agents-hooks（Agent 挂钩）

Claude Code Hooks — Meta Reference

This skill provides the definitive reference for creating Claude Code hooks. Use this when building automation that triggers on Claude Code events.

When to Use This Skill

• Building event-driven automation for Claude Code
• Creating PreToolUse guards to block dangerous commands
• Implementing PostToolUse formatters, linters, or auditors
• Adding Stop hooks for testing or notifications
• Setting up SessionStart/SessionEnd for environment management
• Integrating Claude Code with CI/CD pipelines (headless mode)

Quick Reference

Hook Structure

.claude/hooks/
├── pre-tool-validate.sh
├── post-tool-format.sh
├── post-tool-audit.sh
├── stop-run-tests.sh
└── session-start-init.sh

Configuration

settings.json

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Edit|Write",
        "hooks": [
          {
            "type": "command",
            "command": "$CLAUDE_PROJECT_DIR/.claude/hooks/post-tool-format.sh"
          }
        ]
      }
    ],
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "$CLAUDE_PROJECT_DIR/.claude/hooks/pre-tool-validate.sh"
          }
        ]
      }
    ],
    "Stop": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "$CLAUDE_PROJECT_DIR/.claude/hooks/stop-run-tests.sh"
          }
        ]
      }
    ]
  }
}

Execution Model (Jan 2026)

• Hooks receive a JSON payload via stdin (treat it as untrusted input) and run with your user permissions (outside the Bash tool sandbox).
• Default timeout is 60s per hook command; all matching hooks run in parallel; identical commands are deduplicated.

Hook Input (stdin)

{
  "hook_event_name": "PreToolUse",
  "tool_name": "Bash",
  "tool_input": {
    "command": "ls -la"
  }
}

Environment Variables (shell)

Exit Codes

Stdout injection note: for UserPromptSubmit, SessionStart, and Setup, non-JSON stdout (exit 0) is injected into Claude’s context; most other events show stdout only in verbose mode.

Decision Control + Input Modification (v2.0.10+)

PreToolUse hooks can allow/deny/ask and optionally modify the tool input via updatedInput.

Hook Output Schema

{
  "hookSpecificOutput": {
    "hookEventName": "PreToolUse",
    "permissionDecision": "allow",
    "permissionDecisionReason": "Reason shown to user (and to Claude on deny)",
    "updatedInput": { "command": "echo 'modified'" },
    "additionalContext": "Extra context added before tool runs"
  }
}

Note: older decision/reason fields are deprecated; prefer the hookSpecificOutput.* fields.

See hook-templates.md for full examples: redirect sensitive file edits to /dev/null and strip .env files from git add commands.

Prompt-Based Hooks

For complex decisions, use LLM-evaluated hooks (type: "prompt") instead of bash scripts. They are most useful for Stop and SubagentStop decisions.

Configuration

{
  "hooks": {
    "Stop": [
      {
        "hooks": [
          {
            "type": "prompt",
            "prompt": "Evaluate whether Claude should stop. Context JSON: $ARGUMENTS. Return {\"ok\": true} if all tasks are complete, otherwise {\"ok\": false, \"reason\": \"what remains\"}.",
            "timeout": 30
          }
        ]
      }
    ]
  }
}

Response Schema

• Allow: {"ok": true}
• Block: {"ok": false, "reason": "Explanation shown to Claude"}

Combining Command and Prompt Hooks

Use command hooks for fast, deterministic checks. Use prompt hooks for nuanced decisions:

{
  "Stop": [
    {
      "hooks": [
        { "type": "command", "command": ".claude/hooks/quick-check.sh" },
        { "type": "prompt", "prompt": "Verify code quality meets standards" }
      ]
    }
  ]
}

Permission Prompt Fatigue Reduction Pattern

Use this when frequent approval dialogs slow down repeated safe workflows.

Strategy

1. Identify repetitive, low-risk command prefixes (for example: test runners, read-only diagnostics).
2. Approve narrow prefixes instead of full commands.
3. Keep destructive or broad shells (rm, git reset --hard, generic interpreters with arbitrary input) out of auto-approval rules.
4. Re-check approved prefixes periodically; remove stale ones.

Practical Guardrails

• Allow only task-scoped prefixes (example: npm run test:e2e), not unrestricted executors.
• Keep separate policy for write-outside-workspace actions.
• Pair allow-rules with deny-rules for dangerous patterns.

Outcome

This reduces repeated permission interruptions while preserving high-safety boundaries.

Runtime Preflight Hooks (Mandatory for Tool Reliability)

Add a lightweight runtime preflight hook when workflows depend on specific local tool versions (for example Node for JS REPL, test runners, linters).

Preflight Responsibilities

• Verify required binaries exist.
• Verify minimum version constraints.
• Emit a clear remediation message when requirements are not met.
• Fail fast before expensive task execution starts.

Recommended Trigger

• SessionStart for general runtime checks.
• Setup for repository bootstrap checks.

Minimal Policy

• Keep checks deterministic and fast (<1s target).
• Do not auto-install dependencies silently in hooks.
• Print exact command the operator should run to remediate.

Hook Templates

Copy-paste templates for the five most common hook scenarios: PreToolUse validation, PostToolUse formatting, PostToolUse security audit, Stop test runner, and SessionStart environment check.

See references/hook-templates.md for all scripts.

Matchers

Matchers filter which tool triggers the hook:

• Exact match: Write matches only the Write tool
• Regex: Edit|Write or Notebook.*
• Match all: * (also works with "" or omitted matcher)

Security Best Practices

Hooks run with full user permissions outside the Bash tool sandbox. Key rules: validate all stdin input, quote every variable ("$VAR"), use absolute paths, never eval untrusted data, and set -euo pipefail.

See references/hook-security.md for the full checklist, command injection prevention, path traversal defense, credential protection, and ShellCheck requirements.

Hook Composition

Multiple Hooks on Same Event

{
  "PostToolUse": [
    {
      "matcher": "Edit|Write",
      "hooks": [
        { "type": "command", "command": ".claude/hooks/format.sh" },
        { "type": "command", "command": ".claude/hooks/audit.sh" },
        { "type": "command", "command": ".claude/hooks/notify.sh" }
      ]
    }
  ]
}

All matching hooks run in parallel. If you need strict ordering (format → lint → test), make one wrapper script that runs them sequentially.

Debugging Hooks

# Test a PostToolUse hook manually (stdin JSON)
export CLAUDE_PROJECT_DIR="$(pwd)"
echo '{"hook_event_name":"PostToolUse","tool_name":"Edit","tool_input":{"file_path":"'"$(pwd)"'/src/app.ts"}}' \
  | bash .claude/hooks/post-tool-format.sh

# Check exit code
echo $?

Navigation

Resources

• references/hook-templates.md — Copy-paste hook scripts
• references/hook-patterns.md — Common patterns
• references/hook-security.md — Security guide
• references/runtime-preflight-hooks.md — Runtime/version preflight patterns for SessionStart and Setup
• data/sources.json — Documentation links
• assets/template-preflight-runtime-hook.sh — Shell hook template for runtime/tool version checks

Related Skills

• ../agents-subagents/SKILL.md — Agent creation
• ../agents-skills/SKILL.md — Skill creation
• ../ops-devops-platform/SKILL.md — CI/CD integration

Fact-Checking

• Use web search/web fetch to verify current external facts, versions, pricing, deadlines, regulations, or platform behavior before final answers.
• Prefer primary sources; report source links and dates for volatile information.
• If web access is unavailable, state the limitation and mark guidance as unverified.