""" CodeSensei — Inference Script (OpenEnv Hackathon) =================================================== MANDATORY ENV VARS: API_BASE_URL The API endpoint for the LLM. MODEL_NAME The model identifier to use for inference. HF_TOKEN Your Hugging Face / API key. LOCAL_IMAGE_NAME Docker image name for from_docker_image() STDOUT FORMAT (must be exact): [START] task= env= model= [STEP] step= action= reward=<0.00> done= error= [END] success= steps= score= rewards= """ import asyncio import os import re import textwrap from typing import List, Optional from openai import OpenAI # --------------------------------------------------------------------------- # Import our OpenEnv environment (mirrors: from my_env_v4 import ...) # --------------------------------------------------------------------------- from env import CodeSenseiEnv, CodeSenseiAction # --------------------------------------------------------------------------- # Env vars (submission contract — matches sample exactly) # --------------------------------------------------------------------------- IMAGE_NAME = os.getenv("LOCAL_IMAGE_NAME") API_KEY = os.getenv("HF_TOKEN") or os.getenv("API_KEY") API_BASE_URL = os.getenv("API_BASE_URL", "https://api.groq.com/openai/v1") MODEL_NAME = os.getenv("MODEL_NAME", "llama-3.3-70b-versatile") TASK_NAME = os.getenv("CODESENSEI_TASK", "code-debug") BENCHMARK = os.getenv("CODESENSEI_BENCHMARK", "codesensei") MAX_STEPS = 6 TEMPERATURE = 0.7 MAX_TOKENS = 512 SUCCESS_SCORE_THRESHOLD = 0.5 # Max total reward: best case +2.0 per step, 6 steps MAX_TOTAL_REWARD = MAX_STEPS * 2.0 # --------------------------------------------------------------------------- # System prompt # --------------------------------------------------------------------------- SYSTEM_PROMPT = textwrap.dedent("""\ You are CodeSensei, an expert Python debugger. TASK: You are given a buggy Python function and its failing test cases. Find the bug and return ONLY the corrected function. RULES: 1. Return ONLY the corrected Python function — no explanations. 2. Keep the same function name and signature. 3. Wrap your code in ```python ... ``` markers. 4. Think step-by-step: identify failing tests, find the bug, then fix it. EXAMPLE: Given: ```python def add(a, b): return a - b ``` Return: ```python def add(a, b): return a + b ``` """).strip() # --------------------------------------------------------------------------- # Structured logging — [START], [STEP], [END] (exact format from sample) # --------------------------------------------------------------------------- def log_start(task: str, env: str, model: str) -> None: print(f"[START] task={task} env={env} model={model}", flush=True) def log_step(step: int, action: str, reward: float, done: bool, error: Optional[str]) -> None: error_val = error if error else "null" done_val = str(done).lower() print( f"[STEP] step={step} action={action} reward={reward:.2f} done={done_val} error={error_val}", flush=True, ) def log_end(success: bool, steps: int, score: float, rewards: List[float]) -> None: rewards_str = ",".join(f"{r:.2f}" for r in rewards) print(f"[END] success={str(success).lower()} steps={steps} score={score:.3f} rewards={rewards_str}", flush=True) # --------------------------------------------------------------------------- # Helpers # --------------------------------------------------------------------------- def extract_code(text: str) -> str: """Extract code from markdown fences or raw def statement.""" match = re.search(r"```(?:python)?\s*\n(.*?)```", text, re.DOTALL) if match: return match.group(1).strip() match = re.search(r"(def \w+\(.*?\):.*)", text, re.DOTALL) if match: return match.group(1).strip() return text.strip() def build_user_prompt(step: int, buggy_code: str, feedback: str, history: List[str]) -> str: history_block = "\n".join(history[-4:]) if history else "None" return textwrap.dedent(f"""\ Step: {step} Buggy code: ```python {buggy_code} ``` Last feedback: {feedback} Previous attempts: {history_block} Send your corrected function. """).strip() def get_model_message(client: OpenAI, step: int, buggy_code: str, feedback: str, history: List[str]) -> str: user_prompt = build_user_prompt(step, buggy_code, feedback, history) try: completion = client.chat.completions.create( model=MODEL_NAME, messages=[ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": user_prompt}, ], temperature=TEMPERATURE, max_tokens=MAX_TOKENS, stream=False, ) text = (completion.choices[0].message.content or "").strip() return text if text else "hello" except Exception as exc: print(f"[DEBUG] Model request failed: {exc}", flush=True) return "hello" # --------------------------------------------------------------------------- # Main — follows sample structure exactly # --------------------------------------------------------------------------- async def main() -> None: client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY) # OpenEnv standard: start env from Docker image env = await CodeSenseiEnv.from_docker_image(IMAGE_NAME) history: List[str] = [] rewards: List[float] = [] steps_taken = 0 score = 0.0 success = False log_start(task=TASK_NAME, env=BENCHMARK, model=MODEL_NAME) try: result = await env.reset() # OpenEnv.reset() last_feedback = result.observation.feedback or "All tests failing — fix the bug." last_reward = 0.0 buggy_code = result.observation.buggy_code for step in range(1, MAX_STEPS + 1): if result.done: break message = get_model_message(client, step, buggy_code, last_feedback, history) proposed_fix = extract_code(message) result = await env.step(CodeSenseiAction(proposed_fix=proposed_fix)) obs = result.observation reward = result.reward or 0.0 done = result.done error = obs.error_output if obs.error_output else None rewards.append(reward) steps_taken = step last_feedback = obs.feedback or f"{obs.tests_passed}/{obs.tests_total} tests passing" last_reward = reward # Action for log: just the function name fix summary action_str = proposed_fix.replace("\n", "\\n")[:120] log_step(step=step, action=action_str, reward=reward, done=done, error=error) history.append(f"Step {step}: {obs.tests_passed}/{obs.tests_total} tests -> reward {reward:+.2f}") if done: break # Score strictly bounded to (0, 1) per Phase 2 requirements score = sum(rewards) / MAX_TOTAL_REWARD if MAX_TOTAL_REWARD > 0 else 0.01 score = min(max(score, 0.01), 0.99) # strict bounds success = score >= SUCCESS_SCORE_THRESHOLD finally: try: await env.close() except Exception as e: print(f"[DEBUG] env.close() error (container cleanup): {e}", flush=True) log_end(success=success, steps=steps_taken, score=score, rewards=rewards) if __name__ == "__main__": asyncio.run(main())