[Fix]: render tool_reference schema regardless of tool_result part order (#32522)

Signed-off-by: Shijin Zhang <75300765+Dovis01@users.noreply.github.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>

python/sglang/srt/entrypoints/anthropic/serving.py

@@ -297,7 +297,7 @@ class AnthropicServing:
 
         def _convert_tool_result_content(
             content: Any,
-        ) -> tuple[Union[str, list[dict]], str]:
+        ) -> tuple[list[Union[str, list[dict]]], str]:
             if isinstance(content, list):
                 tool_content_parts = []
                 tool_text_parts = []
@@ -342,17 +342,29 @@ class AnthropicServing:
                             )
 
                 tool_text = "\n".join(tool_text_parts)
-                if (
+                # GLM templates expand references only at the start of a tool
-                    len(tool_content_parts) == 1
+                # message, so isolate reference runs without changing part order.
-                    and tool_content_parts[0]["type"] == "text"
+                tool_content_groups: list[list[dict]] = []
-                ):
+                for part in tool_content_parts:
-                    return tool_content_parts[0]["text"], tool_text
+                    is_reference = part["type"] == "tool_reference"
-                if tool_content_parts:
+                    if (
-                    return tool_content_parts, tool_text
+                        not tool_content_groups
-                return "", tool_text
+                        or (tool_content_groups[-1][0]["type"] == "tool_reference")
+                        != is_reference
+                    ):
+                        tool_content_groups.append([])
+                    tool_content_groups[-1].append(part)
+
+                tool_contents: list[Union[str, list[dict]]] = []
+                for group in tool_content_groups:
+                    if len(group) == 1 and group[0]["type"] == "text":
+                        tool_contents.append(group[0]["text"])
+                    else:
+                        tool_contents.append(group)
+                return tool_contents or [""], tool_text
 
             tool_text = str(content) if content else ""
-            return tool_text, tool_text
+            return [tool_text], tool_text
 
         def _convert_assistant_thinking_blocks(
             blocks: list[AnthropicContentBlock],
@@ -484,7 +496,7 @@ class AnthropicServing:
                     tool_calls.append(tool_call)
 
                 elif block.type == "tool_result":
-                    tool_content, tool_text = _convert_tool_result_content(
+                    tool_contents, tool_text = _convert_tool_result_content(
                         block.content
                     )
 
@@ -497,13 +509,14 @@ class AnthropicServing:
                     # block must come AFTER that text in OpenAI form too).
                     if msg.role == "user":
                         _emit_user_message(content_parts)
-                        openai_messages.append(
+                        for tool_content in tool_contents:
-                            {
+                            openai_messages.append(
-                                "role": "tool",
+                                {
-                                "tool_call_id": tool_call_id,
+                                    "role": "tool",
-                                "content": tool_content,
+                                    "tool_call_id": tool_call_id,
-                            }
+                                    "content": tool_content,
-                        )
+                                }
+                            )
                     else:
                         content_parts.append(
                             {

test/registered/unit/entrypoints/anthropic/test_serving.py

@@ -8,6 +8,7 @@ from sglang.test.test_utils import maybe_stub_sgl_kernel
 maybe_stub_sgl_kernel()  # must precede imports that may pull in sgl_kernel
 
 from fastapi.responses import JSONResponse  # noqa: E402
+from jinja2 import Environment  # noqa: E402
 
 from sglang.srt.entrypoints.anthropic.protocol import (  # noqa: E402
     AnthropicMessage,
@@ -140,6 +141,22 @@ class TestAnthropicServing(unittest.TestCase):
         "{{- message.role }}: {{ message.content }}\n"
         "{%- endfor %}"
     )
+    GLM_TOOL_RESULT_TEMPLATE = """
+{%- for message in messages if message.role == "tool" -%}
+{%- if loop.first -%}<|observation|>{%- endif -%}
+{%- if message.content is string -%}
+<tool_response>{{ message.content }}</tool_response>
+{%- elif message.content.0.type == "tool_reference" -%}
+<tool_response><tools>
+{%- for reference in message.content -%}
+{%- for tool in tools if tool.function.name == reference.name -%}
+{{ tool.function.name }}
+{%- endfor -%}
+{%- endfor -%}
+</tools></tool_response>
+{%- endif -%}
+{%- endfor -%}
+"""
 
     def _serving(self, stream_lines=None, chat_template=None):
         return AnthropicServing(_FakeOpenAIServingChat(stream_lines, chat_template))
@@ -154,6 +171,26 @@ class TestAnthropicServing(unittest.TestCase):
         data.update(overrides)
         return AnthropicMessagesRequest.model_validate(data)
 
+    def _tool_result_request(self, content, tools=None):
+        overrides = {
+            "stream": False,
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "tool_result",
+                            "tool_use_id": "call_1",
+                            "content": content,
+                        }
+                    ],
+                }
+            ],
+        }
+        if tools is not None:
+            overrides["tools"] = tools
+        return self._anthropic_request(**overrides)
+
     def test_stream_closes_tool_block_before_text_delta(self):
         serving = self._serving(
             [
@@ -373,6 +410,74 @@ class TestAnthropicServing(unittest.TestCase):
         self.assertIn("https://docs.sglang.ai", tool_message["content"])
         self.assertIn("Anthropic API notes", tool_message["content"])
 
+    def test_mixed_tool_reference_content_preserves_part_order(self):
+        request = self._tool_result_request(
+            [
+                {"type": "text", "text": "Tool loaded: Bash"},
+                {"type": "tool_reference", "tool_name": "Bash"},
+                {"type": "text", "text": "Ready"},
+            ]
+        )
+
+        chat_request = self._serving()._convert_to_chat_completion_request(request)
+        messages = chat_request.model_dump(exclude_none=True)["messages"]
+
+        self.assertEqual([message["role"] for message in messages], ["tool"] * 3)
+        self.assertEqual(
+            [message["tool_call_id"] for message in messages], ["call_1"] * 3
+        )
+        self.assertEqual(messages[0]["content"], "Tool loaded: Bash")
+        self.assertEqual(
+            messages[1]["content"],
+            [{"type": "tool_reference", "name": "Bash"}],
+        )
+        self.assertEqual(messages[2]["content"], "Ready")
+
+    def test_mixed_tool_reference_content_renders_text_and_schema(self):
+        template = Environment().from_string(self.GLM_TOOL_RESULT_TEMPLATE)
+        request = self._tool_result_request(
+            [
+                {"type": "text", "text": "Tool loaded: Bash"},
+                {"type": "tool_reference", "tool_name": "Bash"},
+            ],
+            tools=[
+                {
+                    "name": "Bash",
+                    "description": "Run a shell command",
+                    "input_schema": {"type": "object", "properties": {}},
+                    "defer_loading": True,
+                }
+            ],
+        )
+
+        chat_request = self._serving()._convert_to_chat_completion_request(request)
+        payload = chat_request.model_dump(exclude_none=True)
+        prompt = template.render(messages=payload["messages"], tools=payload["tools"])
+
+        self.assertIn("<tool_response>Tool loaded: Bash</tool_response>", prompt)
+        self.assertIn("<tools>Bash</tools>", prompt)
+        self.assertEqual(prompt.count("<|observation|>"), 1)
+
+    def test_reference_only_tool_result_remains_one_message(self):
+        request = self._tool_result_request(
+            [
+                {"type": "tool_reference", "tool_name": "Bash"},
+                {"type": "tool_reference", "tool_name": "Read"},
+            ]
+        )
+
+        chat_request = self._serving()._convert_to_chat_completion_request(request)
+        messages = chat_request.model_dump(exclude_none=True)["messages"]
+
+        self.assertEqual(len(messages), 1)
+        self.assertEqual(
+            messages[0]["content"],
+            [
+                {"type": "tool_reference", "name": "Bash"},
+                {"type": "tool_reference", "name": "Read"},
+            ],
+        )
+
     def test_builtin_web_search_tool_without_schema_is_skipped(self):
         request = AnthropicMessagesRequest.model_validate(
             {