Filter Agent Input with a Guardrail

Implement guardrail_check(request, allowed_tools, blocked_patterns) that validates an agent's tool call request before execution. request is a dict: {"tool": str, "args": dict}. allowed_tools is a list of strings (tool names that are permitted). blocked_patterns is a list of strings (substrings that must not appear in any argument value). Rules (apply in this order): 1. If request["tool"] is not in allowed_tools, return {"allowed": False, "reason": "tool not allowed", "tool": request["tool"]}. 2. If any argument value (converted to string via str()) contains any blocked pattern (case-insensitive check), return {"allowed": False, "reason": "blocked pattern detected", "tool": request["tool"], "pattern": <matched_pattern>}. Return the first blocked pattern found. 3. If all checks pass, return {"allowed": True, "reason": "allowed", "tool": request["tool"]}. Edge cases: - If request is None or missing "tool" key, return {"allowed": False, "reason": "invalid request"}. - If allowed_tools is empty, no tool is allowed. - If blocked_patterns is empty, no pattern filtering is applied.