diff --git a/service/app/agents/graph_builder.py b/service/app/agents/graph_builder.py
index d5234917..c0726c15 100644
--- a/service/app/agents/graph_builder.py
+++ b/service/app/agents/graph_builder.py
@@ -334,15 +334,21 @@ async def _build_llm_node(self, config: GraphNodeConfig) -> NodeFunction:
         async def llm_node(state: StateDict | BaseModel) -> StateDict:
             logger.info(f"[LLM Node: {config.id}] Starting execution")
 
-            # Convert state to dict (handles both dict and Pydantic BaseModel)
+            # Get messages BEFORE converting state to dict to preserve BaseMessage types
+            # model_dump() loses tool_call_id and other message-specific fields
+            if isinstance(state, BaseModel):
+                messages: list[BaseMessage] = list(getattr(state, "messages", []))
+            else:
+                messages = list(state.get("messages", []))
+
+            # Convert state to dict for template rendering (but we already have messages)
             state_dict = self._state_to_dict(state)
-            messages: list[BaseMessage] = list(state_dict.get("messages", []))
 
             # Render prompt template
             prompt = self._render_template(llm_config.prompt_template, state_dict)
 
             # Build messages for LLM
-            llm_messages = messages + [HumanMessage(content=prompt)]
+            llm_messages = list(messages) + [HumanMessage(content=prompt)]
 
             # Invoke LLM (using pre-created configured_llm)
             response = await configured_llm.ainvoke(llm_messages)
diff --git a/service/app/core/chat/history.py b/service/app/core/chat/history.py
index 88906d50..25b4c11d 100644
--- a/service/app/core/chat/history.py
+++ b/service/app/core/chat/history.py
@@ -63,8 +63,11 @@ async def load_conversation_history(db: AsyncSession, topic: "TopicModel") -> li
                     history.append(tool_messages[0])
             # Skip unknown roles
 
-        logger.info(f"Length of history: {len(history)}")
-        return history
+        # Validate and filter messages before returning
+        validated_history = _validate_and_filter_messages(history)
+
+        logger.info(f"Loaded {len(history)} messages, {len(validated_history)} after validation")
+        return validated_history
 
     except Exception as e:
         logger.warning(f"Failed to load DB chat history for topic {getattr(topic, 'id', None)}: {e}")
@@ -180,9 +183,16 @@ def _build_tool_messages(
                 return None, num_tool_calls + 1
 
         elif formatted_content.get("event") == ChatEventType.TOOL_CALL_RESPONSE:
+            tool_call_id = formatted_content.get("toolCallId")
+
+            # Validate tool_call_id - must be a non-empty string for LangChain
+            if not tool_call_id or not isinstance(tool_call_id, str):
+                logger.warning(f"Skipping tool response with invalid tool_call_id: {tool_call_id!r}")
+                return None, num_tool_calls
+
             message = ToolMessage(
-                content=formatted_content["result"],
-                tool_call_id=formatted_content["toolCallId"],
+                content=formatted_content.get("result", ""),
+                tool_call_id=tool_call_id,
             )
             return message, num_tool_calls - 1
 
@@ -190,3 +200,57 @@ def _build_tool_messages(
         logger.warning(f"Failed to parse tool message content: {e}")
 
     return None, num_tool_calls
+
+
+def _validate_and_filter_messages(messages: list[BaseMessage]) -> list[BaseMessage]:
+    """
+    Validate and filter messages to ensure LangChain compatibility.
+
+    This function:
+    1. Removes ToolMessages with invalid tool_call_id
+    2. Removes orphaned ToolMessages without matching AIMessage tool_calls
+    3. Logs warnings for filtered messages
+
+    Args:
+        messages: List of LangChain messages loaded from history
+
+    Returns:
+        Filtered list of valid messages
+    """
+    # Collect all valid tool_call_ids from AIMessages
+    valid_tool_call_ids: set[str] = set()
+    for msg in messages:
+        if isinstance(msg, AIMessage) and hasattr(msg, "tool_calls") and msg.tool_calls:
+            for tc in msg.tool_calls:
+                tc_id = tc.get("id")
+                if tc_id:
+                    valid_tool_call_ids.add(tc_id)
+
+    filtered: list[BaseMessage] = []
+    skipped_count = 0
+
+    for msg in messages:
+        if isinstance(msg, ToolMessage):
+            tool_call_id = getattr(msg, "tool_call_id", None)
+
+            # Check 1: tool_call_id must be a non-empty string
+            if not tool_call_id or not isinstance(tool_call_id, str):
+                logger.warning(f"Filtering out ToolMessage with invalid tool_call_id: {tool_call_id!r}")
+                skipped_count += 1
+                continue
+
+            # Check 2: tool_call_id must have a matching AIMessage tool_call
+            if tool_call_id not in valid_tool_call_ids:
+                logger.warning(
+                    f"Filtering out orphaned ToolMessage: tool_call_id={tool_call_id} "
+                    f"not found in any AIMessage.tool_calls"
+                )
+                skipped_count += 1
+                continue
+
+        filtered.append(msg)
+
+    if skipped_count > 0:
+        logger.info(f"Filtered {skipped_count} invalid/orphaned tool messages from history")
+
+    return filtered
diff --git a/service/app/tasks/chat.py b/service/app/tasks/chat.py
index c19208c0..d6c2993a 100644
--- a/service/app/tasks/chat.py
+++ b/service/app/tasks/chat.py
@@ -334,14 +334,24 @@ async def _process_chat_message_async(
                     # Persist tool call response
                     try:
                         resp = stream_event["data"]
+                        tool_call_id = resp.get("toolCallId")
+
+                        # Only persist if toolCallId is valid - skip otherwise
+                        if not tool_call_id or not isinstance(tool_call_id, str):
+                            logger.warning(
+                                f"Skipping persistence of tool response with invalid toolCallId: {tool_call_id!r}"
+                            )
+                            await publisher.publish(json.dumps(stream_event))
+                            continue  # Skip to next event, but still publish to frontend
+
                         tool_message = MessageCreate(
                             role="tool",
                             content=json.dumps(
                                 {
                                     "event": ChatEventType.TOOL_CALL_RESPONSE,
-                                    "toolCallId": resp.get("toolCallId"),
+                                    "toolCallId": tool_call_id,
                                     "status": resp.get("status"),
-                                    "result": resp.get("result"),
+                                    "result": resp.get("result", ""),
                                     "error": resp.get("error"),
                                 }
                             ),