Changes on 2026-06-10

1. **Understand the Goal:** The user wants a concise, insightful analysis of specific documentation changes (single page: `build-with-claude/vision.md`) in Markdown format, focusing on what matters to developers, in Chinese. 2. **Analyze the Input:** * **Source:** `build-with-claude/vision.md` * **Changes:** +96 / -18 (Net addition). * **Diff Content:** * *Token Calculation Logic Change:* The formula for calculating tokens changed from `width * height / 750` to a patch-based approach: `⌈width / 28⌉ × ⌈height / 28⌉` (Visual Tokens). * *Clarification on Resizing/Padding:* Explicit mention that images are viewed in 28x28 patches. Clarification on how images are resized/padded (always padded to multiple of 28). * *Updated Token/Cost Examples:* The table of token costs and examples was updated to reflect the new calculation method. The numbers are generally slightly different (e.g., 200x200 goes from ~54 to 64 tokens). * *High-Resolution Support Section:* Updated table examples to match the new math. Added specific note that 4K images are downscaled. * *New Section:* Added "Working with coordinates and bounding boxes". * *Key Advice:* Claude works best with *absolute pixel coordinates* relative to the resized image, not normalized coordinates (0-1000). * *Coordinate System:* Standard top-left origin (0,0). 3. **Synthesize Key Themes:** * **Technical Accuracy/Precision:** The move from a rough heuristic (`/ 750`) to a precise algorithm (`28x28 patches`) indicates a shift towards more predictable billing and behavior. * **Coordinate Handling:** A strong emphasis on how to ask for bounding boxes (absolute pixels vs. normalized). This is a "Best Practice" addition. * **Cost Predictability:** Updated tables help developers better estimate costs based on the new logic. 4. **Determine Impact Level:** * *Breaking?* No. The API behavior likely didn't change fundamentally, but the documentation is now more accurate. However, if developers were relying on the old `/750` math for cost estimation, their estimates were off. * *Impact:* Medium. The cost calculation logic change is significant for billing estimation. The new guidance on coordinates is significant for vision applications (OCR, object detection). 5. **Draft Action Items:** * Update cost estimation logic in code/tools to use the new `ceil(w/28) * ceil(h/28)` formula. * Adjust prompts to request absolute pixel coordinates instead of normalized ones for better accuracy. * Update client-side coordinate mapping logic to account for the specific resizing/padding rules mentioned (or use pre-resized images to avoid this complexity). 6. **Translate to Chinese and Format (Markdown):** * *Overall Summary:* 文档更新了图像 Token 计算逻辑（从除法改为 28x28 分块），并新增了关于坐标和边界框使用的最佳实践指南。 * *Key Themes:* * 精确的 Token 计算公式。 * 坐标系的明确化（绝对像素 vs 归一化）。 * 图像处理流程的细节（裁剪、填充）。 * *Impact:* Medium. 影响成本估算和视觉任务的提示词工程。 * *Action Items:* * 更新 Token 估算器。 * 修改提示词以要求绝对像素坐标。 * 注意图像预处理建议。 7. **Refine the Output (Self-Correction during drafting):** * Make sure to emphasize the "Visual Token" concept. * Highlight the specific formula change `width * height / 750` -> `ceil(w/28) * ceil(h/28)`. * The coordinate section is a *new* section, so that's a major addition for developers building vision features. 8. **Final Polish (Chinese):** * Summary: 更新了图像 Token 计算模型（从经验公式改为精确的 28x28 像素分块），并新增了关于使用坐标和边界框的最佳实践章节。 * Themes: Token 计算精确化、坐标系统最佳实践、图像处理细节明确化。 * Impact: Medium. 成本估算更准确，且对视觉应用的提示词工程有重要指导意义。 * Actions: 更新成本计算代码、优化 Prompt 获取绝对像素坐标、关注图像预处理建议。 9. **Construct Final Markdown:** (Matches the requested structure). * Check constraints: "concise", "insightful", "markdown format", "Chinese". * *Drafting the response...* * Summary: 文档修正了视觉 Token 的计算逻辑，将其精确化为基于 28x28 像素块的模型，并新增了关于处理坐标和边界框的详细指南。 * Themes: 精确的计费模型、视觉交互最佳实践、图像处理流程透明化。 * Impact: Medium. 对开发者的成本预算和视觉应用（如 OCR/UI 分析）的实现有直接影响。 * Actions: 更新内部 Token 估算工具，调整 Prompt 以要求绝对像素坐标，根据新规则验证图像预处理逻辑。 * *Wait, the specific formula change is the most important technical detail.* * *The coordinate section is new content, not just an edit.* Let's finalize the Chinese text. 1. Overall Summary: 文档更新了图像 Token 的计算算法（从模糊的经验值改为基于 28x28 分块的精确模型），并新增了关于坐标和边界框处理的最佳实践章节。 2. Key Themes: - **算法精确化**: Token 计算从 `width * height / 750` 改为 `⌈width / 28⌉ × ⌈height / 28⌉`，使计费更透明。 - **坐标系指南**: 新增章节明确建议开发者要求 Claude 返回“绝对像素坐标”而非归一化坐标，以提高准确性。 - **预处理细节**: 阐明了图像填充至 28 的倍数及缩放规则，帮助开发者正确处理坐标映射。 3. Impact: **Medium**. 虽然底层 API 行为可能未变，但文档的修正改变了开发者估算成本的方式，且新增的坐标指南对构建视觉应用（如 UI 分析、OCR）至关重要。 4. Action items: - 更新任何基于旧公式 (`/750`) 的本地 Token 成本估算器。 - 修改视觉任务的 Prompt，明确要求返回图像的绝对像素坐标。 - 如果需要坐标映射，确保客户端代码能正确处理 Claude 对图像的缩放和填充（或预缩放图片以避免复杂性）。 Looks good. Matches the prompt's request perfectly.

1. **Analyze the Request:** * **Role:** World-class programmer/developer tools expert. * **Task:** Analyze documentation changes in a single batch. * **Input:** Two diff files (`model-config.md` and `plugins.md`). * **Output Format:** Markdown, concise, insightful. * **Language:** Chinese (requested: "请用中文回复"). * **Structure:** 1. Overall Summary (1-2 sentences). 2. Key Themes (bullets). 3. Impact Level (Low/Medium/High/Breaking) with reasoning. 4. Action Items for developers (bullets, max 5). 2. **Analyze the Changes:** * **Page 1: `model-config.md`** * *Change:* Updated the URL for the "Claude Fable 5" documentation. * *Old URL:* `.../introducing-claude-fable-5` * *New URL:* `.../introducing-claude-fable-5-and-claude-mythos-5` * *Implication:* There is likely a new model called "Mythos 5" or the documentation page now covers both Fable 5 and Mythos 5. The text describing Fable 5 itself didn't change, just the link. This implies a product update (new model or expanded model family documentation). * **Page 2: `plugins.md`** * *Change:* Updated the description of the `claude-plugins-official` marketplace. * *Old Text:* "Available automatically in every Claude Code installation." * *New Text:* "Registered automatically the first time you start Claude Code interactively. A non-interactive script that runs before that first launch must add it explicitly with `claude plugin marketplace add anthropics/claude-plugins-official`." * *Implication:* The behavior of the official plugin marketplace registration has changed. It's no longer just "available automatically" everywhere without context. Now, specifically for *non-interactive scripts* (like CI/CD or setup scripts), explicit registration is required if the script runs *before* the first interactive start. Interactive users are unaffected (it happens on first launch). 3. **Synthesize and Translate to Chinese:** * **Overall Summary:** * *English:* The updates reflect a new model (Mythos 5) and a change in how the official plugin marketplace is registered for non-interactive environments. * *Chinese:* 本次更新主要反映了新模型（Mythos 5）的发布，以及官方插件市场在非交互式环境下的注册行为变更。 * **Key Themes:** * *Theme 1 (Model):* Reference to new model "Claude Mythos 5". * *Theme 2 (Plugin/Mechanism):* Explicit registration required for plugins in automation/CI contexts. * *Chinese:* * **模型生态扩展**：文档链接更新以包含新的 "Claude Mythos 5" 模型。 * **非交互式环境处理**：明确了官方插件市场的加载机制，特别是针对脚本和自动化场景的显式注册要求。 * **Impact Level:** * *Assessment:* Low to Medium. The model link change is cosmetic. The plugin change affects CI/CD or setup scripts specifically. * *Reasoning:* Developers using interactive workflows won't notice the plugin change. Those automating setup might face a "marketplace not found" error if they try to install plugins before the first interactive run. * *Chinese:* **中等 (Medium)** * *Reasoning:* 虽然只是链接和描述更新，但对使用自动化脚本或 CI/CD 流程配置 Claude Code 的开发者来说，插件的默认行为发生了重大变化。如果不做调整，脚本可能会失败。 * **Action Items:** * *Action 1:* Review automation scripts. If they install plugins before running the app interactively, add the command. * *Action 2:* Check out Mythos 5 docs (optional/nice to know). * *Chinese:* * **检查自动化脚本**：如果您的设置脚本或 CI/CD 流程在第一次交互式启动之前尝试安装插件，请务必添加 `claude plugin marketplace add anthropics/claude-plugins-official`。 * **查阅新模型文档**：点击更新后的链接了解 Fable 5 和 Mythos 5 的详细区别。 4. **Refine and Format (Self-Correction):** * Make sure the tone is professional and concise. * Ensure the plugin command is accurate based on the diff. * The prompt asks for "What matters to developers". The plugin change is the critical operational detail here. 5. **Final Output Generation (matching the requested structure):** * *Summary:* Updates reveal the existence of "Claude Mythos 5" and alter the official plugin marketplace registration logic for non-interactive scripts. * *Themes:* Model expansion, Automation/Setup behavior. * *Impact:* Medium (Breaking for some automation). * *Actions:* Update scripts, Read docs. *Drafting the Chinese response:* 1. **Overall Summary**: 本次更新引入了新模型 "Claude Mythos 5" 的参考，并修改了官方插件市场在非交互式脚本中的初始化行为。 2. **Key Themes**: * **模型家族更新**: 文档链接变更暗示了除 Fable 5 外，新增了 Mythos 5 模型相关文档。 * **自动化/CI 集成变更**: 官方插件市场不再是完全“自动可用”，在脚本模式下需要显式注册。 3. **Impact Level**: **中等** * 理由：对于交互式用户无影响。但对于编写安装脚本或 CI 配置的开发者，若未在第一次交互启动前显式添加官方市场，后续插件安装命令将失败。 4. **Action Items**: * **修复安装脚本**：如果您的脚本在首次交互运行前需要使用插件，请显式执行 `claude plugin marketplace add anthropics/claude-plugins-official`。 * **关注模型变化**：查阅更新后的模型文档链接，评估 Mythos 5 是否适用于您的场景。 6. **Final Polish:** Looks good. Ready to output.

1. **Analyze the Request:** * **Role:** World-class programmer analyzing documentation changes. * **Format:** Markdown, concise, insightful. * **Language:** Chinese. * **Input:** A single diff of `CHANGELOG.md.md` showing version `2.1.172`. * **Output Structure:** Overall Summary, Key Themes, Impact Level, Action Items. 2. **Analyze the Diff (`CHANGELOG.md.md`):** * *Version:* 2.1.172. * *Feature Additions:* * Recursive sub-agents (up to 5 levels). * Bedrock config reading (`~/.aws` precedence). * Plugin marketplace search bar. * OTEL metric enhancement (`model` attribute). * *Bug Fixes (Significant):* * Sessions stuck at 1M context without credits (auto-compaction). * Image processing errors (repeating). * Agent view UI issues (busy spinner, stuck status). * Security/Isolation: Background agents reading wrong project settings (`.mcp.json`, trust). * `availableModels` restrictions applied incorrectly (allowlists, overrides). * Model ID issues (doubled suffix, picker bugs). * Permission/Domain rules (wildcards not matching subdomains, mid-pattern wildcards rejected). * Prompt history (subagent tab showing main agent history). * Memory recall (mounted stores in remote sessions). * Workflow validation (false positives on `Date.now()`/`Math.random()`). * *Performance/UX Improvements:* * Long conversation performance (removed redundant normalization). * Reduced idle CPU usage (less re-rendering). * Chrome tool loading (batched calls). * Non-interactive usage policy message. * `/code-review` UI (ultra option visibility). * Remote Control footer. * `/loop` promotion logic. * VSCode PowerShell rendering. 3. **Synthesize Overall Summary:** * This update (v2.1.172) is a heavy maintenance and feature expansion release. * It significantly upgrades agent capabilities (recursion), stabilizes complex scenarios (1M context, background agents), and polishes the UX/UI (model picker, performance). * *Draft (Internal):* Version 2.1.172 introduces recursive sub-agents and fixes numerous bugs related to 1M context handling, model selection, and background agent isolation. It also improves performance and UI responsiveness across various interfaces including VSCode and the CLI. 4. **Identify Key Themes:** * *Agent Architecture:* Recursive sub-agents, isolation fixes for background agents. * *Model & Configuration:* Deep fixes for model pickers, AWS Bedrock config, 1M context handling, and `availableModels` logic. * *Bug Fixes & Stability:* Critical fixes for stuck sessions, image processing, and permission validation (wildcards). * *Performance & UX:* Optimizations for CPU usage and rendering, plus better tool loading in Chrome. 5. **Determine Impact Level:** * *Reasoning:* The fixes regarding "background agents reading another directory's project settings" and "sessions stuck at 1M context" are high-stability issues. Recursive sub-agents unlock new workflows. * *Level:* High. While not "Breaking", the resolution of session-stuck issues and isolation bugs is critical for reliability in multi-agent environments. 6. **Formulate Action Items for Developers:** * Check sub-agent nesting limits (now 5 levels). * Review AWS config if using Bedrock (automatic detection might change behavior). * Re-evaluate wildcard permission rules (they actually work now/mid-pattern support). * Update workflows previously blocked by false positive validation (`Date.now()`). * Test background agent isolation to ensure security posture is updated based on fixed settings reading. 7. **Translate to Chinese & Format:** * *Overall Summary:* 版本 2.1.172 是一次重要的功能增强与维护更新，引入了递归子代理功能，并修复了多项影响稳定性的关键问题（如 1M 上下文会话卡死、后台代理隔离权限）。同时，对 AWS Bedrock 配置读取、模型选择器逻辑及 UI 性能进行了显著优化。 * *Key Themes:* * **代理架构升级**：支持子代理递归生成（最多5层），并修复了后台代理可能读取错误目录项目设置的安全隔离问题。 * **模型与配置修复**：深入修复了 1M 上下文模型 ID 格式错误、`availableModels` 限制未生效、AWS Bedrock 区域读取逻辑以及模型选择器显示错误。 * **稳定性与性能**：解决了会话卡死、重复报错、权限通配符不匹配等 Bug，并优化了长对话性能和空闲 CPU 占用。 * **开发体验改进**：增强了 Chrome 工具加载、VSCode PowerShell 显示及 `/plugin` 市场的交互体验。 * *Impact Level:* **High** (高)。 * *Reasoning:* 修复了可能导致会话永久卡死的严重 Bug，以及涉及多租户/多项目隔离（后台代理读取错误配置）的安全稳定性问题。此外，递归代理是重大功能解锁。 * *Action Items:* * **测试递归代理**：现在可以设计更深层次（最多5层）的代理调用逻辑。 * **验证权限规则**：检查此前可能因通配符（如 `*.domain.com` 或 `secrets-*`）失效而放宽的权限规则，确认新修复后的实际约束。 * **审核 AWS 配置**：如果使用 Bedrock，确认 `~/.aws` 配置中的区域设置是否按预期生效。 * **检查工作流**：重新提交此前因包含 `Date.now()` 等字符串而被错误拒绝的工作流脚本。 8. **Final Review:** Ensure the tone is professional, concise, and "developer-focused". The markdown format looks correct. The Chinese translation is natural.

用户希望我分析文档更改的批次，并提供中文的简明见解。让我分析这些更改： 1. **amazon-bedrock.md, google-vertex-ai.md, microsoft-foundry.md, third-party-integrations.md** - 这些页面都删除了相同的"ContactSalesCard"组件，这是一个销售联系卡片的React组件。这是文档清理工作。 2. **plugins.md** - 插件提交表单的URL更新了，从 `claude.ai/settings/plugins/submit` 变为 `claude.ai/admin-settings/directory/submissions/plugins/new`。还增加了一条说明，解释了权限要求（团队或企业组织需要目录管理权限）。 3. **get-started.md** - Java SDK版本从2.39.0升级到2.40.0。 4. **migration-guide.md** - 迁移指南的更新： - Claude Mythos Preview退役日期更新为2026年6月30日 - 关于从Claude Mythos Preview迁移到Claude Mythos 5的改进文档 - 添加了"Before you migrate"部分，说明Claude Fable 5的定价和数据保留要求 - 结构和可读性改进 5. **model-deprecations.md** - Claude Mythos Preview退役日期更新为2026年6月30日。 6. **service-tiers.md** - 更新了支持模型，明确指出优先级层不支持Claude Mythos 5。 7. **errors.md** - 新增错误类型："Thinking blocks cannot be modified"，解释了当修改思考块时的错误。 8. **overview.md 和 text-generation.md** - 新增功能行： 服务器端回退机制允许在模型拒绝时自动切换备选模型，这提高了系统的容错性和可靠性。回退信用机制则帮助避免重复支付prompt-cache成本，为开发者提供了更经济高效的解决方案。这些功能目前处于测试阶段，主要适用于Claude API和Claude Platform on AWS。 9. **extended-thinking.md** - 新增关于修改思考块错误的说明。 10. **structured-outputs.md** - 结构化输出现在支持Claude Fable 5和Claude Mythos 5。 更新文档以反映对新模型的支持，扩展了结构化输出的可用性。这意味着开发者可以在最新的Claude模型上使用结构化输出功能。 11. 多个页面（Bedrock、Vertex AI、Microsoft Foundry）的Java SDK版本更新。工具文档页面更新，包括支持新模型和一些格式调整。这些变更表明持续改进SDK功能和模型兼容性。

1. **Analyze the Request:** * **Role:** World-class programmer/documentation expert. * **Task:** Analyze documentation changes as a single batch. * **Input:** A diff for `desktop.md` (+1/-1). * **Output Format:** Markdown with specific sections (Overall Summary, Key Themes, Impact Level, Action Items). * **Language:** Chinese (requested: "请用中文回复"). 2. **Analyze the Input Data (Diff):** * **File:** `desktop.md`. * **Context:** A tip block describing permission modes in Plan mode, specifically regarding Cloud sessions. * **Change:** * *Old:* "Cloud sessions support Auto accept edits and Plan mode. Ask permissions is not available because cloud sessions auto-accept file edits by default, and Bypass permissions is not available because the cloud environment is already sandboxed." * *New:* "Cloud sessions support Accept edits, Plan mode, and Auto mode. Accept edits corresponds to `default` mode: cloud sessions pre-approve file edits, so the selector shows Accept edits instead of Ask permissions. Bypass permissions is not available because the cloud environment is already sandboxed." 3. **Interpret the Change:** * **Terminology Update:** It seems the terminology has shifted slightly to align with UI labels or internal modes. * "Auto accept edits" (Old) -> "Auto mode" (New). * Introduction of "Accept edits" corresponding to `default` mode. * **Clarification:** The new text clarifies *why* "Ask permissions" isn't there (because it's replaced by "Accept edits" in the selector due to pre-approval). It explains that "Accept edits" is the `default` mode for cloud sessions. * **Core Meaning:** Cloud sessions don't ask for permission because they pre-approve. The UI shows "Accept edits". They also support Plan mode and Auto mode. 4. **Drafting the Content (Iterative refinement in Chinese):** * **1. Overall Summary (总体概要):** * *Draft:* Updated the description of Cloud session permission modes, clarifying that "Accept edits" is the default state corresponding to auto-approval, and distinguishing it from "Auto mode". * *Refinement (Chinese):* 更新了 Cloud 会话权限模式的描述，阐明了“Accept edits”（接受编辑）作为对应 `default` 模式的默认状态，并区分了其与“Auto mode”（自动模式）的概念。 * **2. Key Themes (核心主题):** * *Theme 1:* Terminology clarification (Mode names). * *Theme 2:* Explanation of UI behavior (Why "Ask permissions" is missing). * *Draft:* * Clarified the relationship between UI labels ("Accept edits") and internal modes (`default`). * Updated the list of supported modes for Cloud sessions (Accept edits, Plan mode, Auto mode). * Explained the default pre-approval behavior in cloud environments. * *Refinement (Chinese):* * **术语规范化与模式映射**：明确了 UI 中的 "Accept edits" 对应内部的 `default` 模式，解释了 Cloud 会话为何不显示 "Ask permissions"。 * **模式列表更新**：将支持的模式更新为 "Accept edits"、"Plan mode" 和 "Auto mode"，并优化了对默认行为的解释。 * **3. Impact Level (影响程度):** * *Level:* Low. It's a documentation clarification, not a code API change. * *Reasoning:* No functional code changes. It just helps users understand the UI better. * *Refinement (Chinese):* **Low (低)**。仅涉及文档澄清和术语解释，未涉及 API 或功能行为的实质性变更。 * **4. Action Items for developers (开发者行动项):** * *Item 1:* If writing tutorials or guides, use the new terms. * *Item 2:* Understand that Cloud sessions pre-approve. * *Refinement (Chinese):* * 更新相关的用户指南或教程，使用 "Accept edits" 和 "Auto mode" 等新术语。 * 向用户解释 Cloud 会话的默认行为（自动预批准）时，引用文档中的解释逻辑。 5. **Final Review:** * Check against constraints: Concise? Yes. Insightful? Yes. Markdown format? Yes. Chinese language? Yes. * Did I catch the nuance? Yes, the shift from "Auto accept edits" to "Accept edits" + "Auto mode" and the clarification of `default` mode is key. 6. **Final Output Generation** (Proceed to generate output based on steps above).