用于构建 Claude Code 钩子的 Python 工具包

原文

A Python utility package for building Claude Code hooks with minimal boilerplate.

What Are Claude Code Hooks?

Claude Code hooks are custom scripts that run at specific points during Claude Code's execution. They allow you to:

Validate tool calls before they execute (PreToolUse)
React to tool results after execution (PostToolUse)
Intercept user prompts before Claude sees them (UserPromptSubmit)
Initialize state when a session starts (SessionStart)

Building Claude Code hooks involves repetitive boilerplate:

Parsing JSON from stdin
Validating input structure
Formatting responses in the correct schema
Handling errors gracefully

claude-hook-utils handles all of this, letting you focus on your validation logic.

One Pattern - Extend HookHandler, override the hooks you need
Type Safety - Typed dataclasses for inputs, builder pattern for responses
Explicit Control - Helper methods on inputs, but you decide when to skip/allow/deny
Multi-Hook Support - One Python program can handle multiple hook types
No Heavy Dependencies - Core package has minimal dependencies; bring your own AI SDK if needed

pip install claude-hook-utils

#!/usr/bin/env python3
"""Validate that Data classes have TypeScript annotation."""

from claude_hook_utils import HookHandler, PreToolUseInput, PreToolUseResponse

class DataClassValidator(HookHandler):
    def pre_tool_use(self, input: PreToolUseInput) -> PreToolUseResponse | None:
        # Skip if not a Data class file
        if not input.file_path_matches('**/app/Data/**/*.php'):
            return None

        # Check for required annotation
        if input.content and '#[TypeScript()]' not in input.content:
            return PreToolUseResponse.deny(
                "Data classes must have #[TypeScript()] annotation for type generation"
            )

        return PreToolUseResponse.allow()

if __name__ == "__main__":
    DataClassValidator().run()

Configure in .claude/settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Write|Edit",
        "hooks": [
          {
            "type": "command",
            "command": "python3 /path/to/data_class_validator.py"
          }
        ]
      }
    ]
  }
}

Hook Type	When It Runs	Use Cases
`PreToolUse`	Before a tool executes	Validate file paths, check content, block dangerous operations
`PostToolUse`	After a tool completes	Log results, trigger follow-up actions, collect metrics
`UserPromptSubmit`	When user submits a prompt	Validate prompts, add context, enforce policies
`SessionStart`	When a Claude Code session begins	Initialize state, set environment variables

Extend this class and override the hooks you need:

from claude_hook_utils import HookHandler

class MyHandler(HookHandler):
    def __init__(self):
        super().__init__()
        # Add any shared state here
        self._cache: dict = {}

    def pre_tool_use(self, input: PreToolUseInput) -> PreToolUseResponse | None:
        """Called before tool execution. Return None to skip."""
        return None

    def post_tool_use(self, input: PostToolUseInput) -> PostToolUseResponse | None:
        """Called after tool execution. Return None to skip."""
        return None

    def user_prompt_submit(self, input: UserPromptSubmitInput) -> UserPromptSubmitResponse | None:
        """Called when user submits a prompt. Return None to skip."""
        return None

    def session_start(self, input: SessionStartInput) -> SessionStartResponse | None:
        """Called when session starts. Return None to skip."""
        return None

if __name__ == "__main__":
    MyHandler().run()

Input for PreToolUse hooks:

@dataclass
class PreToolUseInput:
    # Common fields
    session_id: str
    cwd: str
    hook_event_name: str  # Always "PreToolUse"

    # PreToolUse-specific
    tool_name: str        # "Write", "Edit", "Bash", etc.
    tool_input: dict      # Tool-specific parameters
    tool_use_id: str

    # Helper methods
    def file_path_matches(self, *globs: str) -> bool:
        """Check if tool_input.file_path matches any glob pattern."""

    def file_path_excludes(self, *globs: str) -> bool:
        """Check if tool_input.file_path does NOT match any glob pattern."""

    # Convenience properties
    @property
    def file_path(self) -> str | None:
        """Get file_path from tool_input (for Write/Edit/Read tools)."""

    @property
    def content(self) -> str | None:
        """Get content from tool_input (for Write tool)."""

    @property
    def command(self) -> str | None:
        """Get command from tool_input (for Bash tool)."""

Response builder for PreToolUse hooks:

class PreToolUseResponse:
    @staticmethod
    def allow(reason: str | None = None) -> PreToolUseResponse:
        """Allow the tool to execute."""

    @staticmethod
    def deny(reason: str) -> PreToolUseResponse:
        """Block the tool. Reason is shown to Claude as feedback."""

    @staticmethod
    def ask(reason: str) -> PreToolUseResponse:
        """Request user confirmation before proceeding."""

    def with_updated_input(self, **updates) -> PreToolUseResponse:
        """Modify tool_input before execution (only valid with allow)."""

JSONL-based logging for easy debugging. Logs are organized by namespace (plugin name).

from claude_hook_utils import HookHandler, HookLogger

class MyHandler(HookHandler):
    def __init__(self):
        # Logs to .claude/logs/my-plugin/hooks.jsonl
        super().__init__(
            logger=HookLogger.create_default("MyHandler", namespace="my-plugin")
        )

    def pre_tool_use(self, input: PreToolUseInput) -> PreToolUseResponse | None:
        # Session ID is automatically added from input
        self.logger.info("Checking file", file_path=input.file_path)
        # ... validation logic ...
        self.logger.decision("allow", reason="Validation passed")
        return PreToolUseResponse.allow()

Log format (JSONL - one JSON object per line):

{"ts": "2025-01-04T10:15:23.456+00:00", "level": "INFO", "hook": "MyHandler", "namespace": "my-plugin", "session": "abc123", "msg": "Checking file", "file_path": "/path/to/file.php"}
{"ts": "2025-01-04T10:15:23.458+00:00", "level": "DECISION", "hook": "MyHandler", "namespace": "my-plugin", "session": "abc123", "msg": "decision=allow", "decision": "allow", "reason": "Validation passed"}

Configuration:

Default location: {cwd}/.claude/logs/{namespace}/hooks.jsonl
Without namespace: {cwd}/.claude/logs/hooks.jsonl
Override directory with CLAUDE_HOOK_LOG_DIR env var
Override namespace with CLAUDE_HOOK_LOG_NAMESPACE env var
Session ID is automatically extracted from hook input

Validate Vue Component Structure

class VueValidator(HookHandler):
    def pre_tool_use(self, input: PreToolUseInput) -> PreToolUseResponse | None:
        if not input.file_path_matches('**/*.vue'):
            return None

        content = input.content or ''

        # Check tag order: <script> before <template> before <style>
        script_pos = content.find('<script')
        template_pos = content.find('<template')
        style_pos = content.find('<style')

        if script_pos > template_pos or template_pos > style_pos:
            return PreToolUseResponse.deny(
                "Vue components must have tags in order: <script>, <template>, <style>"
            )

        # Check for setup lang="ts"
        if '<script setup lang="ts">' not in content:
            return PreToolUseResponse.deny(
                "Vue components must use <script setup lang=\"ts\">"
            )

        return PreToolUseResponse.allow()

Validate Controller Location

class ControllerValidator(HookHandler):
    def pre_tool_use(self, input: PreToolUseInput) -> PreToolUseResponse | None:
        if not input.file_path_matches('**/*Controller.php'):
            return None

        # Controllers must be in app/Http/Controllers/
        if not input.file_path_matches('**/app/Http/Controllers/**/*.php'):
            return PreToolUseResponse.deny(
                f"Controllers must be in app/Http/Controllers/. "
                f"Found: {input.file_path}"
            )

        return PreToolUseResponse.allow()

Block FormRequest Usage (Suggest Data Class)

class NoFormRequestValidator(HookHandler):
    def pre_tool_use(self, input: PreToolUseInput) -> PreToolUseResponse | None:
        if not input.file_path_matches('**/*Controller.php'):
            return None

        content = input.content or ''

        if 'FormRequest' in content:
            return PreToolUseResponse.deny(
                "Do not use FormRequest classes. Use Data classes instead. "
                "See: app/Data/ for examples."
            )

        return PreToolUseResponse.allow()

Multi-Hook Handler (Pre + Post)

class FileTracker(HookHandler):
    def __init__(self):
        super().__init__()
        self._pending_writes: set[str] = set()

    def pre_tool_use(self, input: PreToolUseInput) -> PreToolUseResponse | None:
        if input.tool_name == 'Write' and input.file_path:
            self._pending_writes.add(input.file_path)
            self.logger.info(f"Tracking write: {input.file_path}")
        return PreToolUseResponse.allow()

    def post_tool_use(self, input: PostToolUseInput) -> PostToolUseResponse | None:
        if input.tool_name == 'Write' and input.file_path:
            self._pending_writes.discard(input.file_path)
            self.logger.info(f"Write completed: {input.file_path}")
        return None

Claude Code Hook Response Format

This package generates responses in the official hookSpecificOutput format:

{
  "hookSpecificOutput": {
    "hookEventName": "PreToolUse",
    "permissionDecision": "deny",
    "permissionDecisionReason": "Your reason here"
  }
}

Decision	Effect
`allow`	Tool executes immediately, reason shown to user
`deny`	Tool blocked, reason shown to Claude (so it can adapt)
`ask`	User confirmation dialog shown

Use with_updated_input() to modify parameters before execution:

def pre_tool_use(self, input: PreToolUseInput) -> PreToolUseResponse | None:
    # Auto-correct a common mistake
    if input.file_path and '/data/' in input.file_path:
        corrected = input.file_path.replace('/data/', '/Data/')
        return PreToolUseResponse.allow("Auto-corrected path").with_updated_input(
            file_path=corrected
        )
    return PreToolUseResponse.allow()

The package handles errors gracefully:

Invalid JSON input: Returns exit 0 (no output = allow)
Unknown hook type: Returns None (skip)
Exception in handler: Logged to stderr, returns exit 0 (fail open)

This "fail open" approach ensures your hooks don't block Claude Code if something goes wrong.

Claude Code provides these environment variables to hooks:

Variable	Description
`CLAUDE_PROJECT_DIR`	Absolute path to project root
`CLAUDE_CODE_REMOTE`	`"true"` if running in web environment

This package uses:

Variable	Description
`CLAUDE_HOOK_LOG_DIR`	Override default log directory (default: `.claude/logs/{namespace}/`)
`CLAUDE_HOOK_LOG_NAMESPACE`	Override log namespace/subdirectory

Access via input.cwd or os.environ.

Extending for New Hook Types

To add support for a new hook type:

Create input dataclass in inputs/
Create response class in responses/
Add handler method to HookHandler
Add dispatch case in HookHandler._dispatch()

See existing implementations for patterns to follow.

MIT