Changes on 2026-05-21

1. **分析请求：** * **角色：** 世界级程序员，专注于开发工具和文档。 * **任务：** 分析文档变更（具体为 `CHANGELOG.md.md` 的差异）。 * **格式：** Markdown，简洁，富有洞察力，中文回复。 * **结构：** 1. Overall Summary (1-2 sentences)（整体摘要） 2. Key Themes (bullets)（关键主题） 3. Impact Level with reasoning（影响级别及理由） 4. Action items (max 5 bullets)（行动项） 2. **分析输入数据：** * **文件：** `CHANGELOG.md.md` * **变更：** 2 行删除 (`-0 / -2`)。 * **删除的内容：** 1. `Added the 'Workflow' tool for deterministic multi-agent orchestration. It is off by default — set 'CLAUDE_CODE_WORKFLOWS=1' to enable` 2. `Hardened REPL and Workflow tool sandboxes against prototype-pollution and thenable-based escapes` 3. **解读变更：** * *观察：* 变更严格来说是移除。没有新增或修改。 * *语境：* 这看起来像是对变更日志的清理。 * *删除 1：* “Workflow” 工具被提及为新增功能。它处于“默认关闭”状态，需要设置一个环境变量来启用。这通常暗示该功能是实验性的、不稳定的，或者是针对特定受众（可能是测试版用户）。 * *删除 2：* 提到了沙箱安全加固。 * *推断：* 为什么要从变更日志中移除这些？ * *场景 A：* 功能被撤销或完全移除。 * *场景 B：* 功能已毕业，不再需要特别标记为“实验性/可选”（但通常情况下，变更日志会保留历史记录，除非是对*未来*或*未发布*的条目进行编辑）。 * *场景 C（鉴于它是更改日志，这是最可能的）：* 之前针对版本 `2.1.147` 发布了包含这些行的变更日志。开发人员正在*修改发布说明*。为什么？也许功能未准备好发布，有 Bug，或者他们希望不在公告中突出显示。 * *场景 D：* 输入格式 `CHANGELOG.md.md` 暗示可能有拼写错误，但差异是真实的。 * *关键领悟：* 这些条目指的是 *Workflow* 工具。移除它们意味着在版本 2.1.147 中，`Workflow` 工具*不是*面向公众的功能，或者它已被静默撤回。 * *安全角度：* 第二个删除项提到了安全加固（“原型污染”）。移除该条目可能意味着修复实际上并未部署，或者他们不希望过早宣传安全补丁（如果变更日志是预先准备好的）。或者，如果 `Workflow` 工具本身被移除/隐藏，那么修复条目也就不再相关了。 4. **起草分析（内心独白/起草）：** * **Summary：** 版本 2.1.147 的变更日志被修改以移除提及“Workflow”工具及其相关安全修复。这表明该功能被撤回、推迟，或有意在此版本发布中保持未公开状态。 * **Themes：** * 特性回滚/隐藏。 * 发布说明清理。 * 安全/沙箱（被提及后被移除）。 * **Impact：** 中等到高。为什么？如果一个多智能体编排工具被移除/隐藏，依赖它的开发者（或那些遵循 `CLAUDE_CODE_WORKFLOWS=1` 标志来尝试它的人）会感到困惑或失去功能。如果修复没有应用，安全修复的移除也很重要。 * **Action Items：** * 如果你在使用 `CLAUDE_CODE_WORKFLOWS=1`，请注意该功能可能无法正常工作或不存在。 * 检查实际行为与删除的说明是否一致。 * 如果安全补丁被回滚，请注意潜在的沙箱漏洞（虽然对普通用户来说可能性较小，但值得注意）。 5. **针对“世界级程序员”人设进行提炼（中文）：** * *摘要：* 这次更改从 CHANGELOG 中删除了两条关于“Workflow”工具的记录（功能介绍和安全加固）。这暗示该多智能体编排工具可能已被撤回、推迟发布，或者被刻意隐藏在当前版本之外。 * *主题：* 特性撤回/隐藏, 发布说明清理, 实验性功能状态变更。 * *影响级别：* **Medium**（中等）。它移除了一个显著的功能（“Workflow”）。 * *行动项：* * Don't use the flag if you were planning to.（如果你计划使用，请不要使用该标志。） * Check if the feature actually works if you already tried it.（如果你已经尝试过，检查该功能是否真的有效。） 6. **最终润色（Markdown 格式）：** * **Summary:** 针对 `2.1.147` 版本的更新日志进行了修订，删除了关于 `Workflow` 工具的介绍及其安全加固的相关说明。这表明该多智能体编排功能可能已被撤回、推迟，或有意在此版本中保持不可见状态。 * **Themes:** * **功能撤回/隐藏**: 移除了 `Workflow` 工具的发布说明，意味着该实验性功能可能暂未交付给用户。 * **发布说明清理**: 修正了已发布信息的准确性，去除了未实际生效或暂不公开的内容。 * **Impact Level:** **Medium**（中等影响） * **Reasoning:** 虽然代码变动仅仅是删除了日志文本，但这直接关系到 `Workflow` 工具的可用性。原本计划的多智能体编排功能及其安全增强现在确认不在当前版本中，可能会影响期待该特性的开发者。 * **Action items:** * **停用环境变量**: 如果已设置 `CLAUDE_CODE_WORKFLOWS=1`，建议取消设置，因为相关功能已被移除或暂不可用。 * **忽略旧文档**: 忽略之前关于 Workflow 工具启动或沙箱加固的说明，它们当前不适用于 2.1.147 版本。 * **关注后续更新**: 如果您需要多智能体编排能力，请关注未来的版本更新，而非当前版本。 7. **对照约束进行审查：** * 中文？是。 * Markdown？是。 * 简洁？是。 * 结构遵循了吗？是。 *起草过程中的自我修正：* 我最初考虑过安全方面。如果他们移除了关于原型污染修复的日志，这是否意味着修复*不在*代码中？还是仅仅因为工具被隐藏，所以*提及*也被隐藏了？通常，如果安全修复被应用，他们会保留它。移除这两点暗示功能块被移除了。因此，建议不要使用该功能是最稳妥的建议。 8. **最终输出生成。**（基于第 6 步继续生成回复）。

1. **Analyze the Request**: * **Role**: World-class programmer analyzing documentation changes. * **Task**: Analyze 4 changed pages as a single batch. * **Output Format**: Markdown, concise, insightful, Chinese language. * **Structure**: Overall Summary, Key Themes, Impact Level + Reasoning, Action Items (max 5). 2. **Analyze the Diffs**: * **Page 1: `desktop.md`** (+1 / -1) * *Change*: Updated availability text for "Auto mode". * *Old*: "research preview available on Max, Team, Enterprise, and API plans... Pro plans excluded... requires Claude Sonnet 4.6, Opus 4.6, or Opus 4.7..." * *New*: "research preview available to **all users on the Anthropic API**. It is not available on third-party providers. It requires Claude Sonnet 4.6, Opus 4.6, or Opus 4.7." * *Meaning*: Removed the plan restrictions (Max, Team, Enterprise) for API users. It's now open to *all* API users (implies Pro/Individual via API might be included or just simplified language). Crucially, it emphasizes availability on Anthropic API, explicitly excluding third-party providers. * **Page 2: `devcontainer.md`** (+1 / -0) * *Change*: Added a link in the "Related resources" section. * *New Link*: `[Sandbox environments](/en/sandbox-environments): compare dev containers with the built-in Bash sandbox, custom containers, and VMs`. * *Meaning*: Signaling a new page or restructuring to help users choose between isolation methods. * **Page 3: `sandboxing.md`** (+159 / -144) * *Change*: Significant rewrite/restructuring of the Sandboxing documentation. * *Old*: "Sandboxing" page title, "Overview", "Why sandboxing matters", detailed subsections on "How it works" (Filesystem, Network). * *New*: Title changed to "Configure the sandboxed Bash tool". * *Content Changes*: * Added a "Get started" section with specific steps using `/sandbox`. * Platform support clarification (macOS, Linux, WSL2. Native Windows *not* supported). * Mention of `sandbox.failIfUnavailable`. * Removed some deep theoretical "Why sandboxing matters" text, replaced with actionable "How to". * Links to `sandbox-environments` page (comparison). * Links to `permission-modes`. * The diff cuts off at "Security limitations" but implies a focus on *configuration* rather than just *concept*. * *Meaning*: Shift from conceptual overview to a practical, step-by-step guide on configuring and using the Bash sandbox. * **Page 4: `security.md`** (+6 / -5) * *Change*: Updated "Related resources". * *Change*: Added `[Sandbox environments](/en/sandbox-environments)` at the top of the list. * *Meaning*: Elevates the "Sandbox environments" comparison page as a primary resource for security/isolation decisions. 3. **Synthesize Findings**: * **Theme 1: Feature Availability Expansion**. Auto mode is now more accessible to API users. * **Theme 2: Documentation Restructuring**. There is a clear shift towards organizing content around "How to configure" rather than "What is". The Sandboxing page is now a "How-to" guide. * **Theme 3: New Content Node**. The introduction of `sandbox-environments` (seen in links on pages 2, 3, and 4) suggests a new comparative page for isolation strategies. * **Theme 4: Clarification of Platform Support**. Explicit statement about Native Windows vs. WSL2 in the sandboxing doc. 4. **Drafting the Analysis (Internal Monologue -> Chinese)**: * *Overall Summary*: Documentation updated to expand feature access (Auto mode on API) and restructure security/sandboxing guidance to be more action-oriented, introducing a new comparison page for isolation environments. * *Key Themes*: * Expanded API access for Auto mode. * Refactoring Sandboxing docs into a configuration guide. * Introduction of "Sandbox environments" as a new comparison resource. * Clarification on platform support (Windows/WSL2). * *Impact Level*: Low to Medium. No breaking changes to the code/tool itself, but docs are clearer. * *Action Items*: * Check Auto mode eligibility if using the API. * Review the new `/sandbox` workflow if concerned about security/permissions. * Note the Windows platform requirement (WSL2). 5. **Refining the Output (Formatting)**: * **Summary**: 本次文档更新主要集中在扩大 API 用户的“自动模式”访问权限，并对沙盒及安全文档进行了重组，引入了新的隔离环境对比页面，将技术文档重心从概念介绍转向了配置指南。 * **Themes**: * 功能开放：Auto mode 现在面向所有 Anthropic API 用户开放（不再局限于 Max/Team 等计划）。 * 文档重构：`sandboxing.md` 从概念介绍转变为“配置 Bash 沙盒工具”的实操指南。 * 新增资源：引入 `sandbox-environments` 页面，用于对比 Dev Containers、内置沙盒和 VM。 * 平台明确：明确指出沙盒在 Windows 原生环境下不支持，需使用 WSL2。 * **Impact**: Low。仅文档变更，功能逻辑未变，但降低了配置门槛并明确了可用性。 * **Action Items**: * 如果您是 API 用户，现在可以尝试使用 Auto mode。 * 查阅新的“Sandbox environments”页面以选择最适合的隔离方案。 * 如果在 Windows 上使用沙盒，请确保使用 WSL2。 * 探索 `/sandbox` 命令的新面板功能以管理权限模式。 6. **Final Polish**: Ensure tone is professional and concise. (The user asked for Chinese). * *Self-Correction during drafting*: Make sure to highlight the "Plan mode" mention in desktop.md tip? It's context, not a change, but the *change* is the availability restriction removal. Focus on the change. * *Self-Correction on Sandboxing*: The diff is huge (+159/-144), indicating a major rewrite. I should emphasize the shift to "Get started" / "Steps". 7. **Final Output Generation** (matching the structure requested).

1. **Analyze the Request:** * **Role:** World-class programmer with deep expertise in dev tools/docs. * **Task:** Analyze documentation changes as a single batch. * **Input:** 3 changed pages (`mcp.md`, `settings.md`, `CHANGELOG.md.md`) with diffs provided. * **Output Format:** Markdown, Chinese language. * **Structure:** Overall Summary, Key Themes, Impact Level, Action Items. 2. **Analyze the diffs:** * **Page: `mcp.md`** * *Change:* Massive deletion (-117 lines). It removed the entire "Managed MCP configuration" section. * *Content Removed:* Detailed instructions on `managed-mcp.json`, exclusive control, policy-based control (allowlists/denylists), restriction options (server name, command, URL), and multiple examples/accordions explaining how these restrictions work. * **Page: `settings.md`** * *Changes:* * Link update: `.../en/mcp#managed-mcp-configuration` -> `.../en/managed-mcp` (2 instances). * Context: The references point to the section that was just deleted from `mcp.md` but now points to a new page `managed-mcp`. * *Interpretation:* The content removed from `mcp.md` has been moved to a new dedicated page likely named `managed-mcp.md`. * **Page: `CHANGELOG.md.md`** * *Change:* Version bump from `2.1.146` (content shown at top, though previous context implies it's replacing old content) to `2.1.147`. * *Key New Features:* * Added `Workflow` tool (deterministic multi-agent orchestration), off by default (`CLAUDE_CODE_WORKFLOWS=1`). * Pinned background sessions improvements. * `/code-review` renamed from `/simplify`, new behavior (reports bugs, cleanup behavior removed), optional `--comment` flag for GitHub. * *Bug Fixes & Improvements:* Hardened sandboxes, improved auto-updater, diff rendering, prompt history, PowerShell fixes (Windows), terminal rendering fixes, MCP pagination fixes. * *Impact:* This is a feature-rich update with significant changes to command behavior (`/simplify` is dead, `/code-review` is different). 3. **Synthesize Analysis:** * **Overall Summary:** The documentation restructures Managed MCP configuration into a separate page,伴随着版本 `2.1.147` 的发布，该版本引入了工作流、代码审查等重大功能变更并修复了大量 Bug。 * **Key Themes:** * **Docs Restructuring:** Moving detailed "Managed MCP" settings out of the main MCP doc to a dedicated `managed-mcp` page (inferred from link updates in `settings.md` and deletions in `mcp.md`). * **New Features:** Introduction of the `Workflow` tool and the new `/code-review` command (replacing `/simplify`). * **Enterprise/Admin:** Hardening of sandboxes and improvements to background sessions for enterprise usage. * **Platform Stability:** Heavy focus on fixing Windows PowerShell issues, terminal rendering, and auto-updater reliability. * **Impact Level:** * *Level:* High. * *Reasoning:* * Breaking Change: `/simplify` is renamed to `/code-review` with altered behavior (cleanup logic removed). * New Capability: `Workflow` tool (beta/flagged) changes how multi-agent orchestration works. * Documentation Move: Admins need to look in a new place for Managed MCP settings. * Critical Fixes: PowerShell fixes on Windows are significant for that user base. * **Action Items:** * Update scripts/aliases using `/simplify` to `/code-review`. * Review the new `managed-mcp` page if you rely on centralized MCP control. * Set `CLAUDE_CODE_WORKFLOWS=1` if you want to test the new Workflow tool. * Upgrade to `2.1.147` for the PowerShell and terminal rendering fixes. 4. **Drafting the Response (in Chinese):** * *Summary:* 本次更新包含文档重构（将 MCP 管理配置移至独立页面）及版本 `2.1.147` 的发布，引入了 Workflow 工具并重置了 `/simplify` 命令行为。 * *Themes:* * 文档重构：Managed MCP 配置章节从 `mcp.md` 移至 `managed-mcp`（通过链接变更推断）。 * 命令变更：`/simplify` 被 `/code-review` 替代，且逻辑发生重大变化（不再自动修复代码，专注 Bug 报告）。 * 新功能：新增 `Workflow` 工具（需环境变量开启），支持确定性多代理编排。 * 平台修复：重点修复了 Windows PowerShell 兼容性、自动更新器及终端渲染问题。 * *Impact:* High. * *Action Items:* * 脚本更新：将 `/simplify` 替换为 `/code-review` 并调整预期行为。 * 寻找文档：查看新的 `/en/managed-mcp` 页面获取 MCP 管理配置信息。 * 启用新功能：如需测试 Workflow，设置环境变量 `CLAUDE_CODE_WORKFLOWS=1`。 * 利用改进：更新版本以获得 PowerShell 和 diff 渲染的性能/稳定性提升。 5. **Refining the Output:** * Ensure the connection between `mcp.md` deletion and `settings.md` link update is clear in the summary/themes. * Make sure to highlight the breaking change regarding `/simplify` -> `/code-review`. 6. **Final Polish:** Check against markdown constraints and conciseness. * *Self-Correction during drafting:* I noticed the `CHANGELOG` diff shows `2.1.147` replacing what looks like `2.1.146` content in the snippet (though the diff header says `@@ -1,23 +1,42 @@` which usually means replacing the top of the file). It's clear `2.1.147` is the new version to focus on. 7. **Final Output Generation.** (Proceeding to generate the Chinese response based on step 4).

1. **Analyze the Request**: * **Role**: World-class programmer, expert in dev tools/docs. * **Task**: Analyze documentation changes for 3 pages (`hooks.md`, `hooks-guide.md`, `settings.md`) as a single batch. * **Format**: Markdown (Concise, insightful). * **Output Language**: Chinese (as requested in the prompt instructions "请用中文回复"). * **Structure**: 1. Overall Summary. 2. Key Themes. 3. Impact Level. 4. Action Items. 2. **Analyze the Changes**: * **Page 1: `hooks.md`** (+32 / -26): * *Change 1*: Changed comment `# allow the command` to `# no decision; normal permission flow applies`. * *Change 2*: Updated explanatory text about `exit 0`. Clarifies that `exit 0` (no output) means "no decision to report" and proceeds via normal permission flow. Explicitly states: "The hook can deny the call, but staying silent doesn't approve it." * *Change 3*: Updated comment `# Success: tool call proceeds` to `# No decision: the normal permission flow applies`. * *Change 4*: Updated explanation: "Exit codes only let you block or stay silent" instead of "allow or block". * *Change 5*: Clarified `suppressOutput` behavior: "hides the hook's stdout from the transcript" vs "omits stdout from the debug log". * *Change 6*: Table update for decision patterns. (Seems to be a reformat or minor clarification, though the diff is slightly truncated here, it looks like a table reference). * **Page 2: `hooks-guide.md`** (+3 / -3): * *Change 1*: Same comment change as above (`exit 0` explanation). * *Change 2*: Re-iteration of the `exit 0` behavior: "reports no objection... normal permission flow still applies". * *Change 3*: Re-iteration: "Exit codes only let you block or stay silent". * **Page 3: `settings.md`** (+14 / -0): * *Change 1*: Added new table row for `strictPluginOnlyCustomization`. * *Change 2*: Added new section documenting `strictPluginOnlyCustomization`. * *Content of new setting*: Managed settings only. Blocks skills, agents, hooks, and MCP servers from user/project sources. Only allows them from plugins or managed settings. Works with `strictKnownMarketplaces`. Requires v2.1.82+. 3. **Synthesize Insights**: * **Theme 1: Clarification of Hook Exit Codes (Critical Semantics)**. The docs are actively correcting a misconception. Previously, it might have seemed that `exit 0` in a hook (specifically for tool use) meant "Approved". The new docs make it very clear: `exit 0` means "I don't care / I have no opinion". It does *not* mean "Approved". The "normal permission flow" (user prompts, etc.) still happens. This is a significant mental model shift for hook authors. To approve, you *must* use JSON output. * **Theme 2: Enhanced Security/Control for Enterprise (Managed Settings)**. A new feature `strictPluginOnlyCustomization` allows admins to strictly enforce that customizations (skills, agents, hooks, MCP) only come from approved sources (Plugins or Managed settings), preventing users from injecting local scripts or config into the environment. This completes a supply-chain security model alongside marketplace restrictions. 4. **Drafting the Content (in Chinese)**: * *Overall Summary*: The changes focus on two main areas: clarifying the semantics of hook exit codes (specifically that `exit 0` means "no decision" rather than "approve") and introducing a new managed setting (`strictPluginOnlyCustomization`) for enterprise-grade control over customization sources. * *Key Themes*: * **Hook Semantics Correction**: Explicitly redefining `exit 0` behavior. It signals "no decision" rather than "approval", requiring normal permission flows to still run. To actively approve, JSON output is required. * **Enterprise Lockdown**: Introducing `strictPluginOnlyCustomization` to restrict customizations (skills, hooks, MCP, agents) solely to managed settings or plugins, tightening security in managed environments. * *Impact Level*: **Medium**. For enterprise admins, the new setting is High impact (security). For hook developers, the semantic clarification is Medium impact (correcting potential logic errors where they assumed `exit 0` bypassed permissions). * *Action Items*: * Review existing hooks that rely on `exit 0` to implicitly approve tool calls; they might not be working as intended (bypassing permissions) if that was the goal. * Use JSON output (`{ "decision": "allow" }`) if the goal is to explicitly approve a tool call and bypass user prompts. * Admins: Update Claude Code clients to v2.1.82+ to enforce the new `strictPluginOnlyCustomization` setting. 5. **Refining the Output**: * *Drafting the "What Changed"*: * Hooks: Clarified `exit 0` means "no decision", not "allow". Updated comments and explanatory text. * Settings: Added `strictPluginOnlyCustomization` documentation. * *Drafting "Why it matters"*: * Hooks: Prevents security risks where devs think they are approving but actually just passing it to the user prompt logic. * Settings: Allows stricter enterprise control. 6. **Final Polish (Markdown & Chinese)**: * Summary: 本次更新主要修正了 Hook 脚本的退出码语义（明确 `exit 0` 仅表示“无决策”而非“允许”），并引入了新的托管设置 `strictPluginOnlyCustomization` 以增强企业级环境对自定义来源的管控。 * Themes: * **Hook 语义纠正**：文档明确指出 `exit 0` 不会自动批准操作，而是遵循默认的权限流程。若需绕过用户提示显式批准，必须使用 JSON 输出。 * **企业级合规锁定**：新增设置允许管理员禁止用户或项目源的自定义项，强制仅通过插件或托管策略加载。 * Impact: Medium. (Reasoning: It changes the understanding of existing hook behavior logic and adds new security constraints). * Action Items: 1. Check hook scripts expecting `exit 0` to auto-approve. 2. Switch to JSON output for explicit approvals. 3. Update client versions for the new security setting. 7. **Final Review against constraints**: Concise? Yes. Insightful? Yes. Markdown? Yes. Chinese? Yes. 4 sections? Yes. *Self-Correction during drafting*: Make sure to emphasize that "exit 0" doesn't approve tools. This is the subtle but dangerous trap in the old docs. The new docs say "staying silent doesn't approve it". *Action items refinement*: 1. Review hooks relying on `exit 0`. 2. Use JSON for explicit approval. 3. Update for new security setting. Looks good.