Changes on 2026-05-30

1. **Understand the Goal:** The user wants a concise, insightful analysis of 18 documentation pages regarding Claude Code (CLI). I need to identify what changed, why it matters to developers, and present it in specific Markdown sections (Summary, Key Themes, Impact Level, Action Items). The response must be in Chinese. 2. **Analyze the diffs (Page by Page):** * **cli-reference.md:** Content reformatted/changed. Hard to see logic changes just from the snippet provided, but it looks like table row updates or description tweaks. *Correction:* Looking closer, the diff is truncated, but the line counts (+28/-28) suggest a structural or formatting update, possibly renaming or clarifying commands. Wait, looking at the content, it seems the table content just moved. Let's look for semantic changes. Actually, the specific lines shown are mostly header/structure. Let's hold judgment until looking at others, or assume it's minor refinement. *Self-correction*: It looks like the diff shown is just the start of the table. It's likely a general refresh. I shouldn't over-index on this unless I see a specific command change. * **interactive-mode.md:** Truncated diff showing keyboard shortcuts. Similar to CLI ref, likely a format refresh or minor clarification. * **mcp.md:** Added a note about `headers.Authorization` behavior. If a configured header is rejected, it fails instead of falling back to OAuth. *Impact: Fixes confusing behavior where OAuth might kick in unexpectedly.* * **memory.md:** Changed `autoMemoryDirectory` setting scope logic. Now reads from `settings.json` in various scopes, but *project/local* settings now require "workspace trust" to prevent malicious repos from redirecting memory writes. *Impact: Security hardening.* * **monitoring-usage.md:** Updates to OpenTelemetry (OTel) documentation. Mentions "Agent tool" replacing "Task tool" in some contexts, or clarifying legacy terms. Adds `WorkspaceBash` tool details. Clarifies `tool_decision` events. *Impact: Better observability and terminology updates.* * **plugins.md & plugins-reference.md:** Huge additions. Introduces "Skills-directory plugins" (`@skills-dir`). `claude plugin init` command added. Allows developing plugins directly in skills directories without marketplace installation. Covers scope (personal vs project) and trust requirements for project-scoped plugins. *Impact: Major workflow improvement for plugin developers.* * **settings.md:** Updated `agent` setting description to mention it sets the default for `claude agents`. `autoMemoryDirectory` logic updated (seen in memory.md). * **skills.md:** Added note that adding a `.claude-plugin/plugin.json` makes a skill folder load as a plugin. Live change detection note added. *Impact: Bridges skills and plugins concepts.* * **slash-commands.md:** Same content as skills.md (likely a duplication or shared section in the source, updated in both). * **sub-agents.md:** Warning update: `bypassPermissions` now also skips prompts for `.cargo`. *Impact: Security detail update.* * **terminal-config.md:** Added details about `/terminal-setup` modifying `terminal.integrated.gpuAcceleration` to `"off"` in VS Code to fix garbled text. *Impact: Bug fix documentation.* * **build-with-claude/context-windows.md:** Added note about tool selection with large input documents. * **agents-and-tools/tool-use/overview.md:** Added section "When Claude uses tools". Explains the `auto` mode boundary and how to steer it via system prompts. *Impact: UX/Prompt engineering guidance.* * **agents-and-tools/tool-use/computer-use-tool.md:** Added tip about `enable_zoom: true` with `computer_20251124`. * **agents-and-tools/tool-use/web-fetch-tool.md:** Added section "When Claude fetches". Distinguishes between specific resources (fetch) vs general knowledge (no fetch). Explains "Combined search and fetch". *Impact: Clarifies model behavior.* * **agents-and-tools/mcp-connector.md:** Added section "When Claude uses MCP tools". Distinguishes specific queries vs general knowledge. Mentions "defer_loading" for large toolsets. * **CHANGELOG.md.md:** Mention of "Auto mode" availability on Bedrock/Vertex/Foundry for Opus 4.7/4.8. *Impact: Feature availability update.* 3. **Synthesize Themes:** * **Plugin Development & Discovery:** The biggest change is the introduction of `claude plugin init` and the ability to load plugins directly from the skills directory (`@skills-dir`). This simplifies development. * **Tool Usage Clarification:** Several pages now explicitly explain *when* Claude decides to use tools (Search, Fetch, MCP, General tools) vs answering directly. This helps developers control agent behavior. * **Security & Trust:** Updates regarding `autoMemoryDirectory` and project-scoped plugins requiring workspace trust. Also updates to permission bypass warnings (`.cargo`). * **Observability:** Updates to monitoring/telemetry docs regarding "Agent" vs "Task" terminology and specific tool details. * **Terminal/GUI Fixes:** Documentation fix for VS Code GPU acceleration settings. 4. **Drafting the Response (in Chinese):** * **Summary:** Focus on the streamlined plugin dev process and clarified tool usage logic, plus security hardening. * **Key Themes:** * Plugin Development: `claude plugin init`, local loading from skills dir. * Tool Behavior: Explicit docs on "When Claude uses tools" (Search/Fetch/MCP). * Security: Workspace trust for memory/plugins, updated bypass warnings. * Observability: Updated telemetry schema (Agent vs Task). * Fixes: VS Code GPU acceleration. * **Impact Level:** **Medium**. The plugin workflow is a significant quality-of-life improvement for devs, and understanding tool behavior is crucial for agent building. Security updates are important but mostly tightening things up. * **Action Items:** * Try `claude plugin init` for new tools. * Update prompts based on "When Claude uses tools" docs. * Review memory/project settings if using `autoMemoryDirectory`. * Check MCP tool descriptions for large toolsets. * Note the VS Code terminal setup fix if experiencing garbled text. 5. **Refining the Content (Self-Correction during drafting):** * *Theme 1 (Plugins):* Make sure to distinguish between "marketplace install" and "skills-dir load". The key is "no install step" for development. * *Theme 2 (Tools):* Emphasize the distinction between "specific resource" (use tool) and "general knowledge" (answer directly). This is a common pain point. * *Theme 3 (Security):* The workspace trust requirement for project-local `autoMemoryDirectory` and plugins is a subtle but important breaking change for some workflows (requiring user approval). * *Impact:* Is it High? Maybe Medium-High for tool developers, Medium for general users. Let's stick with Medium/High. I'll say **Medium to High** because of the plugin dev workflow improvement. * *Action Items:* Keep it actionable. 6. **Final Polish (Markdown formatting):** * Use headers. * Use bullet points. * Ensure clear, professional tone. *(Self-Correction on specific term)*: The changelog mentions "Auto mode" on specific clouds. I should mention this in themes or action items as it's a new feature flag. 7. **Final Output Generation (translating thoughts to the final response structure):** * *Summary:* 本次更新主要集中在简化插件开发流程、明确 Claude 使用工具的决策逻辑，以及增强安全性和监控能力。 * *Themes:* 1. **插件开发体验革新**: 引入 `claude plugin init` 和 `@skills-dir` 本地加载机制。 2. **工具调用逻辑透明化**: 新增“何时使用工具”文档，涵盖搜索、抓取和 MCP。 3. **安全与权限加固**: 工作区信任机制扩展到 `autoMemoryDirectory` 和项目插件；更新 `bypassPermissions` 警告范围。 4. **可观测性更新**: 明确 "Agent" 与 "Task" 工具的监控术语。 5. **终端体验修复**: VS Code 终端设置文档更新（GPU 加速）。 * *Impact:* Medium to High. * *Action Items:* * Test `claude plugin init`. * Adjust system prompts based on tool usage docs. * Verify memory settings behavior. * Check VS Code terminal setup. * Review MCP tool descriptions. Looks good. Proceeding to generate output.