diff --git a/services/pack_builder.py b/services/pack_builder.py
index 42735d7..df7c86c 100644
--- a/services/pack_builder.py
+++ b/services/pack_builder.py
@@ -1,6 +1,7 @@
 """
 Context pack builder for generating context bundles within token budgets
 """
+
 from typing import List, Dict, Any, Optional
 from loguru import logger
 
@@ -22,7 +23,7 @@ def build_context_pack(
         focus_paths: Optional[List[str]] = None,
         file_limit: int = DEFAULT_FILE_LIMIT,
         symbol_limit: int = DEFAULT_SYMBOL_LIMIT,
-        enable_deduplication: bool = True
+        enable_deduplication: bool = True,
     ) -> Dict[str, Any]:
         """
         Build a context pack from nodes within budget with deduplication and category limits.
@@ -47,22 +48,16 @@ def build_context_pack(
             logger.debug(f"After deduplication: {len(nodes)} unique nodes")
 
         # Step 2: Sort nodes by score
-        sorted_nodes = sorted(
-            nodes,
-            key=lambda x: x.get("score", 0),
-            reverse=True
-        )
+        sorted_nodes = sorted(nodes, key=lambda x: x.get("score", 0), reverse=True)
 
         # Step 3: Prioritize focus paths if provided
         if focus_paths:
             focus_nodes = [
-                n for n in sorted_nodes
+                n
+                for n in sorted_nodes
                 if any(fp in n.get("path", "") for fp in focus_paths)
             ]
-            other_nodes = [
-                n for n in sorted_nodes
-                if n not in focus_nodes
-            ]
+            other_nodes = [n for n in sorted_nodes if n not in focus_nodes]
             sorted_nodes = focus_nodes + other_nodes
 
         # Step 4: Apply category limits and budget constraints
@@ -80,7 +75,9 @@ def build_context_pack(
                 logger.debug(f"File limit reached ({file_limit}), skipping file nodes")
                 continue
             elif node_type == "symbol" and symbol_count >= symbol_limit:
-                logger.debug(f"Symbol limit reached ({symbol_limit}), skipping symbol nodes")
+                logger.debug(
+                    f"Symbol limit reached ({symbol_limit}), skipping symbol nodes"
+                )
                 continue
             elif node_type not in ["file", "symbol", "guideline"]:
                 # Unknown type, count as file
@@ -93,14 +90,13 @@ def build_context_pack(
                 "title": PackBuilder._extract_title(node.get("path", "")),
                 "summary": node.get("summary", ""),
                 "ref": node.get("ref", ""),
-                "extra": {
-                    "lang": node.get("lang"),
-                    "score": node.get("score", 0)
-                }
+                "extra": {"lang": node.get("lang"), "score": node.get("score", 0)},
             }
 
             # Estimate size (title + summary + ref + overhead)
-            item_size = len(item["title"]) + len(item["summary"]) + len(item["ref"]) + 50
+            item_size = (
+                len(item["title"]) + len(item["summary"]) + len(item["ref"]) + 50
+            )
             estimated_tokens = item_size // chars_per_token
 
             # Check if adding this item would exceed budget
@@ -129,10 +125,7 @@ def build_context_pack(
             "budget_limit": budget,
             "stage": stage,
             "repo_id": repo_id,
-            "category_counts": {
-                "file": file_count,
-                "symbol": symbol_count
-            }
+            "category_counts": {"file": file_count, "symbol": symbol_count},
         }
 
     @staticmethod
@@ -140,13 +133,16 @@ def _deduplicate_nodes(nodes: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
         """
         Remove duplicate nodes based on ref handle.
         If multiple nodes have the same ref, keep the one with highest score.
+        Nodes without a ref are preserved with a unique identifier.
         """
         seen_refs = {}
+        nodes_without_ref = []
 
         for node in nodes:
             ref = node.get("ref")
             if not ref:
-                # No ref, keep it (shouldn't happen normally)
+                # No ref, keep it in a separate list
+                nodes_without_ref.append(node)
                 continue
 
             # Check if we've seen this ref before
@@ -159,21 +155,23 @@ def _deduplicate_nodes(nodes: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
             else:
                 seen_refs[ref] = node
 
-        # Return deduplicated nodes
-        deduplicated = list(seen_refs.values())
+        # Combine deduplicated nodes with nodes without refs
+        deduplicated = list(seen_refs.values()) + nodes_without_ref
         removed_count = len(nodes) - len(deduplicated)
 
         if removed_count > 0:
             logger.debug(f"Removed {removed_count} duplicate nodes")
+        if nodes_without_ref:
+            logger.debug(f"Preserved {len(nodes_without_ref)} nodes without ref")
 
         return deduplicated
-    
+
     @staticmethod
     def _extract_title(path: str) -> str:
         """Extract title from path (last 2 segments)"""
-        parts = path.split('/')
+        parts = path.split("/")
         if len(parts) >= 2:
-            return '/'.join(parts[-2:])
+            return "/".join(parts[-2:])
         return path