diff --git a/litellm/llms/sap/chat/handler.py b/litellm/llms/sap/chat/handler.py
index b4f4d4faddb..21c657d0e3d 100755
--- a/litellm/llms/sap/chat/handler.py
+++ b/litellm/llms/sap/chat/handler.py
@@ -23,15 +23,100 @@ class GenAIHubOrchestrationError(BaseLLMException):
         self.message = message
 
 
-# -------------------------------
-# Stream parsing helpers
-# -------------------------------
+# ----------------------------------
+# Parsing and transformation helpers
+# ----------------------------------
 
 
 def _now_ts() -> int:
     return int(time.time())
 
 
+def normalize_choice(
+    choice: dict[str, object],  # mutable-ok: wire dicts from httpx JSON; callers own them
+) -> dict[str, object]:  # mutable-ok: wire dicts from httpx JSON; callers own them
+    """Normalize a single choice dict (message or delta shape) in isolation."""
+    for key in ("message", "delta"):
+        carrier = choice.get(key)
+        if not isinstance(carrier, dict):
+            continue
+        reasoning_content = carrier.get("reasoning_content")
+        if not isinstance(reasoning_content, list):
+            continue
+        thinking_blocks: list[dict[str, object]] = [  # mutable-ok: consumed immediately below; never escapes as mutable
+            {
+                "type": "thinking",
+                "thinking": item.get("content") or "",  # pyright: ignore[reportUnknownMemberType]  # isinstance guard in comprehension filter; pyright can't narrow
+                "signature": item.get(  # pyright: ignore[reportUnknownMemberType]  # isinstance guard in comprehension filter; pyright can't narrow
+                    "signature"
+                ),
+            }
+            for item in reasoning_content  # pyright: ignore[reportUnknownVariableType]  # isinstance(reasoning_content, list) guard above; pyright can't narrow into comprehension
+            if isinstance(item, dict)
+        ]
+        carrier["thinking_blocks"] = thinking_blocks
+        carrier["reasoning_content"] = (
+            "\n".join(str(block["thinking"]) for block in thinking_blocks if block["thinking"]) or None
+        )
+    return choice
+
+
+def normalize_reasoning_content(
+    raw: dict[str, object],  # mutable-ok: wire dicts from httpx JSON; callers own them
+) -> dict[str, object]:  # mutable-ok: wire dicts from httpx JSON; callers own them
+    """Transform reasoning_content from SAP Orchestration format [{"content": ..., "signature": ...}] to Optional[str] and a thinking_blocks list."""
+    choices: Final[list[object]] = list(  # mutable-ok: wire list from httpx JSON
+        raw.get("choices") or []  # pyright: ignore[reportArgumentType, reportUnknownArgumentType]  # wire JSON; choices is always a list
+    )
+    normalized_choices: Final[list[dict[str, object]]] = [  # mutable-ok: wire list passed to pydantic model_validate
+        normalize_choice(choice)  # pyright: ignore[reportUnknownArgumentType]  # isinstance guard in comprehension filter; pyright can't narrow
+        for choice in choices
+        if isinstance(choice, dict)
+    ]
+    normalized_result: Final[dict[str, object]] = {  # mutable-ok: wire dict consumed by pydantic model_validate
+        **raw,
+        "choices": normalized_choices,
+    }
+    return normalized_result
+
+
+def denormalize_reasoning_content(
+    message: dict[str, object],  # mutable-ok: wire dicts from httpx JSON; callers own them
+) -> dict[str, object]:  # mutable-ok: wire dicts from httpx JSON; callers own them
+    """Inverse of normalize_reasoning_content: rebuild list[ReasoningBlock] from thinking_blocks.
+
+    LiteLLM normalizes SAP's reasoning_content list into a str + thinking_blocks before
+    returning a ModelResponse. When that response is replayed as a history message the dict
+    must be converted back to the SAP shape before SAPAssistantMessage validation.
+
+    If thinking_blocks is absent but reasoning_content is a plain str (e.g. from a
+    non-SAP provider), drop the field: we cannot reconstruct SAP ReasoningBlocks without
+    the original signatures.
+    """
+    if message.get("role") != "assistant":
+        return message
+    thinking_blocks: Final = message.get("thinking_blocks")
+    if isinstance(thinking_blocks, list):
+        reasoning_content: Final[list[dict[str, str | None]]] = [  # mutable-ok: wire list passed to validate_dict
+            {
+                "content": str(
+                    block.get("thinking") or ""  # pyright: ignore[reportUnknownMemberType, reportUnknownArgumentType]  # isinstance guard in comprehension filter; pyright can't narrow
+                ),
+                "signature": block.get(  # pyright: ignore[reportUnknownMemberType]  # isinstance guard in comprehension filter; pyright can't narrow
+                    "signature"
+                ),
+            }
+            for block in thinking_blocks  # pyright: ignore[reportUnknownVariableType]  # isinstance(thinking_blocks, list) guard above; pyright can't narrow into comprehension
+            if isinstance(block, dict)
+        ]
+        return {k: v for k, v in message.items() if k not in ("thinking_blocks", "reasoning_content")} | {
+            "reasoning_content": reasoning_content
+        }
+    if isinstance(message.get("reasoning_content"), str):
+        return {k: v for k, v in message.items() if k not in ("thinking_blocks", "reasoning_content")}
+    return message
+
+
 def _is_terminal_chunk(chunk: OpenAIChatCompletionChunk) -> bool:
     """OpenAI-shaped chunk is terminal if any choice has a non-None finish_reason."""
     try:
@@ -47,33 +132,37 @@ class _StreamParser:
     """Normalize orchestration streaming events into OpenAI-like chunks."""
 
     @staticmethod
-    def _from_orchestration_result(evt: dict) -> OpenAIChatCompletionChunk | None:
+    def _from_orchestration_result(evt: dict[str, object]) -> OpenAIChatCompletionChunk | None:
         """
         Accepts orchestration_result shape and maps it to an OpenAI-like *chunk*.
         """
-        orc: Final = evt.get("orchestration_result") or {}
+        orc: Final[dict[str, object]] = evt.get("orchestration_result") or {}  # pyright: ignore[reportAssignmentType]  # wire JSON; orchestration_result is always a dict
         if not orc:
             return None
 
         return OpenAIChatCompletionChunk.model_validate(
-            {
-                "id": orc.get("id") or evt.get("request_id") or "stream-chunk",
-                "object": orc.get("object") or "chat.completion.chunk",
-                "created": orc.get("created") or evt.get("created") or _now_ts(),
-                "model": orc.get("model") or "unknown",
-                "choices": [
-                    {
-                        "index": c.get("index", 0),
-                        "delta": c.get("delta") or {},
-                        "finish_reason": c.get("finish_reason"),
-                    }
-                    for c in (orc.get("choices") or [])
-                ],
-            }
+            normalize_reasoning_content(
+                {
+                    "id": orc.get("id")
+                    or evt.get("request_id")  # pyright: ignore[reportUnknownMemberType]  # wire JSON string field
+                    or "stream-chunk",
+                    "object": orc.get("object") or "chat.completion.chunk",
+                    "created": orc.get("created") or evt.get("created") or _now_ts(),
+                    "model": orc.get("model") or "unknown",
+                    "choices": [
+                        {
+                            "index": choice.get("index", 0),
+                            "delta": choice.get("delta") or {},
+                            "finish_reason": choice.get("finish_reason"),
+                        }
+                        for choice in orc.get("choices") or []  # pyright: ignore[reportUnknownVariableType, reportGeneralTypeIssues]  # wire JSON; choices is always list[dict]
+                    ],
+                }
+            )
         )
 
     @staticmethod
-    def to_openai_chunk(event_obj: dict) -> OpenAIChatCompletionChunk | None:
+    def to_openai_chunk(event_obj: dict[str, object]) -> OpenAIChatCompletionChunk | None:
         """
         Accepts:
           - {"final_result": <openai-style CHUNK>}   (IMPORTANT: this is just another chunk, NOT terminal)
@@ -89,11 +178,14 @@ class _StreamParser:
 
         # FINAL RESULT IS *NOT* TERMINAL: treat it as the next chunk
         if "final_result" in event_obj:
-            fr: Final = event_obj["final_result"] or {}
-            # ensure it looks like an OpenAI chunk
-            if "object" not in fr:
-                fr["object"] = "chat.completion.chunk"
-            return OpenAIChatCompletionChunk.model_validate(fr)
+            fr_raw: Final = event_obj["final_result"] or {}
+            if not isinstance(fr_raw, dict):
+                return None
+            fr: Final[dict[str, object]] = {  # mutable-ok: wire dict consumed by model_validate
+                **fr_raw,
+                **(({"object": "chat.completion.chunk"}) if "object" not in fr_raw else {}),
+            }
+            return OpenAIChatCompletionChunk.model_validate(normalize_reasoning_content(fr))
 
         # Orchestration incremental delta
         if "orchestration_result" in event_obj:
@@ -101,7 +193,7 @@ class _StreamParser:
 
         # Already an OpenAI-like chunk
         if "choices" in event_obj and "object" in event_obj:
-            return OpenAIChatCompletionChunk.model_validate(event_obj)
+            return OpenAIChatCompletionChunk.model_validate(normalize_reasoning_content(event_obj))
 
         # Unknown / heartbeat / metrics
         return None
@@ -248,7 +340,8 @@ class AsyncSAPStreamIterator:
 # -------------------------------
 class GenAIHubOrchestration(BaseLLMHTTPHandler):
     def _add_stream_param_to_request_body(self, data: dict, provider_config: BaseConfig, fake_stream: bool):
-        if data.get("config", {}).get("stream", None) is not None:
+        config = data.get("config") or {}  # pyright: ignore[reportUnknownVariableType]  # wire JSON; config is always a dict when present
+        if isinstance(config, dict) and config.get("stream") is not None:
             data["config"]["stream"]["enabled"] = True
         else:
             data["config"]["stream"] = {"enabled": True}
diff --git a/litellm/llms/sap/chat/models.py b/litellm/llms/sap/chat/models.py
index f2b31d4f9b0..a1954e86b2a 100644
--- a/litellm/llms/sap/chat/models.py
+++ b/litellm/llms/sap/chat/models.py
@@ -133,11 +133,21 @@ class SAPUserMessage(LiteLLMBaseModel):
     content: str | TextContent | ImageContent | list[TextContent | ImageContent]
 
 
+class ReasoningBlock(  # frozen-ok: adding model_config would override LiteLLMBaseModel's defer_build setting
+    LiteLLMBaseModel
+):
+    content: str | None = None
+    signature: str | None = None
+
+
 class SAPAssistantMessage(LiteLLMBaseModel):
     role: Literal["assistant"] = "assistant"
     content: str | TextContent | list[TextContent] = ""
     refusal: str = ""
-    tool_calls: list[MessageToolCall] = Field(default=[])
+    tool_calls: list[MessageToolCall] = Field(  # mutable-ok: pydantic field; callers append tool calls to this list
+        default=[]
+    )
+    reasoning_content: list[ReasoningBlock] | None = None  # mutable-ok: list of reasoning blocks from SAP wire format
 
 
 class SAPToolChatMessage(LiteLLMBaseModel):
diff --git a/litellm/llms/sap/chat/transformation.py b/litellm/llms/sap/chat/transformation.py
index 0270f814328..3eab22af08f 100755
--- a/litellm/llms/sap/chat/transformation.py
+++ b/litellm/llms/sap/chat/transformation.py
@@ -31,6 +31,8 @@ from .handler import (
     AsyncSAPStreamIterator,
     GenAIHubOrchestrationError,
     SAPStreamIterator,
+    denormalize_reasoning_content,
+    normalize_reasoning_content,
 )
 from .models import (
     ChatCompletionTool,
@@ -106,8 +108,9 @@ def _messages_to_sap_template(messages: list[AllMessageValues]) -> list:
     template: Final = []
     for message in messages:
         folded_message = _fold_message_cache_control(message_with_file_parts_as_image_urls(message))
-        message_role_class = _message_role_mapping(str(folded_message["role"]))
-        validated_message = validate_dict(folded_message, message_role_class)
+        denormalized_message = denormalize_reasoning_content(dict(folded_message))
+        message_role_class = _message_role_mapping(str(denormalized_message["role"]))
+        validated_message = validate_dict(denormalized_message, message_role_class)
         template.append(validated_message)
     return template
 
@@ -229,7 +232,9 @@ class GenAIHubOrchestrationConfig(OpenAIGPTConfig):
             "functions",
             "extra_headers",
             "parallel_tool_calls",
+            "reasoning_effort",
             "response_format",
+            "thinking",
             "timeout",
         ]
         # Remove response_format for providers that don't support it on SAP GenAI Hub
@@ -410,7 +415,9 @@ class GenAIHubOrchestrationConfig(OpenAIGPTConfig):
             original_response=raw_response.text,
             additional_args={"complete_input_dict": request_data},
         )
-        response = ModelResponse.model_validate(raw_response.json()["final_result"])
+        response: Final[ModelResponse] = ModelResponse.model_validate(
+            normalize_reasoning_content(raw_response.json()["final_result"])
+        )
 
         # Strip markdown code blocks if JSON response_format was used with Anthropic models
         # SAP GenAI Hub with Anthropic models sometimes wraps JSON in ```json ... ```
@@ -419,7 +426,7 @@ class GenAIHubOrchestrationConfig(OpenAIGPTConfig):
         response_format: Final = optional_params.get("response_format", {})
         if response_format.get("type") in ("json_object", "json_schema"):
             if model.startswith("anthropic"):
-                response = self._strip_markdown_json(response)
+                return self._strip_markdown_json(response)
 
         return response
 
diff --git a/tests/unit/llms/sap/chat/test_sap_transformation.py b/tests/unit/llms/sap/chat/test_sap_transformation.py
index f95bf9b7845..457419e1074 100644
--- a/tests/unit/llms/sap/chat/test_sap_transformation.py
+++ b/tests/unit/llms/sap/chat/test_sap_transformation.py
@@ -941,3 +941,341 @@ class TestCacheControl:
         content = self._template(body)[0]["content"]
         assert "cache_control" not in content[0]
         assert content[1].get("cache_control") == {"type": "ephemeral"}
+
+class TestNormalizeReasoningContent:
+    """Unit tests for normalize_reasoning_content."""
+
+    from litellm.llms.sap.chat.handler import normalize_reasoning_content
+
+    _normalize_reasoning_content = staticmethod(normalize_reasoning_content)
+
+    def test_list_reasoning_content_mapped_to_thinking_blocks(self):
+        """List-shaped reasoning_content is converted to thinking_blocks."""
+        raw = {
+            "choices": [{
+                "message": {
+                    "role": "assistant",
+                    "content": "Latin.",
+                    "reasoning_content": [
+                        {"content": "Romans spoke Latin.", "signature": "sig1"},
+                        {"content": "That is well known.", "signature": "sig2"},
+                    ],
+                }
+            }]
+        }
+        out = self._normalize_reasoning_content(raw)
+        msg = out["choices"][0]["message"]
+        assert msg["thinking_blocks"] == [
+            {"type": "thinking", "thinking": "Romans spoke Latin.", "signature": "sig1"},
+            {"type": "thinking", "thinking": "That is well known.", "signature": "sig2"},
+        ]
+        assert msg["reasoning_content"] == "Romans spoke Latin.\nThat is well known."
+
+    def test_string_reasoning_content_unchanged(self):
+        """String reasoning_content is left as-is (already the right type)."""
+        raw = {
+            "choices": [{
+                "message": {
+                    "role": "assistant",
+                    "content": "42",
+                    "reasoning_content": "I thought about it.",
+                }
+            }]
+        }
+        out = self._normalize_reasoning_content(raw)
+        msg = out["choices"][0]["message"]
+        assert msg["reasoning_content"] == "I thought about it."
+        assert "thinking_blocks" not in msg
+
+    def test_no_reasoning_content_unchanged(self):
+        """A message without reasoning_content is not modified."""
+        raw = {"choices": [{"message": {"role": "assistant", "content": "Hi."}}]}
+        out = self._normalize_reasoning_content(raw)
+        assert out == raw
+
+    def test_empty_list_reasoning_content_sets_none(self):
+        """An empty list produces None for reasoning_content and empty thinking_blocks."""
+        raw = {"choices": [{"message": {"reasoning_content": []}}]}
+        out = self._normalize_reasoning_content(raw)
+        msg = out["choices"][0]["message"]
+        assert msg["thinking_blocks"] == []
+        assert msg["reasoning_content"] is None
+
+    def test_multiple_choices_all_normalized(self):
+        """All choices in the response are normalized."""
+        raw = {
+            "choices": [
+                {"message": {"reasoning_content": [{"content": "thought A", "signature": None}]}},
+                {"message": {"reasoning_content": [{"content": "thought B", "signature": "s"}]}},
+            ]
+        }
+        out = self._normalize_reasoning_content(raw)
+        assert out["choices"][0]["message"]["reasoning_content"] == "thought A"
+        assert out["choices"][1]["message"]["reasoning_content"] == "thought B"
+
+
+    def test_null_content_in_block_uses_empty_string(self):
+        """Explicit null content value must not leak None into thinking field."""
+        raw = {
+            "choices": [{
+                "message": {
+                    "reasoning_content": [{"content": None, "signature": "s"}],
+                }
+            }]
+        }
+        out = self._normalize_reasoning_content(raw)
+        block = out["choices"][0]["message"]["thinking_blocks"][0]
+        assert block["thinking"] == ""
+        assert out["choices"][0]["message"]["reasoning_content"] is None
+
+    def test_transform_response_normalizes_list_reasoning_content(self):
+        """Production path: transform_response must produce a ModelResponse
+        with thinking_blocks populated when the raw payload carries a
+        list-shaped reasoning_content.
+        """
+        import json
+        from unittest.mock import MagicMock
+
+        from litellm.llms.sap.chat.transformation import GenAIHubOrchestrationConfig
+
+        config = GenAIHubOrchestrationConfig()
+
+        final_result = {
+            "id": "chatcmpl-test",
+            "object": "chat.completion",
+            "created": 1700000000,
+            "model": "gemini-test",
+            "choices": [{
+                "index": 0,
+                "message": {
+                    "role": "assistant",
+                    "content": "The answer is 42.",
+                    "reasoning_content": [
+                        {"content": "Let me think.", "signature": "sig1"},
+                        {"content": "Yes, 42.", "signature": "sig2"},
+                    ],
+                },
+                "finish_reason": "stop",
+            }],
+            "usage": {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15},
+        }
+
+        raw_response = MagicMock()
+        raw_response.text = json.dumps({"final_result": final_result})
+        raw_response.json.return_value = {"final_result": final_result}
+
+        response = config.transform_response(
+            model="gemini-test",
+            raw_response=raw_response,
+            model_response=MagicMock(),
+            logging_obj=MagicMock(),
+            api_key="test",
+            request_data={},
+            messages=[],
+            optional_params={},
+            litellm_params={},
+            encoding=None,
+        )
+
+        choice = response.choices[0]
+        assert hasattr(choice.message, "thinking_blocks"), "thinking_blocks missing from message"
+        assert choice.message.thinking_blocks == [
+            {"type": "thinking", "thinking": "Let me think.", "signature": "sig1"},
+            {"type": "thinking", "thinking": "Yes, 42.", "signature": "sig2"},
+        ]
+        assert choice.message.reasoning_content == "Let me think.\nYes, 42."
+
+
+class TestNormalizeChoice:
+    """Unit tests for normalize_choice (message and delta shapes)."""
+
+    from litellm.llms.sap.chat.handler import normalize_choice, normalize_reasoning_content
+
+    _normalize_choice = staticmethod(normalize_choice)
+    _normalize_reasoning_content = staticmethod(normalize_reasoning_content)
+
+    def test_message_list_reasoning_content_normalized(self):
+        choice = {
+            "index": 0,
+            "message": {"role": "assistant", "content": "hi", "reasoning_content": [{"content": "thought", "signature": "s1"}]},
+            "finish_reason": "stop",
+        }
+        result = self._normalize_choice(choice)
+        msg = result["message"]
+        assert msg["reasoning_content"] == "thought"
+        assert msg["thinking_blocks"] == [{"type": "thinking", "thinking": "thought", "signature": "s1"}]
+
+    def test_delta_list_reasoning_content_normalized(self):
+        choice = {
+            "index": 0,
+            "delta": {"role": "assistant", "reasoning_content": [{"content": "delta thought", "signature": None}]},
+            "finish_reason": None,
+        }
+        result = self._normalize_choice(choice)
+        delta = result["delta"]
+        assert delta["reasoning_content"] == "delta thought"
+        assert delta["thinking_blocks"][0]["thinking"] == "delta thought"
+
+    def test_string_reasoning_content_unchanged(self):
+        choice = {"index": 0, "message": {"reasoning_content": "already a string"}}
+        assert self._normalize_choice(choice) == choice
+
+    def test_no_reasoning_content_unchanged(self):
+        choice = {"index": 0, "delta": {"content": "hello"}}
+        assert self._normalize_choice(choice) == choice
+
+    def test_normalize_reasoning_content_covers_delta_path(self):
+        raw = {
+            "id": "c1",
+            "object": "chat.completion.chunk",
+            "choices": [
+                {"index": 0, "delta": {"reasoning_content": [{"content": "stream thought", "signature": "sig"}]}, "finish_reason": None}
+            ],
+        }
+        result = self._normalize_reasoning_content(raw)
+        delta = result["choices"][0]["delta"]
+        assert delta["reasoning_content"] == "stream thought"
+        assert delta["thinking_blocks"][0]["type"] == "thinking"
+
+
+class TestMessagesToSapTemplateReasoningContent:
+    """_messages_to_sap_template must forward reasoning_content on assistant turns."""
+
+    def test_assistant_message_with_reasoning_content_is_preserved(self):
+        from litellm.llms.sap.chat.transformation import _messages_to_sap_template
+
+        messages = [
+            {"role": "user", "content": "What is 22?"},
+            {
+                "role": "assistant",
+                "content": "4",
+                "reasoning_content": [
+                    {"content": "Simple arithmetic.", "signature": "sig1"}
+                ],
+            },
+            {"role": "user", "content": "Are you sure?"},
+        ]
+        result = _messages_to_sap_template(messages)
+        assistant_msg = result[1]
+        assert assistant_msg["reasoning_content"] == [
+            {"content": "Simple arithmetic.", "signature": "sig1"}
+        ]
+
+    def test_litellm_normalized_assistant_message_round_trips_signature(self):
+        """After a non-streaming SAP call, the ModelResponse has reasoning_content: str
+        and thinking_blocks: list (LiteLLM native format). Replaying it as a history
+        message must not raise and must preserve the signature.
+        """
+        from litellm.llms.sap.chat.transformation import _messages_to_sap_template
+
+        signature = "EtwHCkABGAIqQKNbQn0Uz=="
+        messages = [
+            {"role": "user", "content": "Translate it."},
+            {
+                "role": "assistant",
+                "content": "The text is Lorem Ipsum.",
+                "reasoning_content": "Lorem ipsum is pseudo-Latin...",
+                "thinking_blocks": [
+                    {"type": "thinking", "thinking": "Lorem ipsum is pseudo-Latin...", "signature": signature}
+                ],
+            },
+            {"role": "user", "content": "Summarise it."},
+        ]
+
+        result = _messages_to_sap_template(messages)
+        assistant = next(m for m in result if m.get("role") == "assistant")
+
+        reasoning = assistant.get("reasoning_content")
+        assert isinstance(reasoning, list)
+        assert any(block.get("signature") == signature for block in reasoning)
+
+    def test_streaming_assembled_assistant_message_round_trips_signature(self):
+        """After a streaming SAP call, normalize_reasoning_content runs on every chunk.
+        The assembled assistant message has the same LiteLLM-native shape as the
+        non-streaming case. Verify the signature survives the round-trip.
+        """
+        from litellm.llms.sap.chat.handler import normalize_reasoning_content
+        from litellm.llms.sap.chat.transformation import _messages_to_sap_template
+
+        signature = "EtwHCkABGAIqQKNbQn0Uz=="
+        thinking_text = "Lorem ipsum is pseudo-Latin..."
+        sap_final_result = {
+            "id": "msg_stream_01",
+            "object": "chat.completion",
+            "model": "gpt-4o",
+            "choices": [
+                {
+                    "index": 0,
+                    "message": {
+                        "role": "assistant",
+                        "content": "The text is Lorem Ipsum.",
+                        "reasoning_content": [{"content": thinking_text, "signature": signature}],
+                    },
+                    "finish_reason": "stop",
+                }
+            ],
+        }
+        normalized = normalize_reasoning_content(sap_final_result)
+        assembled_message = {"role": "assistant", **normalized["choices"][0]["message"]}
+
+        messages = [
+            {"role": "user", "content": "Translate it."},
+            assembled_message,
+            {"role": "user", "content": "Summarize it."},
+        ]
+
+        result = _messages_to_sap_template(messages)
+        assistant = next(m for m in result if m.get("role") == "assistant")
+
+        reasoning = assistant.get("reasoning_content")
+        assert isinstance(reasoning, list)
+        assert any(block.get("signature") == signature for block in reasoning)
+
+
+class TestDenormalizeReasoningContent:
+    from litellm.llms.sap.chat.handler import denormalize_reasoning_content
+
+    _denormalize = staticmethod(denormalize_reasoning_content)
+
+    def test_thinking_blocks_rebuilt_into_sap_reasoning_content(self):
+        message = {
+            "role": "assistant",
+            "content": "hello",
+            "thinking_blocks": [{"type": "thinking", "thinking": "step one", "signature": "sig1"}],
+            "reasoning_content": "step one",
+        }
+        result = self._denormalize(message)
+        assert result["reasoning_content"] == [{"content": "step one", "signature": "sig1"}]
+        assert "thinking_blocks" not in result
+
+    def test_string_reasoning_content_without_thinking_blocks_is_dropped(self):
+        message = {
+            "role": "assistant",
+            "content": "hello",
+            "reasoning_content": "some thought from another provider",
+        }
+        result = self._denormalize(message)
+        assert "reasoning_content" not in result
+        assert "thinking_blocks" not in result
+        assert result["content"] == "hello"
+
+    def test_non_assistant_message_is_returned_unchanged(self):
+        message = {"role": "user", "content": "hi", "reasoning_content": "should stay"}
+        assert self._denormalize(message) is message
+
+    def test_assistant_without_reasoning_fields_is_returned_unchanged(self):
+        message = {"role": "assistant", "content": "hi"}
+        assert self._denormalize(message) is message
+
+    def test_validate_dict_accepts_denormalized_message_from_string_reasoning_content(self):
+        from litellm.llms.sap.chat.models import SAPAssistantMessage
+        from litellm.llms.sap.chat.transformation import validate_dict
+
+        message = {
+            "role": "assistant",
+            "content": "done",
+            "reasoning_content": "raw string from openai o-series",
+        }
+        result = validate_dict(self._denormalize(message), SAPAssistantMessage)
+        assert result["role"] == "assistant"
+        assert "reasoning_content" not in result or result.get("reasoning_content") is None