fix(adapters): 流式响应丢失尾块 usage

src/amrita_core/adapters/openai.py

@@ -158,6 +158,14 @@ async def call_api(
                     or getattr(completion, "_request_id", None)
                 )
                 async for chunk in completion:
+                    # The final chunk of an OpenAI-compatible stream carries the
+                    # usage with an empty ``choices`` list, so it has to be read
+                    # before touching ``choices[0]`` — otherwise the IndexError
+                    # below ends the loop and the usage is lost.
+                    if chunk.usage:
+                        uni_usage = UniResponseUsage.model_validate(
+                            chunk.usage, from_attributes=True
+                        )
                     try:
                         if chunk.choices[0].finish_reason is not None:
                             model_name = chunk.model
@@ -171,10 +179,6 @@ async def call_api(
                                     else None
                                 ),
                             )
-                        if chunk.usage:
-                            uni_usage = UniResponseUsage.model_validate(
-                                chunk.usage, from_attributes=True
-                            )
 
                         if (
                             reas_chunk := getattr(
@@ -197,7 +201,9 @@ async def call_api(
                             yield chunk
                             debug_log(chunk)
                     except IndexError:
-                        break
+                        # A chunk without choices (e.g. the usage-only tail) is
+                        # not the end of the stream; keep draining it.
+                        continue
 
         else:
             if isinstance(completion, ChatCompletion):

tests/test_adapter.py

@@ -15,6 +15,7 @@
 from openai.types.chat.chat_completion_message_tool_call import (
     Function as ToolCallFunction,
 )
+from openai.types.completion_usage import CompletionUsage
 
 from amrita_core.adapters import anthropic as anthropic_module
 from amrita_core.adapters.anthropic import AnthropicAdapter
@@ -315,6 +316,52 @@ async def test_call_api_streaming_index_error(self, adapter, mock_messages):
             assert isinstance(results[0], UniResponse)
             assert results[0].content == ""
 
+    @pytest.mark.asyncio
+    async def test_call_api_streaming_usage_in_tail_chunk(self, adapter, mock_messages):
+        """usage 挂在 choices 为空的尾块上时不应丢失"""
+        adapter.preset.config.stream = True
+
+        body_chunk = ChatCompletionChunk(
+            id="chatcmpl-1",
+            choices=[
+                {
+                    "index": 0,
+                    "delta": {"content": "hi"},
+                    "finish_reason": None,
+                }
+            ],
+            created=1,
+            model="gpt-4o",
+            object="chat.completion.chunk",
+        )
+        tail_chunk = ChatCompletionChunk(
+            id="chatcmpl-1",
+            choices=[],
+            created=1,
+            model="gpt-4o",
+            object="chat.completion.chunk",
+            usage=CompletionUsage(
+                prompt_tokens=100, completion_tokens=20, total_tokens=120
+            ),
+        )
+
+        with patch("amrita_core.adapters.openai.openai.AsyncOpenAI") as mock_openai:
+            mock_client = AsyncMock()
+            mock_stream = MockAsyncStream([body_chunk, tail_chunk])
+            mock_client.chat.completions.create.return_value = mock_stream
+            mock_openai.return_value = mock_client
+
+            results = []
+            async for result in adapter.call_api(mock_messages):
+                results.append(result)
+
+        final = results[-1]
+        assert isinstance(final, UniResponse)
+        assert final.usage is not None
+        assert final.usage.prompt_tokens == 100
+        assert final.usage.completion_tokens == 20
+        assert final.usage.total_tokens == 120
+
     @pytest.mark.asyncio
     async def test_call_tools_with_string_tool_choice(self, adapter, mock_messages):
         """Test call_tools with string tool choice to cover client creation path"""