Enhance: Multi-step iterative tasks

inference.py

@@ -44,7 +44,16 @@ BENCHMARK = "cloud_ops_env"
 SYSTEM_PROMPT = textwrap.dedent(
     """
     You are an expert Cloud SRE (Site Reliability Engineer). Your goal is to optimize cloud infrastructure
-    to meet the SLA requirements while minimizing costs.
+    to meet the SLA requirements while minimizing costs through ITERATIVE exploration.
+    
+    IMPORTANT: This is a multi-step optimization task. You must make 3-5 changes to find the optimal configuration.
+    Do NOT expect to get high rewards with just 1-2 changes!
+    
+    Strategy:
+    1. Analyze the current infrastructure and identify overloaded servers
+    2. Make incremental changes - upgrade one server at a time
+    3. Evaluate the result and adjust your next action
+    4. Continue until you find the optimal configuration
     
     Available instance types (cost per month, capacity):
     - t3.nano:   $3.60,  capacity 1.0
@@ -92,7 +101,7 @@ def step_env(message: str) -> dict:
     return resp.json()
 
 
-def build_user_prompt(obs_data: dict) -> str:
+def build_user_prompt(obs_data: dict, step_num: int) -> str:
     inventory = obs_data.get("inventory", [])
     metrics = obs_data.get("metrics", {})
     sla = obs_data.get("sla", {})
@@ -102,7 +111,7 @@ def build_user_prompt(obs_data: dict) -> str:
         for r in inventory
     ])
     
-    prompt = f"""Current Infrastructure:
+    prompt = f"""Current Infrastructure (Step {step_num}):
 {inv_str}
 
 Metrics:
@@ -115,6 +124,7 @@ SLA Requirements:
 
 Task: {obs_data.get('task_name', 'Optimize')} ({obs_data.get('difficulty', 'easy')})
 
+Remember: Make incremental changes. Evaluate result, then adjust. Aim for 3-5 changes total.
 Provide your next command:"""
     
     return prompt
@@ -178,7 +188,7 @@ def run_task(client: OpenAI, task_key: str, verbose: bool = False) -> dict:
             if done:
                 break
 
-            user_prompt = build_user_prompt(obs_data)
+            user_prompt = build_user_prompt(obs_data, step)
             response_text = call_model(client, user_prompt, history)
             history.append({"role": "assistant", "content": response_text})