""" Orchestrator – Main controller for RRA v2.0. Responsibilities: - Interpret structured RRA JSON ("phase", "action_type", etc.). - Route to Omnisearch, TerminalAdapter, UniversalInstaller, HermesMemory. - Implement retries, timeouts, and error handling. - Build all real commands from templates or stored metadata – the LLM never directly controls shell strings unless explicitly allowed via config. This file assumes: - The model outputs {...} with a strict JSON schema. - Tool messages in the dataset mirror the real outputs from these methods. """ import asyncio import json import logging from typing import Any, Dict, Optional from core.omnisearch import Omnisearch from core.vision import VisionProcessor from core.terminal import TerminalAdapter from core.universal_installer import UniversalInstaller from core.hermes_memory import HermesMemory logger = logging.getLogger(__name__) class Orchestrator: def __init__( self, allow_llm_raw_terminal: bool = False, ) -> None: """ :param allow_llm_raw_terminal: If True, the orchestrator may expose a "terminal_raw" action_type that forwards commands directly from RRA JSON to TerminalAdapter. This is unsafe by default and should only be used when the operator explicitly wants that behavior. """ self.omnisearch = Omnisearch() self.vision = VisionProcessor() self.terminal_cache: Dict[str, TerminalAdapter] = {} self.installer = UniversalInstaller() self.memory = HermesMemory() self.allow_llm_raw_terminal = allow_llm_raw_terminal # ------------------------------------------------------------------------- # Low-level tool execution with retries + timeouts # ------------------------------------------------------------------------- async def execute_tool( self, tool_name: str, args: Dict[str, Any], retries: int = 3, timeout: int = 120, ) -> Any: """ Generic tool execution wrapper with retries and timeouts. """ for attempt in range(retries): try: result = await asyncio.wait_for( self._execute_tool_internal(tool_name, args), timeout=timeout, ) return result except asyncio.TimeoutError: logger.warning(f"Tool {tool_name} attempt {attempt + 1} timed out") error = {"status": "error", "reason": "timeout"} except Exception as e: logger.warning(f"Tool {tool_name} attempt {attempt + 1} failed: {e}") error = {"status": "error", "reason": str(e)} await asyncio.sleep(2 * (attempt + 1)) return error async def _execute_tool_internal(self, tool_name: str, args: Dict[str, Any]) -> Any: """ All actual tool calls are implemented here. """ if tool_name == "Omnisearch": # args: {"query": "...", "engines": [...], "max_results": int} results = await self.omnisearch.query(**args) return {"status": "success", "results": results} elif tool_name == "Vision": # args: {"mode": "ocr"/"describe"/"screenshot", ...} mode = args.pop("mode") method = getattr(self.vision, mode) if asyncio.iscoroutinefunction(method): out = await method(**args) else: out = method(**args) return {"status": "success", "result": out} elif tool_name == "Terminal": # args: {"target": {...}, "command": "..." } target = args["target"] command = args["command"] adapter = self._get_or_create_terminal(target) result = adapter.execute(command) result.setdefault("status", "success" if result.get("exit_code", 1) == 0 else "error") return result elif tool_name == "UniversalInstaller": # args: {"target": {...}, "tool": "nmap", "method": "auto"} self.installer.set_target(args["target"]) out = self.installer.install(args["tool"], args.get("method", "auto")) # For better dataset consistency, normalize to structured dict if isinstance(out, str): return {"status": "success", "detail": out} return out elif tool_name == "HermesMemory": # args: {"method": "kv_set"/"kv_get"/..., **method_args} method_name = args.pop("method") meth = getattr(self.memory, method_name) if asyncio.iscoroutinefunction(meth): return await meth(**args) else: return meth(**args) else: logger.error(f"Unknown tool: {tool_name}") return {"status": "error", "reason": f"Unknown tool: {tool_name}"} def _get_or_create_terminal(self, target: Dict[str, Any]) -> TerminalAdapter: """ Cache TerminalAdapter per target key (e.g., IP + OS) to avoid reconnecting every time. """ key = f"{target.get('ip')}|{target.get('os','linux')}" if key not in self.terminal_cache: self.terminal_cache[key] = TerminalAdapter(target) return self.terminal_cache[key] # ------------------------------------------------------------------------- # High-level: handle RRA JSON actions # ------------------------------------------------------------------------- async def handle_rra(self, rra_obj: Dict[str, Any]) -> Dict[str, Any]: """ Main entry for interpreting a single RRA JSON object (already parsed from ...). Returns a dict that should be serialized into a tool message and fed back to the LLM. """ phase = rra_obj.get("phase") if phase == "plan": # In many setups, plan phase does not require an immediate tool call. # You might choose to do a quick memory lookup here. return await self._handle_plan(rra_obj) elif phase == "action": return await self._handle_action(rra_obj) elif phase == "cache": return await self._handle_cache(rra_obj) else: return {"status": "error", "reason": f"Unknown phase: {phase}"} async def _handle_plan(self, rra_obj: Dict[str, Any]) -> Dict[str, Any]: """ Optionally consult HermesMemory on the plan (e.g., see if we've exploited similar targets before). This is a hook – you can expand it. """ # Example: you might embed the "plan" text and do a vector search # but that's optional. For now, we just acknowledge the plan. return {"status": "ok", "message": "plan_received"} async def _handle_action(self, rra_obj: Dict[str, Any]) -> Dict[str, Any]: action_type = rra_obj.get("action_type") if action_type == "search_exploit": query = rra_obj.get("query", "") engines = rra_obj.get("engines") max_results = rra_obj.get("max_results", 10) return await self.execute_tool( "Omnisearch", {"query": query, "engines": engines, "max_results": max_results}, ) if action_type == "install_dependency": target = rra_obj.get("target_info") or rra_obj.get("target") tool_name = rra_obj.get("tool_name") or rra_obj.get("package_name") method = rra_obj.get("method", "auto") return await self.execute_tool( "UniversalInstaller", {"target": target, "tool": tool_name, "method": method}, ) if action_type == "run_tool": return await self._handle_run_tool(rra_obj) if action_type == "terminal": return await self._handle_terminal(rra_obj) if action_type == "terminal_raw": # Unsafe path; only allowed if configured if not self.allow_llm_raw_terminal: return { "status": "error", "reason": "terminal_raw not allowed in this configuration", } return await self._handle_terminal_raw(rra_obj) return {"status": "error", "reason": f"Unknown action_type: {action_type}"} async def _handle_run_tool(self, rra_obj: Dict[str, Any]) -> Dict[str, Any]: """ Run a known exploit/tool stored in HermesMemory's arsenal. rra_obj should contain: - tool_key: logical key for the exploit/tool - run_context: "target" or "local" - target: target info dict if run_context == "target" - arguments: dict of arguments to fill into templates """ tool_key = rra_obj.get("tool_key") run_context = rra_obj.get("run_context", "target") target = rra_obj.get("target") arguments = rra_obj.get("arguments", {}) raw = self.memory.get_arsenal_tool_raw(tool_key) if not raw: return { "status": "error", "reason": f"Tool {tool_key} not found in arsenal", } # Here you parse metadata: path, template, etc. # For now we assume 'raw' is a string repr of a dict with a 'path' import ast try: meta = ast.literal_eval(raw) except Exception as e: return {"status": "error", "reason": f"Invalid arsenal metadata: {e}"} path = meta.get("path") if not path: return { "status": "error", "reason": f"Arsenal entry {tool_key} missing 'path'", } # Build command from template: you can extend this to use a more # sophisticated templating system. cmd = self._build_exploit_command(path, arguments) if run_context == "target": return await self.execute_tool( "Terminal", {"target": target, "command": cmd}, ) else: # local context local_target = {"os": "local"} return await self.execute_tool( "Terminal", {"target": local_target, "command": cmd}, ) def _build_exploit_command(self, path: str, arguments: Dict[str, Any]) -> str: """ Minimal example: assume most exploits are python scripts with args like --url and --cmd. In your real system, you should derive this from stored metadata per tool_key, not hardcode here. """ parts = ["python3", path] for k, v in arguments.items(): # simple --key value parts.append(f"--{k}") parts.append(str(v)) return " ".join(parts) async def _handle_terminal(self, rra_obj: Dict[str, Any]) -> Dict[str, Any]: """ Safe terminal action: use operation/command_name/templates, not raw shell. """ target = rra_obj.get("target") os_hint = rra_obj.get("os_hint", "linux") operation = rra_obj.get("operation") command_name = rra_obj.get("command_name") params = rra_obj.get("parameters", {}) cmd = self._build_terminal_command( os_hint=os_hint, operation=operation, command_name=command_name, params=params, ) return await self.execute_tool("Terminal", {"target": target, "command": cmd}) def _build_terminal_command( self, os_hint: str, operation: str, command_name: Optional[str], params: Dict[str, Any], ) -> str: """ Map (operation, command_name, params) to actual shell command. This is where you define your "terminal templates". For now, examples: - basic_enum + whoami/id => simple commands - file_read => cat/path """ # Example for basic enumeration if operation == "basic_enum": if os_hint.lower().startswith("win"): # Windows basic enum if command_name == "whoami": return "whoami" if command_name == "hostname": return "hostname" if command_name == "id": return "whoami /all" return "whoami" else: # Linux/Unix basic enum if command_name == "whoami": return "whoami" if command_name == "id": return "id" if command_name == "hostname": return "hostname" return "whoami" if operation == "file_read": path = params.get("path", "/etc/passwd") if os_hint.lower().startswith("win"): return f"type {path}" else: return f"cat {path}" # Custom templates can be added here, keyed by operation. # You might also load templates from HermesMemory or config. # Fallback: very conservative default return command_name or "whoami" async def _handle_terminal_raw(self, rra_obj: Dict[str, Any]) -> Dict[str, Any]: """ Unsafe raw terminal: directly execute the 'command' field. Only enabled if allow_llm_raw_terminal=True. """ target = rra_obj.get("target") cmd = rra_obj.get("command", "") return await self.execute_tool("Terminal", {"target": target, "command": cmd}) async def _handle_cache(self, rra_obj: Dict[str, Any]) -> Dict[str, Any]: """ Cache phase: store arsenal tools / exploit relations via HermesMemory. """ cache_type = rra_obj.get("cache_type") if cache_type == "arsenal_tool": tool_key = rra_obj.get("tool_key") path = rra_obj.get("path") sha256 = rra_obj.get("sha256", "") tags = rra_obj.get("tags", []) target_os = rra_obj.get("target_os", "unknown") service = rra_obj.get("service", "unknown") metadata = { "tags": tags, "target_os": target_os, "service": service, } self.memory.store_arsenal_tool( tool_key=tool_key, path=path, sha256=sha256, metadata=metadata, ) return {"status": "success", "message": "arsenal_tool_stored"} if cache_type == "exploit_success": cve = rra_obj.get("cve") service = rra_obj.get("service", "unknown") target_os = rra_obj.get("target_os", "unknown") tool_key = rra_obj.get("tool_key") self.memory.record_exploit_success( cve=cve, service=service, target_os=target_os, tool_key=tool_key, ) return {"status": "success", "message": "exploit_success_recorded"} return {"status": "error", "reason": f"Unknown cache_type: {cache_type}"}