Changes on 2026-09-30

1. **Analyze the Request:** * **Role:** World-class programmer analyzing documentation changes. * **Task:** Analyze a batch of 22 documentation pages (diffs provided). * **Output Format:** Markdown, concise, insightful, focus on what matters to developers. * **Language:** Chinese (as requested in the prompt). * **Structure:** 1. Overall Summary (1-2 sentences). 2. Key Themes (bullets). 3. Impact Level (Low/Medium/High/Breaking) with reasoning. 4. Action Items (bullets, max 5). 2. **Process the Diff Data (Page by Page):** * **Page: `cli-reference.md`** * *Change:* Added `--desktop` flag. Opens Claude Desktop app on the current directory and exits. Supports `--continue` and `--resume` (with session ID). Requires v2.1.285+, macOS/x64 Windows, Claude subscription. * *Change:* Updated `--permission-mode` description regarding `-p` runs and starting modes. * *Change:* Updated `--print` description regarding `--resume` on background sessions. * **Page: `desktop.md`** * *Change:* Added section for `claude --desktop`. Explains usage (open directly, open existing session), platform requirements, and behavior (prints opening message, exits). * **Page: `discover-plugins.md`** * *Change:* Clarified plugin status "Active, but a server needs setup". * *Change:* Updated SSH marketplace instructions to respect `GIT_SSH_COMMAND` or config. * *Change:* Differentiated "Configure options" (manifest `userConfig`) vs "Configure" (bundled MCP server `user_config`) in the plugin details menu. * **Page: `headless.md`** * *Change:* Refined explanation of `--allowedTools` (how `Bash` interacts with `auto` mode). * *Change:* Clarified default permission mode behavior in `-p` runs (can be `auto`, not just Manual). * **Page: `hooks.md`** * *Change:* Updated link text for Bash command auto-background eligibility (changed anchor link from `#background-commands` to `#foreground-commands-that-move-to-the-background`). * **Page: `iam.md`** * *Change:* Added "Restrict which API providers a machine may use". New managed setting `allowedProviders`. Restricts API providers (Anthropic, Bedrock, customEndpoint, etc.). Requires v2.1.285+. * **Page: `llm-gateway.md`** * *Change:* Added info on using `allowedProviders` with `customEndpoint` to enforce gateway usage via `ANTHROPIC_BASE_URL`. * **Page: `monitoring-usage.md`** * *Change:* Updated `managed_settings.trigger` event types (`policy_invalid` now covers provider enforcement, not just login). * *Change:* Added new error type `"provider_not_allowed"` for telemetry. * **Page: `interactive-mode.md`** * *Change:* Updated link text for auto-backgrounding commands (anchor change similar to hooks.md). * *Change:* Refined background command time limit text to reference other docs instead of inline explanation. * **Page: `skills.md`** * *Change:* Updated link text for auto-backgrounding commands (anchor change similar to hooks.md). * **Page: `slash-commands.md`** * *Change:* Updated link text for auto-backgrounding commands (anchor change similar to hooks.md). * **Page: `statusline.md`** * *Change:* Image URL updates (CDN hashes changed). No content change, just asset updates. * **Page: `third-party-integrations.md`** * *Change:* Added note about `allowedProviders` for limiting deployment options. * **Page: `vs-code.md`** * *Change:* Updated plugin management UI logic. * Disabling plugins in project settings: "Disable for me" vs "Disable for everyone". * New prompts for configuration options (sensitive fields masked). * Uninstall logic updates: "Disable for me" vs "Uninstall for everyone", keep/delete data choices. * Marketplace removal warnings (uninstalls plugins from that marketplace). * **Page: `about-claude/models/overview.md`** * *Change:* Removed `Claude Sonnet 4.5` from the "Legacy models" list (moved to active? Or just removed from legacy list?). *Actually*, looking closer, it removed the link for Sonnet 4.5 in the legacy list. * **Page: `about-claude/model-deprecations.md`** * *Change:* Formatting change (added left alignment `:` to table headers). * **Page: `api/models/list.md`** * *Change:* Added `"ce-plugins-2026-09-01"` to beta headers list. * **Page: `api/beta/files/upload.md`** * *Change:* Added `"ce-plugins-2026-09-01"` to beta headers list. * **Page: `api/errors.md`** * *Change:* Refined error message for disabled thinking on Sonnet 5.5. * **Page: `build-with-claude/claude-on-amazon-bedrock.md`** * *Change:* Table formatting updates (alignment). Marked `Claude Sonnet 4.5` as deprecated in the table. * **Page: `build-with-claude/claude-on-vertex-ai.md`** * *Change:* Table formatting updates. Marked `Claude Sonnet 4.5` as deprecated. * **Page: `build-with-claude/claude-in-microsoft-foundry.md`** * *Change:* Table formatting updates. Marked `Claude Sonnet 4.5` as deprecated. 3. **Synthesize Key Themes:** * **New CLI/Integration Feature:** `claude --desktop` flag to bridge CLI and Desktop app seamlessly. * **Enterprise/Admin Controls:** New `allowedProviders` managed setting for restricting API usage (gateway enforcement, specific cloud providers). * **Plugin Management Improvements:** Better distinction between plugin config vs bundled MCP server config, plus smarter UI flows in VS Code for shared settings (Disable for me vs everyone). * **Model Lifecycle Updates:** `Claude Sonnet 4.5` marked as deprecated in cloud provider tables. `Claude Sonnet 5.5` thinking behavior clarification. * **Documentation Maintenance:** Link anchor updates for background commands, CDN image updates, beta headers for "ce-plugins". 4. **Drafting the Response (Iterative refinement in Chinese):** * *Summary:* This update focuses on bridging workflows between the CLI and Desktop app, strengthening enterprise governance over API providers, and refining plugin management, while also updating model availability status. * *Themes:* * **CLI & Desktop Integration**: New `--desktop` flag. * **Enterprise Governance**: `allowedProviders` restriction. * **Plugin Ecosystem**: Improved config distinction and VS Code UX. * **Model Lifecycle**: Sonnet 4.5 deprecation notices. * *Impact:* Medium. The new governance features are big for enterprise admins, and the CLI flag is a nice quality-of-life improvement. The rest is polish. * *Action Items:* * Try `claude --desktop` if you use both CLI and Desktop. * Admins should review `allowedProviders` for stricter gateway enforcement. * Note the deprecation of Sonnet 4.5 on cloud platforms. 5. **Final Polish (formatting and tone):** * **Overall Summary:** 本次更新显著增强了 CLI 与 Desktop 应用的集成，引入了新的企业级 API 提供商限制功能，并优化了插件管理体验和模型生命周期说明。 * **Key Themes:** * **CLI 与 Desktop 深度集成**: 新增 `claude --desktop` 标志，允许从终端直接打开会话到 Desktop 应用。 * **企业管控增强**: 引入 `allowedProviders` 设置，允许管理员限制托管机器只能使用特定的 API 提供商（如仅限 Gateway 或 Bedrock）。 * **插件管理体验优化**: 明确区分了插件配置与内置 MCP 服务器配置，并在 VS Code 中改进了共享设置的禁用/卸载逻辑。 * **模型生命周期更新**: 在各云平台文档中将 `Claude Sonnet 4.5` 标记为已弃用，并澄清了 `Claude Sonnet 5.5` 的 thinking 参数行为。 * **Impact Level:** Medium. (Reasoning: The `allowedProviders` is a significant security/policy feature for enterprise; `--desktop` improves daily workflow for mixed users. The rest are minor doc updates). * **Action Items:** * 探索使用 `claude --desktop` 或 `--desktop --resume <id>` 来快速在 GUI 中打开终端会话。 * 企业管理员应检查是否需要配置 `allowedProviders` 来强制流量通过 Gateway 或限制云服务商。 * 注意到 `Claude Sonnet 4.5` 在各云平台上已标记为 deprecated，需规划迁移。 6. **Review against constraints:** * Markdown format? Yes. * Chinese? Yes. * Concise? Yes. * Action items max 5? Yes. 7. **Output Generation.**

Let me analyze the documentation changes across the three pages. 1. **skills.md** and **slash-commands.md** - Both have identical changes: - Replaced a paragraph about skills stopping influencing behavior with a link to a new dedicated troubleshooting section - Added a new section "Claude stops following a skill" with three cases: - Rule that must hold every time → use hooks - Guidance applied with judgment → word it for the whole task - Conversation was compacted → re-invoke the skill 2. **CHANGELOG.md.md** - Version 2.1.286 with extensive bug fixes and improvements: - Permission prompt improvements - Many security fixes (credential masking, secret redaction) - Remote Control fixes - MCP server fixes - Subagent fixes - UI improvements - Commit guidance improvement (verify skill) - Many more bug fixes Let me organize this into the requested format.

1. **Analyze the Request:** * **Role:** World-class programmer/developer tools expert. * **Task:** Analyze documentation changes (27 pages, 20 provided in the prompt) as a single batch. * **Output Format:** Markdown, concise, insightful, Chinese. * **Structure:** Overall Summary, Key Themes, Impact Level, Action Items. 2. **Scan the Diff Content (Page by Page - High Level):** * `desktop.md`: * "Auto" mode description change (background classifier details). * Skills loading behavior change regarding `disableSideloadFlags`. * Plugins behavior change regarding `disableSideloadFlags`. * *Theme:* Organization policy enforcement (`disableSideloadFlags`). * `hooks.md`: * `session_title` definition updated (custom title vs. generated title). * `UserPromptSubmit` hooks now receive `session_title`. * *Theme:* Hooks extensibility improvement. * `interactive-mode.md`: * Background tasks timeout logic changes (specifically for Bash/PowerShell, memory pressure reap, time limits). * Subagent behavior update. * *Theme:* Background task management/refinement. * `mcp.md`: * MCP server add command success message clarification. * Error handling (not saved/may not have been saved). * OAuth credentials security (HTTPS/localhost only). * Automatic backgrounding of MCP tools now shows progress. * Connectors tool blocking (`blocked` vs `ask`), organization controls. * *Theme:* MCP server management, security, and organization controls. * `skills.md` (and `slash-commands.md` - duplicate change): * Permission syntax clarification (`Skill(name)`). * Detailed table on what `deny` rules block (aliases, unqualified names, synced skills, plugin forms). * *Theme:* Permission system refinement for skills. * `about-claude/model-deprecations.md`: * Table updates (Claude Sonnet 4.5 status changed to Deprecated). * Deprecation history entry for Claude Sonnet 4.5. * *Theme:* Model lifecycle updates. * `api/overview.md`: * Pagination example code updated (more specific SDK examples). * `api/messages.md`, `api/messages/count_tokens.md`, `api/messages/batches/create.md`: * Model list formatting change (moved from collapsed list item to expanded items). * *Theme:* Documentation formatting (low functional impact). * `api/models/list.md`, `api/beta/files/upload.md`: * Model/Beta list formatting change (collapsed to expanded). * *Theme:* Documentation formatting. * `api/rate-limits.md`: * Grammar fix ("SDK's" vs "SDKs'"). * `api/beta-headers.md`: * Grammar/Phrasing changes ("SDKs' beta" -> "SDK's beta"). * More specific code examples for different SDKs (java, go, php, etc.). * `api/errors.md`: * Grammar/Phrasing changes ("SDKs'" -> "SDK's"). * `build-with-claude/streaming.md`: * Streaming helper descriptions updated to be more specific about methods per language. * PHP example updated to use `MessageAccumulator`. * `build-with-claude/files.md`: * Pagination link fix. * `build-with-claude/claude-on-amazon-bedrock.md`: * Phrasing "SDK detects" instead of "SDKs detect". * `build-with-claude/claude-on-vertex-ai.md`: * PHP SDK parameter name change (`location` -> `region`). * `build-with-claude/claude-in-microsoft-foundry.md`: * **Major Change:** Go SDK now supports Foundry natively (previously workaround only). * Updated code examples for Go (native client). * C# specific note about base URL. * `agents-and-tools/tool-use/overview.md`: * "SDK executes" instead of "SDKs execute". * `agents-and-tools/tool-use/bash-tool.md`: * "SDK exposes" instead of "SDKs expose". * `agents-and-tools/tool-use/computer-use-tool.md`: * Grammar fixes ("SDK" vs "SDKs"). * Icon removal from steps. * `agents-and-tools/tool-use/memory-tool.md`: * Phrasing changes ("SDK's beta namespace" vs "SDKs'"). 3. **Synthesize Findings:** * **Major functional changes:** * **Go SDK Foundry Support:** The Go SDK now supports Microsoft Foundry natively, replacing the need for workarounds. * **Vertex AI PHP SDK:** Parameter change (`location` to `region`) - potential breaking change for PHP users. * **Organization Policy Enforcement:** Stricter rules on loading cloud skills/plugins based on `disableSideloadFlags` in desktop environments. * **Background Task Logic:** Specific time limits and memory pressure behavior for background tasks (interactive mode). * **MCP Security:** OAuth credentials only sent to HTTPS or localhost. * **Skills Permissions:** More precise definition of what gets blocked when a `deny` rule is applied. * **Documentation/Grammar changes:** * Widespread correction of pluralization regarding "SDK" (using singular possessive "SDK's" or collective noun "The SDK" instead of plural "SDKs'" in many contexts, or standardizing on "The official SDK"). * Expanding collapsed lists (models, beta headers) into individual items for better readability. 4. **Drafting the Response (in Chinese):** * *Summary:* The updates focus on refining the developer experience (Go SDK Foundry support, PHP SDK parameter updates), tightening organization policy controls (skills/plugins, MCP OAuth), and clarifying background task behavior. There is also a large amount of documentation cleanup and formatting standardization. * *Key Themes:* * **Go SDK Foundry Support:** Native support added. * **Organization Policy & Security:** `disableSideloadFlags` enforcement, MCP OAuth HTTPS restriction. * **Behavioral Changes:** Background task timeouts, hooks receiving session titles. * **Permissions:** Skills blocking logic clarified. * **Documentation Formatting:** Expansion of parameter lists and grammar standardization ("SDK" usage). * *Impact Level:* Medium. * *Reasoning:* While mostly doc changes, the Go SDK Foundry update is significant for that ecosystem. The PHP SDK parameter change in Vertex AI is a minor breaking change. The organization policy changes affect enterprise users. * *Action Items:* * **Go Developers:** Update to use native Foundry client (`foundry.NewClient`) instead of standard client workarounds. * **PHP Developers (Vertex AI):** Check code for `location` parameter and update to `region`. * **Enterprise Admins:** Review the impact of `disableSideloadFlags` on skills/plugins loading in local/desktop sessions. * **MCP Developers:** Ensure token endpoints use HTTPS or localhost; OAuth will fail over HTTP on LAN. * **Hook Developers:** Utilize the new `session_title` field in `UserPromptSubmit` hooks. 5. **Refining the Analysis:** * Check the PHP change in `claude-on-vertex-ai.md`. It says `location: 'global'` became `region: 'global'`. This is a breaking change for PHP users. * Check the Go Foundry change. It moves from `anthropic.NewClient` with base URL tricks to `foundry.NewClient`. This is a major improvement/migration path for Go users on Foundry. * Check the Skills permissions. It clarifies that blocking a skill by name also blocks its plugin form or synced form. This is important for security policy. 6. **Final Polish (Markdown Construction):** * Translate concepts accurately. * Keep it concise. *(Self-Correction during drafting)*: Make sure to distinguish between the formatting changes (low impact) and the actual logic changes (medium impact). The prompt asks for what matters to developers. *Drafting Action Items:* 1. Go SDK Foundry users: Migrate to native client. 2. PHP Vertex AI users: `location` -> `region`. 3. MCP OAuth devs: Use HTTPS. 4. Desktop/Skills users: Note the policy flag behavior. 5. Hook users: New `session_title` available. 7. **Final Review against Constraints:** * Concise? Yes. * Insightful? Yes, highlights the subtle policy changes and SDK support additions. * Markdown? Yes. * Chinese? Yes. **(Self-Correction on `desktop.md`):** The change regarding `disableSideloadFlags` suggests that if an organization sets this flag, local sessions *won't* load cloud skills. This is a security hardening move. **(Self-Correction on `interactive-mode.md`):** Background tasks now have specific timeout rules (30 min default, 2 hour max). This is behavior to be aware of for long-running scripts. Okay, ready to generate output.

Let me analyze the documentation changes across all 5 pages. 1. **iam.md**: Reorganized the authentication troubleshooting link and added a "Log in with multiple accounts" heading. The troubleshooting link was moved from after the multi-account section to before it, making the flow more logical: login issues → multiple accounts. 2. **memory.md**: Significant restructuring of the "Write effective instructions" section. The content was reorganized from a paragraph-based format to a more structured list format with clearer headings. The prompt audit section was extracted into its own subsection "Audit your instruction files" with more detailed explanation. The key information is the same but better organized. 3. **model-config.md**: Updated Fable model consent prompt behavior documentation. Clarified that in the `/model` picker, selecting a Fable model keeps your current model. More importantly, added documentation about Agent SDK behavior - when another application hosts a session through Agent SDK, whether the consent prompt appears is up to that application, with a `dialogExpiry` deadline. In non-interactive mode and Agent SDK apps that don't show the prompt, consent is automatic. 4. **skills.md** and **slash-commands.md**: Identical changes - clarified how to turn off a skill for evaluation baselines. Added distinction between personal/project skills (use `skillOverrides`) and plugin skills (use `claude plugin eval`). This is important because `skillOverrides` doesn't apply to plugin skills. Key themes: - Documentation reorganization for better readability and flow - Clarification of behavior differences between contexts (Agent SDK vs non-interactive vs interactive) - Clarification of feature scoping (skillOverrides vs plugin eval for different skill types) - Better section structure with explicit headings Impact: Low to Medium. No breaking changes, but important clarifications that affect how developers use features.

1. **Understand the Goal**: The user wants a concise analysis of documentation changes (16 pages) from the perspective of a world-class programmer. The output must be in markdown and Chinese. 2. **Analyze the Input**: I need to go through each page's diff and identify the changes. * **amazon-bedrock.md**: Added a cross-reference link explaining how to use the 1M window without changing the pin. * **chrome.md**: Clarified managed MCP policies that disable the Chrome install prompt (added `managed-mcp.json` policy check). * **claude-code-on-the-web.md**: Added a note about macOS/Linux/WSL refusing upload if git attributes can't be followed (e.g., `core.attributesFile`). * **cli-reference.md**: Removed specific version requirement ("Available in Claude Code v2.1.195 and later") for `claude gateway`. This implies the feature is now stable or the baseline version has moved up. * **google-vertex-ai.md**: Similar to Amazon Bedrock, added a cross-reference link about 1M window usage. * **hooks.md**: * Removed outdated notes about hyphenated matchers requiring v2.1.195 (likely cleanup/standardization). * Clarified behavior of `UserPromptSubmit` blocking (doesn't erase prompt context everywhere). * Added new section "What a blocked prompt leaves behind" explaining `suppressOriginalPrompt` and privacy implications (transcript files). * **mcp.md**: * Removed version requirement ("Before v2.1.195...") for path placeholders in `headersHelper`. * Removed version requirement for plugin MCP server working directory. * Removed note about older versions skipping tools with `anyOf`/`oneOf`/`allOf`. * **model-config.md**: * Updated delivery mechanism table for Cowork (remote Cowork sessions check model allowlist at server level). * Updated 1M context window troubleshooting: check pinned models if not seeing the option. * Clarified behavior of `ANTHROPIC_DEFAULT_*_MODEL` variables and the `/model` picker regarding 1M context. * Added links to "default model" definitions for caching env vars. * **monitoring-usage.md**: * Clarified Gateway identity attributes for CLI vs. Desktop/Cowork. * Removed specific Gateway identity explanation from one section and redirected to "Standard attributes". * **network-config.md**: * Refined watchdog timer behavior on Bedrock. * Added note about `CLAUDE_ENABLE_BYTE_WATCHDOG_BEDROCK=1` replacing body idle timeout. * Clarified GitHub Enterprise Server access (SCM connector availability). * **plugin-marketplaces.md**: Link fixes/corrections regarding plugin naming rules and validation. * **sandboxing.md**: Minor clarification on bare git repository protection (existing config/hooks vs generic files). * **skills.md** & **slash-commands.md**: * Added a large new section about the bundled `/claude-api` skill. * Listed subcommands: `migrate`, `upgrade`, `managed-agents-onboard`, `prompt-audit`, `cost-optimize`, `build-eval`, `hillclimb`, `preserved-thinking-migration`. * Minimum versions provided for each. * **sub-agents.md**: Removed outdated note about hyphenated matchers (similar to hooks cleanup). * **third-party-integrations.md**: Updated pricing link/text for Claude Teams. 3. **Synthesize the Themes**: * **Version Cleanup**: Removal of "v2.1.195" specific caveats (gateway, MCP path placeholders, hooks matchers). This suggests v2.1.195 is now the baseline or the quirks of older versions are no longer relevant. * **1M Context Window**: Enhanced documentation around pinning models and using 1M context on third-party providers (Bedrock/Vertex). * **New `/claude-api` Skill**: Introduction of a powerful new skill for managing Claude API projects (migration, optimization, evals). * **Telemetry/Gateway**: Clarified identity attribution for CLI vs. Desktop/Cowork. * **Security/Sandboxing**: Refinements on git attribute handling and prompt blocking privacy. 4. **Determine Impact Level**: * **High**: The new `/claude-api` skill with subcommands for `cost-optimize`, `hillclimb`, `build-eval`, etc., is a major workflow improvement for API developers. The 1M context clarifications are also high value for users on Bedrock/Vertex. * *Correction*: While significant, it's documentation updates. "High" impact usually implies breaking changes or major new features. The *feature* (the skill) likely exists, now it's documented. The 1M context fixes are helpful. Let's go with **Medium-High** or simply **High** because the `/claude-api` skill capabilities (like cost optimization and eval building) are significant developer productivity tools. Let's stick to **Medium-High** (or High). I will categorize it as **High** because the new skill documentation reveals powerful capabilities that developers weren't explicitly aware of how to invoke systematically. 5. **Draft Action Items**: * Try the new `/claude-api` skill commands for API projects (especially `cost-optimize` and `build-eval`). * Review pinned model configurations if 1M context isn't showing up on Bedrock/Vertex. * Update managed MCP policies if Chrome installation prompts are behaving unexpectedly. * Note the cleanup of older version-specific behaviors (no longer need to worry about v2.1.195 limitations for hooks/MCP). 6. **Translate to Chinese**: * *Summary*: 文档更新主要集中在移除旧版本限制说明、增强第三方平台上 1M 上下文窗口的配置说明，以及引入了新的 `/claude-api` 内置技能文档。 * *Themes*: 移除 v2.1.195 版本相关限制说明、1M 上下文窗口配置优化、新增 `/claude-api` 技能文档、身份归属与遥测澄清。 * *Impact*: 高（High）。理由：新增的 `/claude-api` 技能提供了成本优化、评估构建等强大工作流；1M 上下文配置的澄清对云端用户很关键。 * *Action Items*: * 在 API 项目中尝试 `/claude-api` 的子命令（如 `cost-optimize`, `hillclimb`）。 * 如果在 Bedrock/Vertex 上无法看到 1M 上下文选项，检查模型固定配置。 * 如果 Chrome 集成安装提示被屏蔽，检查 MCP 管理策略。 * 利用 `UserPromptSubmit` 的 `suppressOriginalPrompt` 功能保护敏感提示词隐私。 7. **Refine**: Ensure the tone is professional ("world-class programmer") and the markdown is clean. * *Refinement on "High" impact*: Is it really high? It's mostly documentation cleaning and revealing a feature. Let's say **Medium**. It helps developers use existing tools better but doesn't change the API contract or break code. *Correction*: The prompt says "Focus on what matters to developers". The new skill documentation is *very* helpful. But documentation changes are rarely "High" impact compared to breaking changes. I'll settle on **Medium** but emphasize the utility of the new skill. Actually, looking at the "What a blocked prompt leaves behind" section - that's a security/privacy nuance. The 1M context fixes are important. Okay, let's call it **Medium**. * *Re-evaluating the "Version Cleanup"*: The removal of "Requires v2.1.195" implies the current stable version is likely well past that, so maintaining docs for ancient versions is unnecessary. This simplifies the docs. * *Drafting the response in Chinese*: * **Summary**: 此次更新主要清理了过时的版本限制说明（如 v2.1.195），完善了第三方平台（Bedrock/Vertex）上 1M 上下文窗口的配置指南，并重点引入了新的 `/claude-api` 内置技能文档，用于 API 项目的迁移、优化和评估。 * **Key Themes**: * **清理旧版本说明**: 移除了多处针对 v2.1.195 之前的版本限制说明（如 Hooks 中的连字符匹配规则、MCP 路径占位符支持等）。 * **1M 上下文窗口增强**: 详细说明了在 Amazon Bedrock 和 Google Vertex AI 上如何通过固定模型和环境变量启用 1M 上下文，并提供了故障排查指引。 * **新增 `/claude-api` 技能**: 详细记录了用于处理 Claude API 项目的内置技能，包含 `migrate`、`cost-optimize`、`hillclimb`、`build-eval` 等子命令，覆盖从代码迁移到成本分析和模型迭代的全流程。 * **安全与隐私细化**: 明确了 Hook 拦截提示词后的数据残留行为（`suppressOriginalPrompt`），以及针对 Bedrock 的网络超时配置（Byte watchdog）。 * **身份归属澄清**: 区分了 CLI 与 Desktop/Cowork 会话在通过 Gateway 连接时的身份遥测属性。 * **Impact**: **Medium**（中等）。 * **理由**: 虽然主要是文档层面的优化，但新增的 `/claude-api` 技能文档为开发者提供了极具价值的工作流指引（如成本优化、Evals 构建），且 1M 上下文的配置澄清直接解决了企业环境下的常见痛点。 * **Action Items**: * **探索新技能**: 在使用 Anthropic SDK 的项目中尝试 `/claude-api` 技能（如 `/claude-api cost-optimize` 进行成本分析）。 * **排查 1M 上下文**: 若在第三方平台未看到 1M 上下文选项，检查是否手动固定了模型且未添加后缀。 * **审查 Hook 隐私**: 如果 Hook 拦截包含敏感信息的 Prompt，利用 `suppressOriginalPrompt` 防止其被写入磁盘日志。 * **更新策略配置**: 若 Chrome 集成安装提示被抑制，检查 `managed-mcp.json` 策略是否允许。 8. **Final Polish**: Check against the markdown structure required. * Overall Summary: Done. * Key Themes: Done. * Impact: Done. * Action items: Done. (Self-Correction on Action Items): Keep it tight. Max 5 bullets. Let's double check the `skills.md` content. It lists subcommands. This is a big addition for API developers. Let's double check `model-config.md`. The change about Cowork receiving server-managed settings "Remote Cowork sessions: the server checks the model" is a specific behavioral clarification. Ready to generate output.