MLR: How to extract feature Selection inside a benchmark

Question

I want to do a benchmark with different models in mlr via 3 fold cross validation. In every fold i want to do again via a 3 fold cross validation a feature selection for every model and pass the best feature set to the outer cross validation. I noticed that the result of a benchmark in MLR uses always all the features included.

How can i extract from a benchmark the features used in every fold and every model and how do i make sure they are really used for the outer Cross Validation fold?

Here is a sample code:

task_cv <- makeClassifTask(
  id = 'predict future outages',
  data = data, 
  target = 'targetVariable', 
  positive=1
)

vali_strat <- makeResampleDesc(method="CV",iters = 3)

featSelControl<- makeFeatSelControlSequential(same.resampling.instance = T,
                                                        method = "sbs",
                                                        tune.threshold = T,
                                                        alpha = 4,
                                                        beta = 4)

learner_nv <- makeLearner(
  id = 'Naive Bayes',
  cl = 'classif.naiveBayes'
)

learner_knn <- makeLearner(
  id = 'KNN',
  cl = 'classif.kknn'
)

featSel_nv <- makeFeatSelWrapper(learner = learner_nv,
                                          resampling = vali_strat,
                                          control = featSelControl,
                                          measures = acc

featSel_knn <- makeFeatSelWrapper(learner = learner_knn,
                                           resampling = vali_strat,
                                           control = featSelControl,
                                           measures = acc


learners <- list(featSel_nv,
                featSel_knn ) 

benchmark = benchmark(
  learners = learners,
  tasks = task_cv,
  resamplings = validation_strategy,
  measures = acc
)

benchmark$results$`predict future outages`$KNN.featsel$models[[1]]$features

I cannot extract the used features and the last line in the code indicates that always all features are used instead the selected one via featureSelection.