Changes on 2026-08-18

1. **Analyze the Request**: * **Role**: World-class programmer analyzing documentation changes. * **Task**: Analyze a batch of documentation changes (8 pages). * **Output format**: Markdown, concise, insightful, focused on developer impact. * **Language**: Chinese. 2. **Analyze the Diff Data**: * **Page: `interactive-mode.md` (+50 / -0)** * *Content*: Added a new section "Check spelling as you type". * *Details*: * New feature: Underlines misspelled words in the prompt input. * Scope: Only input box, not replies or files. * Prerequisites: Requires `aspell`, `hunspell`, or `ispell` on PATH. * Configuration: Via `spellcheck` setting (enabled, checker, language, color) in user settings, CLI flag, or managed settings. Ignore project `.claude/settings.json`. * Behavior: Checks after typing pause. Skips code-like patterns (commands, file paths, etc.) and CJK text. * Debugging: Can fail if checker missing, crashes, or times out. Debug logs via `claude --debug`. * Version: Requires v2.1.235 or later. * **Page: `mcp.md` (+26 / -0)** * *Content*: Added section "Add a server from setup instructions written for another client". * *Details*: * Context: MCP server instructions often target Claude Desktop or other clients. * Guidance: How to convert these instructions for Claude Code CLI. * Scenarios: * URL -> `claude mcp add --transport http/sse` (or Option 4 for `wss://`). * Launch command (npx/uvx) -> `claude mcp add ... -- npx ...`. * JSON block -> `claude mcp add-json ...`. * Caveats: JSON needs `type` field if it has a `url`, keys must be alphanumeric. * **Page: `model-config.md` (+1 / -1)** * *Content*: Updated link reference in the context limit section. * *Details*: Changed link text/code from `` `Prompt is too long` `` to `` `context-limit` `` pointing to the error docs. * **Page: `settings.md` (+2 / -1)** * *Content*: Two small changes. * *Details*: * Updated link for `/doctor` command from `debug-your-config` to `commands#all-commands`. * Added new setting `spellcheck` in the table, referencing the new feature in `interactive-mode.md`. * **Page: `statusline.md` (+1 / -1)** * *Content*: Updated text regarding clickable links. * *Details*: Removed specific mention of "Windows Terminal" in the context of auto-detection failure to broaden the advice or simply clean up the text. * **Page: `intro.md` (+1 / -1)** * *Content*: Changed "Workbench" to "Playground". * *Details*: Updated the description of the web console tool. * **Page: `api/overview.md` (+1 / -1)** * *Content*: Changed "Workbench" to "Playground". * *Details*: Updated reference to the browser testing tool. * **Page: `build-with-claude/vision.md` (+1 / -1)** * *Content*: Changed "The [Workbench]" to "[Playground]". * *Details*: Terminology consistency update. 3. **Synthesize the Analysis**: * **Overall Summary**: The update introduces a significant new feature for input quality control (spell checking) and improves interoperability for MCP server installation, while standardizing terminology (Workbench -> Playground). * **Key Themes**: * **New Feature**: Interactive spell checking for prompts using external tools (aspell/hunspell/ispell). * **Interoperability**: Easier migration of MCP servers from other clients (Desktop, Cursor) to the CLI. * **Terminology**: Consistent rebranding of "Workbench" to "Playground". * **Maintenance**: Minor documentation navigation and link fixes. * **Impact Level**: **Medium**. * *Reasoning*: The spell check feature adds a new dependency requirement and requires configuration changes, which developers need to know about. The MCP installation guide simplifies cross-platform workflows. The terminology change is low impact but good to note. * **Action Items**: * Install `aspell`, `hunspell`, or `ispell` if you want to use the new spell-check feature. * Update documentation/references if you were using the term "Workbench" to describe the API testing tool. * Review the new `claude mcp add` syntax if importing servers from other MCP clients. * Update `/doctor` command link references in internal guides. 4. **Draft the Response (in Chinese)**: * *Summary*: 本次更新引入了交互模式下的拼写检查功能，增强了从其他 MCP 客户端迁移服务的能力，并将文档中的 "Workbench" 术语统一更改为 "Playground"。 * *Themes*: * 新增拼写检查功能：依赖外部拼写检查器。 * MCP 互操作性：更方便地从 Claude Desktop 等客户端导入服务器配置。 * 术语规范化：Workbench -> Playground。 * *Impact*: Medium. * *Action Items*: * Install spellchecker. * Use new `mcp add` logic. * Update links/names. 5. **Refine Response (Self-Correction during drafting)**: * Make sure to emphasize the `v2.1.235` version requirement for spell check. * Note that spell check settings are ignored in project settings (only user/managed/cmdline). * Mention the specific tools supported (aspell/hunspell/ispell). 6. **Final Polish**: * Structure: 1. Overall Summary, 2. Key Themes, 3. Impact Level, 4. Action Items. * Tone: Professional, concise. **(Self-Correction on "Windows Terminal" change in statusline.md)**: The diff removed "This commonly affects Windows Terminal...". It simplified the sentence to just say "If link text appears...". This is just a cleanup, not a functional change in behavior, but good for accuracy. **(Self-Correction on "Workbench")**: It appears in 3 pages. This is a branding update. **(Self-Correction on MCP)**: The added section is very practical for developers using multiple MCP clients. Proceed to generate output.

用户希望分析文档更改。具体来说，是对单个文件 `sub-agents.md` 的更改。 我需要扮演一名世界级程序员的角色，并提供一份简明扼要、富有洞察力的中文分析。 **1. 分析输入数据：** * **文件：** `sub-agents.md` * **变更统计：** +3 行 / -2 行 * **变更详情：** * **变更 1（新增）：** 在关于禁用子代理的列表中添加了一条新规则：“当会话没有 `general-purpose`（通用）子代理作为后备时，省略 `subagent_type` 的 Agent 工具调用将失败并报错 `subagent_type is required`。”这指明了在没有默认类型的情况下省略 `subagent_type` 的行为。 * **变更 2（编辑）：** 将“每次子代理调用都会创建一个具有全新上下文的新实例。”重述为“每次子代理调用都会创建一个新实例，而不是继续前一个实例。”这澄清了语义，强调缺乏连续性，虽然实际含义大致相同（为什么要重新开始？）。 * **变更 3（编辑）：** 修改了关于 fork 模式的要点。“当 Claude 不请求类型时，它会获得通用子代理……”改为“当 Claude 不请求类型时，它会获得通用子代理，*如果该会话仍然拥有该类型*。”这增加了条件 `if the session still has that type`（如果会话仍然拥有该类型），这对应于变更 1（即如果没有通用类型会发生什么）。 **2. 确定关键主题：** * **严格性/要求：** 对 `subagent_type` 参数的要求更加严格。如果默认的“通用”代理缺失，则显式指定 `subagent_type` 是强制性的。 * **行为澄清：** 阐明了默认回退（通用子代理）存在的依赖关系。 * **上下文隔离：** 重申子代理默认情况下是无状态的/隔离的（除非恢复）。 **3. 确定影响级别：** * **级别：** 中 * **推理：** 这对依赖 `Agent` 工具调用的代码具有潜在的中断性（破坏性）。如果开发者按照第一段提到的说明移除了内置代理（尤其是通用的），而他们的代码/提示依赖隐式的代理调用（即省略 `subagent_type`），这些调用现在将会失败，而不是静默回退到通用代理。这改变了对特定配置的 API 行为要求。 **4. 起草开发者行动项：** * 审查对 `Agent` 工具的调用。 * 如果禁用了内置代理，请确保指定 `subagent_type`。 * 处理新的错误情况 `subagent_type is required`。 **5. 翻译成中文并格式化：** * **Overall Summary:** 文档阐明了 `subagent_type` 参数的严格性：如果 `general-purpose`（通用）子代理被禁用或缺失，省略 `subagent_type` 的 Agent 工具调用将会失败。 * **Key Themes:** * 对 `subagent_type` 参数的严格验证。 * 阐明默认回退行为（依赖于通用子代理的存在）。 * 关于子代理实例生命周期（上下文隔离）的细微措辞更改。 * **Impact Level:** 中。移除默认通用代理会破坏隐式的代理调用，这是一个行为变更，可能会影响现有工作流或集成，这些工作流或集成依赖于即使在禁用其他特定内置代理的情况下仍然存在默认通用代理的假设。 * **Action Items:** * 审查代码或提示，查看是否存在省略 `subagent_type` 的 `Agent` 工具调用。 * 如果使用环境变量移除内置代理，确保显式传递 `subagent_type` 或处理预期会出现的错误。 * 更新错误处理逻辑，以捕获 `subagent_type is required`。 **6. 优化输出（自我修正）：** * 变更 2（重述）对功能的影响小于变更 1 和 3。我应该关注规则的变化。 * 变更 1 和 3 是相关的。变更 1 声明了错误，变更 3 解释了回退条件。 * 务必强调如果 `general-purpose` 消失，*原因*在于之前的配置（禁用代理）。 **7. 最终润色（中文）：** * Summary: 文档更新了对 `subagent_type` 参数的行为说明，明确指出当会话中不存在通用子代理作为后备时，省略该参数的 Agent 调用将直接报错。 * Themes: 参数强制校验, 默认回退机制变更. * Impact: Medium. 逻辑变更可能导致原本依赖默认行为的调用失败。 * Actions: 检查代码, 显式指定类型, 处理报错. 让我们基于这个结构生成最终的 markdown。

1. **Analyze the Request:** * **Role:** World-class programmer analyzing documentation changes. * **Output Format:** Markdown, concise, insightful, Chinese language. * **Input:** Two pages of diffs (`agents-and-tools/agent-skills/overview.md` and `CHANGELOG.md.md`). * **Required Sections:** Overall Summary, Key Themes, Impact Level (with reasoning), Action Items. 2. **Analyze Page 1: `agents-and-tools/agent-skills/overview.md`** * **Diff 1 (Security/Enterprise):** * *Change:* Added a note about "Skill content scanning" for Claude Enterprise organizations on `claude.ai` and Claude Cowork. * *Exclusion:* Clarifies that scanning *doesn't* cover Skills API or Console uploads. * *Implication:* Important for security-conscious enterprises. Highlights a gap in security coverage for API workflows. * **Diff 2 (Audit/Logging):** * *Change:* Added a link to "Audit logging" for Skills API operations in the "Using Agent Skills with the API" section. * *Implication:* Improves compliance and visibility for developers using the API. 3. **Analyze Page 2: `CHANGELOG.md.md` (Version 2.1.235)** * *Context:* This looks like a CLI tool or desktop client changelog (mentions "terminal UI", "native macOS/Linux builds", "VSCode", "vim mode", "prompt input"). * *Key Changes:* * *New Feature:* `spellcheck` setting (using aspell/hunspell/ispell). * *Bug Fixes:* Cache invalidation, list alignment, prompt highlighting, permission prompts (Shift+Tab), Agent tool error messages, notebook dialogs, HTML entities in slash commands, update notice, task list persistence. * *Performance:* Improved memory/CPU for background cloud sessions (`/ultrareview`, `/autofix-pr`). * *UX Improvements:* Permission dialogs, grep performance, context-limit error messaging, Vim mode preservation, dialog interaction fixes. * *API/Compliance:* `SendMessage` size check, `claude rc` enterprise gateway check. * *VSCode:* Fixed focus jumping. 4. **Synthesize and Group (Key Themes):** * *Theme 1: Security & Compliance (Enterprise focus).* The documentation update highlights content scanning limitations and audit logging for skills. The changelog mentions enterprise gateway checks. * *Theme 2: CLI/IDE Tool Stability & UX.* The changelog is packed with bug fixes and UX improvements (Vim mode, spellcheck, dialog handling) for a developer tool (likely the Claude CLI or an extension). * *Theme 3: API & Operations.* `SendMessage` validation, cache invalidation, context limit handling. 5. **Drafting the Content (in Chinese):** * *Overall Summary:* 本次更新主要包括针对 Claude Skills 的企业级安全与合规文档补充，以及 CLI/桌面客户端的大规模功能改进和错误修复。 * *Key Themes:* * **企业治理与合规**: 强调了 Skill 内容扫描的覆盖范围（仅限 claude.ai/Cowork，不含 API）并新增审计日志指引。 * **开发者工具体验**: 新增拼写检查，修复了大量 UI/UX 问题（Vim 模式、列表对齐、提示词高亮）并优化了内存/CPU 性能。 * **稳定性增强**: 修复了缓存失效、大消息处理、对话框逻辑等底层问题。 * *Impact Level:* * *Medium*. * *Reasoning:* Documentation changes affect enterprise security posture awareness (critical for some). The tooling updates (v2.1.235) are significant quality-of-life improvements but aren't breaking changes. They fix specific annoyances and add helpful features. * *Action Items:* * **Enterprise Admins**: 确认 Skill 内容扫描的局限性（API 上传不覆盖），并根据需求调整审计日志配置。 * **API 开发者**: 查看 Skills API 的审计日志文档，确保合规性。 * **CLI 用户**: 更新到 v2.1.235 以利用拼写检查和性能优化，特别是涉及后台会话（如 `/ultrareview`）的场景。 * **Vim 用户**: 享受修复后的 Vim 模式交互体验（切换面板时光标位置保持）。 6. **Refining the Output:** * Ensure the tone is professional and technical. * Check specific terminology (e.g., "Skills", "Cowork", "Vim mode"). * Make the distinction between the documentation update (policy/guidance) and the changelog (tool features) clear but integrated. 7. **Final Polish (Self-Correction during drafting):** * *Correction:* Make sure to mention *why* the content scanning limitation matters. It tells developers that if they upload via API, they need their own scanning process because the platform doesn't do it automatically for that path. * *Correction:* The changelog mentions `SendMessage` now refuses large messages. This is a behavior change (breaking if users relied on silent dropping, but likely better to fail fast). 8. **Final Output Generation:** (Proceed to generate Chinese response based on these points).

1. **Understand the Goal**: The user wants a concise analysis of documentation changes across three pages (`quickstart.md`, `skills.md`, `slash-commands.md`). I need to act as a world-class programmer, focusing on what matters to developers. The output must be in Chinese and markdown format. 2. **Analyze Page 1: `quickstart.md`**: * **Changes**: Massive reduction of redundant `theme={null}` attributes in the code blocks for Native Install, Homebrew, and WinGet tabs. * **Content**: The installation commands themselves (curl, brew, winget) haven't changed. * **Impact**: This is a cleanup/formatting fix. It doesn't affect the actual commands developers run. It makes the source code cleaner. 3. **Analyze Page 2: `skills.md`**: * **Changes**: A URL update in a link. * **Old**: `[Agent SDK](/docs/en/agent-sdk/slash-commands)` * **New**: `[Agent SDK](/docs/en/agent-sdk/skills#discover-available-commands)` * **Context**: The text discusses how `"off"` skills are hidden from command lists advertised to Agent SDK callers. * **Impact**: This improves documentation navigation. It points developers to a more specific, relevant section (`skills#discover-available-commands`) rather than a general `slash-commands` page regarding Agent SDK integration. 4. **Analyze Page 3: `slash-commands.md`**: * **Changes**: Identical to `skills.md`. The same URL update. * **Context**: Same context regarding skill overrides and visibility. * **Impact**: Consistency fix. Ensures developers find the right reference documentation for the Agent SDK. 5. **Synthesize the Batch Analysis**: * **Overall Summary**: The changes consist primarily of a cleanup of redundant attributes in the installation guide and a link correction to point to a more specific Agent SDK section. * **Key Themes**: * *Code Cleanup/Formatting*: Removal of repetitive `theme={null` strings. * *Documentation Navigation*: Updating internal links to be more specific/correct. * **Impact Level**: **Low**. No functional changes to the software or commands. Just source code hygiene and link accuracy. * **Action Items**: * None required for existing users. * If reading the docs, follow the updated link for more precise info on Agent SDK skill discovery. 6. **Drafting the Response (in Chinese)**: * *Summary*: 本次更新主要包含清理安装指南中冗余的代码块属性，以及修正指向 Agent SDK 的内部链接，使其指向更具体的章节。 * *Themes*: * **代码清理**: 移除了 `quickstart.md` 中代码块标签内大量重复的 `theme={null}` 属性，提升了源码整洁度。 * **文档导航优化**: 在 `skills.md` 和 `slash-commands.md` 中，将 Agent SDK 的链接从通用页面更新为关于“发现可用命令”的具体锚点 (`skills#discover-available-commands`)。 * *Impact*: **Low** (低)。这些是文档维护性更新，不涉及产品功能变更或命令行操作的改动。 * *Action Items*: * 无需采取任何操作。 * 如果正在查阅 Agent SDK 相关文档，点击更新后的链接可获得更准确的技能发现机制说明。 7. **Review and Refine**: * Check constraints: Markdown format? Yes. Chinese? Yes. Concise? Yes. Focus on developers? Yes (telling them nothing broke). * The diff for `quickstart.md` looks huge in the raw text but is actually just removing noise. I must emphasize that the *commands* didn't change. * The link change is minor but helpful for developers actually using the SDK. 8. **Final Output Generation** (matches the drafted points).

1. **Analyze the Request:** * **Role:** World-class programmer analyzing documentation changes. * **Task:** Analyze a specific diff for `build-with-claude/files.md` and provide a concise summary in markdown. * **Format:** Chinese. * **Structure:** 1. Overall Summary. 2. Key Themes. 3. Impact Level with reasoning. 4. Action items for developers (max 5). 2. **Analyze the Diff (`build-with-claude/files.md`):** * **Change 1 (Warning block):** Added text advising multi-tenant applications to use separate workspaces for each tenant to ensure hard isolation. Mentions a limit of 100 workspaces. * **Change 2 (Storage limits):** Increased total storage from "500 GB per organization" to "1 TB per organization". This change appears twice (once in the limits list, once in the error handling section). * **Change 3 (Audit logging):** Added a new section explaining that Files API operations (upload, download, delete) are now logged in the Compliance API Activity Feed (if enabled). It specifies which events are tracked and notes that AWS CloudTrail is used for the AWS version. It clarifies that listing/metadata retrieval are *not* recorded. * **Change 4 (Rate limits):** Removed the "During the beta period" phrase. Increased the rate limit from "approximately 100 requests per minute" to "approximately 500 requests per minute". Updated the contact link text. 3. **Synthesize the Findings:** * *Overall Summary:* The Files API documentation has been updated to reflect higher capacity limits (storage and rate), introduce new security/architecture guidance for multi-tenancy, and detail audit logging capabilities via the Compliance API. * *Key Themes:* * **Capacity & Performance:** Limits have been increased (1TB storage, 500 req/min), and the beta status seems effectively lifted (or at least no longer cited as a limiting factor). * **Security & Architecture:** Stronger guidance on multi-tenant architecture using workspaces for isolation. * **Observability/Compliance:** New documentation on audit logging for file operations. * *Impact Level:* * Is it breaking? No. * Is it high impact? Yes, for developers building production apps, the increase in limits is significant, and the audit logging is crucial for compliance. The workspace isolation advice is critical for security. * Verdict: **High** (Non-breaking, but important for production readiness and scaling). * *Action Items:* * Review multi-tenant architecture if applicable (consider separate workspaces). * Enable Compliance API if audit trails are needed. * Update application logic/config to take advantage of increased limits. * No code changes strictly required (improvements are backend/platform side), but architectural review is needed. 4. **Draft the Response (in Chinese):** * *Overall Summary:* 本次更新大幅提升了 Files API 的存储和速率限制，增加了关于多租户应用隔离的安全指导，并新增了审计日志功能说明。 * *Key Themes:* * **容量与性能提升**：组织存储上限从 500GB 提升至 1TB，API 调用速率限制从 100 次/分钟提升至 500 次/分钟，且移除了“测试版”相关限制说明。 * **安全架构指导**：针对多租户应用，明确建议为每个租户创建独立的工作区以实现数据硬隔离。 * **合规与可观测性**：新增审计日志文档，说明启用合规 API 后，文件的上传、下载和删除操作将被记录。 * *Impact Level:* **High** (高)。 * Reasoning: 虽然不是破坏性变更，但限制的大幅放宽直接影响生产环境的扩展能力，而审计日志和多租户隔离指导对于企业级应用的安全性和合规性至关重要。 * *Action Items:* * **架构审查**：如果构建多租户应用，评估并实施“每个租户对应独立工作区”的架构。 * **启用审计**：若需追踪文件操作，请配置 Compliance API 以获取 Activity Feed。 * **更新监控**：利用新的 1TB 存储和 500 RPM 速率限制，重新评估或调整应用的扩容策略。 5. **Review and Refine:** * The response looks accurate and concise. * The tone is professional. * The Chinese translation is natural. * Matches the requested format strictly. 6. **Final Output Generation.** (Proceed to generate output based on step 4).

用户希望我分析三个页面（`amazon-bedrock.md`、`desktop.md`、`sub-agents.md`）的文档更改，并提供一份简洁的中文分析。我需要总结变更、主题、影响级别和行动项。 **分析：** 1. **`amazon-bedrock.md`**: * **变更**：阐明 Sonnet 5 在 Amazon Bedrock 上对 1M 上下文的支持。 * **之前**：“Sonnet 5 通过 [Mantle endpoint](#use-the-mantle-endpoint) 提供服务，并且总是运行在 1M 窗口下...” * **之后**：“Sonnet 5 在 Invoke API 和 [Mantle endpoint](#use-the-mantle-endpoint) 上始终以 1M 窗口运行...” * **原因**：阐明了 1M 支持不仅限于 Mantle endpoint，也适用于标准 Invoke API。这纠正了一个潜在的误解。 2. **`desktop.md`**: * **变更 1**：文本编辑以提升清晰度/精确度。“You can switch modes...” -> “You can switch permission modes...”（你可以切换模式... -> 你可以切换权限模式...）。 * **变更 2**：阐明桌面版的键盘快捷键。“...to cycle modes” -> “...to cycle permission modes”（...以循环切换模式 -> ...以循环切换权限模式）。 * **原因**：一致性和精确性。确保用户在指代权限控制时不会将“模式”与类似“Agent 模式”或“Plan 模式”的其他操作模式混淆。 3. **`sub-agents.md`**: * **变更 1**：阐明子代代理（teammate's subagents）和后台任务的错误条件。 * 添加了关于在 fork 模式关闭且后台任务未禁用的情况下，队友子代理设置 `run_in_background: true` 时会发生什么的细节（它会拒绝并报错）。 * **变更 2**：阐明子代理的延迟影响。 * “Subagents start fresh...” -> “A subagent that isn't a [fork](#fork-the-current-conversation) starts fresh...”（子代理重新开始... -> 非 [fork](#fork-the-current-conversation) 的子代理重新开始...）。 * **原因**：技术精确性。区分 fork 子代理（继承上下文，低延迟）和非 fork 子代理（冷启动，高延迟），以便开发者对性能做出更好的决策。 **综合：** * **总体摘要**：这些更改集中在技术精确性和阐明边缘情况/行为细节上，特别是关于 Amazon Bedrock 的上下文窗口、桌面应用中的权限模式以及子代理的执行行为。 * **关键主题**：阐明 API 覆盖范围（Bedrock），术语精确度（权限模式），以及执行逻辑细节（子代理后台任务/性能）。 * **影响级别**：低到中。这些大多是不影响核心功能的澄清/文档修正。然而，Bedrock 的更改消除了关于 API 使用的潜在困惑，而子代理的更改澄清了后台任务中的错误状态，这对调试至关重要。 * **行动项**： * 更新任何假设 Sonnet 5 1M 上下文仅适用于 Mantle 的集成。 * 使用/解释桌面快捷键时使用精确术语（“权限模式”）。 * 针对延迟性能，将 fork 子代理与非 fork 子代理区分开。 * 需要后台执行时，注意子代代理中 `run_in_background: true` 的错误条件。 **格式化：** 用户要求中文回复。 * **Overall Summary**: 文档主要进行了技术细节的澄清和描述的精确化，涉及 Amazon Bedrock 的上下文窗口支持、桌面端的权限模式术语以及子代理的执行行为和性能影响。 * **Key Themes**: * API 能力澄清 * 术语一致性 * 执行逻辑与错误处理细化 * **Impact**: Low/Medium。主要是澄清性的，帮助开发者避免误解，特别是关于 API 行为和性能特性方面。 * **Action items**: * 确认 Bedrock 集成中 Sonnet 5 的 1M 上下文支持不仅限于 Mantle endpoint。 * 在文档或代码注释中提到“模式”时，使用“权限模式”以保持清晰。 * 评估子代理性能时，区分 fork 与非 fork 子代理在上下文加载上的差异。 * 注意子代代理在特定配置下（如 fork 模式关闭时）设置后台运行的限制。 让我们润色中文措辞，使其显得专业且简洁。 * **Page 1 (Bedrock)**: 之前暗示 Mantle endpoint 是 1M 窗口的关键。现在澄清 Invoke API 也支持。 * **Page 2 (Desktop)**: 将通用的“modes”细化为“permission modes”。 * **Page 3 (Sub-agents)**: 阐明了后台任务的错误拒绝条件以及非 fork 子代理的延迟影响。 最终润色。