fix: make it clearer to the llm that a compaction happened

litellm/responses/litellm_completion_transformation/handler.py

@@ -169,7 +169,6 @@ def _append_compaction_output(
         "type": "compaction",
         "id": "cmp_" + uuid.uuid4().hex,
         "encrypted_content": encoded_content,
-        "created_by": None,
     }
     if existing_output:
         return [existing_output[0], compaction_item] + list(existing_output[1:])

litellm/responses/litellm_completion_transformation/transformation.py

@@ -390,8 +390,13 @@ class LiteLLMCompletionResponsesConfig:
                 else:
                     decoded_content = compaction_item.get("content", "")
                 messages.append(
-                    ChatCompletionUserMessage(
-                        role="user", content=decoded_content
+                    ChatCompletionSystemMessage(
+                        role="system",
+                        content=(
+                            "The following is a summary of the prior conversation history, "
+                            "which was compacted to save context space. Treat this as established "
+                            "context, not as a new user message:\n\n" + decoded_content
+                        ),
                     )
                 )
                 if last_compaction_idx > 0: