Changes on 2026-06-09

1. **Understand the Goal:** The user wants a concise, insightful analysis of documentation changes for a developer tool (Claude Code). I need to process 11 changed pages, summarize the changes, identify key themes, assess impact, and suggest actions. The response must be in Markdown and Chinese. 2. **Analyze the Diffs Page by Page:** * **`cli-reference.md`**: Added `--advisor <model>` flag. Description: Enables server-side advisor tool with aliases (opus, sonnet, fable) or model ID. Requires v2.1.98+ (fable requires v2.1.170+). Overwrites `advisorModel` setting. * *Key takeaway:* New CLI feature for "advisor tool". * **`data-usage.md`**: Updated the data flow diagram URL (image change). Updated text regarding `/feedback`: specifies data goes to Google Cloud Storage (encrypted at rest) in addition to Anthropic. Diagram alt text updated to mention Google Cloud Storage. * *Key takeaway:* Clarification on where feedback data is stored (GCS added). * **`desktop.md`**: Updated "Third-party providers" row in the CLI vs. Desktop table. * *Old:* CLI supports Bedrock, Vertex, Foundry. Desktop supports Anthropic API default, Enterprise can configure Vertex/gateway. * *New:* CLI: "Vertex AI" (refined name). Desktop: Added note about "Cowork on 3P research preview" for running Code tab on Bedrock/Vertex/etc. * *Key takeaway:* Expansion of third-party provider info, specifically mentioning the "Cowork on 3P" preview for Desktop. * **`hooks.md`**: Updated image URL for "hook-resolution". Alt text updated to be more descriptive (denying tool call vs allowing). * *Key takeaway:* Image asset update/alt text refinement. * **`llm-gateway.md`**: Added a paragraph at the top of the requirements section. Clarifies that this page covers CLI gateway config. Mentions Enterprise Desktop uses managed settings. Mentions "Cowork on 3P research preview" for self-hosted gateway on Desktop app with its own config keys. * *Key takeaway:* Documentation scope clarification, mentioning "Cowork on 3P" again for Desktop gateway usage. * **`mcp.md`**: Added a sentence to the "Scale with MCP Tool Search" section: "Claude Code does not impose a fixed per-server tool cap; the practical limit is your context window budget." * *Key takeaway:* Clarification on MCP tool limits (context window based, not fixed). * **`model-config.md`**: Added a note in the "Two-phase planning (Opus Plan)" section. Points to the [advisor tool](/en/advisor) for a hybrid approach where Claude decides when to consult a second model mid-task. * *Key takeaway:* Cross-reference to the new "advisor tool" feature. * **`overview.md`**: Cleaned up code blocks. Removed redundant `theme={null}` attributes. * *Key takeaway:* Markdown cleanup/housekeeping. * **`quickstart.md`**: Cleaned up code blocks. Removed redundant `theme={null}` attributes. * *Key takeaway:* Markdown cleanup/housekeeping. * **`settings.md`**: Added `advisorModel` setting key. Description matches the CLI flag (server-side advisor tool, v2.1.98+ requirement, aliases, unset to disable). * *Key takeaway:* Configuration setting for the new "advisor tool". * **`setup.md`**: Cleaned up code blocks. Removed redundant `theme={null}` attributes. * *Key takeaway:* Markdown cleanup/housekeeping. 3. **Synthesize Key Themes:** * **New Feature: Advisor Tool:** Multiple pages (`cli-reference.md`, `settings.md`, `model-config.md`) introduce and reference a new "server-side advisor tool". This allows a second model to be consulted (hybrid approach). * **Third-Party Integrations (Cowork on 3P):** `desktop.md` and `llm-gateway.md` mention "Cowork on 3P research preview", allowing Desktop app usage with various providers (Bedrock, Vertex, etc.) via a self-hosted gateway. * **Data Privacy/Storage Updates:** `data-usage.md` clarifies that feedback data is stored in Google Cloud Storage. * **Documentation Housekeeping:** Several pages (`overview.md`, `quickstart.md`, `setup.md`) had messy `theme={null}` attributes cleaned up in code blocks. 4. **Assess Impact:** * *Advisor Tool:* High. New functionality that changes how models interact/workflows function. * *Third-Party Integrations:* Medium. Important for enterprise users or those not using Anthropic API directly. * *Data Usage:* Low/Medium. Good to know, but doesn't change code execution logic. * *Cleanup:* Low. Aesthetic/technical debt fix. * *Overall:* **Medium**. It introduces a significant new tool (Advisor) but doesn't seem to break anything. The "Cowork on 3P" is specific to research previews. 5. **Formulate Action Items:** * Try out the new `--advisor` flag. * Update settings to use `advisorModel` if preferred. * Check out the "Cowork on 3P" preview if using Desktop with custom gateways. * Be aware of the updated feedback data storage location (GCS). * Note the MCP tool limit clarification (no hard cap). 6. **Drafting the Response (in Chinese):** * *Overall Summary:* This update introduces the "Advisor Tool" feature for multi-model collaboration, expands third-party provider integration options via the "Cowork on 3P" preview, and clarifies data storage for feedback. * *Key Themes:* * Introduction of the Advisor Tool (CLI flag + Settings). * Expansion of third-party provider support (Cowork on 3P). * Clarification on data usage (Google Cloud Storage). * MCP tool limits clarification (context-based). * Markdown cleanup. * *Impact:* Medium. New capabilities added, mostly additive. * *Action Items:* (List derived in step 5). 7. **Refining the Output:** Ensure the tone is professional ("world-class programmer"). Use markdown structure. * *Translation checks:* * "Advisor Tool" -> 顾问工具 * "Cowork on 3P" -> 3P 协作预览版 * "Context window" -> 上下文窗口 8. **Final Polish:** Check against the prompt constraints (concise, insightful, markdown, Chinese). * *Self-Correction during drafting:* The changes in `overview.md`, `quickstart.md`, `setup.md` are purely removing `theme={null}` noise. I should group this under "Maintenance" or "Cleanup" to avoid cluttering the main functional analysis. * *Refining the Advisor Tool description:* It's about letting a second model consult mid-task or at boundaries. The diff says "hybrid approach... consult a second model mid-task". * *Refining Impact:* "Medium" feels right. It's not "Breaking" (v2.1.98+ is noted but existing config should work), but "High" might be overkill if someone doesn't use the advisor. Let's stick with Medium. 9. **Final Output Generation** (Proceeding to generate response based on these thoughts).

1. **Analyze the Request:** * **Role:** World-class programmer with expertise in dev tools/docs. * **Task:** Analyze a batch of 5 documentation changes. Summarize what changed and why it matters. * **Output Format:** Markdown, concise, insightful, Chinese response. * **Input:** * `hooks.md`: Updated table, clarified `settings.local.json` behavior regarding gitignore. * `hooks-guide.md`: Updated table, clarified `settings.local.json` behavior regarding gitignore. * `settings.md`: Updated table, clarified `settings.local.json` behavior regarding gitignore. * `api/client-sdks.md`: Added a new card for "Apple Foundation Models". * `CHANGELOG.md.md`: Mentioned `post-session` hook, child-process SIGTERM window config, safe mode, `/cd` command, `disableBundledSkills`. 2. **Analyze Each Page:** * **hooks.md & hooks-guide.md:** * *Change:* The "Shareable" column for `.claude/settings.local.json` changed from "No, gitignored" to "No, gitignored when Claude Code creates it". * *Reasoning:* This adds nuance. It implies that if *you* manually create this file, it might *not* be gitignored, but if the tool creates it, it *is*. This is a subtle but important distinction for developers relying on the `.gitignore` behavior to protect sensitive data or local overrides. * **settings.md:** * *Change:* Same update as above regarding `settings.local.json`. * *Reasoning:* Reinforces the consistency across the docs regarding the "Local" scope behavior. * **api/client-sdks.md:** * *Change:* Added "Apple Foundation Models" card. * *Reasoning:* New integration/SDK support. This matters to Apple ecosystem developers. * **CHANGELOG.md.md:** * *Change:* Several new features listed. * `post-session` hook (Self-hosted runner context). * SIGTERM→SIGKILL window configurable. * `--safe-mode` flag (troubleshooting). * `/cd` command (working directory change). * `disableBundledSkills` setting. * *Reasoning:* These are functional feature additions that affect workflow, debugging, and system control. 3. **Synthesize Key Themes:** * **Clarification of Local Settings Behavior:** The most significant "documentation-only" change across 3 pages is the precise definition of when `.claude/settings.local.json` is gitignored. * **New Integration:** Apple support. * **Workflow & Lifecycle Enhancements:** New hooks, commands, and troubleshooting modes found in the changelog. 4. **Drafting the Content (Internal Monologue/Drafting in Chinese):** * *Overall Summary:* 此次更新主要细化了本地配置文件的行为说明，同时引入了 Apple 集成及多项 CLI 增强功能（如安全模式、生命周期钩子等）。 * *Key Themes:* * Docs precision update on `.claude/settings.local.json`. * New Apple SDK support. * Workflow improvements (Hooks, `/cd`, Safe Mode). * *Impact:* Medium. The gitignore nuance is a "gotcha" fix, and the new features are additive. * *Action Items:* * Check `.gitignore` for `settings.local.json` if you created it manually. * Try `--safe-mode` for troubleshooting. * Use `/cd` to change directories. * Check out the Apple SDK if relevant. 5. **Refining the Output (Iterative improvement):** * *Refining the "gitignored when Claude Code creates it" point:* It's crucial. If a developer manually creates it, they might accidentally commit secrets if they assume the tool handles the gitignore. The docs now say "it *is* gitignored *when Claude Code creates it*." This implies manual creation might not trigger the automatic gitignore entry or the tool's specific handling. Wait, usually, `.gitignore` entries are static. If the tool adds the line to `.gitignore` when it creates the file, that's the behavior. If the user creates the file, does the tool add the line to `.gitignore` retroactively? The text says "gitignored when Claude Code creates it". This sounds like a conditional behavior. I should highlight this as a clarification. * *Refining the CHANGELOG:* The `post-session` hook is specifically for "Self-hosted runner". That's a specific niche. The general changes (`--safe-mode`, `/cd`) are more broadly applicable. 6. **Finalizing the Structure (Markdown):** ### 1. Overall Summary 本次文档更新重点修正了本地设置文件（`settings.local.json`）在 Git 忽略策略上的描述细节，并补充了 Apple 生态的 SDK 支持及多项 CLI 生命周期与调试功能。 ### 2. Key Themes * **Git 忽略策略精确化**：在 `hooks.md`、`hooks-guide.md` 和 `settings.md` 中，明确指出 `.claude/settings.local.json` *仅在 Claude Code 创建该文件时*才会被自动加入 `.gitignore`。这解决了关于手动创建文件时是否受保护的不明确之处。 * **生态集成扩展**：新增 Apple Foundation Models 集成，支持 Swift 开发者通过 Apple 的 `LanguageModelSession` API 使用 Claude。 * **工作流与调试增强**：新增 `--safe-mode` 用于无自定义项排查问题，新增 `/cd` 命令用于平滑切换目录，以及针对自托管运行器的 `post-session` 钩子。 ### 3. Impact Level **Medium (中)** * **Reasoning**: 文档层面的修正属于“低”影响，但防止了潜在的敏感配置泄露风险；新增的 CLI 功能（安全模式、目录切换）和 Apple SDK 属于“中”影响，提供了实用的开发和调试新工具。 ### 4. Action Items for Developers * **审查本地配置文件**：如果您手动创建了 `.claude/settings.local.json`，请检查项目的 `.gitignore` 是否包含该条目，以防误提交敏感配置。 * **利用安全模式排查**：遇到插件或技能导致的异常时，尝试使用 `--safe-mode` 启动 Claude Code。 * **体验 `/cd` 命令**：在需要切换工作目录但不想中断 Prompt 缓存的场景下，使用 `/cd` 替代传统方式。 * **关注自托管钩子**：如果您使用自托管 Runner，可利用新增的 `post-session` 钩子在会话结束快照未提交的工作。 * **探索 Apple 集成**：Swift 开发者可查阅新增的 "Apple Foundation Models" 文档以了解集成方案。 7. **Final Review:** Does this sound like a world-class programmer? Yes, it focuses on the nuance (the gitignore condition) and highlights the new functional capabilities clearly. It's in Chinese as requested. *Self-Correction during drafting:* Make sure to mention that the gitignore change implies manual creation might *not* be ignored automatically, which is the key insight. 8. **Generate Output.**

用户希望我分析一批文档更改（16页），并提供一个简洁的、有深刻见解的总结，重点关注对开发者重要的事项。输出需要使用Markdown格式，并用中文回复。 让我回顾一下每个页面的更改： 1. **claude-code-on-the-web.md**: * 更新了云会话的表格，澄清了哪些配置是可用的。 * 添加了关于 `~/.claude/skills/`、`~/.claude/agents/`、`~/.claude/commands/` 不可用的行（本地与仓库作用域）。 * 更新了关于安全特性的文本（虽然被截断了）。 * *重要性*：澄清了云环境与本地环境的配置管理，特别是关于 skills/agents/commands 的作用域。 2. **desktop.md**: * 将 macOS MDM 配置域从 `com.anthropic.Claude` 更改为 `com.anthropic.claudefordesktop`。 * *重要性*：使用 MDM（Jamf, Kandji）管理桌面应用的 IT 团队的重大变更。 3. **headless.md**: * 更新了关于 `-p` 模式下 skills 和自定义命令的说明。它们*现在*可以工作了（通过在提示字符串中包含 `/skill-name`）。交互式命令（如 `/config`，`/login`）仍然被阻止。 * *重要性*：提升了脚本/CI/CD（无头模式）的能力。 4. **hooks.md**: * 阐明了哪些 Hook 事件可以重写内容，而不仅仅是允许/阻止。 * 强调了 `PreToolUse`、`PermissionRequest`、`PostToolUse` 的 `updatedInput`/`updatedToolOutput` 字段。 * *重要性*：高级 Hook 用例（审计、编辑输入/输出）的更强功能。 5. **hooks-guide.md**: * 更新了 `PreToolUse` 决策的优先级顺序：`deny`、`defer`、`ask`、`allow`。（添加了 `defer`）。 * *重要性*：Hook 权限逻辑中的精度变更。 6. **model-config.md**: * 添加了 "Fable" 到模型别名列表中。 * 更新了环境变量以包含 `ANTHROPIC_DEFAULT_FABLE_MODEL`。 * 添加了用于 Fable 的 `DISABLE_PROMPT_CACHING_FABLE`。 * *重要性*：引入了一个新的模型系列 "Fable"。 7. **monitoring-usage.md**: * 更新了 `user.id` 的描述，澄清其是随机的、匿名的，并且是持久化的（不是设备安装 ID？文本说是 "首次运行时生成的随机匿名标识符"）。 * 阐明了分段能力（修正了关于某些指标上 `model` 属性的细节）。 * *重要性*：遥测/可观测性的澄清。 8. **overview.md**: * 代码块中出现了一些看起来像垃圾文本的重复 `theme={null}`（可能是构建产物伪影？），但在 markdown 内容变更的语境下，看起来它们只是添加了更多的 `theme={null}` 字符串。除非这是一个可执行的更改，否则这可能是噪音。等等，仔细观察，这只是代码块语言的重复属性。这可能是一个文档生成器的伪影，我会忽略它，除非它看起来是有意为之的。 * 关于 `claude --teleport` 要求订阅的说明。 * *重要性*：功能可用性澄清（Teleport 需要订阅）。 9. **plugins.md**: * 澄清了迁移行为：项目/用户 `.claude/agents/` 会覆盖同名的插件 agents。 * *重要性*：插件/本地 agent 之间的冲突解决逻辑。 10. **sandboxing.md**: * 澄清了默认的沙盒写入权限包含会话临时目录（`$TMPDIR`），而不仅仅是工作目录。 * 详细说明了 `$TMPDIR` 在沙盒与非沙盒命令中的行为。 * *重要性*：修复了沙盒内使用临时文件的工具的行为。 11. **security.md**: * 重述了 Bash 执行策略（只读命令自动批准）。 * 阐明了网络命令（`curl`、`wget`）不再是默认阻止的（"blocklist" 变为 "not auto-approved"），而是像其他非只读命令一样提示。 * 更新了凭据存储详情（macOS Keychain 对比文件权限）。 * *重要性*：安全行为变更（网络命令现在可允许但需提示，而不是硬性阻止）。 12. **settings.md**: * 澄清了 `.claude/settings.local.json` 的 git 忽略行为（自动 vs 手动）。 * 澄清了权限规则评估（第一个匹配决定结果，无论具体性如何）。 * 更新了设置错误的处理（启动时的设置问题通知 + `/doctor` 命令）。 * *重要性*：配置管理的改进。 13. **setup.md**: * 同样的 `theme={null}` 重复伪影。 * 更新了 Linux 包管理器（apt/dnf/apk）的说明，区分了 `stable` 和 `latest` 通道。`stable` 现在大约滞后一周。 * *重要性*：Linux 发行/升级策略的更新。 14. **sub-agents.md**: * 添加了关于在无头模式/Agent SDK 中禁用内置子代理的说明（`CLAUDE_AGENT_SDK_DISABLE_BUILTIN_AGENTS`）。 * *重要性*：集成的配置控制。 15. **third-party-integrations.md**: * 添加了 `ANTHROPIC_DEFAULT_FABLE_MODEL` 到环境变量列表中。 * *重要性*：与模型配置更改保持一致（引入 Fable）。 16. **CHANGELOG.md.md**: * 版本 2.1.170 * **重大新闻**：引入 Claude Fable 5（一个 "Mythos-class" 模型）。 * 修复：当从集成终端启动时，会话无法保存转录内容。 * *重要性*：高（新模型版本，关键的错误修复）。 **总结分析：** 1. **新模型 (Fable 5)**：最显著的变化是引入了 "Claude Fable 5"。这影响了多个页面（model-config, third-party-integrations），增加了一个新的别名 `fable` 和相关的环境变量。 2. **沙盒与安全细化**：安全文档的变更很重要。网络命令（curl/wget）从“默认阻止”转变为“需要批准”（与标准 Bash 处理一致）。沙盒默认行为现在明确包含临时目录写入。 3. **云/桌面配置**：macOS MDM 领域的更改对企业用户来说是一个破坏性变更（如果他们已部署的话）。关于云会话中本地与仓库作用域 skills/agents 的澄清对开发者来说是实用知识。 4. **开发体验/稳定性**：VS Code 集成终端会话保存的修复是一个重要的稳定性改进。无头模式中的 skill 支持增强了自动化。 **影响等级**：高。 * **理由**：新模型 (Fable) 是重大的功能更新。MDM 领域的更改是 IT 管理的破坏性变更。安全行为的转变（curl/wget）改变了核心交互流程。修复的 bug 影响工作流稳定性。 **行动项**： 1. IT 管理员需要更新 MDM 配置脚本。 2. 开发者如果有资格应该测试 Fable 5，并相应地更新环境变量。 3. 用户需注意关于 `curl`/`wget` 的更改——它们不再被硬性阻止，但需要批准。 4. 无头用户现在可以在脚本中使用 skills。 5. 检查 `.claude/settings.local.json` 的 git 状态。

1. **Understand the Goal:** The user wants a concise, insightful analysis of a batch of documentation changes for "Claude Code" and "Claude API". The analysis should be in Chinese, focus on what matters to developers, and follow a specific markdown structure (Overall Summary, Key Themes, Impact Level, Action Items). 2. **Analyze the Input Data:** I have a list of 34 pages with diffs. I need to scan through them to identify the core changes. * *Page: cli-reference.md*: Large table diff, mostly reformatting or whitespace changes? Actually, looking closer, it's just a snippet of the table. The snippet looks identical but is wrapped in a diff block. It shows flags like `--add-dir`, `--agent`, etc. *Self-correction*: The diff shows `+69 / -69`, suggesting a reformat or whitespace change, or perhaps a massive table update where lines are replaced but look the same in the snippet. Let's look at other pages to see if there's a pattern. * *Page: costs.md*: Added note about "Fable 5" always using extended thinking and cannot be disabled. * *Page: desktop.md*: Added note about "Fable 5" always using extended thinking. Updated model list description. * *Page: github-actions.md*: Removed info box about Claude Opus 4.8 availability (likely outdated or replaced by new info). * *Page: interactive-mode.md*: Added note that "Toggle extended thinking" has no effect on Fable 5. * *Page: mcp.md*: Changed text about tool search support (removed specific version requirements "Sonnet 4 and later", just says "requires a model that supports..."). * *Page: model-config.md*: * **Major Change**: Added `fable` alias. * **Major Change**: New section "Work with Fable 5". Describes it as the most capable model for long sessions. * Updated "best" alias to use Fable 5 if available. * Updated environment variable list to include `ANTHROPIC_DEFAULT_FABLE_MODEL`. * Clarified Fable 5 is not the default. * **Major Change**: New section "Automatic model fallback". Fable 5 falls back to Opus (4.8 or 4.7) if safety classifiers trigger. * *Page: settings.md*: Updated `alwaysThinkingEnabled` description: `MAX_THINKING_TOKENS=0` doesn't disable thinking on Fable 5. * *Page: sub-agents.md*: Added `fable` to the list of available model aliases for subagents. * *Page: intro.md*: * **Major Update**: Introduced "Claude Fable 5" as the most capable widely released model. * Introduced "Claude Mythos 5" via Project Glasswing. * Updated Opus 4.8 description. * *Page: get-started.md*: Updated Java SDK dependency from `2.35.0` to `2.39.0`. * *Page: about-claude/models/overview.md*: * **Major Update**: Added detailed table for Claude Fable 5 and Mythos 5. * Updated pricing: $10/$50 per MTok. * Context window: 1M tokens. * Max output: 128k tokens. * Availability: June 9, 2026 (Future date in prompt context, but I treat it as "now" for the docs). * Extended thinking: No. Adaptive thinking: Yes. * Updated Opus 4.8 description to "most capable *Opus-tier* model". * *Page: about-claude/models/migration-guide.md*: * **Major Update**: Added section "Migrating from Claude Mythos Preview to Claude Mythos 5". * **Major Update**: Added section "Migrating from Claude Opus 4.8 to Claude Fable 5". * Key change: Adaptive thinking is always on. Manual extended thinking rejected. Tokenizer changes (30% more tokens). * *Page: about-claude/pricing.md*: Added pricing rows for Fable 5 and Mythos 5 ($10/$50 input/output). Batch pricing ($5/$25). * *Page: about-claude/model-deprecations.md*: Note about Mythos Preview retirement. * *Page: api/messages.md*: * **API Change**: Added `claude-fable-5` and `claude-mythos-5` to model enums. * **API Change**: Added `reasoning_extraction` to refusal categories (in addition to `cyber` and `bio`). * **API Change**: `ToolSearchToolResultErrorParam` added `error_message` field. * *Page: api/messages/count_tokens.md*: Same model enum updates. * *Page: api/messages/batches/create.md*: Same model enum updates. * *Page: api/models/list.md*: Added new capabilities/betas: `server-side-fallback-2026-06-01`, `fallback-credit-2026-06-01`. * *Page: api/beta/files/upload.md*: Same capability updates. * *Page: api/rate-limits.md*: Added rate limit tiers for Claude Fable 5. * *Page: api/errors.md*: Updated "Prefill not supported" list to include Fable 5 and Mythos 5. * *Page: build-with-claude/vision.md*: Added Fable 5/Mythos 5 to high-res image support list (same max res as Opus 4.8). * *Page: build-with-claude/extended-thinking.md*: * **Breaking/Behavioral Change**: Fable 5 and Mythos 5 *always* have extended thinking on. Manual config (`thinking: {type: "enabled"...}`) returns 400 error. * Adaptive thinking is always on. `thinking: {type: "disabled"}` returns an error. * *Page: build-with-claude/streaming.md*: Added note about server-side fallback events in streaming. 3. **Synthesize the Core Themes:** * **New Models (The Headline):** Introduction of **Claude Fable 5** (widely available) and **Claude Mythos 5** (limited availability, Glasswing). * **Fable 5 Characteristics:** Highest capability, adaptive reasoning (always on), no manual extended thinking support (it's built-in), safety classifiers triggering fallbacks. * **API Changes:** New model IDs, new refusal category (`reasoning_extraction`), new error field (`error_message` in tool search), new beta capabilities (`server-side-fallback`). * **Migration:** Guides provided for moving from Mythos Preview -> Mythos 5, and Opus 4.8 -> Fable 5. Key nuance: Fable 5/Mythos 5 reject manual `thinking` parameters and always use adaptive reasoning. * **Pricing & Limits:** Fable 5 is significantly more expensive ($10/$50 vs Opus $5/$25) but has higher limits in higher tiers. 4. **Determine Impact Level:** * **High**. This isn't just a minor tweak. It introduces the "next generation" models (Fable 5), changes fundamental behavior regarding how thinking works (adaptive vs manual), changes pricing, and introduces new API fields. Developers using the API or configuring models need to know about these changes, especially if they want to opt into the new models or if they rely on manual thinking configuration. 5. **Draft Action Items (in Chinese):** * Update model selection logic to include `claude-fable-5` and `claude-mythos-5`. * Review and remove manual `thinking` configuration if migrating to Fable 5/Mythos 5 (switch to `effort` parameter). * Update budgeting/cost calculations due to the significant price increase of Fable 5 ($10/$50). * Handle new API response fields: `error_message` in tool search errors and `reasoning_extraction` in refusal categories. * Note that Fable 5 includes automatic fallback to Opus based on safety classifiers; monitor or configure this behavior if necessary. 6. **Draft the Response (in Chinese):** * *Overall Summary*: 本次更新主要发布了下一代旗舰模型 **Claude Fable 5**（广泛可用）和 **Claude Mythos 5**（受限可用）。文档全面更新了模型配置、API 接口、定价策略以及迁移指南，重点强调了新模型的固定自适应推理模式及安全分类器回退机制。 * *Key Themes*: * 新模型发布：Fable 5 & Mythos 5。 * 推理模式变更：新模型强制使用自适应推理，禁用手动扩展思考配置。 * API 更新：新增模型 ID、拒绝类别和错误字段。 * 成本与配额：Fable 5 价格翻倍，但高等级限流更高。 * 安全机制：Fable 5 包含自动回退到 Opus 的功能。 * *Impact Level*: High. 涉及核心模型能力、API 行为变更及价格调整。 * *Action Items*: (See step 5). 7. **Refine Language:** Ensure technical terms are handled correctly (e.g., "Adaptive Thinking", "Extended Thinking", "Safety Classifiers"). * *Refining the thinking block point*: It's important to note that Fable 5 rejects the manual `thinking` block. This is a breaking change for code that sends it explicitly. 8. **Final Polish:** Check against the requested format. * *Summary*: 简洁有力。 * *Themes*: 覆盖全面。 * *Impact*: High 理由充分。 * *Actions*: 5 条清晰可执行。 *(Self-Correction on `cli-reference.md`)*: I initially thought it was whitespace. Looking at the snippet again, it seems the table formatting might have changed or it's just noise in the diff provided. Since the content (text) of the snippet shown is identical, I won't focus on it as a "major change" but rather treat the explicit content changes (Fable 5) as the drivers. The `+69/-69` suggests a table reformat or just the diff context is large. I will stick to the clear content changes found in other files. *(Self-Correction on dates)*: The docs mention "June 9, 2026". I should mention these models are "Available" or "Launched" based on the doc state presented, regardless of the real-world date (which might be the prompt's fictional future). I will treat the docs as truth for this analysis. 9. **Final Output Generation** (matches the thought process).

Let me analyze these documentation changes across 25 pages to identify the key themes, impacts, and action items. Major changes I'm seeing: 1. **"Remote" → "Cloud" terminology**: Consistent renaming across multiple pages (claude-code-on-the-web.md, desktop.md, iam.md, vs-code.md) from "remote sessions" to "cloud sessions". This is a branding/naming change. 2. **Fallback model chains**: New feature in model-config.md allowing configuration of fallback models when primary is unavailable. New `--fallback-model` CLI flag and `fallbackModel` settings key. 3. **Managed settings tolerance/validation**: New section in settings.md about how invalid entries in managed settings are handled - stripped individually rather than rejecting the whole file. Security-enforcement fields have specific behaviors. 4. **`--safe-mode`**: New CLI flag referenced in troubleshooting.md and monitoring-usage.md for disabling all customizations to debug issues. 5. **Plugin improvements**: - Single-skill plugins can use root-level SKILL.md - Plugin detail view shows components - `/reload-plugins` warning for MCP cache invalidation, `--force` flag - `diagnostics` option for LSP servers (default true, can suppress) - Reserved marketplace names updated - Git source `ref`+`sha` pinning behavior documented - `pluginSuggestionMarketplaces` managed setting for allowlisting 6. **Bypass permissions behavior change**: Now respects explicit `ask` rules - this is significant! Multiple pages updated (desktop.md, sandboxing.md, sub-agents.md, security.md). 7. **`MAX_THINKING_TOKENS` clarification**: Only applies to fixed-thinking-budget models; adaptive-reasoning models ignore it, use effort levels instead. 8. **Hooks matching clarification**: Updated description for patterns that "specify more than the command name" - they run the hook anyway on $(), backticks, or $VAR. 9. **Fork subagent control**: `CLAUDE_CODE_FORK_SUBAGENT` env var documentation clarified. 10. **Resumable subagents**: Built-in Explore and Plan agents can't be resumed; use general-purpose or custom. 11. **`disableBundledSkills` setting**: New setting to disable bundled skills. 12. **Skill argument escaping**: Clarified doubled backslash behavior. 13. **Security trust verification**: `--worktree` no longer requires trust verification in `-p` mode. 14. **Link fixes**: Various markdown link encoding fixes. 15. **`safe_mode` field in monitoring events**: New field in hook and plugin telemetry events. Impact assessment: - The bypass permissions now respecting ask rules is a behavior change that could affect existing workflows - Fallback model chains is a new feature - `--safe-mode` is a new debugging tool - The rest are mostly documentation clarifications and naming changes Let me structure this properly.