Changes on 2026-05-13

1. **Analyze the Request:** * **Role:** World-class programmer, expert in dev tools and documentation. * **Task:** Analyze a single batch of documentation changes (specifically a `CHANGELOG.md` update). * **Format:** Markdown, concise, insightful. * **Language:** Chinese (as requested in the system prompt "请用中文回复"). * **Input:** A diff of `CHANGELOG.md.md` for version `2.1.141` (additions only). 2. **Analyze the Content (CHANGELOG.md diff for v2.1.141):** * **Total Changes:** 62 lines added. * **Nature:** It's a standard changelog for a CLI tool/IDE extension called "Claude Code" (inferred from context like `claude agents`, VSCode mentions, hooks, etc.). * **Key Additions:** * *New Features:* `terminalSequence` for hooks, env vars (`CLAUDE_CODE_PLUGIN_PREFER_HTTPS`, `ANTHROPIC_WORKSPACE_ID`), `--cwd` for agents, "Summarize up to here" in rewind menu, feedback includes recent sessions. * *UX Improvements:* Spinner feedback (amber color), plugin menu navigation, auto-mode permission dialog explanation, background agent state handling. * *Fixes:* A massive list of fixes (around 30+ items). Topics include: * Model selection (fallbacks for Haiku/Bedrock/Vertex). * CLI behavior (`daemon status`, `/doctor`, `/model`, `/tui`, `/mcp`). * Permission handling (modes, prompts, auto-dismissal). * UI/Terminal issues (markdown tables, statuslines, themes, spinner). * Platform specifics (Windows, VSCode, Linux SDK). * MCP (Model Context Protocol) servers (configs, auth, HTTP/SSE). * Voice mode. * Authentication (Remote Control, AWS credentials). 3. **Drafting the Analysis (Iterative refinement):** * *Overall Summary:* This is a feature-rich and bug-fix-heavy release (v2.1.141). It focuses on stabilizing the "agents" workflow, improving MCP (Model Context Protocol) integration, and fixing a wide array of CLI/UI edge cases, particularly around permissions and cross-platform support. * *Key Themes:* * **Agents & Background Tasks:** Enhancements to `claude agents` (dashboard, state management, `--cwd`) and background shells. * **MCP & Ecosystem:** Fixes for plugin MCP servers, environment variable expansions, and auth flows. New env var for HTTPS preference. * **UX & Polish:** Better spinner feedback, "Summarize" feature, clearer permission dialogs. * **Bug Bash:** Extensive fixes covering Windows, VSCode, AWS Bedrock, voice mode, and CLI input handling. * *Impact Level:* **High**. * *Reasoning:* While not a "breaking change" release, the sheer volume of fixes (over 30) targeting critical stability areas (permissions, agents, MCP, cross-platform auth) significantly improves the developer experience. The new agent features change workflows. * *Action Items:* * Review new environment variables (`CLAUDE_CODE_PLUGIN_PREFER_HTTPS`, `ANTHROPIC_WORKSPACE_ID`) if relevant to your CI/SSH setup. * Check the updated `claude agents --cwd` syntax if managing multiple project directories. * Be aware of the "Summarize up to here" rewind feature for better context management. * Update if you rely on Windows, VSCode voice features, or Bedrock/Vertex integration due to specific fixes. 4. **Translating to Chinese (as requested):** * *Summary:* 2.1.141 版本是一个侧重于稳定性与工作流增强的重要更新。它引入了 Agent 管理的新功能，修复了大量关于 MCP、权限控制及跨平台支持的缺陷。 * *Themes:* * Agent 与后台任务增强（`claude agents` 仪表盘、状态管理）。 * MCP 与生态系统（插件服务器、认证流程修复、新增 HTTPS 环境变量）。 * 用户体验（回退菜单的“总结”功能、权限对话框优化、加载状态反馈）。 * 大规模缺陷修复（涉及 Windows、VSCode、Bedrock、Voice Mode 等）。 * *Impact:* **High** (影响程度高)。大量的修复直接解决了权限、Agent 和跨平台认证中的稳定性痛点。 * *Action Items:* * 检查新增的环境变量，特别是涉及 GitHub HTTPS 克隆和 Workspace ID 的配置。 * 尝试新的 `claude agents --cwd` 命令以更好地管理多目录会话。 * 如果使用 Windows 或 VSCode 语音功能，建议升级以获得修复。 * 熟悉“回退菜单”中的“总结到此处”功能以优化 Token 使用。 5. **Final Review against constraints:** * Markdown format? Yes. * Concise? Yes. * Insightful? Yes. * Focus on what matters to devs? Yes. * Chinese? Yes. 6. **Final Polish (Self-Correction):** * Make sure the "Impact" explanation justifies "High" without being alarmist. It's high because of the stability improvements, not because of breakages. * Ensure the distinction between "Added" and "Fixed" is clear in the themes. *(Self-Correction during drafting Chinese output)*: "TerminalSequence" might be niche, but "Agents" and "MCP" are central themes for power users. I will emphasize those. 7. **Generate Output.**

1. **Analyze the Request:** * **Role:** World-class programmer analyzing documentation changes. * **Input:** 18 pages of diffs (Markdown). * **Output Format:** Concise, insightful analysis in Markdown, focusing on developer impact, in Chinese. 2. **Analyze the Diffs (Page by Page):** * *hooks-guide.md*: Clarifies that `disableAllHooks` in managed settings might need to be set separately if hooks are defined there. * *plugins.md*: Clarifies that `--plugin-dir` cannot override plugins that managed settings force-enable or force-disable. * *settings.md*: Refines how settings scopes interact. Clarifies that scalar values are overridden by higher priority, but arrays (like permissions) are merged/concatenated. * *about-claude/pricing.md*: Removes older models (Claude 3, Haiku 3, Sonnet 3.7) from main tables or marks them as "Deprecated" or "Retired, except on Bedrock/Vertex". Updates pricing tables to be cleaner. * *about-claude/model-deprecations.md*: Adds a note about lifecycle status/dates differing between Anthropic-operated platforms and Partner platforms (Bedrock, Vertex). * *api/rate-limits.md*: Updates model lists (removing Sonnet 3.7, etc.), updates "Retired" labels for Haiku 3.5 to be specific to Bedrock/Vertex. * *build-with-claude/extended-thinking.md*: Removes specific references to Claude Sonnet 3.7. Updates context window behavior explanation (now strict limit, but 4.5+ handles overflow differently). Generalizes thinking behavior descriptions. * *build-with-claude/prompt-caching.md*: Updates pricing tables, removes specific old model references or labels them correctly (e.g., Haiku 3.5 retired except Bedrock/Vertex). * *build-with-claude/batch-processing.md*: Updates pricing tables, similar to prompt caching. * *build-with-claude/context-windows.md*: Removes specific 3.7 references. Clarifies context window overflow behavior (4.5+ stops gracefully vs earlier models erroring). * *build-with-claude/search-results.md*: Updates model availability list. Notes Haiku 3.5 is retired except on Bedrock/Vertex. Removes Sonnet 3.7. * *build-with-claude/files.md*: Minor rewording about file support. * *build-with-claude/claude-on-amazon-bedrock.md*: Updates lifecycle dates and status in the table (e.g., Sonnet 3.7 "Retired April 28, 2026"). Adds a disclaimer that dates are platform-specific. * *build-with-claude/claude-on-vertex-ai.md*: Similar to Bedrock, updates dates (Sonnet 3.7 "Retired May 11, 2026"). Adds disclaimer. * *agents-and-tools/tool-use/overview.md*: Updates tool pricing tables, removes old models. * *agents-and-tools/tool-use/text-editor-tool.md*: Removes reference to an older text editor version specific to Sonnet 3.7. * *agents-and-tools/tool-use/computer-use-tool.md*: Removes beta header references to Sonnet 3.7. Removes code examples specific to 3.7 thinking budgets. Removes specific notes about 3.7 improvements, generalizing to "current models". * *agents-and-tools/tool-use/code-execution-tool.md*: Removes references to Sonnet 3.7 and Haiku 3.5 from the tool version table. 3. **Synthesize Themes:** * **Model Lifecycle Cleanup:** There is a massive cleanup across almost all pricing, rate limit, and tool-related pages. Older models (Claude 3.x, Sonnet 3.7, Haiku 3.5) are being removed from main lists, marked as deprecated/retired, or explicitly noted as "only available on Bedrock/Vertex". * **Platform-Specific Deprecation:** A crucial new distinction is being made: "Anthropic-operated" vs "Partner-operated" (Bedrock/Vertex) platforms have different model lifecycle schedules. This is explicitly added to deprecation pages. * **Settings/Hooks Logic Refinement:** Documentation is tightening up the rules for settings precedence (merge vs override) and plugin/hook override capabilities (managed settings are king). * **Behavior Generalization:** Specific references to "Claude Sonnet 3.7" behavior (especially in extended thinking and computer use) are being removed, suggesting the docs are focusing on the current 4.x generation behavior as the standard. 4. **Determine Impact:** * **Impact Level:** Medium. While mostly documentation cleanup, the specific clarification on *platform-specific deprecation schedules* and *settings merge behavior* has real-world implications for developers planning migrations or configuring complex permissions/hooks. 5. **Draft Action Items:** * Check if your code relies on the specific "merge" behavior of settings (now explicitly documented). * Note the differing deprecation dates for Bedrock/Vertex vs Anthropic API. * Update any hardcoded logic relying on specific 3.7 model IDs. * Verify plugin override logic if using managed settings. 6. **Translate and Format (Chinese):** * *Summary:* 本次更新主要清理了旧版模型（如 Claude 3.7, 3.5 系列）的文档引用，明确区分了 Anthropic 自有平台与合作伙伴平台（AWS Bedrock, GCP Vertex）的模型生命周期差异，并完善了设置合并/覆盖逻辑的说明。 * *Themes:* * 模型生命周期清理与平台差异. * 配置优先级与合并规则. * 行为描述通用化. * *Impact:* Medium. (Reason: 主要是文档清理，但关于平台特定的退役日期和设置合并规则的澄清对长期规划和配置调试很重要). * *Action Items:* * 检查代码中是否硬编码了旧版模型 ID（如 Sonnet 3.7）。 * 若使用 AWS Bedrock 或 Vertex AI，请查阅文档中特定于这些平台的模型退役日期，它们与 Anthropic API 的日期不同。 * 审查权限和插件设置，理解数组类设置（如权限）是“合并”而非“覆盖”。 * 更新有关插件和 Hook 的配置逻辑，明确“托管设置”拥有最高优先级。 7. **Final Review against Constraints:** * Single batch analysis? Yes. * Markdown format? Yes. * Chinese response? Yes. * Concise? Yes. 8. **Construct Final Output.** (Proceeding to generate the Markdown response).

1. **Analyze the Request:** * **Role:** World-class programmer analyzing documentation changes. * **Format:** Markdown, concise, insightful, Chinese language. * **Input:** 6 documentation pages with diffs (`headless.md`, `iam.md`, `legal-and-compliance.md`, `memory.md`, `output-styles.md`, `api/rate-limits.md`). * **Focus:** What matters to developers. * **Output Structure:** Overall Summary, Key Themes, Impact Level + Reasoning, Action Items (max 5). 2. **Analyze the Changes (Page by Page):** * **Page 1: `headless.md`** * *Change:* Removed a note about the CLI being previously called "headless mode." Added a note about a new "Agent SDK credit" coming June 15, 2026. Changed phrasing from "run Claude Code programmatically" to "run Claude Code in non-interactive mode." * *Developer Impact:* Terminology update (programmatic -> non-interactive). Important heads-up about future billing/separation of credits (though it's 2026, so far off, but relevant for long-term planning). * **Page 2: `iam.md`** * *Change:* Added the same Note about the June 15, 2026 Agent SDK credit change. * *Developer Impact:* Reiteration of the billing/limit change. * **Page 3: `legal-and-compliance.md`** * *Change:* Added the same Note about the June 15, 2026 Agent SDK credit change. * *Developer Impact:* Reiteration of the billing/limit change. * **Page 4: `memory.md`** * *Change:* Reordered the table of CLAUDE.md file locations. Swapped "User instructions" and "Project instructions" rows. Updated text to clarify load order (broadest to specific). * *Developer Impact:* Clarifies precedence order of configuration files. "Project instructions" now appear before "User instructions" in the list (visually) and the text clarifies the load order explicitly. * **Page 5: `output-styles.md`** * *Change:* Major cleanup and restructuring. * Simplified explanation of how output styles work (removed detailed technical implementation details about token usage/caching unless necessary). * Restructured "Create a custom output style" into `<Steps>` format. * Clarified `keep-coding-instructions`. * Removed the generic "How output styles work" section which explained system prompt modification details, focusing more on *usage*. * *Developer Impact:* Better UX documentation. Easier to create custom styles. Removes some internal implementation details (token usage details) to focus on configuration. * **Page 6: `api/rate-limits.md`** * *Change:* Updated Tier 1 spending limits from $100 to $500 (in two columns). * *Developer Impact:* Increased limits for Tier 1 users. Positive change. 3. **Synthesize and Structure the Output (Drafting in Chinese):** * **Overall Summary:** Updates involve a future pricing model change for Agent SDK, a restructuring of configuration precedence documentation, an increase in API rate limits, and improvements to custom output style documentation. * *Chinese:* 本次更新主要涉及未来定价模式的预告、配置文件加载顺序的明确、API 速率限制的提升，以及对自定义输出样式文档的结构优化。 * **Key Themes:** * Future Policy Change: The 2026 Agent SDK credit separation (major change coming). * Configuration & Precedence: Clarifying CLAUDE.md load order. * Documentation UX: Simplifying output styles guide. * Capacity: Increasing limits. * *Chinese:* * **政策预告**：多页面添加了关于 2026 年 6 月 Agent SDK 积分与交互式使用分离的说明。 * **配置优先级**：明确了 CLAUDE.md 文件在不同层级（用户、项目、托管）的加载顺序。 * **文档体验**：大幅重构了输出样式的创建指南，使其更易于操作。 * **限额提升**：API 第一层级的支出限制从 $100 提升至 $500。 * **Impact Level:** * The rate limit increase is immediate and beneficial. * The config order clarification is important for correctness. * The 2026 change is informational for now. * The output styles change is just doc improvement. * *Level:* Medium (due to config precedence clarification and rate limit increase). The 2026 thing is "High" impact but far future. * *Reasoning:* Config file loading order affects behavior directly. Rate limits affect usage immediately. Future billing is just a heads-up. * **Action Items:** * Review CLAUDE.md placement if you rely on overlapping configs. * Check your tier limits. * Be aware of future billing changes. * *Chinese:* * 检查并确认项目中 `CLAUDE.md` 和 `~/.claude/CLAUDE.md` 的配置是否符合新的优先级规则。 * 利用提升后的 API 速率限制（Tier 1 上限至 $500）评估是否需要调整调用策略。 * 查阅更新后的自定义输出样式文档，优化您的 Claude Code 使用体验。 4. **Refine and Polish (Self-Correction):** * *Headless/Agent SDK note:* The user asked for analysis. The date 2026 is far away. I should mention it's a "future change". * *Memory.md:* The table reorder means "Project instructions" loads *before* "User instructions" in the broad-to-specific sequence shown, though the specific logic "More specific locations take precedence" still holds. The text says "load order, from broadest scope to most specific". Managed -> Project -> User -> Local. Wait, let me look closer at the diff. * *Diff Check on memory.md:* * Old order: Managed -> Project -> User -> Local. * New order: Managed -> User -> Project -> Local. * *Correction:* The table rows swapped. "User instructions" is now above "Project instructions". The text says "from broadest scope to most specific". User (`~/.claude/`) is usually broader than Project (`./.claude/`). So this makes logical sense. * *Output styles:* The explanation removed "Token usage depends on the style..." details in the "How it works" section, but kept the note about caching in the "Change your output style" section. This is just simplification. 5. **Final Output Generation (Markdown):** * *Summary:* Concise. * *Themes:* Clear bullets. * *Impact:* Medium. * *Actions:* Practical. (Proceed to generate Chinese response.)

1. **分析请求：** * **角色：** 世界级程序员。 * **任务：** 分析文档变更（单个批次）。 * **输入：** 一个文件的 Diff：`build-with-claude/prompt-caching.md`（+9 / -8 行）。 * **输出格式：** Markdown。 * **语言：** 中文（按要求）。 * **要求：** 简洁，有见地，关注开发者的需求。 2. **分析 Diff：** * **文件：** `build-with-claude/prompt-caching.md` * **章节：** “缓存策略与考量” -> “缓存限制”。 * **变更内容：** * *原文：* 仅列出了最小可缓存提示词长度（如 4096, 2048, 1024 等具体数值）以及特定模型。 * *新版：* * 阐明了具体平台：Claude API、AWS 上的 Claude Platform、Vertex AI 和 Microsoft Foundry (beta)。 * *关键数据变更：* * Claude Sonnet 4.6：2048 -> **1024** tokens（原文未明确列出 4.6，但现在归入 1024 类别，尽管原文暗示较低模型为 2048）。等等，让我仔细看看列表。 * 旧版：Claude Sonnet 4.6 (2048)，Claude Sonnet 4.5/Opus 4.1 等 (1024)。 * 新版：Claude Sonnet 4.6 (1,024)。*降低*。 * 新列表将 Claude Sonnet 4.6、4.5 等归为 1024。 * 新增文本：“Model availability varies by platform.”（模型可用性因平台而异）。 * 更新了验证建议（提到检查缓存字段以确认）。 * 新增关于 **Amazon Bedrock** 的注释。这是一个特殊/例外情况。它说 Bedrock 是 AWS 运营的平台，并指向 AWS 文档以获取具体的每模型最小值、失败行为和字段名称。这暗示了 Bedrock 的行为与其他平台 *不同*。 * 更新了故障排查列表以反映“你的模型和平台”（而不仅仅是模型）。 3. **综合变更（头脑风暴）：** * *什么变了？* 缓存限制表的细化。 * *关键点：* Claude Sonnet 4.6 的最小缓存大小从 2048 降至 1024。这对该模型来说是件好事。 * *关键点：* 平台区分。文档明确列出了 API、AWS Platform、Vertex AI、Foundry 的新规则。 * *关键点：* Bedrock 被特别排除/例外化。开发者在 Bedrock 上必须查看 AWS 文档，因为“缓存限制”章节可能不适用于它（或者至少，最小值/字段名称不同）。 * *原因？* 准确性，处理不同云提供商之间的实现差异（特别是 Bedrock 的运作方式似乎不同）。 4. **起草输出（中文迭代完善）：** * **1. Overall Summary（总体摘要）：** * *草稿：* 文档更新了 Prompt Caching 的最小 token 限制，明确了不同平台（AWS Bedrock 除外）的规则，降低了 Claude Sonnet 4.6 的缓存门槛。 * *润色：* 更新了 Prompt Caching（提示词缓存）的最小 token 限制，区分了支持平台与 Amazon Bedrock 的不同行为，并显著降低了 Claude Sonnet 4.6 的缓存门槛。 * **2. Key Themes（关键主题）：** * *Theme 1: 降低阈值。* Claude Sonnet 4.6 降至 1024。 * *Theme 2: 平台碎片化。* API/Vertex/Foundry 遵循一套规则；Bedrock 则不同。 * *Theme 3: 精确性。* 明确说明“可用性因平台而异”。 * *草稿：* * **平台差异化**：明确区分了 Claude API/Vertex AI/Foundry 与 Amazon Bedrock 在缓存限制上的差异，Bedrock 需参考独立文档。 * **缓存门槛降低**：将 Claude Sonnet 4.6 的最小可缓存长度从 2048 tokens 下调至 1024 tokens。 * **行为透明化**：强调了当缓存长度不足时请求会静默成功（无错误），增加了开发者验证建议。 * **3. Impact Level（影响级别）：** * *评级：* Medium（中）。 * *推理：* Sonnet 4.6 的降低是一个积极的功能变更。Bedrock 的分离是一个关键的澄清，防止了那些在 Bedrock 上但遵循此处通用规则的开发者出现 Bug。 * *草稿：* **Medium (中等)**。Claude Sonnet 4.6 的缓存门槛降低有利于成本优化，但对 Amazon Bedrock 用户的特殊说明意味着部分开发者需要查阅特定文档以避免配置错误。 * **4. Action Items for Developers（开发者行动项）：** * *Action 1:* Bedrock 用户检查 AWS 文档。 * *Action 2:* Sonnet 4.6 用户可以利用新的 1024 限制。 * *Action 3:* 检查缓存字段 (`cache_creation_input_tokens`) 以确保实际上达到了最小值。 * *草稿：* * **Bedrock 用户**：请务必查阅 [AWS Bedrock 文档](https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html)，因为本文档列出的限制和行为不适用于 Bedrock。 * **Sonnet 4.6 用户**：利用新降低的 1,024 tokens 门槛优化缓存策略，在较小上下文中也能启用缓存。 * **通用检查**：在调试缓存时，检查响应中的 `cache_creation_input_tokens` 和 `cache_read_input_tokens` 字段，确认缓存是否因长度不足而静默失败。 5. **最终审查：** * 是否覆盖了所有变更？是（Sonnet 4.6 降低，Bedrock 警告，平台列表）。 * 是否简洁？是。 * 是否为中文？是。 * 格式是否正确？是。 6. **最终润色（输出生成）：**（根据起草的要点继续生成回答）。