openclaw - 💡(How to fix) Fix Feature: Exec hook plugin system for command interception

Summary

Add a plugin hook system that allows agents to intercept and optionally block exec commands before they run. This enables policy-level enforcement of tool usage (e.g., blocking manual grep -r when MCP tools should be used instead).

Use Case

Agents following strict workflows (like TDD or MCP-first policies) currently rely on self-discipline to avoid running blocked commands. A pre-exec hook would:

1. Intercept commands before execution
2. Match against configurable patterns (regex or glob)
3. Block with a clear error message redirecting to the correct tool
4. Log violations for audit trails

Example Configuration

plugins:
  exec-hooks:
    pre-exec:
      - pattern: "grep -r"
        action: block
        message: "Use codebase-navigator__semantic_search instead"
        log: true
      - pattern: "npm test|vitest|jest"
        action: block  
        message: "Use test-orchestrator MCP tools instead"

Proposed API

interface ExecHook {
  pattern: string | RegExp;
  action: 'block' | 'warn' | 'allow';
  message?: string;
  log?: boolean;
  alternatives?: string[];
}

Benefits

• True enforcement vs policy documents agents might ignore
• Audit trail of attempted violations
• Graceful redirection with clear guidance to correct tools
• Per-agent configurability via agent policy files

Notes

This came up while implementing MCP tool enforcement. Currently using AGENTS.md policy with self-blocking rules, but technical enforcement would be more reliable.

Submitted by Tico 🌴 on behalf of @dab02