Changes on 2026-09-15

1. **Analyze the Request:** * **Role:** World-class programmer/analyst. * **Input:** Diff for `sandboxing.md`. * **Output:** Markdown format analysis in Chinese. * **Structure:** Overall Summary, Key Themes, Impact Level, Action Items. 2. **Analyze the Diff (`sandboxing.md`):** * *Change 1 (Line 6-7):* Adds "PowerShell, or Monitor" to the list of tools covered by the Bash sandbox enforcement. * *Change 2 (Line 40-46):* Refines the description of network domain handling in auto mode. It now specifies that Claude names the hosts *on the command itself* (referring to a new feature "per-command allowed domains") instead of just sending a request to a classifier. * *Change 3 (Line 139-141):* Updates the `<Info>` box about auto-allow mode. Adds an exception: "in auto mode, for a command that carries per-command allowed domains". * *Change 4 (Line 517-523):* Refines "Domain restrictions" description again. Clarifies that in auto mode, Claude names hosts on the command itself. Splits the previous paragraph into separate bullet points for clarity ("Approval choices", "Pre-allowed domains"). 3. **Synthesize the Changes:** * **Core Topic:** Improvements to sandboxing, specifically regarding network domain handling in "auto mode" and support for additional shell tools. * **New Feature/Behavior:** "Per-command allowed domains in auto mode" (or similar concept). The change suggests that in auto mode, the allowed domains are now explicitly attached to or named on the command execution context rather than just being a separate classification request. * **Scope Expansion:** The sandboxing enforcement now officially supports `PowerShell` and `Monitor` commands in addition to `Bash`. 4. **Draft the Analysis (Mental Draft -> Translation to Chinese):** * *Overall Summary:* The documentation updates clarify how sandboxing handles network domains in "auto mode" (specifically mentioning per-command allowed domains) and expands the scope of sandbox enforcement to include PowerShell and Monitor commands. * *Chinese:* 文档更新阐明了沙箱在“自动模式”下处理网络域的方式（特别是提及“每个命令允许的域”），并将沙箱强制执行的范围扩展到了 PowerShell 和 Monitor 命令。 * *Key Themes:* * Expanded tool support (PowerShell/Monitor). * Refinement of Auto Mode network handling (Self-describing commands). * Clarification of Auto-allow logic exceptions. * *Chinese:* * 扩展了沙箱覆盖范围，增加了 PowerShell 和 Monitor 命令的支持。 * 优化了“自动模式”下的网络权限处理机制，引入了“每个命令允许的域”的概念，允许命令自身声明所需的主机。 * 细化了“自动允许”模式的行为说明，明确了自动模式下的例外情况。 * *Impact Level:* Low/Medium. * *Reasoning:* These are mostly behavioral refinements and expansions of existing features. It clarifies how things work rather than breaking them. Developers using PowerShell or auto-mode will notice improved consistency. * *Chinese:* 中等。主要是行为优化和功能范围扩充，而非破坏性变更。PowerShell 用户和自动模式用户会体验到更一致的行为。 * *Action Items:* * Review auto-mode behavior if you rely on it. * Note that PowerShell/Monitor commands are now subject to the same sandbox rules as Bash. * *Chinese:* * 如果使用 PowerShell 或 Monitor 工具，请注意它们现在也遵循与 Bash 相同的沙箱隔离规则。 * 如果启用自动模式，请留意网络域请求现在会在命令本身进行声明，而非单独的请求流程。 5. **Refine and Format (Final Polish):** * *Summary:* 文档更新将沙箱隔离支持扩展至 PowerShell 和 Monitor 命令，并详细阐明了“自动模式”下网络域的批准机制，特别是引入了“每个命令允许的域”这一新特性。 * *Themes:* PowerShell/Monitor 支持、自动模式下的网络域处理、文档结构优化。 * *Impact:* Low. 文档澄清与功能增强。 * *Action Items:* 1. 注意 PowerShell/Monitor 的沙箱行为。 2. 了解自动模式下的网络权限变更。 6. **Final Output Generation:** (Proceed to generate markdown output in Chinese).

1. **Analyze the Request:** * **Role:** World-class programmer analyzing documentation changes. * **Task:** Analyze a batch of 34 documentation changes (though the input provided only diffs for specific pages: `amazon-bedrock.md`, `checkpointing.md`, `claude-code-on-the-web.md`, `cli-reference.md`, `data-usage.md`, `desktop.md`, `devcontainer.md`, `discover-plugins.md`, `github-actions.md`, `google-vertex-ai.md`, `hooks.md`, `iam.md`, `jetbrains.md`, `mcp.md`, `microsoft-foundry.md`, `model-config.md`, `network-config.md`, `overview.md`, `plugin-marketplaces.md`, `plugins-reference.md`, `sandboxing.md`, `security.md`, `settings.md`, `setup.md`, `skills.md`). * **Output Format:** Markdown, Chinese language, concise, insightful. * **Structure:** 1. Overall Summary (1-2 sentences). 2. Key Themes (bullets). 3. Impact Level (Low/Medium/High/Breaking) with reasoning. 4. Action Items for Developers (bullets, max 5). 2. **Analyze the Diffs:** * *General Pattern:* The most significant change is the rebranding/terminology shift from "Claude Code on the web" to "Claude Code in the cloud" or simply "cloud sessions". This appears in almost every file. * *Minor Fixes:* Typographical fixes (e.g., "Navigate" -> "Go", "bash" -> "Bash", capitalization of specific terms). * **Page-by-Page Analysis:** * `amazon-bedrock.md`: Changed "Navigate to" to "Go to". Added anchor links for "AWS credentials", "IAM permissions", etc. (Minor usability improvement). * `checkpointing.md`: "bash commands" -> "Bash commands". (Capitalization consistency). * `claude-code-on-the-web.md`: **Major Rename**. Title changed from "Use Claude Code on the web" to "Use Claude Code in the cloud". Content updated to emphasize "cloud sessions" running on infrastructure vs "local sessions". Mentions starting sessions from browser, mobile, desktop, terminal, routines. Changed terminology like "web and terminal" to "terminal and cloud". * `cli-reference.md`: "web session" -> "cloud session". * `data-usage.md`: "Claude Code on the web" -> "cloud sessions". Updated wording. * `desktop.md`: Updated section titles to match "cloud sessions". "remotely" -> "in the cloud". * `devcontainer.md`: Updated `DISABLE_AUTOUPDATER` instruction. "Claude.ai" -> "claude.ai" (brand consistency). * `discover-plugins.md`: Clarified error message instructions. * `github-actions.md`: "Claude Code on the web" -> "Claude Code in the cloud". * `google-vertex-ai.md`: "Navigate to" -> "Go to". * `hooks.md`: "Claude Code on the web" -> "cloud sessions". Added new `PostToolUse` hook functionality (`bashEditDiff` for tracking file changes via Bash). This is a new feature/documentation. * `iam.md`: "Claude.ai" -> "claude.ai". "Claude Code on the Web" -> "Cloud sessions". * `jetbrains.md`: "bash execution" -> "Bash execution". * `mcp.md`: Clarified MCP server setup instructions (adding options between `--env` and server name). Updated variable expansion documentation. Added section on how references appear in `/mcp`. * `microsoft-foundry.md`: "Navigate to" -> "Go to". * `model-config.md`: "Claude Code on the web" -> "cloud sessions". Added clarification about server-side rejection. * `network-config.md`: "Claude Code on the web" -> "cloud sessions". * `overview.md`: Updated terminology. Links updated (`from-terminal-to-web` -> `from-terminal-to-cloud`). * `plugin-marketplaces.md`: Formatting changes (tables). Clarified path resolution. Clarified "See below" references. * `plugins-reference.md`: Clarified `defaultEnabled` precedence. * `sandboxing.md`: "narrower path wins" -> "rule with the narrower path applies". "Claude.ai" -> "claude.ai". * `security.md`: "Claude Code on the web" -> "cloud sessions". * `settings.md`: "Claude Code on the web" -> "cloud session". Refined text about settings. * `setup.md`: "Claude.ai" -> "claude.ai". * `skills.md`: "remote session" -> "cloud session". Added significant detail on permission checks for injected commands (especially in Auto mode). 3. **Synthesize Key Themes:** * **Terminology Rebranding:** The shift from "Claude Code on the web" to "cloud sessions" / "in the cloud" is ubiquitous. This aligns the naming with the fact that these sessions can be started from desktop, terminal, mobile, etc., not just the web interface. * **Brand Consistency:** Standardizing on lowercase "claude.ai" instead of "Claude.ai". * **New Features/Documentation:** * `hooks.md`: New `bashEditDiff` in `PostToolUse` to track file changes made via Bash commands. * `skills.md`: Detailed explanation of permission checks for injected commands (specifically how Auto mode handles them vs interactive modes). * **Minor Usability Fixes:** "Navigate to" -> "Go to", specific capitalization fixes ("Bash"), better link anchors. 4. **Determine Impact Level:** * *Rebranding:* Low technical impact, medium documentation impact. Users need to know the new terms to search docs. * *New Features (Hooks/Skills):* Medium to High impact for developers using those specific advanced features. * *Overall:* **Low to Medium**. It's mostly terminology alignment, but the new hook data and skills permission logic are important for power users. 5. **Formulate Action Items:** * Update internal documentation/scripts referencing "Claude Code on the web". * Review `bashEditDiff` if using PostToolUse hooks to track file changes. * Understand new permission behavior for injected commands in Skills (especially regarding Auto mode). * Update links/anchors if relying on old URL fragments (though ID changes aren't explicitly shown as breaking, the text implies reorganization). 6. **Drafting the Response (in Chinese):** * *Summary:* 本次文档更新主要涉及将 "Claude Code on the web" 统一重命名为 "cloud sessions"（云端会话），并统一了品牌术语（如 `claude.ai`）。此外，还对 Hooks（特别是 Bash 文件变更追踪）和 Skills 的权限检查逻辑进行了功能增强和文档细化。 * *Themes:* * **术语重命名**：将 "Claude Code on the web" 全局替换为 "cloud sessions" 或 "in the cloud"，以反映会话可从多端（终端、桌面、移动端）启动的特性。 * **品牌规范**：统一使用小写 `claude.ai`，修正大写 `Claude.ai` 及相关措辞。 * **Hooks 功能增强**：在 `PostToolUse` 中新增 `bashEditDiff` 字段，用于捕获通过 Bash 命令修改的文件列表。 * **Skills 权限逻辑**：详细说明了注入命令在 `auto mode` 和其他模式下的权限检查行为。 * **文档细节优化**：增加了内部链接锚点，修正了 "Navigate to" 等指令措辞，统一了 "Bash" 的大小写。 * *Impact:* **Low to Medium**。大部分是术语重命名，不影响现有代码功能。但对于依赖 Hooks 追踪文件变更或深入使用 Skills 权限控制的开发者，新文档提供了重要的实现细节。 * *Action Items:* * 检查并更新内部文档或脚本中对 "Claude Code on the web" 的引用，改用 "cloud sessions"。 * 如果使用 `PostToolUse` Hooks，利用新增的 `bashEditDiff` 字段来更准确地追踪 Bash 命令带来的文件变更。 * 重新审视 Skills 中的注入命令权限逻辑，特别是在 Auto 模式下的行为变化，以避免意外的调用中断。 * 更新任何可能受 URL 结构影响的文档链接。 7. **Review and Refine:** * Make sure the tone is professional ("World-class programmer"). * Ensure the translation is natural. * Check against the provided diffs to ensure no hallucinated features. (Checked: `bashEditDiff` is in `hooks.md`, Auto mode logic in `skills.md`, MCP var expansion clarifications). 8. **Final Output Generation.** (Matches the drafted content).

1. **Analyze the Request:** * **Role:** World-class programmer with expertise in developer tools/documentation. * **Task:** Analyze a specific documentation diff (single page: `computer-use-tool.md`). * **Format:** Markdown, concise, insightful. * **Language:** Chinese (as per "请用中文回复"). * **Structure:** Overall Summary, Key Themes, Impact Level, Action Items. 2. **Analyze the Diff:** * **File:** `agents-and-tools/tool-use/computer-use-tool.md` * **Context:** Documentation regarding "Computer Use" tool, specifically security features related to prompt injection. * **Changes:** * *Old text:* Classifiers run on *prompts* to flag prompt injections. When identified in *screenshots*, model asks for *user confirmation*. * *New text:* Classifiers automatically scan *what the tools return* (e.g., screenshots) to flag prompt injections. When identified, model *checks whether the instruction really came from you* before acting. * *Ordering:* The paragraph about opting out (contacting support) was moved slightly and the phrasing about "precautions remain important" was rephrased to "precautions above remain important even with these classifiers in place". 3. **Synthesize the Meaning:** * The core change is a refinement of how the "prompt injection defense" works. It's not just scanning the user prompt; it's scanning the *output* of the tools (like screenshots). * The behavior changed from "ask for user confirmation" to "verify the instruction source" (essentially the same safety mechanism but described more generically or accurately as checking source attribution). * The documentation emphasizes that the classifier scans *tool outputs* (screenshots), not just the initial user prompt. 4. **Draft the Analysis (Internal Monologue/Drafting in Chinese):** * **Summary:** 更新了关于对抗提示词注入的安全防护机制描述，明确了分类器扫描的是工具返回的内容（如截图）并验证指令来源。 * **Themes:** * Security/Safety (Prompt Injection). * Clarification of mechanism (Scanning tool outputs vs. prompts). * Opt-out availability. * **Impact:** Low/Medium. It's a documentation update clarifying existing security behavior, not a breaking API change. It helps developers understand *why* the agent might pause or verify instructions. * **Action Items:** * None strictly required (code doesn't change). * Developers should be aware of this automatic behavior when building agents without human-in-the-loop (since the text mentions opting out). * Remind users about the limitations. 5. **Refine the Output (Formatting as requested):** * *Overall Summary:* Anthropic 更新了关于计算机使用工具安全机制的文档，明确指出分类器会自动扫描工具返回的内容（如截图）以检测提示词注入，并在检测到潜在风险时指导模型验证指令来源。 * *Key Themes:* * **安全机制澄清**：明确了扫描对象不仅仅是提示词，还包括工具返回的内容（screenshots）。 * **行为描述优化**：将“询问用户确认”更新为“检查指令是否真的来自用户”，表述更精准。 * **配置选项**：重申了针对无人值守场景的退出机制。 * *Impact Level:* **Low**。这是对现有安全防护行为的描述性更新，不涉及 API 接口变更，开发者无需修改代码。 * *Action Items:* * 了解并预期模型在处理受污染的截图或工具输出时会触发验证行为。 * 如果您的应用场景是无人值守且不需要此额外检查，考虑联系支持以关闭该功能。 6. **Final Polish (Professional Tone):** Ensure the tone is authoritative yet helpful. * *Summary check:* Did I capture the "what tools return" nuance? Yes. * *Impact check:* Is it low? Yes, it's a doc change about an internal model behavior. * *Action items:* Keep it relevant to dev workflow (testing, planning). 7. **Final Output Generation** (translating thoughts to the final response block).

1. **Understand the Goal:** The user wants a concise, insightful analysis of documentation changes provided in a diff. The persona is a world-class programmer, and the output must be in Chinese (as requested: "请用中文回复"). The format is specified: Overall Summary, Key Themes, Impact Level, and Action Items. 2. **Analyze the Diff:** * **Context:** Documentation for "Claude Code" plugin marketplaces, specifically regarding private repositories and organization sync. * **Change 1 (Paragraph):** Refined the explanation of how organization sync handles git credentials. It removed the specific mention of "Claude GitHub App" or "organization's GitHub Enterprise App" as the sole methods and generalized it to "organization's GitHub or GitLab connection on claude.ai". * **Change 2 (Source Rules List):** * Clarified that the marketplace repo must be private/internal on github.com/gitlab.com. * Detailed the specific connection method per host (github.com -> Claude GitHub App, GHE -> Enterprise App, GitLab -> Access Token). * Expanded the cases where a plugin source can be private. It added a specific case for GitLab (url/git-subdir source on the same host, matching group/user). * Updated the "must be public" rule to be more restrictive: "Any other plugin source... organization sync rejects plugin sources on hosts these rules don't cover." * **Change 3 (New Section):** Added a completely new section: "Sync a GitLab-hosted marketplace". * Explains that an Owner must add a GitLab configuration in Org Settings. * Notes GitLab integration is in "public beta". * Specifies URL requirements (HTTPS). * Mentions nested subgroups work. * Clarifies sync triggers (default branch only). 3. **Synthesize the Summary:** * *What changed?* The documentation now formally supports GitLab-hosted plugin marketplaces for organizations (public beta), clarifies authentication mechanisms for different Git hosts (GitHub vs. GitLab), and refines rules regarding which plugin sources can be private. * *Why it matters?* Developers can now use GitLab for their internal plugin distribution, and there's clearer guidance on permissions and private source access across different platforms. 4. **Identify Key Themes:** * **GitLab Support:** Major new feature (Public Beta) for syncing marketplaces from GitLab. * **Authentication Clarification:** Explicitly distinguishes how credentials are handled for GitHub vs. GitLab (App vs. Access Token). * **Privacy Rules:** Stricter/More explicit rules on which plugin sources can be private vs. public. 5. **Determine Impact Level:** * *Level:* Medium to High (let's go with Medium for general docs, but High for users relying on GitLab). The prompt asks for Low/Medium/High/Breaking. * *Reasoning:* It's not "Breaking" because it adds functionality and clarifies existing rules rather than removing them. However, it's significant for organizations using GitLab. Let's classify it as **Medium** (feature addition + clarification) or **High** if the user is a GitLab shop. Given the "public beta" tag, Medium-High feels right. I'll stick with **Medium** because it's documentation for a new beta feature and clarification, not a forced migration, but the unlock of GitLab is valuable. Actually, for GitLab users, this is *High* impact. I'll rate it **Medium** generally but note the significance for GitLab users. Let's settle on **Medium** because it requires configuration changes by admins to utilize, but it's a net positive feature addition. 6. **Draft Action Items:** * Configure GitLab connections if they want to sync from GitLab. * Review privacy rules for plugin sources if they are using mixed hosts (e.g., private GitHub repo with public plugins elsewhere). * Check URLs for GitLab marketplaces (use HTTPS). * Ensure nested subgroup support is leveraged if needed. 7. **Translate and Format (Chinese):** * *Summary:* 文档更新正式增加了对 GitLab 托管插件市场的支持（公开测试版），并详细阐明了在不同 Git 托管平台上的身份验证机制和私有源访问规则。 * *Themes:* * GitLab 集成与配置 * 多平台身份验证明确化 * 私有源访问权限规则细化 * *Impact:* **Medium**. Reasoning: 这是一个功能增强（支持 GitLab）和文档澄清，虽然对 GitLab 用户影响重大，但不会破坏现有工作流。 * *Action Items:* * 如果使用 GitLab：管理员需在组织设置中添加 GitLab 配置。 * 检查插件源：确保不支持的私有源已公开化，或符合新的同源规则。 * 更新 URL：确保 GitLab 项目使用 HTTPS URL。 8. **Refine the Output:** Ensure the tone is professional and concise. * *Drafting Action Items:* 1. **配置 GitLab 集成**：如果你的组织使用 GitLab 管理插件市场，需在组织设置中添加 GitLab 配置以启用同步。 2. **审查私有源权限**：检查插件依赖项，确保不符合特定同源规则的源已设为公开（特别是混合使用 GitHub 和 GitLab 的场景）。 3. **使用 HTTPS URL**：在添加 GitLab 托管的市场时，确保使用项目的 HTTPS URL。 * *Drafting Impact:* **Medium**. 此次更新主要增加了 GitLab 支持并澄清了认证逻辑，属于功能增强而非破坏性变更，但对需要混合使用 Git 托管服务的团队至关重要。 * *Drafting Themes:* * 新增 **GitLab 托管市场同步**功能（目前为公开测试版）。 * 细化了不同托管平台 的认证方式区分（GitHub App vs 访问令牌）。 * 明确了插件源在不同场景下保持私有的具体条件。 * *Drafting Summary:* 文档更新正式引入了对 GitLab 托管插件市场的支持，并细化了组织同步过程中针对 GitHub 和 GitLab 的认证机制与私有仓库访问规则。 9. **Final Polish:** Check against the prompt constraints. "Concise, insightful", "Focus on what matters", "Chinese". * *Result looks good.* (Self-Correction during drafting): Make sure to emphasize that this is for *Organizations*. The diff is all about "Organization settings". Also, note the specific constraint about GitLab sources needing to be under the same top-level group/user to be private. That's a technical nuance worth noting. *Revised Action Items:* 1. **配置 GitLab**：若通过 GitLab 分发插件，管理员需在 Claude 组织设置中添加 GitLab 配置。 2. **检查 GitLab 命名空间**：若要在 GitLab 上使用私有插件源，确保其在同一顶级组或用户命名空间下。 3. **审查混合托管源**：对于不符合同源规则的插件源，确保仓库为公开状态，否则组织同步将拒绝。 This looks solid. 10. **Final Output Generation.** (Proceeding to generate response).

1. **Analyze the Request:** * **Role:** World-class programmer analyzing documentation changes. * **Input:** A batch of 15 documentation diffs (markdown format). * **Output Format:** Markdown analysis in Chinese. * **Structure:** 1. Overall Summary (1-2 sentences). 2. Key Themes (bullets). 3. Impact Level (Low/Medium/High/Breaking) + reasoning. 4. Action Items (bullets, max 5). 2. **Analyze the Diffs (Page by Page):** * `claude-code-on-the-web.md`: Cloud sessions now support `/fast` toggling with specific version requirements (`v2.1.271`). Removed `/fast` from the list of commands that behave like terminal pickers (it toggles now). * `costs.md`: Clarifies pricing multipliers (discount vs markup). Adds markup support requirement (`v2.1.271`). * `desktop.md`: * Multi-monitor support (pop out panes). * Background computer use on macOS. * Better integration with CLI sessions (resume CLI sessions from Desktop using `/resume`). * Updated "Unhide apps" logic based on background mode. * `hooks.md`: Clarifies behavior when `SessionStart` hooks are running (prompt wait, `Esc` to cancel). Applies to resumed sessions too. * `interactive-mode.md`: UI improvement for selected lines in the panel (can delete selection using Backspace). Requires `v2.1.271`. * `mcp.md`: * *Security*: Clarifies credential variable behavior. Remote servers' `url` and `headers` now intentionally read certain credential variables (like `ANTHROPIC_API_KEY`) as **empty** for security. They do *not* expand. * Explains how to work around this (use a custom var name). * Debug logging for these empty reads. * `plugin-marketplaces.md`: Adds `--accept-command` flag for non-interactive installs. * `plugins-reference.md`: * Plugin agents now support `omitClaudeMd`. * Synced plugins can't be disabled if required by organization. * Plugin config schema adds `options` field (enum picker support). * `skills.md`: Table update regarding subagent loading context (linking to `sub-agents.md`). Minor clarification. * `slash-commands.md`: Same as `skills.md`. * `sub-agents.md`: * Adds `omitClaudeMd` field support (skip CLAUDE.md files). * Clarifies `omitClaudeMd` behavior in main session vs subagent. * `vs-code.md`: * Prompt cache clock icon (visual indicator of cache expiry). * Auto-attach open file setting (`attachOpenFile`). * `api/messages.md` (and `.../batches/create.md`): Typo fix ("of will" -> "will"). * `CHANGELOG.md.md`: Added entry for `v2.1.272`. 3. **Synthesize Key Themes:** * **Version Requirements:** Multiple mentions of `v2.1.271` (markup pricing, fast mode toggle, interactive mode selection, omitClaudeMd). This is a significant version release. * **Security & Credentials:** Major changes in MCP (Model Context Protocol) regarding how credentials are handled. Preventing leakage of API keys to remote servers by reading specific env vars as empty. * **Agent & Context Control:** Introduction of `omitClaudeMd` for subagents and plugins, allowing finer control over what context/CLAUDE.md files are loaded. * **Desktop & UI Improvements:** Better multi-monitor support, background computer use, prompt cache clock in VS Code, and CLI/Desktop interoperability. * **Plugin System:** Enhanced plugin configuration (options/picker) and installation flags. 4. **Determine Impact Level:** * **Reasoning:** * The MCP credential security change is high impact for anyone configuring remote MCP servers. It breaks existing setups that relied on passing `ANTHROPIC_API_KEY` directly to a remote server URL/headers (intentionally). * `omitClaudeMd` is a medium/low impact feature improvement. * The `v2.1.271` features are feature additions, mostly low/medium impact (quality of life). * The API typo fix is low impact. * Overall, the MCP security change stands out as the most critical "breaking" change for specific workflows, while others are improvements. * Verdict: **Medium** (trending towards High for MCP users). The security change is technically breaking for those specific configurations, but likely intended behavior. However, compared to a full API overhaul, it's specific. Let's stick with Medium but emphasize the security aspect. 5. **Formulate Action Items:** * Check MCP configurations if using remote servers with credential variables. * Update CLI/Desktop apps to v2.1.271 for new features. * Review subagent/plugin definitions if you want to exclude CLAUDE.md. * Check costs settings if using markup pricing. * Update provisioning scripts for plugins if using non-interactive installs. 6. **Draft the Response (in Chinese):** * *Overall Summary:* 此次更新重点提升了安全性（MCP凭证保护）、上下文控制能力以及桌面/Web端的用户体验，并明确了多个新功能所需的 v2.1.271 版本要求。 * *Key Themes:* * **MCP 安全增强**：为防止凭证泄露，远程 MCP 服务器的 URL 和 headers 中引用的特定凭证变量（如 `ANTHROPIC_API_KEY`）现在会被读取为空字符串。 * **精细化上下文管理**：新增 `omitClaudeMd` 字段，允许子代理和插件加载时跳过 CLAUDE.md 文件。 * **桌面与 UI 体验**：增强了多显示器支持（弹出面板）、后台计算机使用、VS Code 中的提示缓存倒计时显示，以及 CLI 与 Desktop 会话的无缝互通。 * **插件系统增强**：插件配置支持枚举选择器，安装命令增加了 `--accept-command` 参数。 * **版本依赖**：多项新功能（如价格加价、快速模式切换、行选择改进）要求 Claude Code v2.1.271 或更高版本。 * *Impact Level:* **Medium** (主要影响配置了远程 MCP 服务器的用户，需调整凭证传递方式；其他为功能增强和体验优化)。 * *Action Items:* * 检查并更新使用了 `${ANTHROPIC_API_KEY}` 等凭证变量的远程 MCP 服务器配置，改用自定义变量名以避免被拦截读取为空。 * 更新至 Claude Code v2.1.271 以使用 `omitClaudeMd`、价格加价及 VS Code 的新 UI 功能。 * 如果需要在特定子代理或插件中排除 CLAUDE.md 上下文，请在定义中添加 `omitClaudeMd: true`。 * 在使用 Desktop 端时，尝试使用 `/resume` 命令从 CLI 历史中恢复会话。 * 更新非交互式插件安装脚本，利用 `--accept-command` 标志提高自动化效率。 7. **Review and Refine:** * Translation looks natural. * Technical terms preserved correctly. * Key point about MCP credential masking is highlighted. * Action items are actionable. 8. **Final Output Generation.** (Proceed to generate markdown).