blocked before hanging, redirecting users to `search` or `hg biggrep`. Spawned subprocess rejections are now explained in plain language instead of displaying internal enum codes. When an MCP (a protocol connecting AI to external tools) server fails to respond during startup liveness checks, it now reports `the server did not respond during startup` instead of a generic timeout. Retraction rejections in the TypeScript and Python SDKs are no longer mistakenly marked as retracted or dropped. The former's `StaleDroppableMethod` now incorporates `turn/retracted`, requiring an additional branch for exhaustive matching. The built-in Three.js skill now switches to working WebGPU post-processing instead of referencing removed legacy files."
              },
              "original": "Spawn rejection hints now explain the reason in plain words instead of internal enum names\nRepository-wide Sapling file listings (`hg files`, `hg locate`, `hg manifest`) on EdenFS checkouts are refused before they hang, pointing at `search` or `hg biggrep` instead\nAn MCP server whose process stays alive but never responds during startup now fails with `the server did not respond during startup` instead of the generic operation timeout\nThe TypeScript and Python SDKs no longer mark a turn retracted, or drop its completion, when its retract is declined; in the TypeScript SDK, `StaleDroppableMethod` gains `turn/retracted`, so code that matches it exhaustively needs a new case\nThe bundled Three.js skill now gives a working WebGPU post-processing example instead of one that imports a file current three.js no longer ships"
            }
          ]
        }
      ]
    },
    {
      "id": "usage",
      "name": "Usage",
      "versions": [
        {
          "version": "0.34.6",
          "period": "2026-10-23",
          "items": [
            {
              "title": {
                "zh-TW": "用量報表按派工額度分塊顯示，清楚列出對話消耗明細",
                "en": "Usage report splits dispatch quotas into separate blocks with per-conversation details"
              },
              "body": {
                "zh-TW": "當你頻繁指派 AI 處理各類任務時，常常需要確認不同模型的配額到底耗費在哪些對話上。派工帳本原本採用七欄表格，現在改為「派工額度花在哪（最近 7 天）」，替 Codex 與每個 Antigravity 模型群組各自獨立分塊。⟦跟翻閱專案的請款帳本一樣，以前所有項目的開銷全擠在一張大表格裡，既難以對照個別上限，加總甚至還會莫名破表；現在各預算項目獨立成頁，清楚列出金額最高的前幾筆花費，對帳時便能逐筆核實。⟧每個區塊都會標出本週已用額度，並列出消耗最多的前五個派工對話，其餘對話寫成「另外 N 段」，剩下的用量歸入「其他（你直接用的，或沒對到派工紀錄）」。Antigravity 每個群組各自計算週額度，不再合併相加，避免舊表格加總超過 100% 的情形；若 Codex 用量全在額度重置前發生，該區塊也會以一句話說明。",
                "en": "When you frequently assign AI different tasks, you often need to check which conversations have used up each model’s quota. The dispatch ledger’s seven-column table has now been replaced with “Where your dispatch quota went (last 7 days),” with separate blocks for Codex and each Antigravity model group. ⟦It works like reviewing a project’s expense ledger: previously, every item’s spending was crammed into one large table, making it hard to check against individual limits, and the totals could even inexplicably exceed the budget. Now each budget category has its own page listing the largest expenses, so you can verify them individually when reconciling the accounts.⟧ Each block shows the quota used this week and lists the top five conversations whose dispatched work used it. Remaining conversations appear as “N more,” and the remaining usage goes under “Other (your own use, or no dispatch record found).” Each Antigravity group’s weekly quota is calculated separately instead of being added together, preventing the old table’s totals from exceeding 100%. If all Codex usage occurred before the quota reset, its block also explains that in a sentence."
              },
              "original": "**The usage report shows where each dispatch quota went.** The dispatch ledger's seven-column table is now \"Where your dispatch quota went (last 7 days)\": one block per quota, for Codex and each Antigravity model group, with how much of it is used this week, the conversations whose dispatched work used it (top five, then \"N more\"), and the rest as \"Other (your own use, or no dispatch record found)\". Antigravity groups each have their own weekly quota and are no longer added together; the old table could total more than 100%. When all of Codex's usage came before its reset, the block says so in one line."
            },
            {
              "title": {
                "zh-TW": "Muse 週額度用盡正確歸屬，卡片依重置時間標記上限",
                "en": "Muse weekly quota exhaustion is correctly identified, with cards marking limits by reset time"
              },
              "body": {
                "zh-TW": "長時間使用 Muse 容易撞上用量限制，此時介面回報哪種額度見底，決定了它告訴你要等多久。Muse 的五小時額度與週額度用完時，都會回傳相同的「Subscription quota exhausted」（訂閱額度已耗盡）錯誤訊息。⟦像在手機門市查詢多個方案用量，以前不管哪種配額見底，簡訊都只通報短週期方案已滿，明明是週方案用盡卻標註三天後重置；現在若重置間隔拉長，介面便會精準指出真正用滿的是長週期配額，並如實標上恢復時間。⟧舊版 v0.34.5 一律將該錯誤判定為五小時額度耗盡，導致週額度用完時，卡片顯示「工作階段 100%、3 天後重置」，而週額度那列仍停在舊數字。現在只要重置時間超過 5 小時，卡片就會把週額度記錄為 100%，並帶上正確的重置時間。",
                "en": "Extended use of Muse can hit usage limits, and which quota the interface reports as exhausted determines how long it tells you to wait. Muse returns the same “Subscription quota exhausted” error message whether its five-hour or weekly quota runs out. ⟦Imagine checking usage across several plans at a mobile phone store. Previously, whichever quota ran out, the text message only reported that the short-cycle plan was full. Even when the weekly plan was exhausted, it showed a reset in three days. Now, if the reset interval is longer, the interface accurately identifies the longer-cycle quota as the one that is full and shows its actual recovery time.⟧ The old v0.34.5 always treated this error as five-hour quota exhaustion. When the weekly quota ran out, the card therefore showed “Session 100%, resets in 3 days,” while the weekly quota row still displayed its previous value. Now, whenever the reset is more than 5 hours away, the card records the weekly quota as 100% and includes the correct reset time."
              },
              "original": "**The Muse card records a used-up weekly allowance as the weekly limit.** Muse sends the same \"Subscription quota exhausted\" error for both limits, and v0.34.5 always recorded it as the 5-hour window. A used-up week showed \"Session 100%, resets in 3 days\" while the weekly row kept an old reading. A reset more than 5 hours away now marks the weekly row as 100% with that reset time."
            },
            {
              "title": {
                "zh-TW": "Wrapped 說明文字改為近一年，反映滾動 53 週數據",
                "en": "Wrapped description now says “over the past year” to reflect rolling 53-week data"
              },
              "body": {
                "zh-TW": "翻閱 Wrapped 的個人 AI 用量摘要時，介面標註的時間範圍會直接影響你對數據的理解。Wrapped 實際統計最近連續 53 週，說明文字現在寫成「近一年」，讓你知道這份回顧不是從今年一月一日起算。⟦好比查看運動手環的年累積步數，若沒看清楚統計期間，容易以為紀錄只從一月一日起算；標記成近一年，就能明白數據呈現的是最近一整年的連續走勢。⟧",
                "en": "When reviewing your personal AI usage summary in Wrapped, the time range shown in the interface directly affects how you interpret the data. Wrapped actually covers the most recent 53 consecutive weeks. Its description now says “over the past year,” making it clear that the review does not start on January 1 of the current year. ⟦Think of checking the yearly step total on a fitness tracker. If you overlook the reporting period, you might assume the record starts on January 1. Labeling it “over the past year” makes it clear that the data shows a continuous trend across the most recent full year.⟧"
              },
              "original": "**Wrapped captions say \"over the past year\".** The data covers a rolling 53 weeks, not the calendar year."
            }
          ]
        },
        {
          "version": "0.34.5",
          "period": "2026-10-21",
          "items": [
            {
              "title": {
                "zh-TW": "Muse 額度用完即使未給數字，卡片也會標示 5 小時已用滿",
                "en": "Cards mark the 5-hour quota as full even when Muse provides no numbers upon exhaustion"
              },
              "body": {
                "zh-TW": "當你連續使用 Muse 處理工作時，一旦額度耗盡，往往會想確認何時才能繼續。以往 5 小時額度一用完，Muse 通常會直接拒絕額度查詢，只拋出「Subscription quota exhausted」錯誤訊息並告知重置時間，導致卡片拿不到具體數字。⟦跟運動場地的租借機一樣，以前名額一被搶光，票機只印出額滿通知與下一場時間，看板卻因抓不到剩餘數值而沒能更新；現在螢幕會直接標示目前時段已全滿，並貼上下次開放進場的時刻。⟧現在卡片遇到這種狀況，會主動把 5 小時額度顯示為 100%，並附上該重置時間。因為這項錯誤回報裡沒有週額度數字，所以卡片上的週額度百分比會沿用上次的讀數，卡片也保留「上次查詢」的時間提示，讓你在等待額度恢復時依然能掌握先前的用量狀態。",
                "en": "When you use Muse continuously for work, you often want to confirm when you can resume once your quota runs out. Previously, as soon as the 5-hour quota was exhausted, Muse would typically reject quota queries outright, only throwing a \"Subscription quota exhausted\" error message alongside the reset time, leaving the card without concrete numbers. ⟦Much like a sports venue reservation kiosk, where the ticket machine previously printed only a full notice and the next session time once slots filled up, leaving the display unupdated because it could not fetch the remaining count, the screen now directly marks the current slot as completely full and displays the time for the next admission.⟧ Now when encountering this situation, the card proactively displays the 5-hour quota at 100% and attaches that reset time. Because this error response contains no weekly quota figures, the weekly quota percentage on the card retains its previous reading, and the card also preserves its \"last checked\" badge, allowing you to keep track of prior usage status while waiting for the quota to recover."
              },
              "original": "**The Muse card shows a used-up 5-hour window even when Muse sends no numbers.** Once the window runs out, Muse usually refuses the quota check with a \"Subscription quota exhausted\" error that only says when the window resets. The card now shows the window as 100% with that reset time. The error has no weekly figure, so the weekly percentage stays from the last reading and the card keeps its \"last checked\" badge."
            }
          ]
        },
        {
          "version": "0.34.4",
          "period": "2026-10-21",
          "items": [
            {
              "title": {
                "zh-TW": "Muse 的 5 小時額度用完後持續更新，用量卡片顯示已用滿",
                "en": "Muse continues updating after 5-hour quota is exhausted, with usage card showing full"
              },
              "body": {
                "zh-TW": "平時頻繁使用 Muse 協作時，如果額度見底，你可能會打開 `usage` 查看剩餘狀況與重置時間。⟦跟看停車場入口的剩餘車位看板一樣，以前場內車輛停滿甚至超額借停時，計數器因為讀到破表數字而停止刷新，螢幕凍結在用滿前一刻並掛著過期提示；現在就算車位全滿，看板也會持續更新並清楚顯示客滿。⟧Muse 在 5 小時額度耗盡時，回報的使用率常會超過 100%（例如 104%）。以前 `usage` 會直接把這種大於 100% 的讀數丟掉。這導致卡片一直停留在低於 100% 的最後一次讀數，並掛著過期標籤（stale-age badge，標示資料多久未更新的標籤），直到 5 小時週期重置。現在卡片在 5 小時額度用完後仍會持續刷新，並清楚標示為已用滿，方便你隨時掌握最新狀態。",
                "en": "When frequently collaborating with Muse, if your quota runs low, you might open `usage` to check the remaining status and reset time. ⟦Think of the vacancy board at a parking lot entrance. Previously, when the lot was full or even overfilled with parked cars, the counter stopped updating because it read an off-the-charts number, freezing the screen right before full capacity with an expired notice. Now, even when all spaces are full, the board continues to update and clearly displays that the lot is full.⟧ Muse exhausts its 5-hour quota, the reported usage often exceeds 100% (such as 104%). Previously, `usage` would directly discard readings greater than 100%. This caused the card to remain stuck at the last reading below 100% with a stale-age badge (a label indicating how long data has not been updated) until the 5-hour cycle reset. Now, the card continues to refresh after the 5-hour quota is exhausted and clearly indicates that it is full, making it easy for you to track the latest status at any time."
              },
              "original": "**The Muse card stays current after a 5-hour window runs out.** Muse reports an exhausted window as more than 100% used (for example 104%), and `usage` threw those readings away. The card kept the last reading under 100% with a stale-age badge until the window reset; it now shows the window as full."
            }
          ]
        }
      ]
    },
    {
      "id": "gh_cli",
      "name": "GitHub CLI",
      "versions": [
        {
          "version": "2.103.0",
          "period": "2026-10-21",
          "items": [
            {
              "title": {
                "zh-TW": "`gh` 可改用會自動續期的登入憑證（預覽）",
                "en": "`gh` can sign in with auto-refreshing credentials (preview)"
              },
              "body": {
                "zh-TW": "一般用 `gh` 登入 GitHub，拿到的 token（證明你身分的授權令牌）發下來就不會過期。現在可以改選會過期的 OAuth 憑證：`gh` 會存一把短效的 access token，另配一把會輪替的 refresh token（專門用來換新令牌的憑證）。access token 需要時會自動換新，你不必重新登入。⟦公司門禁卡改成每天自動換發的臨時卡，以前一張卡用到離職，現在卡片天天換新，你照樣每天刷卡進門。⟧這是預覽功能，要自己選擇開啟。只想這一次登入用它，打：\n\n```shell\ngh auth login --expiring\n```\n\n想以後每次登入都優先用它，打：\n\n```shell\ngh config set expiring_credentials enabled\n```",
                "en": "Normally, signing in to GitHub with `gh` gives you a token (the pass that proves who you are) that never expires. You can now opt into expiring OAuth credentials instead: `gh` stores a short-lived access token alongside a rotating refresh token (a credential used only to get a new pass). The access token is refreshed automatically whenever needed, so you stay signed in without authenticating again. ⟦An office swapped its permanent badges for temporary ones reissued every day: you used to keep one badge until you left the company, and now it renews daily while you still just tap in each morning.⟧ This is a preview you have to opt into. To use it for a single sign-in, run:\n\n```shell\ngh auth login --expiring\n```\n\nTo prefer expiring credentials for every sign-in, run:\n\n```shell\ngh config set expiring_credentials enabled\n```"
              },
              "original": "## Expiring, auto-refreshing OAuth tokens (preview)\n\n`gh` can now sign in with expiring OAuth credentials that it refreshes automatically. When you opt in, `gh` stores a short-lived access token alongside a rotating refresh token and refreshes the access token whenever it is needed, so you stay signed in without having to authenticate again.\n\nTo opt in for a single sign-in, run:\n\n```shell\ngh auth login --expiring\n```\n\nTo prefer expiring credentials for every sign-in, run:\n\n```shell\ngh config set expiring_credentials enabled\n```"
            },
            {
              "title": {
                "zh-TW": "用會過期的 `gh` 憑證時，查狀態、接 git 和寫腳本要注意的事",
                "en": "Living with expiring `gh` credentials: status, git and scripts"
              },
              "body": {
                "zh-TW": "開了自動續期後，想知道憑證還剩多久，就打 `gh auth status`：它會顯示憑證能不能續期、access token 何時到期。它和 `gh auth token` 都會在需要時自動續期。github.com 和 GHE.com 支援這種 token，其他主機會退回一般不過期的 token。git 走 HTTPS 的話，建議用 git 2.46 以上，並執行 `gh auth setup-git`，讓 `gh` 當 git 的 credential helper（替 git 提供登入資料的程式）。這樣 git 不會把憑證留到過期之後。⟦跟鮮奶一樣，以前能囤一箱慢慢喝，現在得隨喝隨買。⟧腳本和擴充別把 `gh auth token` 的輸出存起來，每次要用都重新執行它；透過 `GH_TOKEN` 或 `GITHUB_TOKEN` 給的 token 則永遠不會續期。",
                "en": "Once expiring credentials are on, you may want to know how long yours has left. `gh auth status` shows whether a credential is refreshable and when the current access token expires, and both `gh auth status` and `gh auth token` refresh the access token automatically when needed. Expiring tokens are available on github.com and GHE.com; on hosts that do not issue them, `gh` falls back to a regular non-expiring token. If git talks to GitHub over HTTPS, the advice is to use git 2.46 or newer and run `gh auth setup-git` so `gh` acts as git's credential helper (the program that hands git its sign-in details). That way git does not cache the credential past its lifetime. ⟦Treat it like fresh milk: you could once stock up a crate and drink it slowly, and now you buy it as you need it.⟧ So scripts and extensions should not store the output of `gh auth token`; run it again each time a token is needed. Tokens supplied through `GH_TOKEN` or `GITHUB_TOKEN` are never refreshed."
              },
              "original": "`gh auth status` shows whether a credential is refreshable and when the current access token expires. Both `gh auth status` and `gh auth token` automatically refresh the access token when needed. Expiring tokens are available on github.com and GHE.com; on hosts that do not issue them, `gh` falls back to a regular non-expiring token.\n\nIf you use HTTPS as your git protocol, we advise using git 2.46 or newer and letting `gh` act as your git credential helper by running `gh auth setup-git`. This way `gh` serves and refreshes the expiring credential for git, and git does not cache it past its lifetime.\n\n> [!IMPORTANT]\n> Expiring access tokens should not be cached. Scripts and extensions that store the output of `gh auth token` will eventually hold an expired token, so run `gh auth token` again each time a token is needed. Tokens supplied through `GH_TOKEN` or `GITHUB_TOKEN` are never refreshed.\n\nFor more information, see [Authenticating with GitHub CLI](https://docs.github.com/github-cli/github-cli/authenticating-with-github-cli#using-expiring-refreshable-tokens)."
            },
            {
              "title": {
                "zh-TW": "新指令 `gh issue artifact`，在終端機管理 issue 的附件文件（私人預覽）",
                "en": "New `gh issue artifact` manages issue artifacts from the terminal (private preview)"
              },
              "body": {
                "zh-TW": "規劃一個新功能時，你可能想把規劃書、流程圖或網頁原型直接掛在 GitHub issue（記錄待辦與問題的討論串）上。`gh issue artifact` 把 issue artifacts 帶到命令列，提供 `list`、`view`、`create`、`edit`、`download`、`delete` 六個指令，例如：\n\n```shell\n# Create a Markdown document with an attached image\ngh issue artifact create 142 signin-plan.md --attach ./signin-flow.png\n\n# Create an HTML document\ngh issue artifact create 142 menu-prototype.html\n\n# List or view artifacts\ngh issue artifact list 142\ngh issue artifact view 142 2\n\n# Update a document or restore an earlier version\ngh issue artifact edit 142 2 --body-file signin-plan-v2.md\ngh issue artifact edit 142 2 --restore-version 5\n```\n\n⟦替每張工單配一個專屬資料夾，規劃書和原型都收在裡面，還記得每一版，想退回第 5 版也行。⟧artifacts 也能用 GitHub 網址指定，Markdown 文件可以用 `--attach` 附上圖片和影片。這是私人預覽，只有加入預覽的使用者和組織能用，而且只對加入預覽的儲存庫有效；申請方式在官方的預覽公告討論串。",
                "en": "When planning a new feature, you may want to hang the plan, a flow chart or a web prototype right on a GitHub issue (the thread that tracks a task or problem). `gh issue artifact` brings issue artifacts to the command line with six commands, `list`, `view`, `create`, `edit`, `download` and `delete`, for example:\n\n```shell\n# Create a Markdown document with an attached image\ngh issue artifact create 142 signin-plan.md --attach ./signin-flow.png\n\n# Create an HTML document\ngh issue artifact create 142 menu-prototype.html\n\n# List or view artifacts\ngh issue artifact list 142\ngh issue artifact view 142 2\n\n# Update a document or restore an earlier version\ngh issue artifact edit 142 2 --body-file signin-plan-v2.md\ngh issue artifact edit 142 2 --restore-version 5\n```\n\n⟦Give every work ticket its own folder that holds the plan and the prototype and remembers every version, so rolling back to version 5 is easy.⟧ Artifacts can also be addressed by their GitHub URLs, and Markdown documents take image and video attachments with `--attach`. This is a private preview: only users and organizations enrolled can use it, and the commands work only for repositories in the preview; access instructions are in the official preview announcement discussion."
              },
              "original": "## Manage issue artifacts with `gh issue artifact` (private preview)\n\n`gh issue artifact` brings issue artifacts to the command line. Users and organizations enrolled in the private preview can use the `list`, `view`, `create`, `edit`, `download`, and `delete` commands to manage issue artifacts.\n\n```shell\n# Create a Markdown document with an attached image\ngh issue artifact create 142 signin-plan.md --attach ./signin-flow.png\n\n# Create an HTML document\ngh issue artifact create 142 menu-prototype.html\n\n# List or view artifacts\ngh issue artifact list 142\ngh issue artifact view 142 2\n\n# Update a document or restore an earlier version\ngh issue artifact edit 142 2 --body-file signin-plan-v2.md\ngh issue artifact edit 142 2 --restore-version 5\n```\n\nArtifacts can also be addressed by their GitHub URLs, and Markdown documents support image and video attachments with `--attach`.\n\n> [!IMPORTANT]\n> Issue artifacts are in private preview. These commands only work for repositories enrolled in the preview. See the [Issue Artifacts private preview announcement and feedback discussion](https://github.com/orgs/community/discussions/205881) for access instructions and to share feedback."
            }
          ]
        },
        {
          "version": "2.102.0",
          "period": "2026-09-30",
          "items": [
            {
              "title": {
                "zh-TW": "修復四處安全漏洞，防範非預期檔案寫入與驗證被繞過",
                "en": "Fix Four Security Vulnerabilities to Prevent Unexpected File Writes and Verification Bypasses"
              },
              "body": {
                "zh-TW": "下載、安裝技能或驗證產物時不再受惡意名稱所害，建議盡快升到 `v2.102.0`。`gh release download`、`gh run download`、`gh repo read-file --output` 與 `gh attestation download` 遇目的地含符號連結（指向其他位置的捷徑檔）時，以前可能寫入非預期的本機檔案。`gh attestation verify` 以前比對 `--source-ref` 不分大小寫，且 `--signer-workflow` 只比憑證身分開頭，僅大小寫不同的分支，或同儲存庫內路徑開頭相同的另一個工作流程，都可能通過驗證。互動式 `gh skill search` 傳路徑給 `gh skill install` 時缺少選項分隔符號，搜尋結果可能注入安裝選項並改寫檔案位置。⟦就像同一棟樓四扇門的鎖被一次補好。⟧",
                "en": "Downloading artifacts, installing skills, or verifying attestations is no longer vulnerable to malicious names, and upgrading to `v2.102.0` as soon as possible is recommended. When the destination contains a symlink, `gh release download`, `gh run download`, `gh repo read-file --output`, and `gh attestation download` previously risked writing to unintended local files. `gh attestation verify` previously compared `--source-ref` case-insensitively and only matched the prefix of certificate identities for `--signer-workflow`, allowing branches differing only by case or another workflow with a matching path prefix in the same repository to pass verification. Interactive `gh skill search` lacked an option delimiter when passing paths to `gh skill install`, which allowed search results to inject installation options and alter file locations. ⟦Picture locks on four doors of the same building being fixed all at once.⟧"
              },
              "original": "## Security\n\nFour security vulnerabilities have been identified, and fixed, in this release. Users are advised to update gh to version `v2.102.0` as soon as possible.\n\n`gh release download`, `gh run download`, `gh repo read-file --output`, and `gh attestation download` could write remote content to an unintended local file when the destination contained symbolic links.\n\nSee https://github.com/cli/cli/security/advisories/GHSA-39wj-f2f4-978v for more information.\n\n`gh attestation verify` compared the `--source-ref` value case-insensitively, so an attestation built from a branch whose name differs only in case could satisfy a policy that named a different branch.\n\nSee https://github.com/cli/cli/security/advisories/GHSA-4mq3-hpgx-9cx8 for more information.\n\nInteractive `gh skill search` passed repository paths from search results to `gh skill install` without an option separator, so a search result could inject installer options and change where skill files were written.\n\nSee https://github.com/cli/cli/security/advisories/GHSA-qcwj-mr2r-2cx7 for more information.\n\n`gh attestation verify` matched the `--signer-workflow` value against only the start of the signing certificate's identity, so an attestation signed by a different workflow in the pinned repository could pass verification when its path began with the pinned value.\n\nSee https://github.com/cli/cli/security/advisories/GHSA-wjmr-j3rp-mh2g for more information."
            }
          ]
        },
        {
          "version": "2.101.0",
          "period": "2026-09-15",
          "items": [
            {
              "title": {
                "zh-TW": "Linux 上用 APT 或 RPM 裝的 gh 要認新金鑰",
                "en": "gh installed via APT or RPM on Linux needs the new signing key"
              },
              "body": {
                "zh-TW": "你在 Linux 上更新 `gh`，卻突然裝不起來、報簽章驗證失敗？多半是金鑰換了。GitHub CLI 的 APT 與 RPM 套件庫（Linux 下載與更新軟體的官方來源），連同個別的 RPM 套件檔，現在只用新的 PGP 金鑰簽署，指紋是 `7F38BBB59D064DBCB3D84D725612B36462313325`。如果你在 2026 年 4 月 8 日前從官方 APT 或 RPM 套件庫裝了 `gh`，又還沒更新金鑰圈（電腦裡記著哪些簽章可以信任的清單），安裝或更新就可能失敗。⟦就像快遞改用新封條，你家警衛只認得舊封條，包裹就會被擋在門外，得先把新封條的樣子告訴警衛。⟧照官方的 PGP 簽章金鑰輪換說明檢查、更新金鑰圈就能解決。用 Windows、macOS，或透過 Homebrew、GitHub Releases、其他套件管理器、原始碼安裝的人不受影響。",
                "en": "Updating `gh` on Linux and suddenly getting a signature verification failure? The signing key most likely changed. GitHub CLI's APT and RPM repositories, along with individual RPM packages, are now signed only with the new PGP key, fingerprint `7F38BBB59D064DBCB3D84D725612B36462313325`. If you installed `gh` from the official APT or RPM repositories before April 8, 2026 and haven't refreshed your keyring, installs and updates may fail. ⟦It's like a courier switching to a new seal: your building's guard only knows the old one, so packages get turned away until you show the guard what the new seal looks like.⟧ Follow the official PGP signing key rotation guidance to check and update your keyring. Users on Windows or macOS, and anyone who installed through Homebrew, GitHub Releases, another package manager, or from source, are not affected."
              },
              "original": "## Linux package repository signing key rotation\n\nGitHub CLI's APT and RPM repositories, along with individual RPM packages, are now signed only with the new PGP key (fingerprint: `7F38BBB59D064DBCB3D84D725612B36462313325`)\n\n> [!IMPORTANT]\n> Linux users who installed `gh` from the official APT or RPM repositories before April 8, 2026, and have not refreshed their keyring, may now see package installation or update failures. Follow the [PGP signing key rotation guidance](https://github.com/cli/cli/issues/13118) to check and update your keyring. Users on Windows or macOS, and users who installed `gh` through Homebrew, GitHub Releases, another package manager, or from source, are not affected."
            },
            {
              "title": {
                "zh-TW": "gh 登入時自動幫你複製驗證碼",
                "en": "gh now copies your login code for you"
              },
              "body": {
                "zh-TW": "用 `gh auth login` 登入 GitHub，或用 `gh auth refresh` 更新權限時，畫面會給你一組 OAuth 裝置代碼（要貼到瀏覽器授權頁、證明是你本人在登入的短碼）。以前你得自己反白、複製，再切到瀏覽器貼上；現在這兩個指令預設會直接把代碼放進剪貼簿，切過去按貼上就好。⟦櫃檯人員直接把號碼牌塞到你手上，你就不必再拿筆抄一次號碼。⟧不想要這個行為，可以用下面這行永久關掉：\n\n```shell\ngh config set clipboard disabled\n```\n\n某一次執行時如果你另外指定了剪貼簿選項，那一次就照你指定的來。",
                "en": "When you sign in with `gh auth login` or refresh permissions with `gh auth refresh`, the screen gives you an OAuth device code to paste into the browser's authorization page. You used to highlight it, copy it, then switch to the browser and paste. Now both commands put the code on your clipboard by default, so you just switch over and paste. ⟦Think of a service desk where the clerk hands you the ticket instead of making you copy the number down with a pen.⟧ To turn this off for good, run:\n\n```shell\ngh config set clipboard disabled\n```\n\nIf you pass an explicit clipboard option on a single run, that run follows your choice."
              },
              "original": "## Copy authentication codes to the clipboard by default\n\n`gh auth login` and `gh auth refresh` now copy OAuth device codes to the clipboard by default, saving a manual copy step during authentication.\n\nTo persistently opt out, run:\n\n```shell\ngh config set clipboard disabled\n```\n\nExplicit clipboard options still take precedence for an individual invocation."
            }
          ]
        }
      ]
    }
  ]
}