Matrix parsing in kyber

How can $\operatorname{SHAKE-128}$ be used to generate the (possibly non-terminating) bytes stream of indeterminate length required by the parse function?

$\operatorname{SHAKE-128}$ has the property that for any message $M$ and any output lengths $d$ and $d'$ with $d<d'$, $\operatorname{SHAKE-128}(M,d)$ is the beginning of $\operatorname{SHAKE-128}(M,d')$.

The internal structure of $\operatorname{Keccak}$ and it's extension to $\operatorname{SHAKE-128}$ makes it possible to make all the computations involving $M$, then an output of length $d$, then dynamically extend the output to whatever $d'>d$ without redoing any of the previous sponge rounds already performed.

Granted, not all APIs to $\operatorname{SHAKE-128}$ allow that. E.g. I don't see how to reuse Python's shake_128 to ask for more output without causing some duplicate work. We don't need to rehash $M$, but we need to recompute the whole output.