hooks-configuration

Claude Code Hooks Configuration

Expert knowledge for configuring and developing Claude Code hooks to automate workflows and enforce best practices.

Core Concepts

What Are Hooks? Hooks are user-defined shell commands that execute at specific points in Claude Code's lifecycle. Unlike relying on Claude to "decide" to run something, hooks provide deterministic, guaranteed execution.

Why Use Hooks?

• Enforce code formatting automatically
• Block dangerous commands before execution
• Inject context at session start
• Log commands for audit trails
• Send notifications when tasks complete

Hook Lifecycle Events

Event	When It Fires	Key Use Cases
SessionStart	Session begins/resumes	Environment setup, context loading
SessionEnd	Session terminates	Cleanup, state persistence
UserPromptSubmit	User submits prompt	Input validation, context injection
PreToolUse	Before tool execution	Permission control, blocking dangerous ops
PostToolUse	After tool completes	Auto-formatting, logging, validation
PostToolUseFailure	After tool execution fails	Retry decisions, error handling
PermissionRequest	Claude requests permission for a tool	Auto approve/deny without user prompt
Stop	Main agent finishes responding	Notifications, git reminders
SubagentStart	Subagent (Task tool) is about to start	Input modification, context injection
SubagentStop	Subagent finishes	Per-task completion evaluation
WorktreeCreate	New git worktree created via EnterWorktree	Worktree setup, dependency install
WorktreeRemove	Worktree removed after session exits	Cleanup, uncommitted changes alert
TeammateIdle	Teammate in agent team goes idle	Assign additional tasks to teammate
TaskCompleted	Task in shared task list marked complete	Validation gates before task acceptance
PreCompact	Before context compaction	Transcript backup
Notification	Claude sends notification	Custom alerts
ConfigChange	Claude Code settings change at runtime	Audit config changes, validation

Stop vs SubagentStop: Stop fires at the session level when the main agent finishes a response turn. SubagentStop fires when an individual subagent (spawned via the Task tool) completes. Use Stop for session-level notifications; use SubagentStop for per-task quality gates.

For full schemas, examples, and timeout recommendations for each event, see .claude/rules/hooks-reference.md.

Configuration

File Locations

Hooks are configured in settings files:

• ~/.claude/settings.json - User-level (applies everywhere)
• .claude/settings.json - Project-level (committed to repo)
• .claude/settings.local.json - Local project (not committed)

Claude Code merges all matching hooks from all files.

Frontmatter Hooks (Skills and Commands)

Hooks can also be defined directly in skill and command frontmatter using the hooks field:

---
name: my-skill
description: A skill with hooks
allowed-tools: Bash, Read
hooks:
  PreToolUse:
    - matcher: "Bash"
      hooks:
        - type: command
          command: "echo 'Pre-tool hook from skill'"
          timeout: 10
---

This allows skills and commands to define their own hooks that are active only when that skill/command is in use.

Basic Structure

{
  "hooks": {
    "EventName": [
      {
        "matcher": "ToolPattern",
        "hooks": [
          {
            "type": "command",
            "command": "your-command-here",
            "timeout": 30
          }
        ]
      }
    ]
  }
}

Matcher Patterns

• Exact match: "Bash" - matches exactly "Bash" tool
• Regex patterns: "Edit|Write" - matches either tool
• Wildcards: "Notebook.*" - matches tools starting with "Notebook"
• All tools: "*" - matches everything
• MCP tools: "mcp__server__tool" - targets MCP server tools

Input Schema

Hooks receive JSON via stdin with these common fields:

{
  "session_id": "unique-session-id",
  "transcript_path": "/path/to/conversation.json",
  "cwd": "/current/working/directory",
  "permission_mode": "mode",
  "hook_event_name": "PreToolUse"
}

PreToolUse additional fields:

{
  "tool_name": "Bash",
  "tool_input": {
    "command": "npm test"
  }
}

PostToolUse additional fields:

{
  "tool_name": "Bash",
  "tool_input": { ... },
  "tool_response": { ... }
}

SubagentStart additional fields:

{
  "subagent_type": "Explore",
  "subagent_prompt": "original prompt text",
  "subagent_model": "claude-opus"
}

Output Schema

Exit Codes

• 0: Success (command allowed)
• 2: Blocking error (stderr shown to Claude, operation blocked)
• Other: Non-blocking error (logged in verbose mode)

JSON Response (optional)

PreToolUse (wrapped in hookSpecificOutput):

{
  "hookSpecificOutput": {
    "hookEventName": "PreToolUse",
    "permissionDecision": "allow|deny|ask",
    "permissionDecisionReason": "explanation",
    "updatedInput": { "modified": "input" }
  }
}

Stop/SubagentStop:

{
  "decision": "block",
  "reason": "required explanation for continuing"
}

SubagentStart (input modification):

{
  "updatedPrompt": "modified prompt text to inject context or modify behavior"
}

SessionStart (wrapped in hookSpecificOutput):

{
  "hookSpecificOutput": {
    "hookEventName": "SessionStart",
    "additionalContext": "Information to inject into session"
  }
}

Common Hook Patterns

Block Dangerous Commands (PreToolUse)

#!/bin/bash
INPUT=$(cat)
COMMAND=$(echo "$INPUT" | jq -r '.tool_input.command // empty')

# Block rm -rf /
if echo "$COMMAND" | grep -Eq 'rm\s+(-rf|-fr)\s+/'; then
    echo "BLOCKED: Refusing to run destructive command on root" >&2
    exit 2
fi

exit 0

Auto-Format After Edits (PostToolUse)

#!/bin/bash
INPUT=$(cat)
FILE=$(echo "$INPUT" | jq -r '.tool_input.file_path // empty')

if [[ "$FILE" == *.py ]]; then
    ruff format "$FILE" 2>/dev/null
    ruff check --fix "$FILE" 2>/dev/null
elif [[ "$FILE" == *.ts ]] || [[ "$FILE" == *.tsx ]]; then
    prettier --write "$FILE" 2>/dev/null
fi

exit 0

Remind About Built-in Tools (PreToolUse)

#!/bin/bash
INPUT=$(cat)
COMMAND=$(echo "$INPUT" | jq -r '.tool_input.command // empty')

if echo "$COMMAND" | grep -Eq '^\s*cat\s+[^|><]'; then
    echo "REMINDER: Use the Read tool instead of 'cat'" >&2
    exit 2
fi

exit 0

Load Context at Session Start (SessionStart)

#!/bin/bash
GIT_STATUS=$(git status --short 2>/dev/null | head -5)
BRANCH=$(git branch --show-current 2>/dev/null)

CONTEXT="Current branch: $BRANCH\nPending changes:\n$GIT_STATUS"
jq -n --arg ctx "$CONTEXT" '{
  "hookSpecificOutput": {
    "hookEventName": "SessionStart",
    "additionalContext": $ctx
  }
}'

Inject Context for Subagents (SubagentStart)

#!/bin/bash
INPUT=$(cat)
SUBAGENT_TYPE=$(echo "$INPUT" | jq -r '.subagent_type // empty')
ORIGINAL_PROMPT=$(echo "$INPUT" | jq -r '.subagent_prompt // empty')

# Add project context to Explore agents
if [ "$SUBAGENT_TYPE" = "Explore" ]; then
    PROJECT_INFO="Project uses TypeScript with Bun. Main source in src/."
    cat << EOF
{
  "updatedPrompt": "$PROJECT_INFO\n\n$ORIGINAL_PROMPT"
}
EOF
fi

exit 0

Desktop Notification on Stop (Stop)

#!/bin/bash
# Linux
notify-send "Claude Code" "Task completed" 2>/dev/null

# macOS
osascript -e 'display notification "Task completed" with title "Claude Code"' 2>/dev/null

exit 0

Audit Logging (PostToolUse)

#!/bin/bash
INPUT=$(cat)
TOOL=$(echo "$INPUT" | jq -r '.tool_name')
COMMAND=$(echo "$INPUT" | jq -r '.tool_input.command // "N/A"')

echo "$(date -Iseconds) | $TOOL | $COMMAND" >> ~/.claude/audit.log
exit 0

Auto-Approve Safe Operations (PermissionRequest)

#!/bin/bash
INPUT=$(cat)
TOOL=$(echo "$INPUT" | jq -r '.tool_name')
COMMAND=$(echo "$INPUT" | jq -r '.tool_input.command // empty')

# Auto-approve read-only git operations
if [ "$TOOL" = "Bash" ] && echo "$COMMAND" | grep -Eq '^git (status|log|diff|branch|remote)'; then
  echo '{"decision": "approve", "reason": "Read-only git operation"}'
  exit 0
fi

# Auto-deny destructive operations on root
if [ "$TOOL" = "Bash" ] && echo "$COMMAND" | grep -Eq 'rm\s+(-rf|-fr)\s+/'; then
  echo '{"decision": "deny", "reason": "Destructive root operation blocked"}'
  exit 0
fi

exit 0

Set Up Worktree Environment (WorktreeCreate)

#!/bin/bash
INPUT=$(cat)
WORKTREE_PATH=$(echo "$INPUT" | jq -r '.worktree_path')

if [ -f "$WORKTREE_PATH/package.json" ]; then
  (cd "$WORKTREE_PATH" && bun install --frozen-lockfile) 2>/dev/null
fi

exit 0

Gate Task Completion (TaskCompleted)

#!/bin/bash
INPUT=$(cat)
TASK_TITLE=$(echo "$INPUT" | jq -r '.task_title')

if echo "$TASK_TITLE" | grep -qi 'implement\|add\|fix\|refactor'; then
  if ! npm test --bail 2>/dev/null; then
    echo '{"decision": "block", "reason": "Tests must pass before task is accepted."}'
    exit 0
  fi
fi

exit 0

Configuration Examples

Anti-Pattern Detection

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "bash $CLAUDE_PROJECT_DIR/hooks-plugin/hooks/bash-antipatterns.sh",
            "timeout": 5
          }
        ]
      }
    ]
  }
}

Auto-Format Python Files

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": "bash -c 'FILE=$(cat | jq -r \".tool_input.file_path\"); [[ \"$FILE\" == *.py ]] && ruff format \"$FILE\"'"
          }
        ]
      }
    ]
  }
}

Git Reminder on Stop

{
  "hooks": {
    "Stop": [
      {
        "matcher": "*",
        "hooks": [
          {
            "type": "command",
            "command": "bash -c 'changes=$(git status --porcelain | wc -l); [ $changes -gt 0 ] && echo \"Reminder: $changes uncommitted changes\"'"
          }
        ]
      }
    ]
  }
}

Prompt-Based and Agent-Based Hooks

In addition to command hooks, Claude Code supports LLM-powered hooks for decisions that require judgment.

Hook Types

Type	How It Works	Default Timeout	Use When
command	Runs a shell command, reads stdin, returns exit code	600s	Deterministic rules (regex, field checks)
http	Sends hook data to an HTTPS endpoint, reads JSON response	30s	Remote/centralized policy enforcement
prompt	Single-turn LLM call (Haiku), returns {ok: true/false}	30s	Judgment on hook input data alone
agent	Multi-turn subagent with tool access, returns {ok: true/false}	60s	Verification needing file/tool access

Additional Hook Handler Fields

• async: true: Fire-and-forget for command hooks (non-blocking, exit code ignored)
• once: true: Run only once per session; subsequent triggers are skipped

HTTP Hook Configuration

{
  "type": "http",
  "url": "https://hooks.example.com/pre-tool-use",
  "headers": {
    "Authorization": "Bearer ${HOOKS_API_KEY}"
  },
  "timeout": 30
}

Only HTTPS URLs are allowed. Header values support ${ENV_VAR} expansion.

Supported Events

Prompt and agent hooks work on: PreToolUse, PostToolUse, PostToolUseFailure, PermissionRequest, Stop, SubagentStop, TaskCompleted, UserPromptSubmit.

CLAUDE_ENV_FILE (SessionStart)

SessionStart hooks can write environment variables that persist for the session via CLAUDE_ENV_FILE:

if [ -n "$CLAUDE_ENV_FILE" ]; then
  echo "NODE_ENV=development" >> "$CLAUDE_ENV_FILE"
fi

All other events (SessionStart, SessionEnd, PreCompact, etc.) support only command hooks.

Prompt Hook Configuration

{
  "type": "prompt",
  "prompt": "Evaluate whether all tasks are complete. $ARGUMENTS",
  "model": "haiku",
  "timeout": 30,
  "statusMessage": "Checking completeness..."
}

Agent Hook Configuration

{
  "type": "agent",
  "prompt": "Check for TODO/FIXME comments and debugging artifacts in changed files. $ARGUMENTS",
  "model": "haiku",
  "timeout": 60,
  "statusMessage": "Checking implementation quality..."
}

Note: Prefer command hooks over agent hooks when the logic is deterministic. For example, test verification (detect changed files → find test runner → run tests) is better as a bash script than an agent hook — it eliminates LLM latency on every invocation.

Response Schema (both types)

{"ok": true}
{"ok": false, "reason": "Explanation of what's wrong"}

Stop Hook Loop Prevention

Stop hooks fire every time Claude finishes responding, including after acting on stop hook feedback. Include this check in Stop hook prompts:

First: if stop_hook_active is true in the input, respond with {"ok": true} immediately.

For the full decision guide on when to use each hook type, see .claude/rules/prompt-agent-hooks.md.

Handling Blocked Commands

When a PreToolUse hook blocks a command:

Situation	Action
Hook suggests alternative	Use the suggested tool/approach
Alternative won't work	Ask user to run command manually
User says "proceed"	Still blocked - explain and provide command for manual execution

Critical: User permission does NOT bypass hooks. Retrying a blocked command will fail again.

When command is legitimately needed:

1. Explain why the command is required
2. Describe alternatives considered and why they won't work
3. Provide exact command for user to run manually
4. Let user decide

Example response:

The hook blocked `git reset --hard abc123` because it's usually unnecessary.

I considered:
- `git pull`: Won't work because we need to discard local commits, not merge
- `git restore`: Only handles file changes, not commit history

If you need to proceed, please run manually:
git reset --hard abc123

This is needed because [specific justification].

Best Practices

Script Development:

1. Always read input from stdin with cat
2. Use jq for JSON parsing
3. Quote all variables to prevent injection
4. Exit with code 2 to block, 0 to allow
5. Write blocking messages to stderr
6. Keep hooks fast (< 5 seconds)

Configuration:

1. Use $CLAUDE_PROJECT_DIR for portable paths
2. Set explicit timeouts (default: 10 minutes / 600s as of 2.1.50)
3. Use specific matchers over wildcards
4. Test hooks manually before enabling

Security:

1. Validate all inputs
2. Use absolute paths
3. Avoid touching .env or .git/ directly
4. Review hook code before deployment

Debugging

Verify hook registration:

/hooks

Enable debug logging:

claude --debug

Test hooks manually:

echo '{"tool_input": {"command": "cat file.txt"}}' | bash your-hook.sh
echo $?  # Check exit code

Available Hooks in This Plugin

• bash-antipatterns.sh: Detects when Claude uses shell commands instead of built-in tools (cat, grep, sed, timeout, etc.)

See hooks/README.md for full documentation.