Changes on 2026-09-25

Let me analyze all the documentation changes across these 16 pages and provide a comprehensive summary. Let me go through each page's changes: 1. **cli-reference.md**: `--bg`/`--background` flag now checks workspace trust before starting a background agent session. 2. **costs.md**: Added clarification that subagent requests draw on your usage; tips to reduce costs by choosing smaller models. Added "Subagents and workflows" to the list of idle-context consumers. 3. **headless.md**: SIGTERM behavior changed - now leaves interrupted turn as-is instead of continuing it on resume (opt-in via `CLAUDE_CODE_RESUME_INTERRUPTED_TURN=1`). Added section about working directory being deleted mid-session. Plugin error tracking now includes `path` field for `--plugin-dir` failures. 4. **hooks.md**: Added link for scratchpad_dir. Updated deferred session resume in plan mode with a cross-reference. Added info about certain launch flags preventing plan mode restoration on resume. 5. **interactive-mode.md**: Commands that timeout now automatically move to background instead of stopping (unless `sleep`). 6. **mcp.md**: - `"type": "sdk"` servers now only work in SDK host applications, skipped in `.mcp.json`/settings - URL endpoint deduplication now considers case-insensitive scheme/host, default ports, trailing slashes - MCP tool image results (PNG, JPEG, GIF, WebP) now saved to files and available for Claude to process - MCP elicitation URL mode now asks before opening browser - Elicitation client capabilities declared on newer protocol revision - MCP Apps UI resources (`ui://` URIs) described - they don't appear in resource listings 7. **model-config.md**: Major changes: - `deniedModels` and `availableModelsMatch` settings introduced to block specific models/versions - `availableModels: []` behavior changed - now blocks named selections AND default model - Organization model restrictions consolidated - New section on blocking specific models or versions 8. **monitoring-usage.md**: - Debug verification changed from `--debug` to `--debug-file <path>` - `request_id` field description changed to match event correlation attribute - `tool.output` span event now supported for MCP tools, WebFetch, WebSearch - Background-moved WebFetch/WebSearch calls excluded from tool.output - New attributes added: `url`, `status_code`, `method` 9. **overview.md/quickstart.md/setup.md**: Installation instructions improved - terminal guide reference, `claude --version` verification step, PATH troubleshooting link. 10. **plugins-reference.md**: `claude plugin validate` now checks `outputStyles`, `lspServers`, `monitors`, and `themes` paths (previously didn't). 11. **sandboxing.md**: "Session temp directory" renamed to "per-user temp directory" across multiple places. This is a terminology change. 12. **skills.md/slash-commands.md**: - Skills can now use frontmatter `name` field to override directory name as command name - Name resolution clarified for frontmatter `name` field - Synced skill name collision behavior updated 13. **sub-agents.md**: Subagents' own requests count toward usage limits. Warning updated to mention token spending. Key themes: 1. **Cost/usage transparency**: Multiple pages clarify that subagents/workflows consume their own tokens 2. **Model access control**: New `deniedModels`/`availableModelsMatch` settings, changed `availableModels: []` behavior 3. **MCP enhancements**: Image handling, SDK-type restriction, URL dedup, elicitation improvements 4. **Installation UX**: Better onboarding with terminal guide and PATH troubleshooting 5. **Behavior changes**: SIGTERM resume behavior, timeout→background, `availableModels: []` semantics 6. **Debugging improvements**: `--debug-file`, plugin validation expanded, plugin error paths Impact: Medium - Several behavior changes but most are additive or have opt-in/out mechanisms. The `availableModels: []` change and SIGTERM resume change could surprise existing users.

1. **Analyze the Request:** * **Role:** World-class programmer/developer tools expert. * **Task:** Analyze a batch of documentation changes (14 pages). * **Format:** Concise, insightful markdown analysis in Chinese. * **Focus:** What matters to developers. * **Specific Output:** Overall Summary, Key Themes, Impact Level, Action Items. 2. **Analyze the Diffs (Page by Page):** * **Page: `discover-plugins.md`** * *Change:* Clarified "Project scope" plugin installation. Added note that committing the `.claude/settings.json` entry enables it for collaborators but doesn't download it; they still need to run `claude plugin install`. * *Theme:* Plugin installation/Collaboration workflow. * *Impact:* Clarifies behavior, prevents confusion about missing downloads. * **Page: `iam.md`** * *Change:* Expanded how to set `CLAUDE_CODE_OAUTH_TOKEN`. Changed from a single bash export example to a tabbed interface including macOS/Linux/WSL, Windows PowerShell, and Windows CMD. * *Theme:* Cross-platform configuration/Accessibility. * *Impact:* Improves DX for Windows users. * **Page: `interactive-mode.md`** * *Change:* Reformatted a table (General controls) to use a specific markdown structure (likely syntax highlighting or better formatting). The content description seems similar, but the rendering/formatting changed (e.g., `Ctrl+C` descriptions). *Correction:* Looking closely, it's a table structure change. The descriptions look identical. * *Theme:* Documentation formatting/Readability. * *Impact:* Low (visual improvement). * **Page: `mcp.md`** * *Change:* Added a note that prompts from an MCP server named `anthropic-skills` don't appear because that name is reserved for synced skills. Tools still work. Suggests renaming. * *Theme:* MCP/Skills naming conflicts/Reservations. * *Impact:* Important for MCP server developers avoiding conflicts. * **Page: `skills.md`** (and `slash-commands.md`, likely duplicate content or similar context) * *Change:* Updated "Reserved name" section. Added `anthropic-skills` as a reserved name (outside plugins). Updated table row for synced skills to specify "short name". * *Theme:* Skills naming/Namespacing. * *Impact:* Medium (prevents naming collisions). * **Page: `terminal-config.md`** * *Change:* Added a section on capping response width in wide terminals via `maxProseWidth` setting. * *Theme:* Terminal UI customization/Readability. * *Impact:* Medium (UX improvement for wide screens). * **Page: `troubleshooting.md`** * *Change:* Added entry for "disk quota full" errors in the troubleshooting table. * *Theme:* Error handling/Troubleshooting. * *Impact:* Low/Medium (helps debugging). * **Page: `build-with-claude/overview.md` & `text-generation.md` & `files.md`** (Platform docs) * *Change:* Updated availability badges. `Agent Skills`, `Files API`, `MCP connector` moved from `Beta` to `GA` (General Availability) on specific platforms (AWS/Azure). Updated footnotes regarding Microsoft Foundry limitations (Skill version download endpoint not supported). Updated model support in `web-fetch-tool.md` (e.g., `claude-opus-5-5` replacing `4-8`). * *Theme:* Feature maturity/GA releases/Platform support updates. * *Impact:* High (Production readiness). * **Page: `web-fetch-tool.md`** * *Change:* Updated supported models list for dynamic filtering to "Claude 4.6 and later". Updated code examples to use `claude-opus-5-5` instead of `claude-opus-4-8`. * *Theme:* Model versions/Feature support. * *Impact:* Medium (Code snippets/models need updating). * **Page: `agent-skills/overview.md`** * *Change:* Added limitation note about Microsoft Foundry not supporting the Skill version download endpoint. * *Theme:* Platform limitations. * *Impact:* Medium (for Foundry users). * **Page: `CHANGELOG.md.md`** (Version 2.1.283) * *Change:* Massive list of features, fixes, and improvements. * *Highlights:* * `x-claude-code-prompt-id` header for gateways. * `availableModelsMatch` and `deniedModels` settings for model management. * `/doctor prompt-audit` command. * Fixes for MCP, plugins, Vim mode, Remote Control, telemetry. * Improvements to UI lists (`/tasks`, `/mcp`, etc.). * *Theme:* Feature additions/Bug fixes/CLI improvements. * *Impact:* High (New management capabilities, stability fixes). 3. **Synthesize and Structure (Drafting the Analysis):** * *Overall Summary:* This update primarily focuses on advancing platform features to General Availability (GA), clarifying naming conflicts for Skills/MCP, improving cross-platform documentation (specifically Windows), and releasing significant CLI updates (v2.1.283) including new model management controls and UI improvements. * *Key Themes:* * **Platform Maturity:** Agent Skills, Files API, and MCP connectors moved to GA on AWS and Azure. * **Naming & Namespacing:** strict reservation of `anthropic-skills` and `synced` for Skills/MCP prompts to prevent conflicts. * **CLI & Configuration:** New settings for model control (`deniedModels`, `maxProseWidth`), Windows-specific config examples, and enhanced list UI. * **Model Updates:** Code examples updated to use newer model versions (e.g., `claude-opus-5-5`). * **Bug Fixes:** A large changelog focusing on MCP stability, Vim mode, and plugin management. * *Impact Level:* **High**. * *Reasoning:* Multiple features moving to GA indicates production readiness. New configuration options for model management (`deniedModels`) are critical for enterprise/admin control. Large number of bug fixes suggests stability improvements. Naming conflicts can break existing custom Skills/MCP servers. * *Action Items:* * Review custom Skill/MCP server names to ensure they don't conflict with the reserved `anthropic-skills` or `synced` names. * Update code examples/model strings to use the latest Opus 5.5 models if applicable. * Test project-scoped plugin installation workflows with collaborators, ensuring they run the install command locally. * Consider utilizing the new `/doctor prompt-audit` command to check legacy prompting patterns. * Update configuration documentation for Windows users using the new PowerShell/CMD examples. 4. **Refine Language (Chinese):** * Translate concepts accurately: * "General Availability" -> 正式发布 (GA). * "Reserved name" -> 保留名称. * "Project scope" -> 项目作用域. * "Changelog" -> 更新日志. 5. **Final Polish (Markdown construction):** * Ensure the tone is professional and concise. * Check against the "max 5 action items" constraint. *(Self-Correction during drafting)*: The `interactive-mode.md` diff was just a table reformatting, I'll minimize its emphasis. The CHANGELOG is huge, so I will pick the highest impact items (model management, doctor command, MCP fixes). 6. **Final Output Generation:** (Matches the provided good response).

用户要求我分析 15 个页面的文档变更，提供简洁、深入的中文 markdown 格式分析。 让我先快速浏览这些变更： 1. **plugin-marketplaces.md**: 将 "community marketplace" 改为 "directory" 2. **plugins.md**: 将多处 "community marketplace" 改为 "directory"，添加了关于 MCP 服务器的链接 3. **plugins-reference.md**: 主要是格式调整（空格变化） 4. **migration-guide.md**: API release notes 链接修正 5. **api/overview.md**: Token Counting API URL 修正 6. **api/messages.md**: 大量格式调整（minLength/maxLength 顺序），模型列表重新排序，添加了新模型和弃用警告 7. **api/messages/count_tokens.md**: 类似 messages.md 的格式调整 8. **api/messages/batches/create.md**: 类似格式调整 9. **api/models/list.md**: 参数描述格式调整 10. **api/beta/files/upload.md**: 格式调整 11. **build-with-claude/overview.md**: MCP 链接修正为 modelcontextprotocol.io 12. **build-with-claude/text-generation.md**: 同上 13. **build-with-claude/batch-processing.md**: 批处理端点链接修正 14. **build-with-claude/claude-on-amazon-bedrock.md**: Bedrock 模型表格重新格式化 15. **resources/glossary.md**: MCP 链接修正 总结主要变更： 术语变更主要体现为从 "marketplace" 到 "directory" 的转变，这可能反映了产品定位的调整。模型列表的重新排序意味着文档结构的优化，可能是为了提高信息的可读性和清晰度。这些变更虽然看似细微，但可能代表了产品的战略性调整。 API 文档的细节修正包括链接更新和参数顺序的标准化，这些改进有助于提升开发者使用文档的准确性和一致性。同时，新增了 "claude-opus-5-5"、"claude-fable-5-1" 等模型，显示出模型家族的持续扩展。 变更对开发者影响中等，主要是术语变化和文档优化。建议开发者注意新产品名称，并根据文档更新调整工作流程，以保持与最新发展同步。

1. **Analyze the Request:** * **Role:** World-class programmer, expert in dev tools and docs. * **Task:** Analyze documentation changes (Changelog). * **Input:** One page (`CHANGELOG.md.md`) with a specific diff (2 additions, 1 deletion). * **Language:** Chinese (requested in the prompt). * **Output Format:** Markdown, specifically: Summary, Key Themes, Impact Level, Action Items. 2. **Analyze the Input (Diff):** * *Change 1 (Line ~231):* `+ Changed AGENTS.md support to also work on Amazon Bedrock, Google Vertex AI, Microsoft Foundry, LLM gateways, and sessions with telemetry disabled` * *Meaning:* The AGENTS.md feature (project instructions) was previously limited, but now supports major cloud providers (Bedrock, Vertex, Foundry) and specific configurations (gateways, no telemetry). * *Change 2 (Line ~396):* `- Added AGENTS.md support... (not yet on Bedrock, Vertex or Foundry)` -> `+ Added AGENTS.md support...` (removed the limitation note). * *Meaning:* This confirms the removal of the restriction mentioned in Change 1. The documentation now reflects that the limitation is gone. 3. **Synthesize the Content:** * **Overall Summary:** The update expands the compatibility of the `AGENTS.md` feature to include major cloud providers (AWS Bedrock, Google Vertex, Microsoft Foundry) and specific runtime configurations, removing previous limitations. (关键更新：扩展了 AGENTS.md 功能的兼容性，支持主要的云服务商并移除了之前的限制。) * **Key Themes:** * *Feature Parity:* Bringing AGENTS.md support to non-default environments (Bedrock, Vertex, Foundry). * *Documentation Correction:* Updating docs to reflect removed restrictions. * *Broader Compatibility:* Support for LLM gateways and telemetry-disabled sessions. * **Impact Level:** * *Level:* Medium. * *Reasoning:* It's not "Breaking" (no code breaks). It's not "Low" because it opens up a specific feature (`AGENTS.md` custom instructions) to enterprise/cloud users who were previously excluded. It's a significant feature enablement. * **Action Items:** * *Developers using Bedrock/Vertex/Foundry:* You can now use `AGENTS.md` for project-specific instructions. * *Developers using Gateways/Telemetry-off:* You can now use `AGENTS.md`. * *General:* Update docs/READMEs if they previously warned about incompatibility. 4. **Drafting the Response (in Chinese):** * *Summary:* 本次更新主要移除了 `AGENTS.md` 功能对特定云平台（Amazon Bedrock、Google Vertex AI、Microsoft Foundry）及特定配置（网关、无遥测）的限制，实现了功能的全平台兼容。 * *Themes:* 功能扩展、文档更新、云平台兼容性。 * *Impact:* Medium. Reason: 为使用企业级/第三方 LLM 服务的开发者解锁了项目自定义指令功能。 * *Action Items:* 尝试在这些新平台上配置 AGENTS.md；移除关于不支持 AGENTS.md 的旧文档备注。 5. **Refining the Output (World-class Programmer Persona):** * Make it sharp and professional. * Focus on "Why it matters": It means users can standardize project instructions across *all* supported backends, not just the standard Anthropic API. 6. **Final Polish (Markdown Structure):** * **1. Overall Summary**: 本次更新消除了 `AGENTS.md` 功能对特定云服务商和配置的限制，将其支持范围扩展到了 Amazon Bedrock、Google Vertex AI、Microsoft Foundry 以及使用 LLM 网关或禁用遥测的会话中。 * **2. Key Themes**: * **功能扩展与统一**：实现了项目级指令（`AGENTS.md`）在不同后端提供商之间的一致性体验。 * **企业级支持增强**：专门针对企业常用的 Foundry 和网关架构进行了适配。 * **文档修正**：删除了关于不支持 Bedrock/Vertex/Foundry 的过时备注。 * **3. Impact Level**: * **Medium** (中等) * **Reasoning**: 此变更不会破坏现有代码，但对依赖 Bedrock、Vertex 或 Foundry 等非 Anthropic 直连环境的开发者至关重要，它解锁了在这些平台上配置自定义项目指令的能力，提升了开发体验的一致性。 * **4. Action items for developers**: * 如果您正在使用 Bedrock、Vertex AI 或 Foundry，现在可以在项目中创建 `AGENTS.md` 来定义项目特定的指令。 * 检查现有项目文档，移除关于“不支持 Bedrock/Vertex/Foundry”的说明。 * 如果您禁用了遥测或使用 LLM 网关，现在也可以安全使用 `AGENTS.md` 功能。 7. **Review against constraints:** * Chinese? Yes. * Markdown? Yes. * Single batch analysis? Yes. * Concise? Yes. 8. **Final Output Generation.** (Proceed to generate response).