fix(llm): adapt async tool results for compatible APIs

backend/services/pipecat/service_factory.py

@@ -117,7 +117,7 @@ def create_llm(cfg: AssistantConfig):
         raise ValueError(f"不支持的 LLM 接口类型: {cfg.llm_interface_type}")
     extra_body = cfg.llm_values.get("extraBody")
     extra = {"extra_body": extra_body} if isinstance(extra_body, dict) else {}
-    return OpenAILLMService(
+    service = OpenAILLMService(
         api_key=_require(cfg.llm_api_key, "LLM apiKey"),
         base_url=_require(cfg.llm_base_url, "LLM apiUrl"),
         settings=OpenAILLMService.Settings(
@@ -125,6 +125,15 @@ def create_llm(cfg: AssistantConfig):
             extra=extra,
         ),
     )
+    # Pipecat represents late async-tool results with the newer `developer`
+    # role. Most OpenAI-compatible providers (including DeepSeek) only accept
+    # system/assistant/user/tool. This flag uses Pipecat's built-in adapter to
+    # downgrade developer messages before sending the request. Official or
+    # otherwise compatible endpoints can opt in through the resource values.
+    service.supports_developer_role = (
+        cfg.llm_values.get("supportsDeveloperRole") is True
+    )
+    return service
 
 
 def create_tts(cfg: AssistantConfig):

backend/tests/test_service_factory.py

@@ -4,13 +4,61 @@ import unittest
 from unittest.mock import patch
 
 from models import AssistantConfig
+from pipecat.processors.aggregators.llm_context import LLMContext
 from services.pipecat.service_factory import (
     HTTP_TTS_STOP_FRAME_TIMEOUT_S,
     WEBSOCKET_TTS_STOP_FRAME_TIMEOUT_S,
+    create_llm,
     create_tts,
 )
 
 
+class LLMServiceFactoryTest(unittest.TestCase):
+    def test_openai_compatible_llm_converts_async_tool_developer_result(self):
+        service = create_llm(
+            AssistantConfig(
+                llm_interface_type="openai-llm",
+                model="deepseek-chat",
+                llm_api_key="test-key",
+                llm_base_url="https://llm.example.test/v1",
+            )
+        )
+        context = LLMContext(
+            messages=[
+                {"role": "system", "content": "你是助手"},
+                {
+                    "role": "developer",
+                    "content": '{"type":"async_tool","status":"finished"}',
+                },
+                {"role": "user", "content": "ok"},
+            ]
+        )
+
+        params = service.get_llm_adapter().get_llm_invocation_params(
+            context,
+            convert_developer_to_user=not service.supports_developer_role,
+        )
+
+        self.assertEqual(
+            [message["role"] for message in params["messages"]],
+            ["system", "user", "user"],
+        )
+        self.assertNotIn("developer", str(params["messages"]))
+
+    def test_provider_can_explicitly_enable_developer_role(self):
+        service = create_llm(
+            AssistantConfig(
+                llm_interface_type="openai-llm",
+                model="gpt-compatible",
+                llm_api_key="test-key",
+                llm_base_url="https://llm.example.test/v1",
+                llm_values={"supportsDeveloperRole": True},
+            )
+        )
+
+        self.assertTrue(service.supports_developer_role)
+
+
 class TTSServiceFactoryTest(unittest.TestCase):
     def test_http_tts_keeps_wider_audio_chunk_timeout(self):
         config = AssistantConfig(