fix(anthropic): use model capability flag for response_format routing + eager_input_streaming + citation offsets

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/llms/anthropic/chat/transformation.py

@@ -790,6 +790,26 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
                         raise ValueError("defer_loading must be a boolean")
                     returned_tool["defer_loading"] = _defer_loading_function  # type: ignore[typeddict-item]
 
+        ## check if eager_input_streaming is set in the tool
+        _eager_input_streaming = tool.get("eager_input_streaming", None)
+        _eager_input_streaming_function = tool.get("function", {}).get("eager_input_streaming", None)
+        if returned_tool is not None:
+            tool_type = returned_tool.get("type", "")
+            if tool_type not in (
+                "tool_search_tool_regex_20251119",
+                "tool_search_tool_bm25_20251119",
+                "computer_20241022",
+                "computer_20250124",
+            ):
+                if _eager_input_streaming is not None:
+                    if not isinstance(_eager_input_streaming, bool):
+                        raise ValueError("eager_input_streaming must be a boolean")
+                    returned_tool["eager_input_streaming"] = _eager_input_streaming  # type: ignore[typeddict-item]
+                elif _eager_input_streaming_function is not None:
+                    if not isinstance(_eager_input_streaming_function, bool):
+                        raise ValueError("eager_input_streaming must be a boolean")
+                    returned_tool["eager_input_streaming"] = _eager_input_streaming_function  # type: ignore[typeddict-item]
+
         ## check if allowed_callers is set in the tool
         _allowed_callers = tool.get("allowed_callers", None)
         _allowed_callers_function = tool.get("function", {}).get("allowed_callers", None)
@@ -1411,25 +1431,7 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
                     output_key=param,
                 )
             elif param == "response_format" and isinstance(value, dict):
-                if any(
-                    substring in model
-                    for substring in {
-                        "sonnet-4.5",
-                        "sonnet-4-5",
-                        "opus-4.1",
-                        "opus-4-1",
-                        "opus-4.5",
-                        "opus-4-5",
-                        "opus-4.6",
-                        "opus-4-6",
-                        "opus-4.7",
-                        "opus-4-7",
-                        "sonnet-4.6",
-                        "sonnet-4-6",
-                        "sonnet_4.6",
-                        "sonnet_4_6",
-                    }
-                ):
+                if self._supports_model_capability(model, "supports_output_config"):
                     _output_format = self.map_response_format_to_anthropic_output_format(value)
                     if _output_format is not None:
                         optional_params["output_format"] = _output_format
@@ -1980,6 +1982,7 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
         tool_results: Optional[List[Any]] = None
         compaction_blocks: Optional[List[Any]] = None
         for idx, content in enumerate(completion_response["content"]):
+            text_start = len(text_content)
             if content["type"] == "text":
                 text_content += content["text"]
             ## TOOL CALLING
@@ -2034,6 +2037,8 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
                         {
                             **citation,
                             "supported_text": content.get("text", ""),
+                            "supported_text_start_char_index": text_start,
+                            "supported_text_end_char_index": len(text_content),
                         }
                         for citation in content["citations"]
                     ]

litellm/types/llms/anthropic.py

@@ -50,6 +50,7 @@ class AnthropicMessagesTool(TypedDict, total=False):
     type: Literal["custom"]
     cache_control: Optional[Union[dict, ChatCompletionCachedContent]]
     defer_loading: bool
+    eager_input_streaming: bool
     allowed_callers: Optional[List[str]]
     input_examples: Optional[List[Dict[str, Any]]]
 

tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py

@@ -267,6 +267,8 @@ def test_extract_response_content_with_citations():
                 "start_char_index": 0,
                 "end_char_index": 20,
                 "supported_text": "the grass is green",
+                "supported_text_start_char_index": 28,
+                "supported_text_end_char_index": 46,
             },
         ],
         [
@@ -278,6 +280,8 @@ def test_extract_response_content_with_citations():
                 "start_char_index": 20,
                 "end_char_index": 36,
                 "supported_text": "the sky is blue",
+                "supported_text_start_char_index": 51,
+                "supported_text_end_char_index": 66,
             },
         ],
     ]
@@ -5733,3 +5737,91 @@ def test_top_k_forwarded_at_transform_on_models_that_accept_it():
     )
 
     assert result["top_k"] == 40
+
+
+def test_opus_4_8_uses_native_structured_output():
+    """claude-opus-4-8 has supports_output_config=true in the model map,
+    so response_format should route to the native output_format path
+    rather than the tool-call emulation path"""
+    config = AnthropicConfig()
+
+    response_format = {
+        "type": "json_schema",
+        "json_schema": {
+            "name": "test_schema",
+            "schema": {
+                "type": "object",
+                "properties": {"answer": {"type": "string"}},
+                "required": ["answer"],
+                "additionalProperties": False,
+            },
+        },
+    }
+
+    optional_params = config.map_openai_params(
+        non_default_params={"response_format": response_format},
+        optional_params={},
+        model="claude-opus-4-8",
+        drop_params=False,
+    )
+
+    assert "output_format" in optional_params
+    assert optional_params["output_format"]["type"] == "json_schema"
+    assert "tools" not in optional_params
+    assert "tool_choice" not in optional_params
+    assert optional_params.get("json_mode") is True
+
+
+def test_map_tool_helper_eager_input_streaming_passthrough():
+    """eager_input_streaming on a custom tool must survive the mapping"""
+    config = AnthropicConfig()
+
+    tool = {
+        "type": "function",
+        "function": {
+            "name": "my_tool",
+            "description": "a tool",
+            "parameters": {"type": "object", "properties": {}},
+            "eager_input_streaming": True,
+        },
+    }
+
+    result, _ = config._map_tool_helper(tool)
+    assert result is not None
+    assert result.get("eager_input_streaming") is True
+
+
+def test_map_tool_helper_eager_input_streaming_top_level():
+    """eager_input_streaming set at top level (not nested in function)"""
+    config = AnthropicConfig()
+
+    tool = {
+        "type": "function",
+        "eager_input_streaming": True,
+        "function": {
+            "name": "my_tool",
+            "description": "a tool",
+            "parameters": {"type": "object", "properties": {}},
+        },
+    }
+
+    result, _ = config._map_tool_helper(tool)
+    assert result is not None
+    assert result.get("eager_input_streaming") is True
+
+
+def test_map_tool_helper_eager_input_streaming_rejects_non_bool():
+    """eager_input_streaming must be a boolean"""
+    config = AnthropicConfig()
+
+    tool = {
+        "type": "function",
+        "function": {
+            "name": "my_tool",
+            "parameters": {"type": "object", "properties": {}},
+            "eager_input_streaming": "yes",
+        },
+    }
+
+    with pytest.raises(ValueError, match="eager_input_streaming must be a boolean"):
+        config._map_tool_helper(tool)