from __future__ import annotations import asyncio import os from dataclasses import dataclass from typing import Any from backend.agents.blue_agent import BlueAgent from backend.agents.red_agent import RedAgent from backend.agents.scripted import ScriptedBlueAgent, ScriptedRedAgent from backend.core.event_bus import event_bus from backend.core.models import GOATAction, GOATObservation @dataclass class EpisodeResult: session_id: str winner: str score_red: float score_blue: float registry_delta: int input_layer_attacks: int = 0 output_layer_attacks: int = 0 input_blocks: int = 0 output_blocks: int = 0 class CoEvolutionOrchestrator: def __init__(self, env: Any, session_id: str, task_id: int = 4, red_agent: Any | None = None, blue_agent: Any | None = None): self.env = env self.session_id = session_id self.task_id = task_id self.red_agent = red_agent self.blue_agent = blue_agent async def run_episode(self, use_scripted: bool = True) -> EpisodeResult: turn_delay_s = float(os.getenv("BATTLE_TURN_DELAY_SECONDS", "0.35")) obs = self.env.reset(task_id=self.task_id, episode_id=self.session_id) if self.red_agent is not None and self.blue_agent is not None: red_agent = self.red_agent blue_agent = self.blue_agent elif use_scripted: red_agent = ScriptedRedAgent() blue_agent = ScriptedBlueAgent() else: raise RuntimeError("Non-scripted orchestrator path missing configured agents") while not obs.is_done: red_action = await red_agent.act(obs) obs, _, done, _ = self.env.step(GOATAction(role="red", red_payload=red_action)) if turn_delay_s > 0: await asyncio.sleep(turn_delay_s) if done: break blue_action = await blue_agent.act(obs) obs, _, done, _ = self.env.step(GOATAction(role="blue", blue_payload=blue_action)) if turn_delay_s > 0: await asyncio.sleep(turn_delay_s) else: await asyncio.sleep(0) winner = "red" if obs.red_cumulative_reward >= obs.blue_cumulative_reward else "blue" env_state = self.env.state event_bus.emit( "cycle_episode_done", session_id=self.session_id, winner=winner, input_blocks=env_state.input_blocks, output_blocks=env_state.output_blocks, ) return EpisodeResult( session_id=self.session_id, winner=winner, score_red=obs.red_cumulative_reward, score_blue=obs.blue_cumulative_reward, registry_delta=obs.emergent_count, input_layer_attacks=env_state.input_layer_attacks, output_layer_attacks=env_state.output_layer_attacks, input_blocks=env_state.input_blocks, output_blocks=env_state.output_blocks, )