diff --git a/src/pyob/core_utils.py b/src/pyob/core_utils.py
index 5e75be2..b2581b0 100644
--- a/src/pyob/core_utils.py
+++ b/src/pyob/core_utils.py
@@ -393,6 +393,7 @@ def _stream_single_llm(
         input_tokens = len(prompt) // 4
         first_chunk_received = [False]
         gen_start_time = time.time()
+        is_cloud = os.environ.get("GITHUB_ACTIONS") == "true"
 
         def spinner():
             spinner_chars = ["⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏"]
@@ -420,17 +421,18 @@ def on_chunk():
                 sys.stdout.write("\r\033[K")
                 sys.stdout.flush()
                 source = f"Gemini ...{key[-4:]}" if key else "GitHub Models"
+                if not key and not is_cloud:
+                    source = "Local Ollama"
                 print(f"🤖 AI Output ({source}): ", end="", flush=True)
 
         response_text = ""
         try:
             if key is not None:
                 response_text = self.stream_gemini(prompt, key, on_chunk)
-            elif os.environ.get("GITHUB_ACTIONS") == "true":
-                # In cloud, 'None' key always means use GitHub Models
+            elif is_cloud:
+                # Force GitHub Models in cloud, skip Ollama entirely
                 response_text = self.stream_github_models(prompt, on_chunk)
             else:
-                # Only iMac uses Ollama
                 response_text = self.stream_ollama(prompt, on_chunk)
         except Exception as e:
             first_chunk_received[0] = True
@@ -448,82 +450,59 @@ def get_valid_llm_response(self, prompt: str, validator, context: str = "") -> s
         attempts = 0
         is_cloud = os.environ.get("GITHUB_ACTIONS") == "true"
 
-        logger.info(
-            f"📊 Engine check: Found {len(self.key_cooldowns)} Gemini API keys."
-        )
-
         while True:
             key = None
             now = time.time()
-            available_keys = [
-                k for k, cooldown in self.key_cooldowns.items() if now > cooldown
-            ]
+            available_keys = [k for k, cd in self.key_cooldowns.items() if now > cd]
 
-            # --- 1. ENGINE SELECTION LOGIC ---
+            # 1. Select Engine
             if available_keys:
-                # Use Gemini (Primary)
                 key = available_keys[attempts % len(available_keys)]
                 logger.info(
                     f"Attempting Gemini API Key {attempts % len(available_keys) + 1}/{len(available_keys)}"
                 )
-                response_text = self._stream_single_llm(
-                    prompt, key=key, context=context
-                )
-
             elif is_cloud:
-                # ALL GEMINI KEYS LIMITED -> TRY GITHUB MODELS (Secondary)
                 logger.warning(
                     "⏳ Gemini keys limited. Pivoting to GitHub Models (Phi-4)..."
                 )
-                response_text = self._stream_single_llm(
-                    prompt, key=None, context=context
-                )
-
-                # If GitHub Models ALSO fails or returns an error
-                if not response_text or response_text.startswith("ERROR_CODE_"):
-                    logger.warning(
-                        "🚫 All Cloud AI engines exhausted. Sleeping 5 minutes for cooldown..."
-                    )
-                    time.sleep(300)
-                    continue
-
+                # key remains None, which triggers stream_github_models in _stream_single_llm
             else:
-                # LOCAL IMAC -> FALLBACK TO OLLAMA
                 logger.info("🏠 Using Local Ollama Engine...")
-                response_text = self._stream_single_llm(
-                    prompt, key=None, context=context
-                )
 
-            # --- 2. RESPONSE VALIDATION & ROTATION ---
+            response_text = self._stream_single_llm(prompt, key=key, context=context)
 
-            # Handle standard Gemini Rate Limit (429)
+            # 2. Handle Rate Limits (429)
             if response_text.startswith("ERROR_CODE_429"):
                 if key:
                     self.key_cooldowns[key] = time.time() + 1200
                     logger.warning(f"⚠️ Key {key[-4:]} rate-limited. Rotating...")
+                else:
+                    # If GitHub Models is also rate-limited
+                    logger.warning(
+                        "🚫 All cloud engines limited. Sleeping 5 minutes..."
+                    )
+                    time.sleep(300)
                 attempts += 1
                 continue
 
-            # Handle Empty or Generic Error Responses
+            # 3. Handle Empty or Error Responses (STOPS THE INFINITE LOOP)
             if not response_text or response_text.startswith("ERROR_CODE_"):
                 logger.warning(
-                    f"⚠️ LLM Error detected ({response_text[:20]}...). Retrying in 10s..."
+                    "⚠️ API Error/Empty Response. Sleeping 10s before retry..."
                 )
-                time.sleep(10)
+                time.sleep(10)  # MANDATORY SLEEP to prevent tight-looping
                 attempts += 1
                 continue
 
-            # Check if the AI's content matches our XML/Format rules
+            # 4. Final Validation
             if validator(response_text):
-                # SUCCESS BREATHER: Stay under RPM limits
                 if is_cloud:
-                    time.sleep(2)
+                    time.sleep(2)  # Success breather
                 return response_text
-            else:
-                logger.warning("LLM response failed internal validation. Retrying...")
-                if is_cloud:
-                    time.sleep(5)
-                attempts += 1
+
+            logger.warning("LLM response failed internal validation. Retrying in 5s...")
+            time.sleep(5)
+            attempts += 1
 
     def _get_user_prompt_augmentation(self, initial_text: str = "") -> str:
         import tempfile