fix(responses-bridge): join folded system blocks on newlines; keep a guardrail's rewritten system item in place

Three follow-ups from review of the list-content fold:

- Folded blocks join on newlines with empty blocks dropped, as the
  Anthropic-to-Responses adapter joins the same shape, so a multi-block
  system prompt keeps its section boundaries.
- The Responses guardrail handler converts one rewritten chat row through
  the bridge to patch its input item in place. A lone list-content system
  row now folds into `instructions` and returns no item, so a guardrail
  rewriting a system item abandoned the patch and the request lost its
  system item (an empty `input` for a system-only request). The row now
  converts as a developer row, whose content maps to the same input parts.
- The system-only comment no longer cites non-string content, and the
  mid-conversation list test is dropped: main already covers that case.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

litellm/completion_extras/litellm_responses_transformation/transformation.py

@@ -155,7 +155,8 @@ def _system_text_for_instructions(message: Mapping[str, object]) -> str | None:
         return None
     blocks: Final = cast(list[object], content)  # cast-ok: isinstance confirms a list of content blocks
     texts: Final = tuple(text for text in map(_foldable_system_text, blocks) if text is not None)
-    return " ".join(texts) if len(texts) == len(blocks) else None
+    # Newline-joined, empty blocks dropped, as the Anthropic-to-Responses adapter joins the same shape.
+    return "\n".join(filter(None, texts)) if len(texts) == len(blocks) else None
 
 
 def _pending_audio_breakpoint(pending: object, part: object) -> object:
@@ -791,8 +792,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
         )
         # OpenAI's Responses API rejects an empty input. For a system-only
         # request, carry the system message as a system-role input item instead
-        # of instructions, mirroring how non-string system content is already
-        # handled in convert_chat_completion_messages_to_responses_api.
+        # of instructions.
         is_system_only_request: Final = not converted_input_items and converted_instructions is not None
         target_input_items: Final = (
             _without_audio_input_parts(converted_input_items)

litellm/llms/openai/responses/guardrail_translation/handler.py

@@ -265,7 +265,12 @@ def _rewritten_input_item(item: Mapping[str, object], rewritten: object) -> Mapp
     rewritten_content: Final = rewritten.get("content")
     if isinstance(item.get(field), str) and isinstance(rewritten_content, str):
         return {**item, field: rewritten_content}
-    rewritten_row: Final = cast("AllMessageValues", rewritten)  # cast-ok: guardrails hand back chat-shaped rows
+    # A lone system row with content would fold into `instructions` and leave no item to patch; as a developer
+    # row its content converts to the same input parts and stays an item.
+    row: Final = cast("Mapping[str, object]", rewritten)  # cast-ok: isinstance confirms a mapping, keyed by field name
+    folds: Final = row.get("role") == "system" and row.get("content") is not None
+    chat_row: Final = {**row, "role": "developer"} if folds else row
+    rewritten_row: Final = cast("AllMessageValues", chat_row)  # cast-ok: guardrails hand back chat-shaped rows
     converted_items, _ = LiteLLMResponsesTransformationHandler().convert_chat_completion_messages_to_responses_api(
         [rewritten_row]
     )

tests/unit/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py

@@ -4840,26 +4840,42 @@ def test_leading_system_messages_mixed_str_and_list_concatenate_in_order():
     ]
 
 
-def test_mid_conversation_system_list_content_stays_in_input_after_a_user_turn():
-    """Only the leading run folds (#40269): a list-content system message after the first
-    non-system message stays a positioned input item."""
+def test_leading_system_list_blocks_join_on_newlines_without_empty_blocks():
+    """A multi-block system prompt keeps its section boundaries: blocks join on newlines and
+    empty blocks drop, as the Anthropic-to-Responses adapter joins the same shape."""
     handler: Final = LiteLLMResponsesTransformationHandler()
 
-    input_items, instructions = handler.convert_chat_completion_messages_to_responses_api(
+    _, instructions = handler.convert_chat_completion_messages_to_responses_api(
         [
-            {"role": "user", "content": "Read the file."},
             {
                 "role": "system",
-                "content": [{"type": "text", "text": "<total_tokens>14982391 tokens left</total_tokens>"}],
+                "content": [
+                    {"type": "text", "text": "You are Claude Code."},
+                    {"type": "text", "text": ""},
+                    {"type": "text", "text": "# Tone\nBe brief.", "cache_control": {"type": "ephemeral"}},
+                ],
             },
+            {"role": "user", "content": "hi"},
         ]
     )
 
-    assert instructions is None
-    assert input_items == [
-        {"type": "message", "role": "user", "content": [{"type": "input_text", "text": "Read the file."}]},
-        _system_input_item("<total_tokens>14982391 tokens left</total_tokens>"),
-    ]
+    assert instructions == "You are Claude Code.\n# Tone\nBe brief."
+
+
+def test_system_only_request_with_list_content_stays_a_system_input_item():
+    """Responses rejects an empty input, so a request holding only a list-format system
+    message still goes out as one system input item, as a string one does."""
+    request: Final = LiteLLMResponsesTransformationHandler().transform_request(
+        model="gpt-4o",
+        messages=[{"role": "system", "content": [{"type": "text", "text": "You are terse."}]}],
+        optional_params={},
+        litellm_params={},
+        headers={},
+        litellm_logging_obj=Mock(),
+    )
+
+    assert "instructions" not in request
+    assert request["input"] == [_system_input_item("You are terse.")]
 
 
 def test_leading_system_list_with_a_non_text_block_stays_an_input_item():

tests/unit/llms/openai/responses/test_openai_responses_guardrail_handler.py

@@ -3105,6 +3105,29 @@ class TestPatchEdgeBranches:
         user_items = [item for item in result["input"] if item.get("role") == "user"]
         assert _texts(user_items[0]) == [COMPRESSED_MARKER]
 
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize("trailing", [[{"role": "user", "content": "What is the codename?"}], []])
+    async def test_list_content_system_item_rewrite_patches_in_place(self, trailing):
+        """A rewritten list-content system item stays an input item, never moved into instructions."""
+        handler = OpenAIResponsesHandler()
+        data = {
+            "model": "gpt-5.6",
+            "input": [
+                {"role": "system", "content": [{"type": "input_text", "text": "Answer from the memo only."}]},
+                *trailing,
+            ],
+        }
+
+        result = await handler.process_input_messages(
+            data, SystemRewriteGuardrail(rewritten_content=[{"type": "text", "text": COMPRESSED_MARKER}])
+        )
+
+        assert "instructions" not in result
+        assert result["input"] == [
+            {"role": "system", "content": [{"type": "input_text", "text": COMPRESSED_MARKER}]},
+            *trailing,
+        ]
+
     def test_item_rewrite_field_ignores_non_string_type(self):
         from litellm.llms.openai.responses.guardrail_translation.handler import _item_rewrite_field