fix(anthropic-adapter): translate stop_sequences and disabled thinking for non-Claude targets

Claude Code's auto-mode classifier sends stop_sequences and thinking:
{type: disabled} on /v1/messages. The Anthropic adapter passed
stop_sequences through unchanged instead of mapping it to OpenAI's stop,
which Fireworks' OpenAI-compatible endpoint rejects with HTTP 400. It also
dropped disabled thinking instead of mapping it to reasoning_effort: none,
so the model spent its output budget on reasoning it was told to skip.

Resolves LIT-4798

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

@@ -331,6 +331,7 @@ class LiteLLMAnthropicMessagesAdapter:
             "thinking",
             "output_format",
             "output_config",
+            "stop_sequences",
         ]
 
     def _is_web_search_tool(self, tool: Dict[str, Any]) -> bool:
@@ -615,7 +616,7 @@ class LiteLLMAnthropicMessagesAdapter:
         thinking_type = thinking.get("type", "disabled")
 
         if thinking_type == "disabled":
-            return None
+            return "none"
         elif thinking_type == "enabled":
             return reasoning_effort_from_thinking_budget(thinking.get("budget_tokens", 0))
         elif thinking_type == "adaptive":
@@ -919,6 +920,18 @@ class LiteLLMAnthropicMessagesAdapter:
             tool_choice=cast(AnthropicMessagesToolChoice, tool_choice)
         )
 
+    def _translate_stop_sequences_to_openai(
+        self,
+        anthropic_message_request: AnthropicMessagesRequest,
+        new_kwargs: ChatCompletionRequest,
+    ) -> None:
+        if "stop_sequences" not in anthropic_message_request:
+            return
+        stop_sequences = anthropic_message_request["stop_sequences"]
+        if not stop_sequences:
+            return
+        new_kwargs["stop"] = stop_sequences
+
     def _translate_tools_to_openai(
         self,
         anthropic_message_request: AnthropicMessagesRequest,
@@ -1098,6 +1111,11 @@ class LiteLLMAnthropicMessagesAdapter:
             anthropic_message_request=anthropic_message_request,
             new_kwargs=new_kwargs,
         )
+        ## CONVERT STOP_SEQUENCES
+        self._translate_stop_sequences_to_openai(
+            anthropic_message_request=anthropic_message_request,
+            new_kwargs=new_kwargs,
+        )
         ## CONVERT OUTPUT_FORMAT to RESPONSE_FORMAT
         self._translate_output_format_to_openai(
             anthropic_message_request=anthropic_message_request,

tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py

@@ -1582,6 +1582,34 @@ def test_thinking_still_translated_to_reasoning_effort_for_non_claude_model():
     assert new_kwargs["reasoning_effort"] == "low"
 
 
+def test_thinking_disabled_translated_to_reasoning_effort_none_for_non_claude_model():
+    adapter = LiteLLMAnthropicMessagesAdapter()
+    thinking = {"type": "disabled"}
+
+    new_kwargs = {"model": CACHE_CONTROL_NON_ANTHROPIC_MODEL}
+    adapter._translate_thinking_to_openai(cast(Any, {"thinking": thinking}), cast(Any, new_kwargs))
+
+    assert "thinking" not in new_kwargs
+    assert new_kwargs["reasoning_effort"] == "none"
+
+
+def test_stop_sequences_translated_to_stop_for_non_claude_model():
+    from litellm.types.llms.anthropic import AnthropicMessagesRequest
+
+    anthropic_request = AnthropicMessagesRequest(
+        model=CACHE_CONTROL_NON_ANTHROPIC_MODEL,
+        max_tokens=1024,
+        messages=[{"role": "user", "content": "hi"}],
+        stop_sequences=["</block>"],
+    )
+
+    adapter = LiteLLMAnthropicMessagesAdapter()
+    openai_request, _ = adapter.translate_anthropic_to_openai(anthropic_message_request=anthropic_request)
+
+    assert openai_request["stop"] == ["</block>"]
+    assert "stop_sequences" not in openai_request
+
+
 def test_cache_control_preserved_in_image_content_for_claude():
     """Cache control should be preserved in image content for Claude models."""
     anthropic_messages = [