mirror of
https://github.com/luongnv89/claude-howto.git
synced 2026-09-22 01:10:42 +02:00
Luong NGUYEN
8aeb5a72dc
* docs: sync to Claude Code v2.1.278 Type: docs Audit of the tutorial against Claude Code 2.1.278 (repo baseline was 2.1.263). Fixes claims the repo asserts that are now false; new-feature coverage is English-only. P0 - the repo stated the opposite of current behavior: - AGENTS.md: the claude-md skill said Claude Code "does not read AGENTS.md directly" in all 5 languages. It has since v2.1.277, when no CLAUDE.md exists at or above the working directory. - /output-style: listed as "Removed in v2.1.91" in 5 languages and in 09-advanced-features. Re-added in v2.1.269; row moved out of the Deprecated Commands table. - Permission rules: setup-auto-mode-permissions.py seeded Write(*), Glob(*) and NotebookEdit(*), which Claude Code accepts but never consults and warns about at startup. Removed; same fix in the config-examples.json profile and the 06/09 category tables. - taskOutputMaxChars: documented as live. The TaskOutput tool was removed in v2.1.278 and the key is inert. P1: - Dynamic workflow "medium" guideline is fewer than 10 agents, not 15; the default is small on Pro plans since v2.1.271. - Todo-tool availability restated as the official allowlist (Claude 3.x, Opus 4-4.7, Sonnet 4-4.6, Haiku 4.5) instead of a denylist; 23 sites across 5 languages. - Memory hierarchy now notes the AGENTS.md fallback. - Auto-mode classifier cost qualified: not charged on Enterprise plans and Claude API accounts since v2.1.278. P3 (English only): full AGENTS.md section in 02-memory; omitClaudeMd subagent field; plugin eval / --json / --accept-command / --marketplace; maxEffortLevel, bashEditDiffEnabled, syncClaudeAi*, 4 env vars, --system-prompt-snapshot, --plugin-dir folders; send-now key and Monitor deadlines; checkpoints subagent-edit limitation. Learning skills audited clean: 100 quiz questions and 100 Review pointers all valid, 20 self-assessment options correctly mapped. Cross-file anchors: 15 broken before, 15 after, 0 introduced. * docs: address review — stale version stamps, /output-style re-add, permission tables
Hooks 参考
Hooks 是在 Claude Code 事件发生时自动执行的 shell 命令，用来做格式化、校验、通知、审计等自动化工作。
概览
Hooks 是事件驱动的自动化机制。它们会在 Claude Code 发生某些动作时自动运行，不需要你手动触发。
常见用途：
- 写文件前自动格式化
- 提交前运行测试
- 扫描安全问题
- 记录 bash 命令
- 校验用户提示词
- 发送团队通知
Hook 类型
Claude Code 支持 33 个 hook 事件，按用途分为以下 4 组。事件的 handler 类型共 5 种：command、http、mcp_tool、prompt、agent。
- Tool Hooks：
PreToolUse、PostToolUse、PostToolUseFailure、PostToolBatch、PermissionRequest、PermissionDenied - Session Hooks：
SessionStart、Setup、SessionEnd、Stop、StopFailure、SubagentStart、SubagentStop - Task Hooks：
UserPromptSubmit、UserPromptExpansion、MessageDisplay、TaskCompleted、TaskCreated、TeammateIdle（TaskCompleted和TaskCreated仅在启用 todo 工具时触发 —— 默认仅在 Claude 3.x 模型、Opus 4 至 4.7、Sonnet 4 至 4.6 及 Haiku 4.5 上可用；CLAUDE_CODE_ENABLE_TODO_TOOLS=1可恢复） - Lifecycle Hooks：
ConfigChange、CwdChanged、DirectoryAdded、FileChanged、PreCompact、PostCompact、PreModelSwitch、PostModelSwitch、WorktreeCreate、WorktreeRemove、Notification、InstructionsLoaded、Elicitation、ElicitationResult
安装
mkdir -p ~/.claude/hooks
cp 06-hooks/*.sh ~/.claude/hooks/
chmod +x ~/.claude/hooks/*.sh
然后在 ~/.claude/settings.json 里配置：
{
"hooks": {
"PreToolUse": [
{
"matcher": "Write",
"hooks": [
{
"type": "command",
"command": "~/.claude/hooks/format-code.sh"
}
]
}
],
"PostToolUse": [
{
"matcher": "Write",
"hooks": [
{
"type": "command",
"command": "~/.claude/hooks/security-scan.sh"
}
]
}
]
}
}
使用方法
Hooks 会在匹配到事件时自动执行。你可以把它理解成 Claude Code 的事件回调。
常见示例
format-code.sh- 写入前自动格式化pre-commit.sh- 提交前跑测试security-scan.sh- 做安全扫描log-bash.sh- 记录 bash 命令validate-prompt.sh- 校验输入notify-team.sh- 发通知
最佳实践
- 把 hooks 保持短小明确
- 只做单一职责
- 先在本地测试
- 不要在 hook 里放复杂业务逻辑
- 对副作用保持谨慎
故障排查
- 检查文件路径和权限
- 确认脚本可执行
- 检查 settings.json 语法
- 查看 Claude Code 版本兼容性
相关概念
更多资源
最后更新: 2026 年 9 月 19 日 Claude Code 版本: 2.1.278 来源: