fix: drain OpenAI compat streams for usage

backend/internal/pkg/apicompat/anthropic_responses_test.go

@@ -434,6 +434,45 @@ func TestStreamingTextOnly(t *testing.T) {
 	assert.Equal(t, "message_stop", events[1].Type)
 }
 
+func TestResponsesEventToAnthropicEvents_ResponseDone(t *testing.T) {
+	state := NewResponsesEventToAnthropicState()
+	state.Model = "gpt-4o"
+
+	events := ResponsesEventToAnthropicEvents(&ResponsesStreamEvent{
+		Type: "response.done",
+		Response: &ResponsesResponse{
+			Status: "completed",
+			Usage:  &ResponsesUsage{InputTokens: 12, OutputTokens: 4},
+		},
+	}, state)
+	require.Len(t, events, 2)
+	assert.Equal(t, "message_delta", events[0].Type)
+	assert.Equal(t, "end_turn", events[0].Delta.StopReason)
+	assert.Equal(t, 12, events[0].Usage.InputTokens)
+	assert.Equal(t, 4, events[0].Usage.OutputTokens)
+	assert.Equal(t, "message_stop", events[1].Type)
+	assert.Nil(t, FinalizeResponsesAnthropicStream(state))
+}
+
+func TestResponsesEventToAnthropicEvents_ResponseDoneIncomplete(t *testing.T) {
+	state := NewResponsesEventToAnthropicState()
+	state.Model = "gpt-4o"
+
+	events := ResponsesEventToAnthropicEvents(&ResponsesStreamEvent{
+		Type: "response.done",
+		Response: &ResponsesResponse{
+			Status:            "incomplete",
+			IncompleteDetails: &ResponsesIncompleteDetails{Reason: "max_output_tokens"},
+			Usage:             &ResponsesUsage{InputTokens: 12, OutputTokens: 4},
+		},
+	}, state)
+	require.Len(t, events, 2)
+	assert.Equal(t, "message_delta", events[0].Type)
+	assert.Equal(t, "max_tokens", events[0].Delta.StopReason)
+	assert.Equal(t, "message_stop", events[1].Type)
+	assert.Nil(t, FinalizeResponsesAnthropicStream(state))
+}
+
 func TestStreamingCachedTokensUseAnthropicInputSemantics(t *testing.T) {
 	state := NewResponsesEventToAnthropicState()
 	ResponsesEventToAnthropicEvents(&ResponsesStreamEvent{

backend/internal/pkg/apicompat/chatcompletions_responses_test.go

@@ -720,6 +720,49 @@ func TestResponsesEventToChatChunks_Completed(t *testing.T) {
 	assert.Equal(t, 30, chunks[1].Usage.PromptTokensDetails.CachedTokens)
 }
 
+func TestResponsesEventToChatChunks_ResponseDone(t *testing.T) {
+	state := NewResponsesEventToChatState()
+	state.Model = "gpt-4o"
+	state.IncludeUsage = true
+
+	chunks := ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type: "response.done",
+		Response: &ResponsesResponse{
+			Status: "completed",
+			Usage:  &ResponsesUsage{InputTokens: 13, OutputTokens: 7},
+		},
+	}, state)
+	require.Len(t, chunks, 2)
+	require.NotNil(t, chunks[0].Choices[0].FinishReason)
+	assert.Equal(t, "stop", *chunks[0].Choices[0].FinishReason)
+	require.NotNil(t, chunks[1].Usage)
+	assert.Equal(t, 13, chunks[1].Usage.PromptTokens)
+	assert.Equal(t, 7, chunks[1].Usage.CompletionTokens)
+	assert.Nil(t, FinalizeResponsesChatStream(state))
+}
+
+func TestResponsesEventToChatChunks_ResponseDoneIncomplete(t *testing.T) {
+	state := NewResponsesEventToChatState()
+	state.Model = "gpt-4o"
+	state.IncludeUsage = true
+
+	chunks := ResponsesEventToChatChunks(&ResponsesStreamEvent{
+		Type: "response.done",
+		Response: &ResponsesResponse{
+			Status:            "incomplete",
+			IncompleteDetails: &ResponsesIncompleteDetails{Reason: "max_output_tokens"},
+			Usage:             &ResponsesUsage{InputTokens: 13, OutputTokens: 7},
+		},
+	}, state)
+	require.Len(t, chunks, 2)
+	require.NotNil(t, chunks[0].Choices[0].FinishReason)
+	assert.Equal(t, "length", *chunks[0].Choices[0].FinishReason)
+	require.NotNil(t, chunks[1].Usage)
+	assert.Equal(t, 13, chunks[1].Usage.PromptTokens)
+	assert.Equal(t, 7, chunks[1].Usage.CompletionTokens)
+	assert.Nil(t, FinalizeResponsesChatStream(state))
+}
+
 func TestResponsesEventToChatChunks_CompletedWithToolCalls(t *testing.T) {
 	state := NewResponsesEventToChatState()
 	state.Model = "gpt-4o"

backend/internal/pkg/apicompat/responses_to_anthropic.go

@@ -212,7 +212,9 @@ func ResponsesEventToAnthropicEvents(
 		return resToAnthHandleReasoningDelta(evt, state)
 	case "response.reasoning_summary_text.done":
 		return resToAnthHandleBlockDone(state)
-	case "response.completed", "response.incomplete", "response.failed":
+	// response.done 是 Realtime/WS 与项目透传路径使用的终止别名；
+	// 普通 Responses HTTP SSE 的公开终止事件仍以 response.completed 为主。
+	case "response.completed", "response.done", "response.incomplete", "response.failed":
 		return resToAnthHandleCompleted(evt, state)
 	default:
 		return nil

backend/internal/pkg/apicompat/responses_to_chatcompletions.go

@@ -160,7 +160,9 @@ func ResponsesEventToChatChunks(evt *ResponsesStreamEvent, state *ResponsesEvent
 		return resToChatHandleReasoningDelta(evt, state)
 	case "response.reasoning_summary_text.done":
 		return nil
-	case "response.completed", "response.incomplete", "response.failed":
+	// response.done 是 Realtime/WS 与项目透传路径使用的终止别名；
+	// 普通 Responses HTTP SSE 的公开终止事件仍以 response.completed 为主。
+	case "response.completed", "response.done", "response.incomplete", "response.failed":
 		return resToChatHandleCompleted(evt, state)
 	default:
 		return nil

backend/internal/pkg/apicompat/types.go

@@ -314,7 +314,7 @@ type ResponsesOutputTokensDetails struct {
 type ResponsesStreamEvent struct {
 	Type string `json:"type"`
 
-	// response.created / response.completed / response.failed / response.incomplete
+	// response.created / response.completed / response.done / response.failed / response.incomplete
 	Response *ResponsesResponse `json:"response,omitempty"`
 
 	// response.output_item.added / response.output_item.done