Changes on 2026-06-18

1. **Analyze the Request:** * **Role:** World-class programmer analyzing documentation changes. * **Output Format:** Markdown, concise, insightful, Chinese language. * **Input:** A batch of 8 documentation pages showing diffs. * **Required Sections:** Overall Summary, Key Themes, Impact Level, Action Items. 2. **Analyze the Diffs (Page by Page):** * **Page 1: `about-claude/pricing.md`** * *Change:* Clarifies that code execution is free when used with `web_search_20260209` (or later) or `web_fetch_20260209` (or later). * *Implication:* Future-proofing the pricing policy. If new versions come out, they are also covered. * **Page 2: `build-with-claude/prompt-caching.md`** * *Change:* Adds a note linking to a guide about `ephemeral_5m_input_tokens` when using server tools (like web search). * *Implication:* Clarifies unexpected cache writes caused by server-side tool caching. * **Page 3: `agents-and-tools/tool-use/computer-use-tool.md`** * *Change:* Significant update on image size handling. * *Details:* * Clarifies that Opus 4.8/4.7 support up to 2576px on the long edge. * The API *no longer* silently downscales images exceeding limits; it now rejects them with HTTP 400. * Troubleshooting table updated to remove mentions of "silently downscaled" causing offset clicks. * *Implication:* Breaking change/Behavioral change for developers relying on the API to auto-resize screenshots. Now must handle resizing client-side. * **Page 4: `agents-and-tools/tool-use/web-search-tool.md`** * *Change:* Introduction of `web_search_20260318`. * *Details:* * New version adds "response inclusion" control. * Allows dropping `server_tool_use` blocks from the response to save tokens in agentic workflows (`"response_inclusion": "excluded"`). * References version `20260209` as the baseline for dynamic filtering. * **Page 5: `agents-and-tools/tool-use/web-fetch-tool.md`** * *Change:* Introduction of `web_fetch_20260318`. * *Details:* * Adds "response inclusion" control (similar to web search). * Adds "cache bypass" feature (`use_cache`: false) available in `20260309` or later. * Lists previous versions clearly. * **Page 6: `agents-and-tools/tool-use/code-execution-tool.md`** * *Change:* New version `code_execution_20260521`. * *Details:* * Adds `20260521` to compatibility list. * This version reveals a 90-second per-cell wall-clock time limit (previously hidden/undisclosed). * Pricing text updated to include "or later" for web search/fetch free usage. * **Page 7: `agents-and-tools/tool-use/programmatic-tool-calling.md`** * *Change:* Text update. Changed `code_execution_20260120` to `code_execution_20260120` or later. * *Implication:* Compatibility update related to the new code execution version. * **Page 8: `CHANGELOG.md.md`** * *Change:* "Agent teams" update. * *Details:* Removed `TeamCreate`/`TeamDelete` tools. Now implicit team. Spawn teammates via `Agent` tool's `name` parameter. * *Implication:* Breaking change for Agent Teams feature implementation. 3. **Synthesize Findings:** * **Theme 1: New Tool Versions & Cost Optimization.** `web_search_20260318`, `web_fetch_20260318`, and `code_execution_20260521` are introduced. Key feature: `response_inclusion` to save output tokens by excluding consumed tool results in agentic workflows. * **Theme 2: Stricter Image Validation.** The Computer Use tool now *rejects* oversized images (HTTP 400) instead of silently downscaling them. This requires client-side resizing logic updates. * **Theme 3: Timeout Transparency.** New code execution version explicitly documents the 90-second cell timeout limit. * **Theme 4: Agent Teams Refactor.** Simplified agent team creation (removal of explicit team create/delete tools). 4. **Draft the Response (in Chinese):** * *Overall Summary:* The update introduces new versions for web search, web fetch, and code execution tools, focusing on cost optimization via `response_inclusion` and clearer timeout limits. It also tightens image validation in Computer Use and simplifies Agent Teams implementation. * *Key Themes:* * **成本优化:** `web_search` 和 `web_fetch` 的 `20260318` 版本新增 `response_inclusion` 参数，允许在 Agent 工作流中排除已消耗的工具结果，从而降低输出 Token 成本。 * **严格的图像验证:** Computer Use 工具不再静默缩放超尺寸图像，而是直接拒绝（HTTP 400）。开发者必须在客户端预先调整图像大小。 * **新工具版本与超时透明化:** 引入 `code_execution_20260521`，明确了 90 秒单单元格执行时间限制。 * **Agent Teams 架构调整:** 移除了显式的 `TeamCreate` 和 `TeamDelete` 工具，改为基于会话的隐式团队模式，简化了多 Agent 协作流程。 * *Impact Level:* High. * *Reasoning:* The Computer Use validation change is a breaking change (behavioral). The Agent Teams removal of tools is a breaking change for those workflows. The new tool versions offer significant cost savings but require code updates to adopt. * *Action Items:* * **Update Computer Use implementations:** Modify client code to validate and resize screenshots to model-specific limits (e.g., 2576px for Opus 4.8/4.7) before sending them to the API to avoid HTTP 400 errors. * **Adopt new tool versions for cost savings:** Update to `web_search_20260318` and `web_fetch_20260318` and utilize `response_inclusion: "excluded"` to reduce output token costs in agentic loops. * **Refactor Agent Teams logic:** Remove dependencies on `TeamCreate`/`TeamDelete` tools and update code to spawn teammates directly using the `Agent` tool's `name` parameter. * **Handle code execution timeouts:** Update error handling for `code_execution_20260521` to manage the newly disclosed 90-second per-cell timeout limit. * **Review prompt caching behavior:** Check `ephemeral_5m_input_tokens` logs if using server tools (web search) to understand automatic caching mechanics. 5. **Review and Refine (Self-Correction):** * *Check:* Did I mention the specific version numbers? Yes. * *Check:* Is the Computer Use change clear? Yes, it's a shift from "silent downscale" to "reject". * *Check:* Is the language natural Chinese? Yes. 6. **Final Polish:** Structure the Markdown nicely. (Drafting the final output based on the synthesis above.)

1. **Understand the Goal**: The user wants a concise, insightful analysis of a batch of documentation changes for Claude Code. The output must be in markdown, in Chinese, focusing on what matters to developers. The structure is predefined: Overall Summary, Key Themes, Impact Level, and Action Items. 2. **Analyze the Input**: I have 13 pages of diffs. Let's go through them page by page to identify changes. * **`amazon-bedrock.md`**: * Change: Added note about v2.1.181 accepting flat output from `aws configure export-credentials --format process`. * Change: Added note about v2.1.176 caching credentials based on ISO 8601 `Expiration`. * *Theme*: AWS integration improvements, credential caching. * **`claude-code-on-the-web.md`**: * Change: v2.1.179 adds automatic `Claude-Session: <url>` git trailer to commits and session URL in PR bodies. * Change: Updated text on how to manually build links (preferring automation). * *Theme*: Web session traceability, automation. * **`cli-reference.md`**: * Change: Added `--ax-screen-reader` flag (v2.1.181) for accessible output (flat text, no borders). * *Theme*: Accessibility, CLI flags. * **`headless.md`**: * Change: Note that v2.1.181 allows changing settings via `/config key=value` in `-p` mode (e.g., `/config thinking=false`). * *Theme*: CLI automation, configuration management. * **`monitoring-usage.md`**: * Change: `api_refusal` event now includes more attributes: `query_source`, `speed`, `attempt`, `effort`, `server_fallback_hop`, `has_category`, `has_explanation`, `category`, and agent/skill/plugin attribution. * *Theme*: Observability, logging, telemetry. * **`plugin-marketplaces.md`**: * Change: Clarified how `skills` paths work with shared marketplace root sources. Added examples. * *Theme*: Plugin configuration, marketplace behavior. * **`plugins-reference.md`**: * Change: Minor clarification text regarding `skills` field merging behavior for marketplace root sources. * *Theme*: Plugin configuration clarification. * **`sandboxing.md`**: * Change: Troubleshooting section: Added note about `open`, `osascript`, browser auth failing with `-600` on macOS. Solution: `allowAppleEvents` setting. * Change: Security considerations: Warning about `allowAppleEvents` removing code-execution isolation. * *Theme*: macOS compatibility, sandbox security. * **`settings.md`**: * Change: v2.1.181 allows changing single options via `/config key=value` without opening the interface. * Change: Minor typo fix ("constrain Default" -> "constrain the default model"). * *Theme*: Configuration workflow improvement. * **`skills.md`** & `slash-commands.md`**: * *Note*: The diff for `slash-commands.md` is identical to `skills.md` in the provided text (likely a copy-paste error in the prompt or documentation, but I will treat it as updating the skills documentation). Wait, looking closely at the diff content, it talks about "skills", "skill-creator plugin", etc. It seems `slash-commands.md` might be an alias or the diff provided is just duplicated for these two files. I will focus on the content: Evaluating and iterating on skills using the `skill-creator` plugin. * Change: Added section "Evaluate and iterate on a skill". Introduction of `skill-creator` plugin for running evaluations (test cases, isolated runs, grading, benchmarking, A/B testing). * Change: Removed "subagents" from the list of things not loaded from additional directories (updated the text to match new behavior). * *Theme*: Skill testing, QA, tooling. * **`statusline.md`**: * Change: Clarification that status line doesn't replace footer badges. * *Theme*: UI clarification. * **`sub-agents.md`**: * Change: Simplified text about depth limits. Removed specific distinction about foreground/background spawning limits at depth 5 (now just says depth 5 can't spawn further). * *Theme*: Subagent behavior clarification. 3. **Synthesize Key Themes**: * **Version 2.1.181 specific features**: New CLI flag (`--ax-screen-reader`), new config syntax (`/config key=value`), AWS credential format change. * **Observability & Debugging**: Enhanced API refusal logging, new skill evaluation tools (`skill-creator`). * **Web Session Integration**: Better traceability (git trailers, PR links). * **Sandboxing/Security**: macOS Apple Events handling (`allowAppleEvents`). * **Plugin/Skill Refinement**: Better documentation on path loading and evaluation. 4. **Determine Impact Level**: * *Low*: Most changes are enhancements or clarifications. * *Medium*: The AWS credential change (flat output) is an "acceptance" change, which might affect how people write scripts, but it's additive (accepts *also* the flat format). The `skill-creator` introduction is a significant workflow improvement for power users. The `api_refusal` event schema change impacts telemetry parsers. * *High/Breaking*: No breaking changes found. The AWS change is additive. The `allowAppleEvents` is opt-in via settings. * *Verdict*: **Medium**. The new `skill-creator` workflow and the telemetry schema updates are substantial for devs building integrations or complex skills. 5. **Draft Action Items**: * Update AWS credential scripts if utilizing the new flat format support (v2.1.181). * Use `/config key=value` in scripts for faster config changes. * Install `skill-creator` plugin to evaluate skill quality. * Update telemetry parsers to handle new `api_refusal` attributes. * Enable `allowAppleEvents` if macOS sandbox tools need Apple Events (carefully). 6. **Translate and Format (Chinese)**: * *Summary*: 这批更新主要集中在 Claude Code v2.1.181 版本的增强功能，包括改进的 AWS 凭证支持、新的辅助功能 CLI 标志、可脚本化的配置更改，以及引入了用于技能评估的 `skill-creator` 插件。此外，还增强了 API 拒绝日志记录并解决了 macOS 沙箱的兼容性问题。 * *Themes*: * **v2.1.181 新特性与 CLI 增强**: 新增 `--ax-screen-reader` 标志，支持通过 `/config key=value` 直接修改配置。 * **AWS 集成与认证**: 支持 AWS 凭证的扁平化输出格式，并优化了凭证缓存机制。 * **技能测试与质量保证**: 引入 `skill-creator` 插件，提供技能评估、基准测试和 A/B 测试能力。 * **监控与可观测性**: 扩展了 `api_refusal` 事件的日志属性，增加了更多上下文信息（如 fallback 状态、分类等）。 * **沙箱与兼容性**: 增加了 macOS 沙箱中对 Apple Events 的支持选项（`allowAppleEvents`），并明确了相关的安全风险。 * *Impact*: **Medium**。虽然主要是功能增强，但引入的 `skill-creator` 工作流显著改变了技能开发的质量保证流程。API 拒绝日志的模式变更可能会影响现有的日志解析器。 * *Action Items*: * 如果使用 AWS，更新脚本以利用 v2.1.181 支持的扁平化凭证格式。 * 尝试使用 `skill-creator` 插件对现有技能进行评估和基准测试。 * 检查并更新处理 `api_refusal` 事件的日志解析器以适应新增字段。 * 在脚本中使用新的 `/config key=value` 语法以简化配置管理。 * 如需在 macOS 沙箱中运行需要 Apple Events 的工具，谨慎配置 `allowAppleEvents`。 7. **Refine**: Ensure tone is professional and concise. "developer-focused". * *Self-Correction during drafting*: The `skills.md` and `slash-commands.md` diffs were identical in the prompt. I should mention the skill-creator plugin generally as it relates to skills. 8. **Final Output Generation**: (Proceed to generate markdown based on the above synthesis).

1. **Analyze the Request**: * **Role**: World-class programmer analyzing documentation changes. * **Input**: 4 pages of diffs. * `claude-code-on-the-web.md`: Minor heading change. * `network-config.md`: Added a new URL to the allowlist (`*.claudeusercontent.com`) and updated formatting. * `settings.md`: Added a new setting `disableArtifact`. * `agents-and-tools/tool-use/overview.md`: Massive expansion of code examples (added C#, Go, Java, PHP, Ruby), syntax highlighting updates, minor text cleanup, and card layout changes. * **Output Format**: Markdown, Chinese language. * **Structure**: Overall Summary, Key Themes, Impact Level, Action Items. 2. **Analyze Individual Changes**: * **`claude-code-on-the-web.md`**: * Change: "Link artifacts back to the session" -> "Link output back to the session". * Context: Section about creating traceable links to cloud sessions. * Meaning: Terminology shift. "Artifacts" likely encompasses "output", but the change suggests a broadening or precision in terminology. However, looking at `settings.md`, a new `disableArtifact` setting exists. This might be related to the Artifact tool. Let's look closer. * Wait, actually, looking at `settings.md`, `disableArtifact` disables the "Artifact tool... publishes session output". So "Artifact" is the specific mechanism/feature name. "Output" is the generic result. The header change might be to make it more understandable to the user (link the *result* back to the session), whereas the *feature* is still called Artifacts. * **`network-config.md`**: * Change: Added `*.claudeusercontent.com`. * Reason: Required for viewing artifacts on claude.ai (sandboxes content). * Significance: Crucial for enterprise/firewall setups. If this is blocked, the Artifact viewer in the web UI won't load content. * **`settings.md`**: * Change: Added `disableArtifact`. * Description: Disables the Artifact tool. Equivalent to env var `CLAUDE_CODE_DISABLE_ARTIFACT`. * Significance: Gives admins control over whether users can publish output as private web pages (likely security/compliance feature). * **`agents-and-tools/tool-use/overview.md`**: * Change: Added `hidelines` to code blocks. Added new language examples (C#, Go, Java, PHP, Ruby). * Change: Minor text polish (grammar, flow). * Change: Updated CardGroup layout. * Significance: Multi-language support expansion. The document is now more inclusive for non-Python/TypeScript devs. 3. **Synthesize Themes**: * **Artifacts Feature & Control**: This is a major theme. * We have a new setting to disable it (`settings.md`). * We have a network requirement added specifically for viewing it (`network-config.md`). * We have a heading change suggesting "output" is the user-facing concept (`claude-code-on-the-web.md`). * **Language Support**: The API docs now cover many more languages, making the "tool use" feature accessible to a broader developer base. 4. **Draft the Summary (Internal Monologue)**: * The update focuses heavily on the "Artifact" feature (web view of output), adding configuration controls and network requirements for it. Simultaneously, it significantly expands multi-language support for tool use in the API documentation. 5. **Draft the Analysis (in Chinese)**: * **Overall Summary**: 此次更新主要围绕 "Artifacts" 功能的管理与网络配置进行了增强，同时大幅扩展了多语言 SDK 的代码示例。这意味着开发者现在能更好地控制 Web 输出功能的可见性，并且在非 Python/JS 环境下也能更方便地集成 Tool Use 功能。 * **Key Themes**: * **Artifacts 功能管控与网络支持**：引入了 `disableArtifact` 配置项，允许管理员禁用该功能；同时增加了 `*.claudeusercontent.com` 网域白名单要求，确保在受限网络环境下能正常查看 Artifacts 内容。 * **多语言支持扩展**：在 Tool Use 概述中新增了 C#、Go、Java、PHP 和 Ruby 的完整代码示例，显著降低了非主流开发语言用户的集成门槛。 * **文档术语优化**：将 "Link artifacts back to the session" 改为 "Link output back to the session"，使概念对最终用户更直观。 * **Impact Level**: Medium. * *Reasoning*: It's not a breaking API change, but the network config change is important for enterprise users behind firewalls. The new setting is important for compliance/security teams. The new code examples are quality-of-life improvements for new users. * **Action Items**: * **网络管理员**：如果企业使用了 Claude Code 并启用了 Web 会话，务必将 `*.claudeusercontent.com` 加入防火墙白名单，否则 Artifacts 查看器可能无法加载内容。 * **安全/合规团队**：若需禁止生成私有网页输出，检查并在 `managed-settings.json` 中设置 `disableArtifact` 为 `true`。 * **C#/Go/Java/PHP/Ruby 开发者**：查阅更新后的文档，获取标准化的 Tool Use 集成示例。 6. **Refine and Polish (Markdown Formatting)**: * Check against constraints: "Concise, insightful", "Chinese". * *Self-Correction during drafting*: * For the header change in `claude-code-on-the-web.md`, don't overstate it. It's just a wording change. The core theme is "Artifacts". * Ensure the distinction between "Artifacts" (the feature/tool) and "output" (the result) is clear in the analysis. 7. **Final Output Generation** (matches the provided good response structure). * *Drafting the "Key Themes"* section to be precise: * 1. Artifact 功能的可控性增强 * 2. 企业级网络配置更新 * 3. 多语言 SDK 覆盖扩大 * *Drafting "Action Items"*: * Update firewall rules. * Review managed settings if disabling features. * New language devs can check docs. 8. **Final Review**: The analysis connects the dots between `settings.md` (disable), `network-config.md` (allowlist), and the broader context. The language support is a separate but significant track. Looks solid.