test_that("optimizing works", {
  skip("Backwards compatibility")

  code <- "
    parameters {
      real y;
    } model {
      y ~ normal(0,1);
    }
  "
  m <- stan_model(model_code = code)
  o <- optimizing(m, hessian = TRUE)
  expect_equal(o$par[1], 0, tolerance = 0.1, ignore_attr = "names")
  expect_equal(o$hessian[1,1], -1, tolerance = 0.1)

  mc <- "
    parameters {
      simplex[2] a;
      real y;
    }
    model {
      y ~ normal(0,1);
      target += -square(a[1]) - square(a[2]);
    }
    generated quantities {
      matrix[4, 5] B;
      for (i in 1:4) for (j in 1:5) B[i,j] = i * 10 + j;
    }
  "
  m2 <- stan_model(model_code = mc)
  set.seed(1287)
  o2 <- optimizing(m2, hessian = TRUE, seed = 4)
  # Check the names of the returns.
  expect_equal(o2$par["B[3,4]"], 34, ignore_attr = "names")
  expect_equal(o2$par["B[1,4]"], 14, ignore_attr = "names")
  expect_equal(o2$par["B[4,5]"], 45, ignore_attr = "names")
  expect_equal(o2$par["B[4,1]"], 41, ignore_attr = "names")

  set.seed(1287)
  o3 <- optimizing(m2, as_vector = FALSE, seed = 4)
  s <- list(a = 1:2, y = 3, B = array(0, dim = c(4,5)))
  attr(s$a, "dim") <- 2
  expect_equal(o3$par, rstan_relist(o2$par, s))
  expect_equal(o2$par[3], 0, tolerance = 0.1, checkNames = FALSE)
})