Shiny no longer accepts reactive function for xreg value - r

Question

I have been working on a shiny app which uses auto.arima() from the package, along with an xreg parameter. For some reason, the xreg function (which takes a data frame or matrix) has stopped accepting the output of a reactive function - which is a subsetted dataframe. The work-around sucks, and so I'm hoping someone knows a good way to fix this.
Shiny produces this ERROR:
xreg should be a numeric matrix or vector
I've tried:
re-installing: forecast, shiny, dplyr, and completing wiping our R and Rstudio, and re-installing from scratch.
I've tried rolling back the forecast package to the build from 6 months ago (when this was working)
I've tried rolling back R to the build from 6 months ago (when this was working)
I've tried using other forecast packages; bsts for example seems to work just fine
This is some minimally viable shiny code. Create a .csv with 3 columns of data, be sure to name the columns, like y, x1 and x2.
library(shiny)
library(forecast)
library(dplyr)
ui <- fluidPage(
# Application title
titlePanel("arima test"),
# Sidebar with a slider input for number of bins
sidebarLayout(
sidebarPanel(
# Input: Select a file ----
fileInput("uploaded_file", "Choose CSV File",
multiple = TRUE,
accept = c("text/csv",
"text/comma-separated-values,text/plain",
".csv")),
uiOutput("dv_dropdown"),
uiOutput("iv_checkbox")
),
# Show a plot of the generated distribution
mainPanel(
verbatimTextOutput("model"),
tableOutput('data')
)
)
)
server <- function(input, output) {
# Read .csv file ----
df <- reactive({
req(input$uploaded_file)
read.csv(input$uploaded_file$datapath,
header = TRUE)
})
# Dynamically generate a list of the dependent variables for the user to choose from the uploaded .csv ----
output$dv_dropdown <- renderUI({
selectInput(inputId = "select_dependents",
label = "dependent time series:",
choices = names(df()[,c(-1)])
)
})
# Dynamically generate a list of the independent variables for the user to choose ----
output$iv_checkbox <- renderUI({
checkboxGroupInput(inputId = "select_ivs",
label = "Select Independent variables:",
choices = setdiff(names(df()[,c(-1)]), input$select_dependents),
selected = setdiff(names(df()), input$select_dependents))
})
####### CREATE REACTIVE DATA OBJECTS BASED ON FILTERS #################
# this creates a dataframe of the dependent time series
df_sel <- reactive({
req(input$select_dependents)
df_sel <- df() %>% select(input$select_dependents)
})
# this creates a dataframe of the dependent time series
df_indep <- reactive({
req(input$select_ivs)
df_indep <- df() %>% select(input$select_ivs)
})
#proof that the independent variables are in fact in a matrix/dataframe
output$data <- renderTable({
df_indep()
})
#This is the TEST ARIMA Model
modela <- reactive({
modela <- auto.arima(df_sel(), xreg = df_indep()) #model with regressors
#modela <- auto.arima(df_sel()) #model without regressors to demonstrate that it actually works without the xreg argument
})
#this is the solution to the xreg problem above, but there's no reason I can think of that xreg shouldn't be able to take a reactive data object like it was doing before...
#x = as.matrix(as.data.frame(lapply(df_indep(), as.numeric)))
# then set: xreg = x
output$model <- renderPrint({
summary(modela())
})
}
# Run the application
shinyApp(ui = ui, server = server)