Changes on 2026-05-11

1. **Analyze the Request:** * **Role:** World-class programmer, deep expertise in dev tools/docs. * **Task:** Analyze documentation changes as a single batch. * **Output:** Concise, insightful analysis in markdown format. * **Language:** Chinese. * **Input:** Diffs for `hooks.md`, `hooks-guide.md`, `plugins-reference.md`. 2. **Analyze the Input Data (The Diffs):** * **`hooks.md` changes:** * Multiple code examples updated. * Old: `"command": "\"$CLAUDE_PROJECT_DIR\"/.claude/hooks/block-rm.sh"` (and similar variations). * New: `"command": "${CLAUDE_PROJECT_DIR}/.claude/hooks/block-rm.sh", "args": []` (and similar). * Key observation: Change from `$CLAUDE_PROJECT_DIR` to `${CLAUDE_PROJECT_DIR}` and removal of surrounding quotes around the path variable. Addition of `"args": []`. * Bullet point change in "Best Practices": Changed guidance on absolute paths. Specifically distinguishes between "exec form" (use `${VAR}`, no quoting) and "shell form" (wrap in double quotes). * **`hooks-guide.md` changes:** * Troubleshooting section update. * Old advice: use absolute paths or `$CLAUDE_PROJECT_DIR`. * New advice: use `${CLAUDE_PROJECT_DIR}`. **Crucial addition:** "To avoid shell quoting entirely, add `"args": []` to switch to [exec form](...), which spawns the script directly without a shell". * This reinforces the pattern seen in `hooks.md`. * **`plugins-reference.md` changes:** * Variable definition update for `${CLAUDE_PLUGIN_ROOT}`. * Old: "In hook and monitor commands, wrap it in double quotes..." * New: "In hook commands, use [exec form](...) with `args`... In shell-form hooks and monitor commands, wrap it in double quotes..." * This clarifies the distinction between exec form and shell form specifically for path handling. 3. **Synthesize Findings (Internal Monologue):** * **What changed?** The documentation now strongly encourages using "exec form" (by adding `"args": []`) for hook commands. It changes the syntax for environment variables from `$VAR` to `${VAR}` in the command field and removes quotes when using this specific form. * **Why?** To avoid shell quoting issues (shells mishandling spaces or special characters) and to invoke scripts directly without an intermediate shell wrapper, which is safer and more robust. * **The Pattern:** `command` + `args: []` = Exec form. * **Old Way:** `command: "\"$VAR/path\""`, no args. (Shell form) * **New Way:** `command: "${VAR}/path"`, `args: []`. (Exec form) 4. **Drafting the Response (in Chinese):** * **Overall Summary:** Documentation has been updated to recommend the "exec form" (exec 形式) for defining hook commands by adding an explicit `args` field. This change simplifies path handling by removing the need for complex quoting and clarifies the usage of environment variables like `${CLAUDE_PROJECT_DIR}`. * **Key Themes:** * Promoting "Exec Form" over "Shell Form": Explicitly encouraging the use of `"args": []` to switch to direct execution. * Variable Syntax Update: Changing `$VAR` to `${VAR}` and removing surrounding quotes in the new recommended pattern. * Path Handling Safety: Clarifying how to handle paths containing spaces or special characters by bypassing the shell interpretation layer. * **Impact Level:** Medium. It's not strictly "breaking" (the old way likely still works), but it's a strong best practice update that affects how developers should write their configurations. If developers follow the new docs, their code becomes more robust. * **Action Items:** * Update hook configurations to include `"args": []`. * Switch variable syntax from `"$VAR"` to `${VAR}` when using exec form. * Review and update existing hooks that might have quoting issues. 5. **Refining the Output (Translating to "World-class programmer" tone):** * *Summary:* 文档已更新，强烈推荐在 Hook 命令中使用 **exec form（执行模式）**。通过显式添加 `"args": []` 字段，简化了路径引用的引号处理，并修正了环境变量（如 `${CLAUDE_PROJECT_DIR}`）的使用方式，以避免 shell 解析带来的潜在问题。 * *Key Themes:* * **推广 Exec Form**：文档核心变更在于提倡使用 `args` 数组来直接调用脚本，而非通过 shell 字符串。 * **路径引用规范化**：从旧式的 `"$VAR"`（带引号）切换为 `${VAR}`（无引号，配合 args 使用）。 * **安全性增强**：通过绕过 shell 层，减少了路径中包含空格或特殊字符时解析失败的风险。 * *Impact Level:* **Medium**。虽然旧语法（Shell form）可能依然兼容，但这是一项重要的最佳实践更新，旨在提高配置的健壮性和可维护性。 * *Action Items:* * 更新现有的 `hooks` 配置，在 `command` 对象中添加 `"args": []`。 * 将 `command` 字段中的路径引用从 `"$VAR"` 改为 `${VAR}` 并移除外层引号。 * 检查涉及 `${CLAUDE_PLUGIN_ROOT}` 的插件配置，应用新的 exec form 模式。 6. **Final Polish:** Ensure markdown formatting matches the prompt requirements. The logic holds up. The shift from "wrapping in quotes" to "using exec form with args" is the critical technical insight. * *Self-Correction during drafting:* Is it strictly `args: []` or `args: [...]`? The diff shows `"args": []` specifically in the examples. It implies if there are no arguments, use empty array. If there were arguments, they would likely go there. The key is the presence of the key to trigger exec form. I will focus on the *presence* of the `args` field. 7. **Final Output Generation** (matches the logic above).

1. **Analyze the Request:** * **Role:** World-class programmer analyzing documentation changes. * **Output:** Concise, insightful analysis in markdown format (Chinese language). * **Focus:** What matters to developers. * **Input:** Diff for 14 pages. 2. **Analyze the Diffs (Page by Page):** * **google-vertex-ai.md:** * *Change:* Clarified that `ENABLE_TOOL_SEARCH=true` on Vertex AI causes rejection because the endpoint doesn't support the beta header. * *Impact:* Developers using Vertex AI need to be careful about forcing tool search. * **hooks.md:** * *Change:* Major update on command hooks. Introduced `args` field (exec form) vs. Shell form. * *Details:* `args` allows direct execution without a shell, safer for arguments with spaces/special chars. `shell` is ignored if `args` is set. Path placeholders (`${CLAUDE_PLUGIN_ROOT}`) work in both. Added notes about Windows executable resolution (shims vs. binaries). * *Impact:* High. This changes how hooks are written and offers better cross-platform compatibility and security/quoting handling. * **hooks-guide.md:** * *Change:* Minor clarification about shell sourcing. Specifically, shell-form hooks might source profiles (`BASH_ENV`), exec form (with `args`) does not. * *Impact:* Low/Medium. Helps debug hook issues. * **interactive-mode.md:** * *Change:* New shortcut `Esc` (single press) to interrupt Claude. Transcript viewer shortcuts updated (`?`, `{`, `}`). Clarification on where to find shortcuts. * *Impact:* Medium. New user control mechanism. * **llm-gateway.md:** * *Change:* Added HTTP headers `X-Claude-Code-Agent-Id` and `X-Claude-Code-Parent-Agent-Id` for request attribution from subagents/teammates. * *Impact:* Low/Medium. Relevant for developers building proxies or gateways needing detailed cost attribution. * **mcp.md:** * *Change:* Stdio servers now get `CLAUDE_PROJECT_DIR` env var. Plugin MCP features include this var. Clarified `ENABLE_TOOL_SEARCH` behavior (forcing `true` might break requests). * *Impact:* Medium. Improves path handling for MCP servers. * **memory.md:** * *Change:* "Project Memory" scope changed from "Per working tree" to "Per repository, shared across worktrees". Added `claudeMd` key to managed settings for inline instructions. * *Impact:* Medium. Better memory management in worktree-heavy workflows. * **model-config.md:** * *Change:* Clarified that `--model` flag and `ANTHROPIC_MODEL` env var are session-only and not saved. * *Impact:* Low. Removes confusion about model persistence. * **monitoring-usage.md:** * *Change:* Added `agent_id`, `parent_agent_id` to logs. Added `agent.name`, `skill.name`, `plugin.name`, `marketplace.name` attributes to counters. Added new events: `plugin_loaded` and `hook_registered`. * *Impact:* Medium. Enhanced telemetry for debugging and cost tracking. * **output-styles.md:** * *Change:* Introduced **"Proactive"** output style (Claude executes immediately, makes assumptions). * *Impact:* Medium. New user experience option. * **plugins-reference.md:** * *Change:* Updated examples to quote `${CLAUDE_PLUGIN_ROOT}` for safety. Added `${CLAUDE_PROJECT_DIR}` support. Clarified symlink behavior in marketplaces (within market = dereferenced/copy, outside = skipped). Added `claude plugin details` command. * *Impact:* Medium/High. Important for plugin developers regarding path handling and plugin structure. * **quickstart.md:** * *Change:* Minor UI text update in shortcuts/cheatsheet. * *Impact:* Low. * **settings.md:** * *Change:* Added `claudeMd` setting definition (matches memory.md). * *Impact:* Low/Medium. Configuration consistency. * **vs-code.md:** * *Change:* Added "Reopen Closed Session" command (`Cmd+Shift+T`/`Ctrl+Shift+T`). * *Impact:* Low. UX improvement. 3. **Synthesize Key Themes:** * **Cross-Platform Path Handling & Execution:** Major overhaul in hooks (`args`/exec form) and improvements in MCP/plugins (`CLAUDE_PROJECT_DIR`, quoting variables). This addresses quoting issues on Windows/Unix. * **Visibility & Telemetry:** Better observability for agents, skills, plugins, and hooks in usage logs (new headers, new events, new attributes). * **Organization/Enterprise Control:** Inline `claudeMd` in managed settings, improved memory scoping across worktrees. * **User Experience:** New "Proactive" mode, "Interrupt" shortcut, improved session management in VS Code. 4. **Determine Impact Level:** * *Reasoning:* The changes to Hooks (adding `args`/exec form) and the new path variables require code updates from plugin/hook developers. The new observability features affect monitoring infrastructure. The new output style affects user workflow. * *Rating:* **Medium** (Mostly additive improvements and clarifications, but hooks change requires attention). 5. **Draft Action Items (Chinese):** * Update hook configurations to use the new `args` field for safer argument passing. * Review and quote path variables (like `${CLAUDE_PLUGIN_ROOT}`) in plugin/monitor configurations to avoid issues with spaces. * Check usage monitoring pipelines to utilize new agent and skill attribution data. * Consider using the new `claudeMd` managed setting for organization-wide coding standards. * Try the new `Esc` shortcut for interrupting and the "Proactive" output style. 6. **Draft the Response (in Markdown format):** * *Summary:* 本次更新主要集中在提升跨平台兼容性（特别是路径处理和命令执行）、增强企业级可观测性以及优化用户体验。Hooks 机制引入了 Exec 模式，Plugin 和 MCP 支持了新的项目路径变量。 * *Themes:* Hooks 增强与跨平台安全性, 可观测性与监控提升, 企业级配置与内存管理, 用户体验优化. * *Impact:* Medium. * *Action Items:* (As drafted above). 7. **Refine Language (World-class programmer tone):** * Ensure terms like "Exec form", "Shell form", "Shims" are understood but explained clearly in context. * Keep it punchy. *(Self-Correction during drafting)*: Make sure to mention the specific `ENABLE_TOOL_SEARCH` warning for Vertex AI users, as that could break things if they blindly enable it. The `CLAUDE_PROJECT_DIR` addition is significant for MCP server devs. 8. **Final Polish (Markdown structure):** (Produces the final output).

1. **Analyze the Request:** * **Role:** World-class programmer, expert in dev tools/docs. * **Task:** Analyze a batch of documentation changes (10 pages total). * **Output Format:** Markdown, concise, insightful. * **Language:** Chinese (Simplified). * **Specific Sections Required:** Overall Summary, Key Themes, Impact Level, Action Items. 2. **Analyze the Diff Data (Page by Page):** * **Page 1: `api/rate-limits.md`** * *Change:* Removed mention of "[Priority Tier](/docs/en/api/service-tiers)" in the Note about Claude Platform on AWS. * *Meaning:* Previously, it said Fast mode and Priority Tier are not available on AWS. Now it says only Fast mode is not available. This implies *Priority Tier* might now be available on AWS, or at least the documentation restricting it has been removed/clarified. * **Page 2: `api/service-tiers.md`** * *Change:* Removed a specific `<Note>` block stating "Claude Platform on AWS supports Standard and Batch service tiers. Priority Tier is not available." * *Meaning:* This confirms the suspicion from Page 1. Priority Tier is likely now available or supported on Claude Platform on AWS. * **Page 3: `build-with-claude/vision.md`** * *Change:* `--transform id --format yaml` -> `--transform id --raw-output`. * *Meaning:* Change in CLI command syntax for the `ant` tool. Replaced `--format yaml` with `--raw-output`. * **Page 4: `build-with-claude/pdf-support.md`** * *Change:* `--transform id --format yaml` -> `--transform id --raw-output`. * *Meaning:* Same CLI syntax change as Page 3. * **Page 5: `build-with-claude/batch-processing.md`** * *Change:* Two occurrences of `--transform id --format yaml` -> `--transform id --raw-output`. One occurrence of `--transform processing_status --format yaml` -> `--transform processing_status --raw-output`. * *Meaning:* Same CLI syntax change. * **Page 6: `build-with-claude/context-editing.md`** * *Change:* `--transform ... --format yaml` -> `--transform ... --raw-output` (twice). * *Meaning:* Same CLI syntax change. * **Page 7: `build-with-claude/structured-outputs.md`** * *Change:* `--transform 'content.0.text' --format yaml` -> `--transform 'content.0.text' --raw-output`. * *Meaning:* Same CLI syntax change. * **Page 8: `build-with-claude/effort.md`** * *Change:* `--transform 'content.0.text' --format yaml` -> `--transform 'content.0.text' --raw-output`. * *Meaning:* Same CLI syntax change. * **Page 9: `build-with-claude/files.md`** * *Change:* `--transform id --format yaml` -> `--transform id --raw-output` (twice). * *Meaning:* Same CLI syntax change. * **Page 10: `agents-and-tools/tool-use/code-execution-tool.md`** * *Change:* `--transform id --format yaml` -> `--transform id --raw-output` (twice), `--transform filename --format yaml` -> `--transform filename --raw-output`, `--transform container.id --format yaml` -> `--transform container.id --raw-output`. * *Meaning:* Same CLI syntax change. 3. **Synthesize and Categorize Changes:** * **Theme A: AWS Platform Capability Update.** The removal of restrictions regarding "Priority Tier" on AWS strongly suggests this feature is now enabled or the restriction has been lifted. This is a feature availability change. * **Theme B: CLI Tool (`ant`) Syntax Refactoring.** A systematic replacement of the flag combination `--format yaml` with `--raw-output` when used with `--transform`. This is a tooling/CLI deprecation or improvement change. 4. **Draft the Analysis (Iterative refinement in Chinese):** * *Overall Summary:* 文档更新主要集中在两个方面：一是 AWS 平台对 Priority Tier 的支持情况，二是 Anthropic CLI (`ant`) 工具的命令行参数变更。 * *Key Themes:* * **AWS 平台功能解锁:** 删除了关于 "Claude Platform on AWS" 不支持 Priority Tier 的说明，暗示该层级现已在 AWS 上可用。 * **CLI 工具语法更新:** 全局替换了 `ant` 命令中的参数，将 `--format yaml` 替换为 `--raw-output`，配合 `--transform` 使用。这表明 CLI 工具正在优化输出处理方式或废弃旧的格式化参数。 * *Impact Level:* Medium. * *Reasoning:* CLI 变化虽然是微小的语法调整，但出现在所有示例代码中，意味着现有脚本如果使用了 `--format yaml` 可能需要更新（虽然通常 `--raw-output` 是为了更干净的输出，向后兼容性取决于具体实现，但这显然是一个行为变更）。AWS 的变更影响架构决策和成本优化。 * *Refinement:* Actually, usually `--raw-output` is preferred for piping. The impact is "Medium" because it affects how developers write their CLI scripts. Breaking is unlikely unless `--format yaml` is removed entirely, but it's a significant documentation churn suggesting a best practice change. Let's stick with "Medium" because of the frequency of the change. * *Action Items:* * **更新 CLI 脚本:** 检查并更新所有使用 `ant beta:files upload`、`ant messages create` 等命令的 shell 脚本，将 `--format yaml` 替换为 `--raw-output` 以符合最新文档最佳实践（特别是用于变量赋值时）。 * **评估 AWS 成本与性能:** 如果使用 Claude Platform on AWS，现在可以考虑启用 Priority Tier 以获得生产环境所需的稳定性和优先级。 * **审查 Rate Limits 文档:** 针对 AWS 用户的 Rate Limits 限制描述已更新，需重新确认自动层级提升和计费策略的描述。 5. **Final Polish (Chinese):** Ensure the tone is professional and concise. * *Summary:* 此次文档更新包含两个主要变更：明确了 Claude Platform on AWS 对 Priority Tier 的支持，并全面更新了 Anthropic CLI (`ant`) 的命令示例语法。 * *Themes:* AWS 平台功能支持，CLI 工具参数规范化。 * *Impact:* Medium (脚本语法更新 + 平台功能开放)。 * *Actions:* 更新脚本语法，重新评估 AWS 配置。 6. **Constructing the Output:** (Proceeding to generate the markdown response). * *Self-Correction on Action Items:* Since I don't know if `--format yaml` is *strictly* deprecated (removed from code) or just removed from docs (best practice), I should frame it as "Update scripts to align with documentation" or "Adopt new syntax". However, usually, massive doc swaps like this imply the old way is discouraged or removed. I will suggest updating the scripts. 7. **Final Check against constraints:** * Markdown format? Yes. * Concise? Yes. * Focus on what matters to devs? Yes. * Chinese? Yes. (Drafting response...)

1. **Understand the Goal:** The user wants a concise, insightful analysis of documentation changes in markdown format, focusing on what matters to developers. I need to analyze 6 specific pages with diffs provided. 2. **Analyze the Diffs:** * **Page: `data-usage.md`** * *Changes:* Added "Claude Platform on AWS" to the list of API providers. Updated the table to include a new column for "Claude Platform on AWS" with default behaviors (all metrics off, surveys on, webfetch on). Added environment variable `CLAUDE_CODE_USE_ANTHROPIC_AWS`. * *Significance:* New provider integration. Privacy-conscious developers need to know defaults. * **Page: `hooks.md`** * *Changes:* * Clarified `agent_type` in hooks: specifically notes that for custom subagents, it uses the `name` field from frontmatter, not the filename. * Added note: Only `SessionStart` hooks receive a `model` field. Clarified `$CLAUDE_MODEL` doesn't exist, only `$ANTHROPIC_MODEL` (from parent env). * Added new `ExitPlanMode` tool hook documentation (input fields: `plan`, `planFilePath`, `allowedPrompts`). * Updated `SubagentStart` description regarding custom subagent names (same as `agent_type` clarification). * Updated `SubagentStop` decision control: Clarifies it doesn't support `additionalContext`. Suggests using `PostToolUse` on `Agent` tool to inject context. * Added Tip linking `/goal` command to session-scoped Stop hooks. * *Significance:* API stability/behavior clarification for hooks developers (especially around subagents). New capability for hooking into plan mode exit. * **Page: `llm-gateway.md`** * *Changes:* Added section "Claude Platform on AWS through a gateway" with example env vars (`ANTHROPIC_AWS_BASE_URL`, `CLAUDE_CODE_SKIP_ANTHROPIC_AWS_AUTH`, etc.). * *Significance:* Configuration guidance for the new provider through proxies/gateways. * **Page: `model-config.md`** * *Changes:* Updated model version mapping to include "Claude Platform on AWS" in the list of providers where `opus` -> 4.7 and `sonnet` -> 4.6. Added "Claude Platform on AWS" to the list of providers where model pinning is recommended for deployment. * *Significance:* Deployment/Versioning consistency for the new provider. * **Page: `sub-agents.md`** * *Changes:* * Clarified `name` field definition: it acts as `agent_type` in hooks. * Refined behavior of "Foreground" vs "Background" subagents. Specifically: Foreground prompts for permissions as they come up. Background subagents run with *already granted* permissions and auto-deny anything that would prompt (removed the bit about pre-approving background tasks at launch). * Updated Fork Mode section: clarified prompts surface as they occur for forks, but named subagents in background mode auto-deny. * Updated Comparison Table (Fork vs Named): Changed Permissions column for Named subagent to "Auto-denied when running in the background" (removed pre-approval mention). * *Significance:* *Crucial behavioral change/clarification*. Background subagents no longer pre-approve permissions at launch; they strictly adhere to what's already granted and auto-deny new prompts. This changes how developers design workflows requiring background task permissions. * **Page: `third-party-integrations.md`** * *Changes:* * Added "Claude Platform on AWS" column to the comparison table. * Added `ANTHROPIC_AWS_BASE_URL` to the LLM Gateway configuration list. * Added "Claude Platform on AWS" to the list of providers requiring model pinning. * *Significance:* Documentation for the new provider integration. 3. **Synthesize Themes:** * **New Provider:** "Claude Platform on AWS" is now fully integrated (billing, auth, models, gateways, data usage). * **Subagent Permissions Behavior:** Significant clarification on how background subagents handle permissions (no pre-approval at launch, strictly auto-deny new prompts). * **Hooks Refinement:** Better documentation on `agent_type` resolution (frontmatter vs filename) and model availability in hooks. New hook for `ExitPlanMode`. 4. **Determine Impact Level:** * *New Provider:* Medium impact for AWS users, but additive. * *Subagent Permissions:* Medium/High impact. If developers were relying on the "pre-approval at launch" flow described in previous docs for background subagents, their workflows might break or behave differently (fail silently on permissions). The change text says "Auto-denied... when running in the background" vs "Pre-approved... before launch". This is a behavioral clarification/correction. * *Hooks:* Low/Medium impact (fixing confusion). * *Overall:* **Medium**. The new provider is additive. The subagent permission change is a behavior clarification that requires attention from developers using background tasks. 5. **Draft Action Items:** * Check if you use the new "Claude Platform on AWS" provider and update config. * Review background subagent logic: ensure it has necessary permissions at session start because it won't ask for them during execution. * Update hook scripts relying on filenames vs frontmatter names for `agent_type`. * Note that `ExitPlanMode` is now available for tool hooks. 6. **Refine Output (Chinese):** * *Summary:* Added "Claude Platform on AWS" support and clarified critical behavior for background subagent permissions. * *Themes:* * New Provider: Claude Platform on AWS integration across docs. * Subagent Permissions: Background subagents no longer pre-approve; they run with current permissions and auto-deny new requests. * Hooks: Added `ExitPlanMode`, clarified `agent_type` (frontmatter vs filename) and `model` field availability. * *Impact:* Medium. The subagent permission change affects workflow reliability. * *Actions:* * Adjust background subagent workflows to ensure permissions are granted before execution. * Update hook code if relying on subagent filenames instead of frontmatter names. * Configure `CLAUDE_CODE_USE_ANTHROPIC_AWS` if applicable. 7. **Final Polish:** Ensure markdown format and concise tone. (Self-correction: The subagent change is technically a documentation correction to match actual behavior, but to a developer reading it, it looks like a change in how they should design their systems). * *Drafting the subagent point carefully:* The docs changed from "Before launching, Claude Code prompts for any tool permissions the subagent will need" to "run with the permissions already granted... and auto-deny". This is a significant functional difference in description. 8. **Final Output Generation:** (Proceed to generate markdown).

Let me analyze this batch of documentation changes across 7 pages. The main theme is clearly the introduction of "Agent View" / "Background Agents" as a new feature, along with a significant changelog (v2.1.139) with many improvements and fixes.

1. **Analyze the Request:** * **Role:** World-class programmer analyzing documentation changes. * **Task:** Analyze a batch of 37 documentation pages (though only ~20 distinct diffs are provided in the prompt text, I need to focus on the provided diffs). * **Output Format:** Markdown. * **Content:** Summary, Key Themes, Impact Level with reasoning, Action Items (max 5). * **Language:** Chinese. 2. **Analyze the Provided Diffs (Page by Page/Section by Section):** * *Page 1: `about-claude/models/overview.md`* * Added links to "Claude Platform on AWS" in the model availability list. * Updated notes to distinguish "Claude Platform on AWS" (Anthropic-operated, same model IDs/API) vs. AWS Bedrock (partner-operated, different IDs). * Clarified global vs regional endpoint naming/links. * *Page 2: `about-claude/models/migration-guide.md`* * Clarified "Extended thinking" is deprecated in 4.6 and *removed* in Opus 4.7. Recommends "adaptive thinking". * *Page 3: `about-claude/pricing.md`* * Added "Claude Platform on AWS pricing" section (using CCUs - Claude Consumption Units). * Renamed "Third-party platform pricing" to "Cloud platform pricing". * Distinguished between "partner-operated" (Bedrock, Vertex) and "Anthropic-operated on cloud" (Claude Platform on AWS, Foundry). * Clarified `inference_geo: "us"` pricing multiplier (1.1x). * *Page 4: `about-claude/model-deprecations.md`* * Minor table header changes (capitalization consistency). * *Page 5: `api/overview.md`* * Added authentication info for cloud platforms (IAM integration). * Renamed "Third-Party Platform APIs" to "Cloud platform APIs". * Split table: Added "Claude Platform on AWS" (Anthropic-operated) vs Bedrock/Vertex (Partner-operated). * Clarified request size limits (Bedrock/Vertex vs Claude API). * Added note about "Claude Platform on AWS" request IDs (`x-amzn-requestid`). * *Page 6: `api/client-sdks.md`* * Updated platform support table to include "Claude Platform on AWS". * Clarified "Anthropic-operated" vs "Partner-operated". * *Page 7: `api/rate-limits.md`* * Added note specific to "Claude Platform on AWS": No spend limits, no auto-tier advancement, no Fast mode/Priority Tier support. * *Page 8: `api/service-tiers.md`* * Added note: Claude Platform on AWS supports Standard and Batch, but *not* Priority Tier. * *Page 9: `api/errors.md`* * Updated error descriptions for 401/402 to mention AWS Marketplace/Claude Platform on AWS. * Added section on "Request ID" for Claude Platform on AWS (dual IDs: AWS `x-amzn-requestid` + Anthropic `request-id`). * Added code examples for Python/TS to read headers on AWS Platform. * Added specific error for "Outbound web identity federation disabled". * *Page 10 & 11: `build-with-claude/overview.md` & `build-with-claude/text-generation.md`* (Diffs look very similar/identical in the snippet) * Updated feature availability tables to include "Claude Platform on AWS" (`claudePlatformAws`). * Updated "Platform labels" to distinguish Anthropic-operated vs Partner-operated. * Updated supported model lists for features (e.g., Sonnet 4.6 added to some contexts). * *Page 12: `build-with-claude/vision.md`* * Clarified request size limits on "partner-operated platforms" (Bedrock/Vertex). * Minor token count adjustments (1568 vs 1600). * *Page 13: `build-with-claude/pdf-support.md`* * Confirmed PDF support on Claude Platform on AWS. * Updated section title "Amazon Bedrock PDF Support" to lowercase. * *Page 14: `build-with-claude/extended-thinking.md`* * Clarified "Interleaved Thinking" behavior on Claude Platform on AWS (ignores header safely). * Clarified deprecation of `interleaved-thinking-2025-05-14` header on newer models (Opus 4.7). * *Page 15: `build-with-claude/streaming.md`* * Updated code examples to use `Model.ClaudeOpus4_7`. * Switched from explicit "extended thinking" to "adaptive thinking" in examples. * Minor grammatical fixes. * *Page 16: `build-with-claude/prompt-caching.md`* * Updated code examples (C#) to use `Model.ClaudeOpus4_7`. * Added "Claude Platform on AWS" to the list of platforms supporting automatic caching. * Clarified workspace-level isolation (now active as of Feb 5, 2026) for specific platforms. * *Page 17: `build-with-claude/batch-processing.md`* * Confirmed "Extended output" (300k tokens) is available on Claude Platform on AWS. * *Page 18: `build-with-claude/context-windows.md`* * Minor text refinements ("that is" vs "i.e.", grammar). * *Page 19: `build-with-claude/context-editing.md`* * Updated SDK notes (Go, Java, PHP) regarding `tool_runner` support. * Fixed a math calculation in the example. * *Page 20: `build-with-claude/search-results.md`* * Minor code cleanup (removed duplicate line). * *Page 21: `build-with-claude/structured-outputs.md`* * Confirmed availability on Claude Platform on AWS. * Updated Java annotations notes (Swagger Core OpenAPI 3). * *Page 22: `build-with-claude/effort.md`* * Clarified "High effort" description (removed "default" tag which seems inconsistent with text elsewhere, clarified "complex reasoning"). * Updated model support list (Sonnet 4.6). * Updated Opus 4.7 behavior regarding adaptive thinking. * *Page 23: `build-with-claude/files.md`* * Confirmed Files API availability on Claude Platform on AWS. * *Page 24: `build-with-claude/working-with-messages.md`* * Minor code fix (printing message object). * *Page 25: `build-with-claude/claude-on-amazon-bedrock.md`* * Added links to "Claude Platform on AWS" as an alternative/migration path. * Changed language "AWS Bearer Token" default region to `us-west-2`. * Updated regional endpoint prefix usage (`us.` instead of just removing `global.`). 3. **Synthesize Findings:** * **Major Theme 1: Claude Platform on AWS Launch/Integration.** This is the biggest change. The docs are heavily updated to distinguish between: * **Claude API (Direct):** Anthropic billing. * **Claude Platform on AWS:** Anthropic-operated, runs on AWS infra, uses CCUs (Claude Consumption Units) for billing via AWS Marketplace, uses standard Anthropic API model IDs. * **Amazon Bedrock / Vertex AI:** Partner-operated, provider billing, specific model IDs. * **Major Theme 2: Model Updates & Deprecations.** * Extended thinking is *removed* in Opus 4.7 (replaced by Adaptive Thinking). * Adaptive thinking is the standard for Opus 4.7. * Sonnet 4.6 support added/clarified in various feature matrices. * **Major Theme 3: Billing & Pricing Nuances.** * Introduction of CCUs for the AWS platform. * Clarification on `inference_geo` pricing (1.1x). * Workspace-level isolation for prompt caching (active Feb 5, 2026). * **Major Theme 4: API Specifics.** * Dual request IDs for the AWS platform (AWS ID + Anthropic ID). * Rate limits differ on the AWS platform (no auto-tier, no Priority tier). 4. **Draft the Response (Internal Monologue/Rough Draft in Chinese):** * *Summary:* Docs updated to support the launch of "Claude Platform on AWS". Distinguishes it from Bedrock. Also updates for Opus 4.7 (removing extended thinking, adaptive thinking). * *Key Themes:* * New platform: Claude Platform on AWS (Anthropic-run, AWS infra, CCU billing). * Platform categorization: Anthropic-operated vs Partner-operated (Bedrock/Vertex). * Opus 4.7 specifics: Adaptive thinking mandatory, extended thinking removed. * Billing/Pricing: CCUs, data residency multipliers. * Feature parity/matrix updates across many pages. * *Impact:* High. It's a new deployment option with different billing/limits, plus a breaking change in thinking logic for Opus 4.7. * *Action Items:* * If on Bedrock, consider migrating to "Claude Platform on AWS" for feature parity. * Update code for Opus 4.7: switch from `extended thinking` to `adaptive thinking`. * Check billing logic if using AWS Platform (CCUs vs raw tokens). * Update error handling for dual request IDs if on AWS Platform. * Note prompt caching isolation changes (workspace level). 5. **Refine the Output (Markdown & Formatting):** * Translate concepts accurately. * "Claude Platform on AWS" -> 保持英文 (专有名词). * "CCU" -> Claude Consumption Unit. * "Extended thinking" -> 扩展思考. * "Adaptive thinking" -> 自适应思考. 6. **Final Polish (Self-Correction during drafting):** * *Correction:* Make sure to emphasize *why* Claude Platform on AWS matters—it gives Anthropic control on AWS infra, meaning faster feature rollout than Bedrock but AWS billing. * *Correction:* The migration guide explicitly says Extended Thinking is *removed* in 4.7, not just deprecated. This is a breaking change for those moving to 4.7. 7. **Final Output Generation** (Matches the provided structure).