Changes on 2026-10-01

1. **Analyze the Request**: * **Role**: World-class programmer analyzing documentation changes. * **Task**: Analyze a batch of documentation changes (37 pages total, but snippets provided for specific pages). * **Output Format**: Markdown, concise, insightful analysis, focus on what matters to developers. * **Language**: Chinese. 2. **Scan the Diffs (Page by Page Analysis)**: * *amazon-bedrock.md*: Adds detail about guardrail behavior (streaming interruption). Small but clarifies API behavior. * *cli-reference.md*: Updates `claude auth status` output info (new `authMethod` values). Clarifies system prompt flag usage (combining flags, file order). Clarifies version requirements. * *costs.md*: Minor text fix (removing a confusing sentence about fresh sessions). * *desktop.md*: Small UI instruction update (where to manage cloud environments). * *discover-plugins.md*: Grammar/phrase tweak ("the plugins" vs "plugins"). * *headless.md*: **Significant change**. Clarifies `--bare` mode limitations (MCP servers, system reminders, background tasks) starting in v2.1.286. This impacts scripting/SDK usage significantly. * *hooks.md*: Simplifies the `/hooks` menu description. Mentions managed hooks only. Removes detail list about sources. * *hooks-guide.md*: Updates instructions to match the new simplified `/hooks` menu behavior. * *memory.md*: Adds detail about worktree discovery. Clarifies default state of auto memory (off by default outside Claude Tag/self-hosted sessions). * *plugin-marketplaces.md*: Simplifies explanation of entry fields vs `plugin.json`. * *plugins.md*: Phrasing tweak regarding marketplaces. * *plugins-reference.md*: Adds new fields for directory listings (icon, documentationUrl, etc.) that Claude Code ignores but the directory reads. Clarifies validation behavior. * *sandboxing.md*: Adds note about Mods running outside the sandbox. * *security.md*: **Major rewrite/refactor**. Restructures permission mode section (Auto vs Manual). Removes some "marketing" claims about context-aware analysis/input sanitization that might have been misleading or overly generic. Refines "Built-in protections" and "Additional safeguards". * *skills.md* & *slash-commands.md* (Identical content): Adds a new feature: Auto-running `verify`/`simplify` skills before commits (v2.1.286). Adds section on `allowManagedPermissionRulesOnly` ignoring `allowed-tools` in skills. * *sub-agents.md*: Improves behavior when viewing fork/subagent transcripts (confirmations for dangerous commands, model changes affecting main conversation). Adds feature to interrupt background work with `Ctrl+Enter` (v2.1.286). * *troubleshooting.md*: Adds specific troubleshooting step for "Prompt is too long" after clearing. * *get-started.md*: Updates Java SDK version from 2.66.0 to 2.67.0. * *api/messages.md*, *api/messages/count_tokens.md*, *api/messages/batches/create.md*: **Model deprecations and reordering**. `claude-sonnet-4-5` and `claude-sonnet-4-5-20250929` marked as deprecated (EOL Nov 2026). `claude-mythos-preview` deprecated (EOL June 2026). Reordering the list. * *api/models/list.md* & *api/beta/files/upload.md*: Adds a new header `anthropic-workspace-id` (wait, checking the diff... actually `spend-limit-reads-2026-09-26` is added to headers list, not a new request header itself but a date header). * *build-with-claude/overview.md*: Description update for "Thinking" feature (now "optional summary" instead of just "transparency"). 3. **Synthesize Themes**: * **New Features (v2.1.286 focus)**: Pre-commit skill execution (`verify`/`simplify`), refined `--bare` mode behavior, interrupting background work in subagents. * **Security & Permissions**: Detailed description of managed permissions (`allowManagedPermissionRulesOnly`), refinement of security documentation (removing vague claims), sandbox limits for Mods. * **API Changes**: Model deprecations (Sonnet 4.5, Mythos Preview), Java SDK update. * **CLI/User Experience**: Refined system prompt flags, updated `/hooks` menu. 4. **Determine Impact**: * *High*: API model deprecations (EOL dates set), `--bare` mode breaking changes for scripts (behavior change in v2.1.286), Security model changes (managed permissions ignoring local skill rules). * *Medium*: New features (auto-verify, subagent interrupt). * *Low*: Documentation cleanup, minor version bumps, UI phrasing. 5. **Formulate Action Items**: * Update model usage plans due to EOL dates. * Review scripts using `--bare` mode for v2.1.286 compatibility. * Check managed security settings if using `allowed-tools` in skills. * Upgrade Java SDK. * Leverage new pre-commit skill workflow. 6. **Draft the Response (in Chinese)**: * *Summary*: Focus on v2.1.286 features, API model deprecations, and security refinements. * *Themes*: API lifecycle (deprecations), Security (managed controls), CLI updates (bare mode, skills), Documentation polish. * *Impact*: High (because of model EOLs and behavior changes in bare mode/managed permissions). * *Action Items*: Migrate models, check scripts, update SDK, verify security policies. 7. **Refine the Draft**: * *Self-Correction during drafting*: Make sure to distinguish between "API changes" (deprecated models) and "CLI changes" (bare mode behavior). The "security.md" changes are significant because they change the tone and structure of the permissions explanation, clarifying that some previous protections (like "input sanitization") aren't what the user might think. * *Specifics*: Mention `claude-sonnet-4.5` EOL (Nov 2026) and `claude-mythos-preview` (June 2026). Mention `allowManagedPermissionRulesOnly` explicitly. 8. **Final Polish (Markdown)**: Ensure structure matches the prompt requirements. * *Overall Summary*: 这次更新主要集中在 API 模型生命周期管理、CLI 脚本行为的严格化以及企业级安全控制的细化。 * *Key Themes*: * 模型弃用: Sonnet 4.5 和 Mythos Preview 标记了 EOL 日期。 * 行为变更: `--bare` 模式在 v2.1.286 后更加严格，限制了 MCP 和后台任务。 * 安全管控: 强化了 `allowManagedPermissionRulesOnly` 对 Skill 权限覆盖的说明。 * 新特性: Git 提交前自动运行 verify/simplify skills。 * *Impact*: High. * *Action Items*: * 迁移计划。 * 脚本审计。 * SDK 更新。 * 策略检查。 9. **Output Generation** (proceeding to generate output based on these points).

Let me analyze the documentation changes across all 6 pages. 1. **hooks.md** - Added info about plugins registering hooks as JavaScript functions (mods), which can draw in the UI and act on events. Links to mods docs. 2. **hooks-guide.md** - Clarified that hooks in settings files and plugin's `hooks/hooks.json` can tighten but not loosen restrictions. Added that a mod's `tool.check` hook can approve a call that `PreToolUse` blocked, unless the hook is in managed settings. 3. **plugins.md** - Added "hooks module" as a new plugin component type. A plugin with a hooks module is called a "mod". Updated diagram description. 4. **plugins-reference.md** - Added `types` field to plugin manifest (for `.d.ts` file declaring `$.state` values and `$` nouns of a mod). Added plugin name validation rules - reserved names starting with `claude-`, `anthropic-`, etc. cause errors or warnings. 5. **api/overview.md** - Minor clarification: single-workspace API keys run in that workspace when you omit the `anthropic-workspace-id` header. 6. **CHANGELOG.md.md** - Version 2.1.287 with extensive changes including Claude Mods, "You should know" mod, various fixes for hooks, permissions, screen reader mode, MCP, etc. Key themes: - **Claude Mods** - Major new feature: plugins can now register JavaScript function hooks that run in-process, can draw UI panes, and add commands. This is the "mod" concept. - **Permission hierarchy clarification** - Clear rules about what hooks vs mods can do regarding tightening/loosening permissions. Managed settings hooks override mods. - **Plugin name validation** - Reserved names for Anthropic's own plugins. - **New plugin manifest field** - `types` for TypeScript declarations in mods. - **API workspace header clarification** - Minor doc fix. - **Massive bug fix release** - 2.1.287 has tons of fixes. Impact level: This is High. Claude Mods is a significant new capability that changes the plugin architecture fundamentally - plugins can now run JavaScript in-process with UI capabilities, which is a major evolution from the previous shell/HTTP/MCP-only hook model. Action items: - Learn about Mods if you build plugins - new JS function hooks with UI capabilities - Check plugin names don't conflict with new reserved naming rules - Understand the permission hierarchy: managed settings hooks > mods > regular hooks - Add `types` field to plugin manifest if building a mod - Review the extensive changelog for relevant fixes

用户希望我以世界级程序员的角度，分析一批 Claude Code 文档的变更。我需要用中文回复，提供简洁、有洞察力的 Markdown 格式分析。 让我先快速浏览一下这9个页面的变更： 1. **amazon-bedrock.md**: 新增了关于模型访问记忆和会话期间模型被禁用时的回退行为。增加了环境变量 `CLAUDE_CODE_SKIP_MODEL_ACCESS_MEMORY` 和 `CLAUDE_CODE_DISABLE_MODEL_ACCESS_FALLBACK`。 2. **costs.md**: 更新了 "Misses" 的定义，移除了具体的百分比阈值（5%和2000 tokens），并更新了版本要求说明。 3. **google-vertex-ai.md**: 类似于 Amazon Bedrock 的变更，增加了模型访问记忆和会话期间模型被禁用时的回退行为。 4. **hooks.md**: 移除了两个关于 Bash hook 匹配的示例（`cat` 和 `git` 的例子）。 5. **interactive-mode.md**: 更新了后台任务的行为，当禁用后台任务或在裸模式下时，命令会在超时停止而不是移到后台。还更新了 `CLAUDE_CODE_DISABLE_BACKGROUND_TASKS` 的说明。 6. **mcp.md**: 多处小改动：移除了版本特定的说明，更新了插件服务器的信息（隐藏敏感值），移除了关于根级组合器重写的说明。 7. **model-config.md**: 重要更新！关于上下文窗口的处理，特别是在 LLM Gateway 后面时。现在 Sonnet 5.5 和 Sonnet 5 在 Gateway 后面也默认使用 1M 窗口。 8. **network-config.md**: 移除了对已弃用设置 `disableArtifact` 的引用。 9. **vs-code.md**: 简化了状态栏访问 Claude Code 的描述，移除了对 `preferredLocation` 设置的依赖提及。 现在我需要以 Markdown 格式提供分析： 1. 整体摘要 2. 关键主题 3. 影响级别（低/中/高/破坏性） 4. 开发者行动项 让我组织一下： **整体摘要**： 这次更新主要增强了模型可用性处理机制（特别是在 Amazon Bedrock 和 Google Vertex AI 上），调整了上下文窗口在 Gateway 代理下的默认行为，并清理了一些过时的文档说明。核心改进集中在提升用户体验的稳定性和一致性上。 **关键主题**： - 模型可用性管理：新增"访问记忆"和会话中模型降级切换机制，避免重复检查已被拒绝的模型 - 上下文窗口策略调整：Sonnet 5.x 在 LLM Gateway 后默认启用 1M 上下文，不再强制选择 `[1m]` 变体 - 后台任务行为明确化：禁用后台任务或裸模式下超时命令将直接终止而非转后台 - 文档清理：移除过时版本说明、弃用设置引用及冗余示例 **影响级别**： - **中等**（Medium） - 上下文窗口行为变更可能影响通过 Gateway 部署的用户（但通常是改进） - 模型回退行为变更增强了系统稳定性 - 主要是非破坏性改进，向后兼容 **开发者行动项**： - 如果使用 LLM Gateway，验证新的默认 1M 上下文行为是否符合预期 - 对于需要严格 200K 限制的环境，设置 `CLAUDE_CODE_DISABLE_1M_CONTEXT=1` - 如需禁用模型自动回退，配置 `CLAUDE_CODE_DISABLE_MODEL_ACCESS_FALLBACK=1` - 更新依赖旧 `disableArtifact` 设置的配置改为 `enableArtifact` - 检查后台任务超时行为是否符合工作流程需求