Merge pull request #6736 from qiu121/codex/fix/openai-stream-failure-observability

fix(openai): 记录流式失败的上游诊断信息

backend/internal/service/openai_capacity_shed_test.go

@@ -298,6 +298,59 @@ func TestOpenAIStreamCapacityShedAfterOutputRewritesCodeForClient(t *testing.T)
 	require.True(t, logSink.ContainsFieldValue("upstream_request_id", "rid-shed-after-output"))
 }
 
+func TestOpenAIStreamProcessingFailureAfterOutputIsRecorded(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	stream := strings.Join([]string{
+		"event: response.created",
+		`data: {"type":"response.created","response":{"id":"resp_processing_failure"}}`,
+		"",
+		"event: response.output_text.delta",
+		`data: {"type":"response.output_text.delta","delta":"partial"}`,
+		"",
+		"event: response.failed",
+		`data: {"type":"response.failed","response":{"id":"resp_processing_failure","status":"failed","error":{"code":"server_error","message":"An error occurred while processing your request. Please include the request ID rid-processing-failure in your message."}}}`,
+		"",
+	}, "\n")
+
+	for _, passthrough := range []bool{false, true} {
+		t.Run(map[bool]string{false: "native", true: "passthrough"}[passthrough], func(t *testing.T) {
+			svc := &OpenAIGatewayService{cfg: &config.Config{Gateway: config.GatewayConfig{MaxLineSize: defaultMaxLineSize}}}
+			rec := httptest.NewRecorder()
+			c, _ := gin.CreateTestContext(rec)
+			c.Request = httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+			resp := &http.Response{
+				StatusCode: http.StatusOK,
+				Body:       io.NopCloser(strings.NewReader(stream)),
+				Header:     http.Header{"X-Request-Id": []string{"rid-processing-failure"}},
+			}
+			account := &Account{ID: 42, Platform: PlatformOpenAI, Type: AccountTypeOAuth, Name: "codex-account"}
+
+			var err error
+			if passthrough {
+				_, err = svc.handleStreamingResponsePassthrough(c.Request.Context(), resp, c, account, time.Now(), "gpt-5.6-terra", "gpt-5.6-terra")
+			} else {
+				_, err = svc.handleStreamingResponse(c.Request.Context(), resp, c, account, time.Now(), "gpt-5.6-terra", "gpt-5.6-terra")
+			}
+
+			require.Error(t, err)
+			var failoverErr *UpstreamFailoverError
+			require.False(t, errors.As(err, &failoverErr))
+			require.NotEmpty(t, rec.Body.String())
+			require.Contains(t, rec.Body.String(), "response.failed")
+			require.NotNil(t, c)
+			rawEvents, ok := c.Get(OpsUpstreamErrorsKey)
+			require.True(t, ok)
+			events, ok := rawEvents.([]*OpsUpstreamErrorEvent)
+			require.True(t, ok)
+			require.NotEmpty(t, events)
+			event := events[len(events)-1]
+			require.Equal(t, "stream_failed", event.Kind)
+			require.Equal(t, "rid-processing-failure", event.UpstreamRequestID)
+			require.Contains(t, event.Message, "An error occurred while processing your request")
+		})
+	}
+}
+
 // helper 单测：只有降载码被改写，其余错误码（尤其 rate_limit_exceeded，客户端
 // 依赖其原码解析重试延时）必须原样保留。
 func TestSanitizeOpenAICapacityShedErrorCodeForClient(t *testing.T) {

backend/internal/service/openai_gateway_passthrough.go

@@ -2079,6 +2079,11 @@ func (s *OpenAIGatewayService) handleStreamingResponsePassthrough(
 						s.handleOpenAIStreamTerminalAccountSideEffects(c, account, dataBytes, failedMessage, resp.Header, mappedModel)
 						bareErrorAccountSideEffectsPending = false
 					}
+					if eventType == "response.failed" {
+						// The stream cannot be replayed after semantic output. Preserve the
+						// terminal event, while making the upstream failure queryable.
+						s.recordOpenAIStreamUpstreamError(c, account, true, upstreamRequestID, "stream_failed", dataBytes, failedMessage)
+					}
 				}
 				if !outputStarted {
 					shouldFailover := false

backend/internal/service/openai_gateway_response_handling.go

@@ -561,6 +561,12 @@ func (s *OpenAIGatewayService) handleStreamingResponseWithReasoning(ctx context.
 						s.handleOpenAIStreamTerminalAccountSideEffects(c, account, dataBytes, failedMessage, resp.Header, mappedModel)
 						bareErrorAccountSideEffectsPending = false
 					}
+					if eventType == "response.failed" {
+						// Once semantic output is committed, failover replay is unsafe. Keep
+						// the terminal event on the existing stream, but retain the upstream
+						// request ID and payload for operations diagnostics.
+						s.recordOpenAIStreamUpstreamError(c, account, false, upstreamRequestID, "stream_failed", dataBytes, failedMessage)
+					}
 				}
 				if !outputStarted {
 					shouldFailover := false