agent-governance（Agent 治理）

Agent Governance Skill

Implement security and governance controls for custom agents using hooks.

Purpose

Design and implement hook-based governance that controls agent permissions, blocks dangerous operations, and provides audit trails.

When to Use

• Building agents with security requirements
• Need to block access to sensitive files/operations
• Require audit logging of agent actions
• Implementing permission policies

Hook Architecture

Hook Types

Documentation Verification: Hook event types (PreToolUse, PostToolUse, etc.) are Claude Code internal types. For authoritative current types, verify via hook-management skill → docs-management.

Hook Function Signature

async def hook_function(
    input_data: dict,     # Tool call information
    tool_use_id: str,     # Unique tool call ID
    context: HookContext  # Session context
) -> dict:
    # Return empty dict to allow
    # Return with permissionDecision to block
    pass

Design Process

Step 1: Identify Security Requirements

Questions to answer:

• What files should be blocked? (e.g.,.env, credentials)
• What commands should be blocked? (e.g., rm -rf)
• What operations need logging?
• What tool access needs validation?

Step 2: Design Hook Matchers

from claude_agent_sdk import HookMatcher

hooks = {
    "PreToolUse": [
        # Match specific tool
        HookMatcher(matcher="Read", hooks=[block_sensitive_files]),

        # Match all tools
        HookMatcher(hooks=[log_all_tool_usage]),
    ],
    "PostToolUse": [
        HookMatcher(hooks=[audit_tool_results]),
    ],
}

Step 3: Implement Hook Functions

Security Hook (Block Pattern):

BLOCKED_PATTERNS = [".env", "credentials", "secrets", ".pem", ".key"]

async def block_sensitive_files(
    input_data: dict,
    tool_use_id: str,
    context: HookContext
) -> dict:
    tool_name = input_data.get("tool_name", "")
    tool_input = input_data.get("tool_input", {})

    # Only check file operations
    if tool_name not in ["Read", "Write", "Edit"]:
        return {}

    file_path = tool_input.get("file_path", "")

    # Check for blocked patterns
    for pattern in BLOCKED_PATTERNS:
        if pattern in file_path.lower():
            return {
                "hookSpecificOutput": {
                    "hookEventName": "PreToolUse",
                    "permissionDecision": "deny",
                    "permissionDecisionReason": f"Security: Access to {pattern} files blocked",
                }
            }

    return {}  # Allow

Audit Hook (Log Pattern):

async def log_all_tool_usage(
    input_data: dict,
    tool_use_id: str,
    context: HookContext
) -> dict:
    tool_name = input_data.get("tool_name", "")
    tool_input = input_data.get("tool_input", {})
    session_id = input_data.get("session_id", "unknown")

    log_entry = {
        "timestamp": datetime.now().isoformat(),
        "session_id": session_id,
        "tool": tool_name,
        "input": tool_input,
    }

    # Write to audit log
    log_file = Path("audit_logs") / f"{session_id}.jsonl"
    log_file.parent.mkdir(exist_ok=True)
    with open(log_file, "a") as f:
        f.write(json.dumps(log_entry) + "\n")

    return {}  # Always allow (logging only)

Validation Hook (Conditional Pattern):

async def validate_bash_commands(
    input_data: dict,
    tool_use_id: str,
    context: HookContext
) -> dict:
    tool_name = input_data.get("tool_name", "")

    if tool_name != "Bash":
        return {}

    command = input_data.get("tool_input", {}).get("command", "")

    DANGEROUS_PATTERNS = [
        r"rm\s+-rf\s+/",
        r"sudo\s+rm",
        r":(){ :|:& };:",  # Fork bomb
    ]

    for pattern in DANGEROUS_PATTERNS:
        if re.search(pattern, command):
            return {
                "hookSpecificOutput": {
                    "hookEventName": "PreToolUse",
                    "permissionDecision": "deny",
                    "permissionDecisionReason": f"Security: Dangerous command blocked",
                }
            }

    return {}

Step 4: Configure Agent with Hooks

hooks = {
    "PreToolUse": [
        HookMatcher(matcher="Read", hooks=[block_sensitive_files]),
        HookMatcher(matcher="Bash", hooks=[validate_bash_commands]),
        HookMatcher(hooks=[log_all_tool_usage]),
    ],
    "PostToolUse": [
        HookMatcher(hooks=[audit_tool_results]),
    ],
}

options = ClaudeAgentOptions(
    system_prompt=system_prompt,
    model="opus",
    hooks=hooks,
)

Common Governance Patterns

File Access Control

ALLOWED_DIRECTORIES = ["src/", "docs/", "tests/"]

async def restrict_file_access(input_data, tool_use_id, context) -> dict:
    file_path = input_data.get("tool_input", {}).get("file_path", "")

    if not any(file_path.startswith(d) for d in ALLOWED_DIRECTORIES):
        return deny_response("Access restricted to allowed directories")

    return {}

Rate Limiting

tool_call_counts = defaultdict(int)
RATE_LIMITS = {"WebFetch": 10, "Bash": 50}

async def rate_limit_tools(input_data, tool_use_id, context) -> dict:
    tool_name = input_data.get("tool_name", "")

    if tool_name in RATE_LIMITS:
        tool_call_counts[tool_name] += 1
        if tool_call_counts[tool_name] > RATE_LIMITS[tool_name]:
            return deny_response(f"Rate limit exceeded for {tool_name}")

    return {}

Content Filtering

BLOCKED_CONTENT = ["api_key", "password", "secret"]

async def filter_output_content(input_data, tool_use_id, context) -> dict:
    tool_output = input_data.get("tool_output", "")

    for blocked in BLOCKED_CONTENT:
        if blocked.lower() in tool_output.lower():
            return deny_response("Output contains sensitive content")

    return {}

Output Format

When designing governance:

## Governance Design

**Agent:** [agent name]
**Security Level:** [low/medium/high]

### Requirements

- [ ] Requirement 1
- [ ] Requirement 2

### Hooks

**PreToolUse:**
| Matcher | Hook | Purpose |
| --- | --- | --- |
| Read | block_sensitive_files | Block .env, credentials |
| Bash | validate_commands | Block dangerous commands |
| * | log_usage | Audit all tool calls |

**PostToolUse:**
| Matcher | Hook | Purpose |
| --- | --- | --- |
| * | audit_results | Log tool outputs |

### Implementation

[Hook function implementations]

### Test Scenarios

| Scenario | Expected | Actual |
| --- | --- | --- |
| Read .env file | Blocked |  |
| Read src/main.py | Allowed |  |
| rm -rf / | Blocked |  |

Design Checklist

• Security requirements identified
• File access controls defined
• Command validation rules defined
• Audit logging implemented
• Hook matchers configured
• Test scenarios documented
• Error messages are helpful

Key Insight

"Hooks enable governance and permission checks in custom agents."

Hooks work for both main agent and subagents spawned via Task tool.

Cross-References

• @custom-agent-design skill - Agent design workflow
• @core-four-custom.md - Governance in Core Four
• @hook-management skill - Hook management patterns

Version History

• v1.0.0 (2025-12-26): Initial release

Last Updated

Date: 2025-12-26 Model: claude-opus-4-5-20251101