Changes on 2026-06-24

1. **Analyze the Request**: * **Role**: World-class programmer/doc expert. * **Task**: Analyze a specific documentation change (CHANGELOG.md.md) in a single batch. * **Format**: Markdown, concise, insightful, Chinese language. * **Content Requirements**: 1. Overall Summary (1-2 sentences). 2. Key Themes (bullets). 3. Impact Level (Low/Medium/High/Breaking) with reasoning. 4. Action Items for developers (max 5). 2. **Analyze the Diff Content**: * **Page**: `CHANGELOG.md.md` (Changelog for version `2.1.191`). * **Type**: Release notes / Bug fixes / Performance improvements. * **Specific Changes**: * *New Feature*: `/rewind` support to resume after `/clear`. * *UX/UI Fixes*: Scroll position jumping, agent panel jumping, splash art overflow, Cmd+click on links (Ghostty), welcome splash art overflow, permissions denial persistence. * *Functionality Fixes*: Background agents resurrecting, `/voice` message genericity, `/login` URL truncation, sending builtin commands as prompt text to background agents, image path display, hooks with comma-separated matchers not firing, managed settings refresh. * *Improvements*: Sandbox network permission dialog (remembering "Yes"), MCP server reliability (retries), MCP OAuth (retries + headless flow), MCP error messages (404s), vim mode search hint. * *Performance*: Reduced CPU usage (~37%) by coalescing text updates, reduced memory growth from output cache. 3. **Synthesize the Analysis**: * **Overall Summary**: This release (v2.1.191) focuses heavily on stability, performance, and bug fixes, particularly around MCP (Model Context Protocol) integration, background agents, and user interface responsiveness. It also introduces a `/rewind` feature and significantly reduces CPU usage during streaming. * **Key Themes**: * *Performance & Stability*: Significant CPU reduction (37%) during streaming and memory leak fixes. * *MCP & Integration*: Reliability improvements for MCP servers, OAuth, and error messaging. * *Agent Management*: Fixes regarding background agents resurrecting and command handling. * *User Experience (UX)*: Fixes for scrolling, terminal specific issues (Ghostty, Windows Terminal), and permission dialogs. * *Developer Tools*: Hooks fixes (`Bash,PowerShell` matcher) and managed settings enforcement. * **Impact Level**: **Medium**. * *Reasoning*: While mostly bug fixes and polish, the CPU reduction (37%) is a tangible performance gain for developers using the tool heavily. The fix for hooks silently failing with comma-separated matchers is critical for users relying on that configuration. The MCP reliability improvements are significant for plugin/extension developers. It's not "High" because there are no breaking changes or massive new architectural shifts, but it's not "Low" because of the performance and stability boosts. * **Action Items for Developers**: * *Verify Hooks*: Check if your hooks were using comma-separated matchers (e.g., `"Bash,PowerShell"`), as they should now work correctly. * *Review MCP Configs*: If you encountered HTTP 404 errors or connection timeouts with MCP servers, check the improved error messages for debugging. * *Update Managed Settings*: If using MDM or file policies for `forceRemoteSettingsRefresh`, verify it now takes effect immediately. * *Test Background Agents*: Confirm that previously persistent background agents now actually stop when commanded. * *Utilize `/rewind`*: Try the new command to recover conversation context after accidental clears. 4. **Drafting the Response (in Chinese)**: * *Summary*: 版本 2.1.191 主要聚焦于稳定性提升、性能优化（CPU 使用率降低 37%）以及错误修复，特别是针对 MCP 集成、后台代理和终端用户体验的改进。 * *Themes*: * 性能与资源优化：显著降低了流式响应时的 CPU 占用并修复了内存增长问题。 * MCP 集成增强：改进了服务器可靠性、OAuth 流程及错误提示。 * Bug 修复与稳定性：解决了后台代理“复活”、Hooks 配置失效及 UI 跳动等问题。 * 用户体验改善：增加了 `/rewind` 命令，优化了权限对话框和终端兼容性。 * *Impact*: Medium. 理由：虽然没有破坏性变更，但 CPU 性能提升和 Hooks 配置修复对重度用户影响显著，MCP 的稳定性改进对开发者友好。 * *Action Items*: * 检查使用了逗号分隔匹配器（如 `"Bash,PowerShell"`）的 Hooks 配置是否现已生效。 * 利用新的错误提示排查之前的 MCP HTTP 404 连接问题。 * 验证通过 MDM 设置的 `forceRemoteSettingsRefresh` 是否能立即生效。 * 体验新的 `/rewind` 命令以恢复被清除的对话上下文。 * 确认后台代理在停止后是否真正终止，不再意外重启。 5. **Refinement**: Ensure the tone is professional and concise. The structure matches the prompt. * *Draft check*: * Summary: Concise? Yes. * Themes: Bullets? Yes. * Impact: Level + Reasoning? Yes. * Action Items: Max 5 bullets? Yes. 6. **Final Output Generation**. (Matches the drafted Chinese text).

1. **Analyze the Request:** * **Role:** World-class programmer analyzing documentation changes. * **Input:** Two documentation diffs (`discover-plugins.md` and `sub-agents.md`). * **Language:** Chinese. * **Format:** Markdown with specific sections (Overall Summary, Key Themes, Impact Level, Action Items). 2. **Analyze `discover-plugins.md`:** * *Change:* Added a sentence at the end of a paragraph about the "Not used recently" feature. * *Content:* "The group and the **Last used** line are both hidden when your organization restricts marketplaces with [`strictKnownMarketplaces`]." * *Context:* This is related to the Installed tab and identifying unused plugins. * *Developer Impact:* Developers in organizations with strict marketplace settings won't see the "Not used recently" group or "Last used" line. This is important to know so they don't wonder why these features are missing. 3. **Analyze `sub-agents.md`:** * *Change:* Added a paragraph explaining depth behavior for background subagents when resuming. * *Content:* "As of Claude Code v2.1.187, a background subagent's depth is fixed when it is first spawned, and resuming it later does not change that depth." It gives an example: Main -> A -> B (depth 2, background). Resuming B directly from Main keeps it at depth 2. It cannot spawn more levels just because it's being resumed from a shallower context. * *Context:* Subagent depth limits and resuming background subagents. * *Developer Impact:* Developers need to understand that the depth calculation is static upon creation. Resuming a deep subagent from a higher level doesn't "reset" or "shallow" its depth constraints. This prevents unexpected behavior where one might think they could bypass depth limits by resuming. 4. **Synthesize Key Themes:** * Both changes are clarifications or edge-case handling related to version `v2.1.187`. * Both deal with *hidden* or *fixed* behaviors that might not be obvious. * One is about UI visibility based on policy (`strictKnownMarketplaces`). * One is about logic persistence (depth calculation) based on initialization time. 5. **Drafting the Analysis (in Chinese):** * **Overall Summary:** Two pages received updates related to Claude Code v2.1.187, clarifying specific behaviors regarding plugin management visibility under restricted policies and the immutable nature of background subagent depth limits upon resumption. * *Chinese translation thought:* 这两个页面的更新均与 Claude Code v2.1.187 相关，主要阐明了在受限策略下插件管理的可见性规则，以及恢复后台子代理时其深度限制的不可变性。 * **Key Themes:** * Feature Visibility Restrictions: Hiding specific UI elements ("Not used recently", "Last used") based on organization settings. * Static State Logic: Clarifying that subagent depth is determined at spawn time and does not change based on the context from which it is resumed. * *Chinese translation thought:* * 特性可见性与权限限制：说明组织策略（`strictKnownMarketplaces`）如何影响插件管理 UI 的显示。 * 状态固化逻辑：阐明后台子代理的深度限制在生成时即已固定，恢复操作不会改变这一属性。 * **Impact Level:** Low to Medium. * *Reasoning:* These are clarifications that fix potential misunderstandings or edge cases. They aren't breaking API changes, but understanding the subagent depth rule is crucial for designing complex multi-agent workflows. * *Chinese translation thought:* 低至中。虽然不是破坏性 API 变更，但对于设计复杂多代理工作流的开发者来说，理解深度固化规则至关重要，以免预期出现偏差。 * **Action Items:** * Check organization settings if the "Not used recently" plugin group is missing. * Design subagent workflows knowing that depth limits are calculated at creation, not resumption. * Review v2.1.187 release notes if applicable for these behavioral changes. * *Chinese translation thought:* * 如果在插件管理界面看不到“最近未使用”分组，请检查组织的 `strictKnownMarketplaces` 设置。 * 在设计深度嵌套的子代理工作流时，需记住深度在生成时锁定，直接恢复深层子代理不会赋予其创建新层级的能力。 6. **Review and Refine:** * The tone is professional and technical. * The formatting matches the requested markdown structure. * The language is natural Chinese. *Refinement on "Impact Level":* Actually, knowing that you can't bypass depth limits by resuming is a "gotcha" fix. I'll keep it "Low to Medium" but emphasize the logic fix aspect. 7. **Final Output Generation:** (Proceeding to generate the response based on the drafted points).

1. **Analyze the Request:** * **Role:** World-class programmer, expert in dev tools/docs. * **Task:** Analyze a specific documentation change (interactive mode search behavior). * **Input:** A single diff from `interactive-mode.md`. * **Format:** Markdown, Chinese language. * **Structure:** 1. Overall Summary, 2. Key Themes, 3. Impact Level + Reasoning, 4. Action Items (max 5). 2. **Analyze the Input Data:** * **File:** `interactive-mode.md` * **Change:** Modified a sentence describing the search functionality in interactive mode. * **Old Text:** "The search displays matching commands with the search term highlighted..." * **New Text:** "The search loads the 100 most recent unique prompts in the selected scope, with duplicates collapsed to the newest occurrence. Matching prompts display with the search term highlighted..." * **Key Insight:** The documentation now clarifies the *scope* and *limit* of the search history. It's not just a fuzzy search of everything; it's limited to the 100 most recent *unique* prompts, handling duplicates by keeping the newest. 3. **Draft the Content (Internal Monologue/Rough Draft):** * *Summary:* The docs for interactive mode search now explain exactly how history is managed. It limits the search to the top 100 unique prompts and handles duplicates. * *Themes:* Search behavior, History management, Specific limits/constraints. * *Impact:* Low. It's a clarification of behavior, not a functional change in the code itself (implied, usually docs follow code or clarify intent). However, knowing the limit is useful for users. * *Action Items:* None really, except knowing that old prompts might disappear if you have more than 100 unique ones. Be aware of the "newest occurrence" rule for duplicates. 4. **Refine and Translate to Chinese (Target Language):** * *Overall Summary:* * Draft: 文档更新了交互模式搜索功能的说明，明确指出搜索仅加载选中范围内最近 100 条唯一的提示词，重复项会保留最新的一条。 * Refinement: 使其更紧凑。更新了交互模式搜索功能的文档说明，明确限制搜索范围为最近 100 条唯一的提示词，重复项仅保留最新记录。 * *Key Themes:* * 搜索历史限制 (Search history limit). * 去重逻辑 (Deduplication logic). * 用户体验澄清 (UX clarification). * *Impact Level:* * Level: Low (低). * Reasoning: 这是一个文档说明的改进，阐明了现有机制，而非破坏性的代码变更。开发者需要了解此限制以管理历史记录预期。 * *Action Items:* * 了解旧的唯一提示词超过 100 条后将不再出现在搜索中。 * 重复的指令在搜索结果中将仅显示最新的执行版本。 5. **Final Review against Constraints:** * Markdown format? Yes. * Chinese language? Yes. * Concise? Yes. * Action items max 5? Yes. 6. **Final Output Generation:** (Proceeding to generate the response based on the drafted points).

1. **Analyze the Request:** * **Role:** World-class programmer analyzing documentation changes. * **Task:** Analyze 15 changed pages as a single batch. * **Output:** Markdown format (Chinese language requested at the start: "请用中文回复"). * **Sections Required:** 1. Overall Summary (1-2 sentences). 2. Key Themes (bullets). 3. Impact Level (Low/Medium/High/Breaking) + Reasoning. 4. Action Items for Developers (bullets, max 5). 2. **Analyze the Diff Content (Page by Page):** * **`amazon-bedrock.md`**: * *Change:* Updated example for `availableModels`. Replaced `"haiku"` with `"claude-haiku-4-5"`. Clarified that listing a specific ID removes bare aliases. * *Theme:* Model configuration/aliasing specificity. * **`claude-code-on-the-web.md`**: * *Change:* Updated table about cloud sessions. Added mention of "server-managed settings" arriving separately. Refined wording on availability. * *Theme:* Cloud session behavior, managed settings. * **`desktop.md`**: * *Change:* Clarified "Managed settings". Refined descriptions for `managedMcpServers`. Explained where managed settings apply (Desktop vs CLI vs IDE). * *Theme:* Managed settings, enterprise configuration. * **`discover-plugins.md`**: * *Change:* New section in "Installed" tab: "Not used recently". Logic: 2 weeks, 10 sessions. Added "Last used" line. Exceptions: Managed plugins, LSP servers, themes, etc. * *Theme:* Plugin management, performance optimization (cleaning up unused plugins). * **`github-actions.md`**: * *Change:* Updated `/install-github-app` command description. Added "Skip for now" option in v2.1.187 to install just the App without workflows immediately. * *Theme:* CI/CD setup, version-specific features (v2.1.187). * **`interactive-mode.md`**: * *Change:* New keybind: `Left` / `Right` to navigate between `/btw` answers. Version: 2.1.187+. * *Theme:* UI interaction, version-specific features. * **`mcp.md`**: * *Change:* New behavior in v2.1.187: MCP tool calls abort after 5 minutes of no response/progress (idle timeout). Env var `CLAUDE_CODE_MCP_TOOL_IDLE_TIMEOUT` to control/disable. Stdio servers exempt. * *Theme:* MCP timeout handling, performance/reliability. * **`model-config.md`**: * *Change:* *Major Update.* `availableModels` logic clarification. Now applies to skills, commands, advisor, background agents. Added "Surface coverage" table explaining where settings apply (CLI, IDE, Desktop, Web, Cloud, etc.). Clarified that listing full IDs removes bare aliases. * *Theme:* Enterprise model governance, settings precedence, surface coverage. * **`monitoring-usage.md`**: * *Change:* Clarification on `attempt` attribute in `claude_code.api_error` event. It equals max_retries + 1 (e.g., 11 if max is 10). * *Theme:* Monitoring/Observability accuracy. * **`sandboxing.md`**: * *Change:* New feature: `sandbox.credentials`. Protects files and environment variables. Explicit `mode: "deny"`. v2.1.187+. * *Theme:* Security, sandboxing, credential protection. * **`settings.md`**: * *Change:* Updated "Available scopes" table with more detail on Managed settings (Server-managed, registry, etc.). Updated `availableModels` description to match `model-config.md` (skills, etc.). * *Theme:* Settings structure, scope clarity. * **`skills.md`** & **`slash-commands.md`**: * *Change:* Updated documentation for frontmatter fields (likely wrapping/formatting changes based on the truncated output in the prompt, but looking closely at the provided snippets: The snippets are truncated in the prompt itself, showing the same table start. I'll assume consistency updates or formatting based on other changes). * *Theme:* Documentation consistency. * **`sub-agents.md`**: * *Change:* Added note about `availableModels` allowlist checking for subagent model parameters. If excluded, falls back to inherited model. * *Theme:* Model governance enforcement. * **`CHANGELOG.md.md`**: * *Change:* Added v2.1.190 (bug fixes). v2.1.187 mentioned in context of other changes. * *Theme:* Version history. 3. **Synthesize Key Themes:** * **Governance & Control:** Significant updates to `availableModels`, managed settings (`desktop.md`, `model-config.md`), and credential protection (`sandboxing.md`). * **Version 2.1.187 Features:** Several new features explicitly tied to this version (Plugin cleanup, MCP idle timeout, `/btw` navigation, GitHub App skip option). * **Enterprise/Cloud Context:** Clarification on how settings propagate to cloud sessions, web sessions, and desktop SSH environments. * **Refinements:** Updates to API error logging, model aliasing behavior. 4. **Determine Impact Level:** * *Reasoning:* The changes to `availableModels` are significant for enterprise admins configuring model access across all surfaces. The new `sandbox.credentials` feature impacts security compliance. Version 2.1.187 introduces noticeable UX/behavior changes (idle timeouts, cleanup UI). While not "Breaking" in the code-breaking sense, it changes operational behavior significantly for enterprises. * *Level:* Medium to High (leaning Medium for general devs, High for Admins). Let's settle on **Medium** overall as it adds features and clarifies existing behavior without removing core functionality, but specific aspects are high impact. 5. **Draft Action Items for Developers:** * Check for unused plugins using the new "Not used recently" feature. * Review `availableModels` configuration if managing enterprise settings (mind the alias behavior). * Update sandbox configs to use `sandbox.credentials` for better secret management. * Monitor MCP tool calls for new 5-minute idle timeout if using long-running remote tools. * Update local Claude Code to v2.1.187+ to access new keybinds and timeout controls. 6. **Translate to Chinese (Target Language):** * *Summary:* 此次更新主要集中在增强企业级管理能力（如模型选择限制、凭证保护）、优化资源管理（插件清理、MCP 超时）以及改进 v2.1.187 版本的交互体验。 * *Themes:* * 企业级管控与配置 (`availableModels` 扩展, Managed Settings 传输). * 安全与沙箱增强 (新增 `sandbox.credentials`). * 性能与维护 (插件未使用检测, MCP 空闲超时). * 版本特性 (v2.1.187 功能: 导航键, GitHub App 设置). * *Impact:* Medium (重要配置澄清和新安全功能，但主要是增强型更新). * *Action Items:* * Review `availableModels` settings (alias behavior change). * Implement `sandbox.credentials` for secrets. * Check MCP remote tool timeout limits. * Update CLI for new shortcuts. * Clean up plugins. 7. **Final Polish (Formatting):** Ensure Markdown structure is clean. * *Self-Correction during translation:* Make sure to highlight that `availableModels` now affects Skills and Subagents, not just the main session. Make sure to mention the specific MCP idle timeout (5 mins) as it might break long-running tasks if not configured. 8. **Final Output Generation** (Proceed to generate Chinese response).