fix: pass together_ai replay/streaming tests under the thinking contract

Controlled base-vs-head run of the full llms test shard exposed two
together_ai anthropic-messages tests broken by the thinking_disabled
contract change in this PR:

- test_anthropic_messages_replays_tool_loop: an unsigned thinking block
  in replayed history was being dropped entirely. Unsigned thinking texts
  now map to the provider-facing reasoning_content field (keeping the
  signature-400 defense: they still stay out of thinking_blocks).
- replay + streaming tests asserted provider reasoning is surfaced
  without a thinking param on the request; per the PR contract that is
  suppressed. Updated both to assert the suppression (sending
  thinking=enabled would fail together's parameter validation via
  reasoning_effort).

litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py

@@ -495,6 +495,7 @@ class LiteLLMAnthropicMessagesAdapter:
             has_cache_control_in_text = False
             tool_calls: list[ChatCompletionAssistantToolCall] = []
             thinking_blocks: list[ChatCompletionThinkingBlock | ChatCompletionRedactedThinkingBlock] = []
+            unsigned_thinking_texts: list[str] = []
             if m["role"] == "assistant":
                 if isinstance(m.get("content"), str):
                     assistant_message_str = str(m.get("content", ""))
@@ -555,6 +556,15 @@ class LiteLLMAnthropicMessagesAdapter:
                                         signature=content.get("signature") or "",
                                     )
                                     thinking_blocks.append(thinking_block)
+                                else:
+                                    # Unsigned text is NOT dropped: it is
+                                    # replayed as the flat reasoning_content field
+                                    # below (the provider-visible form), while
+                                    # staying out of thinking_blocks, so the
+                                    # signature-400 stays avoided.
+                                    unsigned_text = str(content.get("thinking") or "")
+                                    if unsigned_text:
+                                        unsigned_thinking_texts.append(unsigned_text)
                             elif content.get("type") == "redacted_thinking":
                                 redacted_thinking_block = ChatCompletionRedactedThinkingBlock(
                                     type="redacted_thinking",
@@ -587,6 +597,9 @@ class LiteLLMAnthropicMessagesAdapter:
                 if len(thinking_blocks) > 0:
                     assistant_message["thinking_blocks"] = thinking_blocks
                 reasoning_content = reasoning_content_from_thinking_blocks(thinking_blocks)
+                if unsigned_thinking_texts:
+                    unsigned = "\n".join(unsigned_thinking_texts)
+                    reasoning_content = f"{reasoning_content}\n{unsigned}" if reasoning_content else unsigned
                 if reasoning_content:
                     assistant_message["reasoning_content"] = reasoning_content
                 new_messages.append(assistant_message)

tests/test_litellm/llms/together_ai/chat/test_together_ai_chat_transformation.py

@@ -1013,7 +1013,11 @@ def test_anthropic_messages_replays_tool_loop_and_maps_reasoning_to_thinking_blo
     assert tool_turn["content"] == "Sunny, 18C"
 
     blocks = {block["type"]: block for block in response["content"]}
-    assert blocks["thinking"]["thinking"] == "Tool said sunny."
+
+    # Contract (thinking param absent): the provider's reasoning is NOT
+    # surfaced as an Anthropic thinking block. The mock response still
+    # carries reasoning, and this asserts it is suppressed.
+    assert "thinking" not in blocks
     assert blocks["text"]["text"] == "Sunny in SF."
     assert response["stop_reason"] == "end_turn"
 
@@ -1031,6 +1035,9 @@ def test_anthropic_messages_streams_together_tool_call_as_input_json_delta():
     captured_requests: list[httpx.Request] = []
     client = _sync_client(captured_requests, _sse_response(*PARALLEL_TOOL_CALL_STREAM))
 
+    # Explicit thinking= would be needed to surface provider reasoning
+    # (the pass-through contract suppresses it when the thinking param is
+    # absent); this test asserts the streaming translation only.
     events = _anthropic_sse_events(
         litellm.anthropic.messages.create(
             model=f"together_ai/{UNMAPPED_MODEL}",
@@ -1067,7 +1074,9 @@ def test_anthropic_messages_streams_together_tool_call_as_input_json_delta():
         for event in events
         if event["type"] == "content_block_delta" and event["delta"]["type"] == "thinking_delta"
     )
-    assert thinking_text == "Need weather and time."
+    # thinking param absent in the request: provider reasoning is
+    # suppressed per the contract, so no thinking block may appear.
+    assert thinking_text == ""
     assert [event["delta"]["stop_reason"] for event in events if event["type"] == "message_delta"] == ["tool_use"]