Update inference.py

inference.py

@@ -1,28 +1,50 @@
 from env import EmailEnv
+import os
+from openai import OpenAI
+
+# ✅ Use validator proxy (MANDATORY)
+client = OpenAI(
+    api_key=os.environ["API_KEY"],
+    base_url=os.environ["API_BASE_URL"],
+)
+
+def call_llm(text):
+    res = client.chat.completions.create(
+        model="openai/gpt-4.1-mini",
+        messages=[
+            {
+                "role": "system",
+                "content": "Classify the email into spam, important, or normal and give a short reply in JSON format: {label, reply}"
+            },
+            {"role": "user", "content": text},
+        ],
+        temperature=0,
+    )
+    return res.choices[0].message.content
+
 
 env = EmailEnv()
 
-# START block
 print("[START] task=email_triage", flush=True)
 
-# Reset environment
 obs = env.reset()
-
-# Simple agent logic
 text = obs["text"]
 
-if "meeting" in text.lower():
-    action = {"label": "important", "reply": "Sure, I will attend the meeting."}
-elif "lottery" in text.lower():
-    action = {"label": "spam", "reply": "This looks like spam."}
-else:
-    action = {"label": "normal", "reply": "Sounds good!"}
+# ✅ CALL LLM (THIS FIXES YOUR ERROR)
+response = call_llm(text)
+
+# simple parsing (safe fallback)
+import json
+try:
+    parsed = json.loads(response)
+    action = {
+        "label": parsed.get("label", "normal"),
+        "reply": parsed.get("reply", "OK"),
+    }
+except:
+    action = {"label": "normal", "reply": "OK"}
 
-# Step
 reward = env.step(action)
 
-# STEP block
 print(f"[STEP] reward={reward}", flush=True)
-
-# END block
 print(f"[END] score={reward}", flush=True)