chore: align inference.py with hackathon checklist

inference.py

@@ -23,9 +23,14 @@ except Exception:
 # Configuration
 # ---------------------------------------------------------------------------
 
-BASE_URL = os.getenv("BASE_URL", "http://localhost:8000")
-API_KEY = os.getenv("API_KEY") or os.getenv("HF_TOKEN")
-MODEL = "gpt-4o-mini"
+# Required checklist variables:
+# - API_BASE_URL and MODEL_NAME have defaults
+# - HF_TOKEN has no default
+# - LOCAL_IMAGE_NAME is optional
+API_BASE_URL = os.getenv("API_BASE_URL", "http://localhost:8000")
+MODEL_NAME = os.getenv("MODEL_NAME", "gpt-4o-mini")
+HF_TOKEN = os.getenv("HF_TOKEN")
+LOCAL_IMAGE_NAME = os.getenv("LOCAL_IMAGE_NAME")
 
 # List of task IDs to evaluate
 TASKS = os.getenv("TASKS", "task_1,task_2,task_3").split(",")
@@ -44,13 +49,13 @@ def _build_action(task_description: str, code_snippet: str) -> Dict[str, Any]:
         "confidence": 0.0,
     }
 
-    if not API_KEY:
+    if not HF_TOKEN:
         return fallback_action
 
     try:
         from openai import OpenAI  # Lazy import to avoid failing at module import time
 
-        client = OpenAI(api_key=API_KEY)
+        client = OpenAI(base_url=API_BASE_URL, api_key=HF_TOKEN)
     except Exception:
         return fallback_action
 
@@ -71,7 +76,7 @@ Respond ONLY with valid JSON, no markdown:
 
     try:
         response = client.chat.completions.create(
-            model=MODEL,
+            model=MODEL_NAME,
             messages=[{"role": "user", "content": prompt}],
             temperature=0.0,
         )
@@ -114,7 +119,7 @@ def run_task(task_id: str) -> float:
     score = 0.0
     steps = 1
 
-    reset_data = _safe_post_json(f"{BASE_URL}/reset", {"task_id": task_id}) or {}
+    reset_data = _safe_post_json(f"{API_BASE_URL}/reset", {"task_id": task_id}) or {}
     obs = reset_data.get("observation", {}) if isinstance(reset_data, dict) else {}
 
     code_snippet = obs.get("code_snippet", "")
@@ -122,9 +127,11 @@ def run_task(task_id: str) -> float:
     action = _build_action(str(task_description), str(code_snippet))
 
     # If stepping fails, we still emit structured output with reward=0.0
-    _safe_post_json(f"{BASE_URL}/step", {"action": action})
+    _safe_post_json(f"{API_BASE_URL}/step", {"action": action})
 
-    grader_data = _safe_get_json(f"{BASE_URL}/grader?task_id={task_id}&episode_id=baseline") or {}
+    grader_data = _safe_get_json(
+        f"{API_BASE_URL}/grader?task_id={task_id}&episode_id=baseline"
+    ) or {}
     if isinstance(grader_data, dict):
         try:
             score = float(grader_data.get("score", 0.0))