inference.py

import asyncio
import os
from typing import List
import httpx
from openai import OpenAI

API_BASE_URL   = os.environ.get("API_BASE_URL", "https://api-inference.huggingface.co/v1")
MODEL_NAME     = os.environ.get("MODEL_NAME", "mistralai/Mistral-7B-Instruct-v0.3")
HF_TOKEN       = os.environ.get("HF_TOKEN", "")
ENV_URL        = os.environ.get("ENV_URL", "http://localhost:7860")
OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY", HF_TOKEN)

TASK_NAME  = "json_repair_all_tasks"
BENCHMARK  = "json-repair-env"
MAX_STEPS  = 5
MAX_TOTAL_REWARD     = 5.0
SUCCESS_SCORE_THRESHOLD = 0.7


# ✅ FIXED: Plain text format exactly as required
def log_start(task, env, model):
    print(f"[START] task={task} env={env} model={model}", flush=True)

def log_step(step, action, reward, done, error):
    # Sanitize action to ensure it's on a single line and doesn't break parsing
    clean_action = str(action).replace("\n", " ").replace("\r", "")[:80]
    print(f"[STEP] step={step} action={clean_action} reward={reward} done={done} error={error}", flush=True)

def log_end(task, success, steps, score, rewards):
    print(f"[END] task={task} success={success} steps={steps} score={score} rewards={rewards}", flush=True)


class EnvClient:
    def __init__(self, base_url):
        self.base_url = base_url.rstrip("/")
        self.http = httpx.AsyncClient(timeout=60.0)
        self.last_result = None

    async def reset(self):
        resp = await self.http.post(f"{self.base_url}/reset")
        resp.raise_for_status()
        self.last_result = resp.json()
        return self

    async def step(self, repaired_json, explanation=""):
        resp = await self.http.post(
            f"{self.base_url}/step",
            json={"repaired_json": repaired_json, "explanation": explanation}
        )
        resp.raise_for_status()
        self.last_result = resp.json()
        return self

    async def close(self):
        await self.http.aclose()

    @property
    def observation(self):
        return self.last_result.get("observation", {})

    @property
    def reward(self):
        return float(self.last_result.get("reward", 0.0))

    @property
    def done(self):
        return bool(self.last_result.get("done", False))


def get_model_repair(client, broken_json, hint, history):
    history_text = "\n".join(history[-3:]) if history else "None"
    prompt = f"""You are a JSON repair expert. Fix the broken JSON below.

BROKEN JSON:
{broken_json}

HINT: {hint}

HISTORY:
{history_text}

Rules:
- Output ONLY the repaired JSON string, nothing else
- No markdown, no explanation, no code blocks
- Must be valid JSON

Repaired JSON:"""

    response = client.chat.completions.create(
        model=MODEL_NAME,
        messages=[{"role": "user", "content": prompt}],
        max_tokens=500,
        temperature=0.1
    )
    return response.choices[0].message.content.strip()


async def main():
    client = OpenAI(api_key=OPENAI_API_KEY or HF_TOKEN, base_url=API_BASE_URL)
    env = EnvClient(ENV_URL)

    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()
        obs = result.observation

        for step in range(1, MAX_STEPS + 1):
            if result.done:
                break

            broken    = obs.get("broken_json", "")
            hint      = obs.get("hint", "")
            task_name = obs.get("task_name", f"task_{step}")

            action = get_model_repair(client, broken, hint, history)
            result = await env.step(repaired_json=action, explanation="")

            obs    = result.observation
            reward = result.reward
            done   = result.done

            rewards.append(reward)
            steps_taken = step

            log_step(step=step, action=action[:80], reward=reward, done=done, error=None)
            history.append(f"Step {step} [{task_name}]: reward={reward:.2f}")

            if done:
                break

        # Clip score to (0.1, 0.9) instead of (0.0, 1.0)
        # This is required by the validator to ensure non-binary scoring
        raw_score = sum(rewards) / MAX_TOTAL_REWARD
        score   = round(0.1 + (raw_score * 0.8), 4)
        success = score >= SUCCESS_SCORE_THRESHOLD

    except Exception as e:
        print(f"[ERROR] {e}", flush=True)
    finally:
        try:
            await env.close()
        except Exception as e:
            pass # Silent cleanup
        log_end(task=TASK_NAME, success=success, steps=steps_taken, score=score, rewards=rewards)


if __name__ == "__main__":
    asyncio.run(main())