fix: Prevent assistant-message prefill rejected by Claude 4.6

strix/llm/llm.py

@@ -188,6 +188,9 @@ class LLM:
         conversation_history.extend(compressed)
         messages.extend(compressed)
 
+        if messages[-1].get("role") == "assistant":
+            messages.append({"role": "user", "content": "<meta>Continue the task.</meta>"})
+
         if self._is_anthropic() and self.config.enable_prompt_caching:
             messages = self._add_cache_control(messages)
 

strix/llm/memory_compressor.py

@@ -91,7 +91,7 @@ def _summarize_messages(
     if not messages:
         empty_summary = "<context_summary message_count='0'>{text}</context_summary>"
         return {
-            "role": "assistant",
+            "role": "user",
             "content": empty_summary.format(text="No messages to summarize"),
         }
 
@@ -123,7 +123,7 @@ def _summarize_messages(
             return messages[0]
         summary_msg = "<context_summary message_count='{count}'>{text}</context_summary>"
         return {
-            "role": "assistant",
+            "role": "user",
             "content": summary_msg.format(count=len(messages), text=summary),
         }
     except Exception: