If your command grows long or needs project-specific scripts,create a shell file and reference it via `${CLAUDE_PROJECT_DIR}`.
`Hook 输入输出怎么工作`?
`
A hook receives event context as JSON.
`command`: read from stdin.
`http`这方面，sent as POST body.
字段名
The typical flow is:
bash
INPUT="$"
TOOL_NAME="$"
COMMAND="$"
Make sure not to parse manually with grep—always use `jq`.
If you want a decision。return a single JSON object via stdout and exit with code `0`.
If stdout is empty or not valid JSON,hook is considered an error and will not block execution .
When multiple hooks fire for same event y run concurrently—only final result counts if y modify same tool input. That’s *** you should design each hook independently and avoid race conditions.
`常用生命周期事件怎么理解`?
`
Select `SessionStart`,`UserPromptSubmit`,`PreToolUse`,`PermissionRequest`,`PostToolUse`,`Stop`。and `PreCompact/ PostCompact` as your starting points.
The key difference:
`PreToolUse`: before tool runs—ideal for blocking dangerous commands or protecting sensitive files.
`PostToolUse`: after tool succeeds—good for formatting or logging but cannot undo a bad write.
`PermissionRequest`: triggered when a tool requires explicit approval—useful for automated approvals in batch agents.
The matcher "permission_prompt" ensures notification only appears when Claude asks for permission—low risk but great visibility.
