import os import json from openai import OpenAI from openenv_core import make # Match the exact format required by the hackathon checklist API_BASE_URL = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1") MODEL_NAME = os.getenv("MODEL_NAME", "google/gemma-4-31B-it") HF_TOKEN = os.getenv("HF_TOKEN") # NO DEFAULT HERE client = OpenAI(base_url=API_BASE_URL, api_key=HF_TOKEN) def run_inference(task_name): env = make("data-privacy-env") obs = env.reset(task_name=task_name) print(f"[START] task={task_name} env=DataPrivacyEnv model={MODEL_NAME}") done = False step = 0 total_reward = 0 rewards = [] while not done and step < 10: step += 1 prompt = f""" You are a Data Privacy Agent. Task Goal: {obs.current_task_goal} Last Observation: {obs.response_data} Respond ONLY with a JSON object: {{"method": "GET/PATCH/DELETE", "endpoint": "/route", "payload": {{...}}}} """ try: response = client.chat.completions.create( model=MODEL_NAME, messages=[{"role": "user", "content": prompt}], temperature=0.0 ) action_json = response.choices[0].message.content.strip() # Clean potential markdown backticks if "```json" in action_json: action_json = action_json.split("```json")[1].split("```")[0].strip() action_dict = json.loads(action_json) obs, reward, done, error = env.step(action_dict) except Exception as e: action_json = json.dumps({"error": f"Failed to parse JSON: {str(e)}"}) reward, done, error = 0.0, False, str(e) total_reward += reward rewards.append(f"{reward:.2f}") print(f"[STEP] step={step} action={action_json} reward={reward:.2f} done={str(done).lower()} error={str(error).lower()}") success = total_reward >= 0.99 print(f"[END] success={str(success).lower()} steps={step} score={total_reward:.3f} rewards={','.join(rewards)}") if __name__ == "__main__": task = os.getenv("TASK_NAME", "easy_log_redaction") run_inference(task)