Multiple StackingEstimators - interpretation?

In running a TPOT job, the best pipeline has multiple StackingEstimators in sequence:

StackingEstimator(estimator=LassoLarsCV(normalize=True)),
StackingEstimator(estimator=AdaBoostRegressor(learning_rate=0.5, loss="square", n_estimators=100)),
StackingEstimator(estimator=RidgeCV()),
SGDRegressor(alpha=0.001, eta0=1.0, fit_intercept=True, l1_ratio=0.75, learning_rate="invscaling", loss="huber", penalty="elasticnet", power_t=0.1)

In this situation, how do I interpret what is actually happening? I see two possibilities:

One possibility is that LassoLarsCV, AdaBoostRegressor, and RidgeCV make predictions where are appended to the data matrix and SGDRegressor trains a model with p+3 data columns.

More likely (I think), LassoLarsCV uses a matrix of p columns and its prediction is appended, AdaBoostRegressor uses a matrix of p+1 (data plus LassoLarsCV prediction), RidgeCV uses a matrix of p+2, and finally SGDRegressor uses a matrix of p+3.

If the second interpretation that a progressive stacking occurs, is it possible for TPOT to generate the first interpretation in which 3 different predictors contribute to a matrix with p+3 columns but each of those has only a p columns to start with?