I think the answer is neither one sadly. If you do a taylor expansion you find
\begin{equation}
\partial_\mu e^{i t^a \phi_a} = \sum_n \left(\frac{i^n}{n!}\right) \partial_\mu (\phi^a t_a)^n = i \partial_\mu (\phi^a t_a) - \frac{1}{2}\left( \partial_\mu (\phi^a t_a) \phi^b t_b + \phi^a t_a \partial_\mu (\phi^b t_b) \right) + \cdots
\end{equation}
The second term on the right hand side already prevents us from factoring the answer into anything nice, since $\partial \phi $ does not commute with $\phi $. At higher order you can get even uglier things like $\phi^n \partial \phi \phi^m$.
If we work in $0+1$ dimensions, then what you are basically asking is for the derivative of $e^{X(t)}$ with respect to $t$, where $X$ is a matrix. There is a (terrible) formula for this on wikipedia (https://en.wikipedia.org/wiki/Matrix_exponential)
\begin{equation}
\frac{d}{dt} e^{X(t)} = \int_0^1 d\alpha e^{\alpha X(t)}\frac{dX}{dt}e^{(1-\alpha)X(t)}
\end{equation}
I suspect this is rarely useful in practice (or, at least in chiral perturbation theory).
I'm assuming this comes up when expanding out the effective action in terms of $\phi$. I think probably the only thing you can really do is work perturbatively and keep track of all the different vertices you get. Physically I think what's going on is basically "identical particles are identical," so if you have a vertex like $\partial \phi \partial \phi \phi \phi$ you must also have a corresponding vertex where you exchange some of the $\phi$'s, like $\partial \phi \phi \partial \phi \phi$.
If you are working with a lagrangian there is presumably a trace, in which case you can use the cyclic property of the trace to simplify things somewhat, for example from ${\rm tr}(\partial_\mu U \partial^\mu U^\dagger)$ you end up with vertices like
\begin{equation}
{\rm tr}(\partial U \partial U^\dagger) \supset {\rm tr}\left( t_a t_b t_c t_d\right)\left( \partial_\mu \phi^a \phi^b + \phi^a \partial_\mu \phi^b \right) \left(\partial^\mu \phi^c \phi^d + \phi^c \partial^\mu \phi^d \right) = 2 {\rm tr}\left(t_a t_b t_c t_d\right) \left(\partial_\mu \phi^a \phi^b \partial^\mu \phi^c \phi^d + \partial_\mu \phi^a \partial^\mu \phi^b \phi^c \phi^d\right)
\end{equation}
where in the last line we used the cyclic property of the trace.