diff --git a/litellm/llms/anthropic/chat/guardrail_translation/handler.py b/litellm/llms/anthropic/chat/guardrail_translation/handler.py index 2ea20143f0c..350cea697c0 100644 --- a/litellm/llms/anthropic/chat/guardrail_translation/handler.py +++ b/litellm/llms/anthropic/chat/guardrail_translation/handler.py @@ -507,7 +507,8 @@ class AnthropicMessagesHandler(BaseTranslation): chat_completion_compatible_request, _tool_name_mapping, ) = LiteLLMAnthropicMessagesAdapter().translate_anthropic_to_openai( - anthropic_message_request=cast(AnthropicMessagesRequest, data.copy()) + anthropic_message_request=cast(AnthropicMessagesRequest, data.copy()), + preserve_midturn_system=True, ) return chat_completion_compatible_request diff --git a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py index 76c56f6ed46..10ba2431bcc 100644 --- a/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py +++ b/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py @@ -422,6 +422,8 @@ class LiteLLMAnthropicMessagesAdapter: self, messages: list[AllAnthropicPassThroughMessageValues], model: str | None = None, + *, + preserve_midturn_system: bool = False, ) -> list: new_messages: Final[list[AllMessageValues]] = [] replayable_messages: Final = strip_encrypted_reasoning_blocks_from_anthropic_messages(messages) @@ -430,8 +432,12 @@ class LiteLLMAnthropicMessagesAdapter: len(replayable_messages), ) ordered_messages: Final = ( - *replayable_messages[:leading_count], - *convert_mid_conversation_system_turns(replayable_messages[leading_count:]), + replayable_messages + if preserve_midturn_system + else ( + *replayable_messages[:leading_count], + *convert_mid_conversation_system_turns(replayable_messages[leading_count:]), + ) ) for m in ordered_messages: user_message: ChatCompletionUserMessage | None = None @@ -1166,6 +1172,7 @@ class LiteLLMAnthropicMessagesAdapter: anthropic_message_request: AnthropicMessagesRequest, *, custom_llm_provider: str | None = None, + preserve_midturn_system: bool = False, ) -> tuple[ChatCompletionRequest, dict[str, str]]: """ This is used by the beta Anthropic Adapter, for translating anthropic `/v1/messages` requests to the openai format. @@ -1187,6 +1194,7 @@ class LiteLLMAnthropicMessagesAdapter: new_messages = self.translate_anthropic_messages_to_openai( messages=messages_list, model=anthropic_message_request.get("model"), + preserve_midturn_system=preserve_midturn_system, ) ## ADD SYSTEM MESSAGE TO MESSAGES self._add_system_message_to_messages(new_messages, anthropic_message_request) diff --git a/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py b/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py index fb6a1c40253..129b5b4f647 100644 --- a/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py +++ b/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py @@ -744,7 +744,8 @@ def _count_effective_tokens( messages=cast( "list[AllAnthropicPassThroughMessageValues]", messages_without_compaction, - ) + ), + preserve_midturn_system=True, ) except Exception as e: verbose_logger.debug( @@ -899,7 +900,8 @@ def _build_summary_messages( messages=cast( "list[AllAnthropicPassThroughMessageValues]", stripped, - ) + ), + preserve_midturn_system=True, ) except Exception as e: verbose_logger.warning(