The first step is to deal with the min-max objective function. If you deal with these often enough, you might derive a general rule, but one way to turn this into a standard linear program is as follows.
- Introduce a new variable $z$, with the intent that in an optimal solution, $$z = \max_{s \in S} \sum_{t \in \delta(s)} x(st).$$
- To enforce this intent, add the inequalities $$z - \sum_{t \in \delta(s)} x(st) \ge0 \qquad \forall s \in S.$$
- Now, the objective function will be to minimize $z$. Although the inequalities from step 2 only guarantee that $$z \ge \max_{s \in S} \sum_{t \in \delta(s)} x(st)$$ in a feasible solution, since we are minimizing $z$, it will actually be equal to this maximum in an optimal solution.
Once you have done this, you can start dealing with the LP in the usual way. Since we have a constraint for every $s \in S$, the dual will have a variable for every $s \in S$, as well as a variable for every $t \in T$ coming from the constraints you already had.