Changes on 2026-06-29

1. **Analyze the Request:** * **Role:** World-class programmer, expert in dev tools/docs. * **Input:** A changelog diff for version 2.1.196. * **Output Language:** Chinese. * **Format:** Markdown, concise, insightful. * **Structure:** Overall Summary, Key Themes, Impact Level, Action Items. 2. **Analyze the Input (CHANGELOG.md.md diff):** * *Version:* 2.1.196. * *New Features:* Organization default models, readable default session names, clickable file attachments. * *Security:* `claude mcp list`/`get` no longer spawns servers from untrusted/repo-approved `.claude/settings.json` (pending approval required). * *Bug Fixes:* * Waking background jobs deleting conversation (critical data loss bug fixed). * Rate limit warning flickering/telemetry over-counting. * Duplicate recap lines. * PowerShell git diff/grep exit code handling (exit 1 is not failure). * `claude agents` side panel issues (focus, subagent types, status). * `--dangerously-skip-permissions` falling back to auto mode. * Mid-turn crash recovery for Remote sessions. * Sessions moved with `/cd` reappearing incorrectly. * `claude plugin validate` skipping local plugins. * `Esc Esc` regression, Ctrl+C for background agents. * MCP OAuth scope issues (GitLab enterprise). * `/context` token count on Bedrock. * `/deep-research` verifier failure messages. * Plugin dependency version pins. * `claude agents` session status labels. * Voice dictation swallowing spaces. * *Improvements:* * Background session reliability (survive restart/kill, Windows support). * Background agents auto-resume after daemon restart. * `/code-review` workflow efficiency (25% token reduction). * TUI rendering performance (skipping no-op subtrees). * Streaming idle watchdog default enabled (5 min timeout). * Remote Control disabled for non-Anthropic hosts. * UI change: opening agents view requires one `←` press. 3. **Synthesize - Overall Summary:** * Version 2.1.196 focuses heavily on stability, reliability (especially for background tasks and Remote sessions), and security hardening regarding MCP servers. It also introduces user experience improvements like organization defaults and UI refinements. 4. **Synthesize - Key Themes:** * *Reliability & Resilience:* Fixes for data loss (waking jobs), crash recovery, background job survival across restarts. * *Security:* Tightening MCP server spawning rules to prevent untrusted execution. * *Tooling & Agents:* Fixes for `claude agents`, PowerShell compatibility, and plugin validation. * *UX & Performance:* Better session naming, TUI performance, `/code-review` efficiency. * *Cloud Integration:* Bedrock token fixes, GitLab OAuth fixes, Remote Control behavior. 5. **Synthesize - Impact Level:** * *Level:* High. * *Reasoning:* It fixes a critical bug where waking a background job could permanently delete a conversation (data loss). It changes security defaults regarding MCP server spawning. It significantly improves the reliability of background processes. 6. **Synthesize - Action Items:** * Review `.claude/settings.json` if using MCP, as untrusted servers now require manual approval. * Note the new environment variable `CLAUDE_ENABLE_STREAM_WATCHDOG=0` if the 5-minute timeout causes issues. * Update user behavior regarding background agents (Ctrl+C to stop, auto-resume functionality). * Be aware of the UI change (single `←` to open agents view). 7. **Drafting the Response (in Chinese):** * *Summary:* 版本 2.1.196 侧重于后台任务的稳定性、数据安全性以及用户体验的优化。它修复了可能导致对话丢失的关键 Bug，加强了 MCP 服务器的安全审查，并提升了远程会话和代理 的恢复能力。 * *Themes:* * **稳定性与可靠性**：修复了唤醒后台任务导致对话永久删除的严重 Bug；大幅改进了后台会话在进程终止或重启后的存活与自动恢复能力。 * **安全性增强**：收紧了 MCP 服务器启动策略，现在不再自动执行仓库通过 `.claude/settings.json` 自批准的 `.mcp.json` 服务器，需手动审批。 * **开发者工具修复**：修复了 PowerShell 中 git 命令退出码处理、插件验证跳过、OAuth 授权范围请求等多个开发者工具层面的兼容性问题。 * **体验与性能优化**：改进了 `/code-review` 工作流的 Token 效率，默认开启流式看门狗超时机制，并优化了终端 UI 渲染。 * *Impact:* High (高)。 * *Reasoning:* 此版本包含了一个针对“唤醒后台任务导致对话永久删除”的关键修复，直接影响数据安全。同时，MCP 服务器启动逻辑的安全变更可能会影响现有的自动化工作流，需要开发者关注。 * *Action Items:* * 检查 MCP 配置：确认仓库中的 `.claude/settings.json` 不会被自动信任，如有需要请手动审批新出现的“Pending approval”服务器。 * 留意超时机制：流式看门狗默认开启（5分钟），如果您的长任务受此影响，请设置环境变量 `CLAUDE_ENABLE_STREAM_WATCHDOG=0`。 * 使用后台代理：了解后台代理现在会在守护进程重启后自动恢复，停止它们请使用 `Ctrl+C`。 * 测试 PowerShell 环境：如果在 Windows 上运行，验证 `git diff`/`grep` 等命令的行为是否符合预期（现已对齐 Bash 行为）。 8. **Final Review:** Does the output match the constraints? Yes. Concise, markdown, Chinese, insightful. * *Self-Correction during drafting:* Ensure the "Esc Esc" fix is mentioned as a regression fix. Ensure the significance of the "waking a background job deleting conversation" fix is highlighted as a major reason for the High impact. Mention the UI change for agents view. 9. **Final Polish:** (Proceed to generate output based on drafted points).

1. **Analyze the Request:** * **Role:** World-class programmer, expert in dev tools/docs. * **Task:** Analyze a batch of 13 documentation changes. * **Output Format:** Markdown, concise, insightful, Chinese language. * **Specific Requirements:** Overall summary, Key themes, Impact level, Action items. 2. **Analyze the Input Data (Diff-by-Diff):** * *Page 1: `about-claude/models/overview.md`* * Diff shows a table comparison. The specific line changes are a bit cut off at the end (`+| [Extended thinking](/docs/en/build-with-claude/extended-thinking)** | No`), but the context implies table formatting/updates. * Wait, looking closer at the snippet provided for Page 1: The diff cuts off abruptly. However, usually, these diffs imply updates to feature flags or metadata. Let's look at the context provided. It shows "Latest models comparison". The diff shows the table headers and rows. * *Correction*: The provided diff for page 1 is actually incomplete (ends mid-line). I need to be careful not to over-interpret. However, usually, updates in the overview table reflect deprecations or feature additions seen elsewhere. Let's look at the other pages to infer the big picture. * *Page 2: `about-claude/pricing.md`* * **Key Change:** "Fast mode" text updated. * **Detail:** Removed "Claude Opus 4.6" from the list of supported models in the intro sentence. * **Detail:** Table removed Opus 4.6, kept 4.7 and 4.8. * **Crucial Detail:** Added a note: "Fast mode for Claude Opus 4.7 is deprecated and will be removed on July 24, 2026." * **Crucial Detail:** "As of June 29, 2026, fast mode is not available on Claude Opus 4.6: requests... run at standard speed and are billed at standard rates." * *Page 3: `api/rate-limits.md`* * **Key Change:** Fast mode rate limits section. * **Detail:** Removed "or Opus 4.6" from the list of models with dedicated rate limits. * **Detail:** Added note stating Fast Mode is not available on Opus 4.6 (runs at standard speed). * *Page 4: `build-with-claude/context-windows.md`* * **Key Change:** Context window sizes. * **Detail:** Microsoft Foundry now supports 1M tokens for Opus 4.8, 4.7, 4.6, and Sonnet 4.6 (previously it only supported 200k for Opus 4.8). * *Page 5: `build-with-claude/structured-outputs.md`* * **Key Change:** Availability on Microsoft Foundry. * **Detail:** Changed "Structured outputs are available in beta" to "generally available" on Microsoft Foundry. * **Detail:** Added requirement: requires a "Hosted on Anthropic deployment". * *Page 6: `build-with-claude/files.md`* * **Key Change:** Files API on Microsoft Foundry. * **Detail:** Added requirement: requires a "Hosted on Anthropic deployment". * *Page 7: `build-with-claude/claude-in-microsoft-foundry.md`* * **Key Change:** "Hosted on Azure" vs "Hosted on Anthropic" comparison table. * **Detail:** "Hosted on Azure" model availability changed from "Opus, Sonnet, and Haiku" to just "Opus and Haiku". (Sonnet seems removed from Hosted on Azure? Or implies Sonnet isn't the "latest" there? It says "The latest models in the Opus and Haiku families"). * **Detail:** Updated deployment instruction text regarding "Default settings" to be more specific about when it auto-configures to Azure. * *Page 8: `agents-and-tools/tool-use/web-search-tool.md`* * **Key Change:** Web search tool on Microsoft Foundry. * **Detail:** Added requirement: requires a "Hosted on Anthropic deployment". * *Page 9: `agents-and-tools/tool-use/web-fetch-tool.md`* * **Key Change:** Web fetch tool on Microsoft Foundry. * **Detail:** Added requirement: requires a "Hosted on Anthropic deployment". * *Page 10: `agents-and-tools/tool-use/code-execution-tool.md`* * **Key Change:** Code execution on Microsoft Foundry. * **Detail:** Added requirement: requires a "Hosted on Anthropic deployment". * *Page 11: `agents-and-tools/tool-use/programmatic-tool-calling.md`* * **Key Change:** Programmatic tool calling on Microsoft Foundry. * **Detail:** Added requirement: requires a "Hosted on Anthropic deployment". * *Page 12: `agents-and-tools/agent-skills/overview.md`* * **Key Change:** Agent Skills on Microsoft Foundry. * **Detail:** Added requirement: requires a "Hosted on Anthropic deployment". * *Page 13: `agents-and-tools/mcp-connector.md`* * **Key Change:** MCP connector on Microsoft Foundry. * **Detail:** Added requirement: requires a "Hosted on Anthropic deployment". 3. **Synthesize the Findings:** * **Theme 1: Fast Mode Deprecation/Restriction.** Fast Mode is being phased out for older models (4.7 deprecated, 4.6 removed/ignored). * **Theme 2: Microsoft Foundry Hosting Requirements.** This is the biggest theme. There is a strong push to differentiate between "Hosted on Azure" and "Hosted on Anthropic". Almost every advanced feature (Files, Web Search, Web Fetch, Code Execution, Tool Calling, Agent Skills, MCP, Structured Outputs) now explicitly requires "Hosted on Anthropic" deployment, implying these features are *not* available or fully supported on the "Hosted on Azure" tier. * **Theme 3: Model Availability Shifts.** Sonnet family removed from "Hosted on Azure" list in the comparison table. * **Theme 4: Context Window Expansion.** Foundry now supports 1M tokens for more models (Opus/Sonnet). 4. **Draft the Analysis (Internal Monologue/Translation to Chinese):** * *Summary:* The updates primarily clarify Microsoft Foundry deployment requirements (pushing towards "Hosted on Anthropic") and deprecate Fast Mode for older Opus models. * *Themes:* * Microsoft Foundry Feature Gating: Most advanced tools/features now require "Hosted on Anthropic" deployment, effectively restricting them from "Hosted on Azure". * Fast Mode Sunset: Opus 4.6 support dropped; 4.7 deprecated. * Context Window Increase: 1M tokens support expanded on Foundry. * GA Status: Structured Outputs on Foundry is now Generally Available (GA). * *Impact Level:* Medium. It's not breaking existing code immediately, but it restricts feature availability based on hosting configurations and forces migration off deprecated features. It's significant for Foundry/Azure users. * *Action Items:* * Check Foundry hosting config if using advanced tools. * Stop using Fast Mode on Opus 4.6/4.7. * Update billing expectations (no more fast mode premium or standard rates). 5. **Refine the Output (Chinese):** * *Overall Summary:* 此次更新主要针对 Microsoft Foundry 的部署模型进行了重大功能限制澄清，明确区分了 "Hosted on Azure" 和 "Hosted on Anthropic" 的能力差异，并废弃了旧版 Opus 模型的 Fast Mode 功能。 * *Key Themes:* * **Microsoft Foundry 功能门槛**：文件 API、代码执行、Web 搜索/抓取、MCP 连接器、Agent Skills 及结构化输出等高级功能，现在明确要求使用 "Hosted on Anthropic" 部署方式。 * **Fast Mode 废弃**：停止支持 Claude Opus 4.6 的 Fast Mode，并宣布 Opus 4.7 的 Fast Mode 将于 2026 年 7 月 24 日废弃。 * **上下文窗口扩展**：Microsoft Foundry 现在支持 Claude Opus 4.8/4.7/4.6 和 Sonnet 4.6 的 1M token 上下文窗口。 * **可用性更新**：Sonnet 系列模型似乎已从 "Hosted on Azure" 的可用列表中移除；结构化输出在 Foundry 上转为正式发布 (GA)。 * *Impact Level:* Medium。主要影响 Microsoft Foundry 和 Azure 用户。虽然代码层面不一定立即报错，但功能可用性受到严格限制，且 Fast Mode 用户需迁移。 * *Action Items:* * 如果你是 Microsoft Foundry 用户，且需要使用文件处理、代码执行、MCP 或 Web 工具，请务必检查并切换至 "Hosted on Anthropic" 部署模式。 * 立即停止在 `claude-opus-4-6` 上使用 `speed: "fast"` 参数（将被忽略并以标准速度/价格运行）。 * 为 Claude Opus 4.7 的 Fast Mode 废弃（2026-07-24）做准备，计划迁移至标准模式或更新后的模型。 * 更新计费逻辑，不再依赖 Fast Mode 的特定费率。 6. **Final Review:** Does this capture all the nuances? * The distinction between Hosted on Azure vs Anthropic is the most critical finding. * The deprecation of Fast Mode is the second most critical. * The context window update is a nice "plus" for Foundry users. The prompt asks for concise, insightful analysis. The summary looks good. *Self-Correction on Page 7:* The change "The latest models in the Opus and Haiku families" for Hosted on Azure vs previous "Opus, Sonnet, and Haiku" implies Sonnet is not available on the "Hosted on Azure" tier anymore or only Opus/Haiku are considered "latest" for that specific hosting path. This reinforces the "Feature Gating" theme. 7. **Final Output Generation** (matching the requested format).

1. **Understand the Goal**: The user wants a concise, insightful analysis of a batch of documentation changes (10 pages) related to a developer tool (Claude Code). I need to focus on what changed and why it matters to developers. The response must be in Chinese and Markdown format. 2. **Analyze the Input Data**: * **Total pages**: 10. * **Page 1 (cli-reference.md)**: Added `claude gateway` command to start a self-hosted server for administrators (SSO/policy on Bedrock/Vertex/Foundry). * **Page 2 (costs.md)**: Updated info on Bedrock/Vertex/Foundry. Previously said it doesn't send metrics. Now says "Claude apps gateway" provides per-user attribution, OTLP metrics, and spend limits. * **Page 3 (data-usage.md)**: Updated feedback upload behavior. On Bedrock/Vertex/Foundry + "Claude apps gateway", feedback writes to a local archive (`~/.claude/feedback-bundles/`) instead of uploading. Also mentions gateway disables specific telemetry. * **Page 4 (iam.md)**: Added "Cloud gateway" auth option (SSO). Added "Claude apps gateway" to the list of authentication types. Added a note that gateway sessions outrank Bedrock/Vertex/Foundry env vars. * **Page 5 (llm-gateway.md)**: Renamed section from "LLM gateways" to "Other LLM gateways". Differentiates between the new "Claude apps gateway" and third-party gateways. * **Page 6 (monitoring-usage.md)**: Updated OTLP metrics to include gateway-specific attributes (IdP subject, email, groups). Mentions `identity.source: gateway-oidc`. * **Page 7 (network-config.md)**: Minor updates. Added "Claude apps gateway" to the list of where traffic goes (replacing api.anthropic.com). Added a note about Node version requirements for OS cert store reading. * **Page 8 (quickstart.md)**: Added "Claude apps gateway" to login options. * **Page 9 (settings.md)**: Updated "Server-managed settings" to include delivery via the gateway. Added `forceLoginMethod` support for "gateway" and new `forceLoginGatewayUrl` setting. * **Page 10 (third-party-integrations.md)**: Added "Claude apps gateway" to the list of authentication options. 3. **Synthesize Findings**: * **Core Change**: Introduction of a new self-hosted feature called "Claude apps gateway". * **Functionality**: It acts as a gateway for administrators to deploy SSO (Single Sign-On) and policies in front of Claude Code, specifically for cloud providers like Bedrock, Vertex AI, or Foundry. * **CLI Impact**: New command `claude gateway`. * **Auth Impact**: New login method ("Cloud gateway") that can override other provider env vars. * **Telemetry/Metrics**: It provides better usage attribution (per-user), OTLP metrics, and spend limits. It also changes how data is handled (local archives for feedback). * **Admin/Settings**: New settings for managed settings delivery and forcing the gateway login method. 4. **Draft the Analysis (Internal Monologue/Translation)**: * *Summary*: The docs update focuses heavily on the introduction of the "Claude apps gateway," a self-hosted solution for enterprise SSO and policy management. * *Key Themes*: * New "Claude apps gateway" feature introduction. * SSO integration for cloud providers (Bedrock, Vertex, Foundry). * Enhanced telemetry and cost tracking (OTLP, per-user limits). * Authentication hierarchy changes (Gateway > Env vars). * *Impact Level*: High for enterprise administrators. Medium for general developers (if their org uses it). It's a new feature, so it's not "breaking" in the sense of destroying existing workflows, but it introduces a major new architectural component. * *Action Items*: * Check if your org is deploying this. * Use `claude gateway --config ...` if you are an admin. * Update auth flows if your org mandates gateway login. * Note the change in telemetry/metrics attribution if using the gateway. 5. **Refine the Output (Chinese)**: * *Overall Summary*: 本次文档更新主要引入了名为“Claude apps gateway”的新功能，这是一个面向企业管理的自托管网关，支持在 Claude Code 与云服务提供商之间实施 SSO 和策略控制。 * *Key Themes*: * **引入 Claude apps gateway**：新增 `claude gateway` 命令和配套文档，允许管理员部署自托管网关。 * **企业级认证集成**：网关支持通过企业 IdP 进行 SSO 登录，且优先级高于现有的 Bedrock/Vertex/Foundry 环境变量。 * **增强的可观测性**：网关提供细粒度的使用归因（OTLP 指标、Token 计数、用户群组）以及每用户消费限制。 * **数据隐私与本地化**：在网关会话下，反馈数据默认存档至本地而非上传，且特定的遥测数据会被禁用。 * **配置管理**：支持通过网关远程下发托管设置，并新增 `forceLoginGatewayUrl` 等配置项。 * *Impact Level*: **High (for Enterprise Admins)** / **Medium (for Developers)** * *Reasoning*: 这是一个重大的架构更新，引入了新的认证路径和管理能力。对于使用 Bedrock/Vertex/Foundry 的企业用户，这将显著改变认证、流量路由和成本追踪的方式。 * *Action Items*: * **管理员**: 若需部署 SSO 或策略控制，请查阅 `claude-apps-gateway` 相关文档并使用 `claude gateway --config gateway.yaml` 启动服务。 * **开发者**: 若组织启用了网关，请使用 `/login` 中的 **Cloud gateway** 选项通过企业 SSO 登录；注意网关会话下的流量不再直接发送至 `api.anthropic.com`。 * **监控**: 利用网关提供的自动 OTLP 指标（如 `user.groups`, `user.email`）替代手动配置的 `OTEL_RESOURCE_ATTRIBUTES` 以进行用户归因。 6. **Final Review**: Does this meet all constraints? Yes. Concise, insightful, markdown, Chinese, developer-focused. 7. **Construct Final Output**: (Proceed to generate markdown).

1. **Understand the Goal**: The user wants a world-class programmer's analysis of a batch of documentation changes. The response must be in Chinese, in markdown format, concise, and insightful. It needs to cover four specific areas: Overall Summary, Key Themes, Impact Level, and Action Items. 2. **Analyze the Source Material (Diffs)**: * **Page 1: `about-claude/models/migration-guide.md`** * *Change*: Removed a note stating Microsoft Foundry has a 200k context window (initial launch limitation). Updated text to say 1M context window applies to Foundry as well (API, Bedrock, GC, Foundry). * *Significance*: Microsoft Foundry now supports the full 1M context window, removing the previous limitation. * **Page 2: `about-claude/pricing.md`** * *Change*: Added a new section "Claude in Microsoft Foundry pricing". Details CCU (Claude Consumption Units) billing ($0.01/CCU). Added "Inference geography" for Foundry (US Data Zone Standard = `inference_geo: "us"`, 1.1x multiplier). Updated text regarding `inference_geo` parameter scope. * *Significance*: Foundry billing is now clearly defined (CCUs), and data residency pricing (1.1x multiplier) is confirmed for US Data Zone deployments. * **Page 3 & 4: `build-with-claude/overview.md` & `build-with-claude/text-generation.md`** (Identical diffs) * *Change*: Updated "Availability" column in feature tables. Changed `<PlatformAvailability ... azureAiBeta />` to `azureAi`. * *Significance*: Microsoft Foundry is now generally available (GA) / stable, out of beta, for features like Context windows, Adaptive thinking, Citations, Effort, and Extended thinking. * **Page 5, 6, 9, 10: `vision.md`, `pdf-support.md`, `files.md`, `code-execution-tool.md`** * *Change*: Changed `import fs from "fs"` or `import fs from "fs/promises"` to `import fs from "node:fs"` or `import fs from "node:fs/promises"`. * *Significance*: Node.js best practice update. Using explicit protocol (`node:`) for built-in modules avoids conflicts with local modules named `fs`. * **Page 7: `build-with-claude/prompt-caching.md`** * *Change*: Removed "(beta)" mentions for Microsoft Foundry in various notes (automatic caching, cache isolation, etc.). * *Significance*: Prompt caching features for Foundry are now GA (matching the "azureAi" change in other docs). * **Page 8: `build-with-claude/claude-in-microsoft-foundry.md`** * *Change*: Massive rewrite. * "Preview" removed/replaced with general availability context. * New section "Hosting options": Distinguishes between "Hosted on Azure" vs "Hosted on Anthropic". * Azure hosting: Inference runs on Azure infra, prompts remain in Azure (metadata/safety flags egress). * Deployment types: Global Standard, US Data Zone Standard. * Updated provisioning steps (new portal flow). * *Significance*: Foundry integration has matured significantly. New hosting option (On Azure) implies data residency compliance benefits. New provisioning flow. 3. **Synthesize the Analysis**: * **Overall Summary**: The primary focus is the graduation of **Microsoft Foundry** from Beta to General Availability (GA), specifically introducing new Azure-hosted deployment options and full 1M token context window support. Additionally, minor code updates were made to use Node.js explicit imports. * **Key Themes**: * **Microsoft Foundry GA**: Removal of "beta" tags across multiple features (caching, context windows, etc.). * **New Hosting Option**: Introduction of "Hosted on Azure" vs "Hosted on Anthropic" within Foundry. * **Data Residency & Pricing**: Clarification of US Data Zone pricing (1.1x multiplier) and CCU billing. * **Context Window Equality**: Foundry now supports the full 1M context window (no longer limited to 200k). * **Node.js Best Practices**: Updating import statements to use `node:` protocol. * **Impact Level**: **High** (or Medium-High). * *Reasoning*: The Foundry changes represent a major platform update (GA status, new hosting model, billing changes). If a developer is using Foundry, they need to know about the new hosting options and the removal of context limits. If they are just using the raw API, the impact is lower, but the Foundry section is substantial. * **Action Items for Developers**: * Review Foundry deployment options: Decide between "Hosted on Azure" and "Hosted on Anthropic". * Update code if using Foundry: Remove any context window limitations (beta headers). * Check billing: Understand the CCU (Claude Consumption Units) billing model. * Consider Data Residency: If using Foundry for US data, note the 1.1x pricing multiplier for US Data Zone. * Code Hygiene: Update Node.js imports to `node:fs`/`node:fs/promises`. 4. **Draft the Response (in Chinese)**: * *Drafting Summary*: 此次文档更新主要集中在 **Microsoft Foundry 平台的重大升级**（从 Beta 转为 GA），包括新的 Azure 托管选项、1M 上下文窗口支持及 CCU 计费模式，同时更新了 Node.js 代码示例中的最佳实践。 * *Drafting Themes*: * Microsoft Foundry 正式发布 (GA)：移除了 Beta 标签，功能完全可用。 * 新增托管选项：区分“Azure 托管”和“Anthropic 托管”。 * 上下文窗口升级：Foundry 现支持完整 1M token。 * 计费与合规：明确了 CCU 计费单位及美国数据区域的定价倍率。 * Node.js 代码优化：统一使用 `node:` 前缀导入核心模块。 * *Drafting Impact*: **High**。对于使用 Foundry 的开发者，这是功能上的重大扩展和配置流程的变化。 * *Drafting Actions*: 1. 评估 Foundry 托管模式（Azure vs Anthropic）。 2. 移除 Foundry 相关的 200k 上下文限制代码。 3. 熟悉 CCU 计费模式。 4. 更新 Node.js 导入语句（`fs` -> `node:fs`）。 5. **Refine and Polish**: Ensure the tone is professional and "world-class". Make sure the markdown formatting is correct. * *Self-Correction during drafting*: The Node.js change is minor but good to mention as a best practice update. The Foundry changes are the "meat" of this update. The migration guide change is important because it affects the actual code if someone was working around the 200k limit. 6. **Final Output Generation** (matches the requested structure).

Let me analyze the documentation changes across all four pages. 1. **Context Windows** (+60/-67): Major restructuring. Adds a summary line, reorganizes content about how context windows work, adds a new "Context window sizes by model" section mentioning new models (Claude Opus 4.8, 4.7, 4.6, Sonnet 4.6, Fable 5, Mythos 5), clarifies that everything counts toward context (system prompt, messages, tool definitions, tool results, images, documents, extended thinking output), and changes the guidance about how previous thinking blocks are handled (now model-dependent rather than always stripped). 2. **Web Fetch Tool** (+1/-1): Small but important clarification about mixing server tools and client tools in parallel tool calls - when Claude calls web fetch AND a client tool simultaneously, the API returns with `stop_reason: "tool_use"` before the fetch completes, then runs the fetch when you send back the client tool_result. 3. **Tool Search Tool** (+109/-95): Significant rewrite. Adds a "Model compatibility" section with a table showing which models support tool search (Fable 5, Mythos 5, Opus 4.8/4.7/4.6, Sonnet 4.6, Opus 4.5, Sonnet 4.5, Haiku 4.5). Clarifies that at least one tool must stay non-deferred. Adds more detail about the workflow. Code examples updated. 4. **Fine-Grained Tool Streaming** (+136/-75): Major rewrite. The `eager_input_streaming` per-tool field replaces the legacy `fine-grained-tool-streaming-2025-05-14` beta header. All code examples are completely rewritten to actually show streaming tool input (printing partial JSON fragments) instead of just showing token usage. Added clearer warning about invalid/partial JSON. Added note about `max_tokens` stop reason cutting parameters off midway. Key themes: - New model introductions (Opus 4.8, 4.7, 4.6, Sonnet 4.6, Fable 5, Mythos 5) - Better documentation of edge cases and behavior nuances - Migration from beta headers to stable per-tool configuration - Improved code examples that actually demonstrate the feature - Clarification on server tool + client tool interaction in parallel calls Impact: Medium-High. The fine-grained tool streaming migration from beta header to per-tool field is a significant API change. The new models and their context window sizes matter. The parallel tool call behavior for web fetch is important to know about. Action items: 1. Migrate from `fine-grained-tool-streaming-2025-05-14` beta header to per-tool `eager_input_streaming` field 2. Review new model compatibility for tool search - older models don't support it 3. Handle the case where web fetch + client tool are called in parallel (stop_reason: tool_use before fetch completes) 4. Update context window expectations for new models (1M tokens for several new models, 200k on Microsoft Foundry) 5. Be aware that thinking block handling in context varies by model (no longer universally stripped)