I have seen two equivalent definitions of the modular sheaf $\omega$. Let $S$ be some base scheme. If $p \colon \mathcal{E} \to X$ is the universal generalized elliptic curve over the modular curve $X$, and $e \colon X \to \mathcal{E}$ is the zero section (so that $e(x)$ is the identity in the fiber $p^{-1}(x)$), then $$ \omega = p_*\Omega^1_{\mathcal{E}/S} = e^*\Omega^1_{\mathcal{E}/S}. $$ Here $\Omega^1_{\mathcal{E}/S}$ is the sheaf of relative 1-forms on $\mathcal{E}$ with logarithmic poles at the cusps.
I believe that "the reason" that these are the same is because $\Omega^1_{E/S}$ is a trivial line bundle for any generalized elliptic curve $E = p^{-1}(x)$, and its sections are all constant. This gives a natural way to identify the sections of $e^*\Omega^1_{\mathcal{E}/S}$ with the sections of $p_*\Omega^1_{\mathcal{E}/S}$: just extend any section of $e^*\Omega^1_{\mathcal{E}/S}$ to all of $\Omega^1_{\mathcal{E}/S}$ by making it constant along the fibers, which we can do because it's a trivial bundle. The property above says that any section of $p_*\Omega^1_{\mathcal{E}/S}$ will be constant along the fibers anyway, so we get all of them.
My question is, in how much generality can we expect a natural isomorphism $e^*\mathcal{F} \cong p_*\mathcal{F}$ where $p \colon T \to X$ is a fiber bundle, $e \colon X \to T$ is a section, and $\mathcal{F}$ is a sheaf on $T$? Does my reasoning from the previous paragraph have to be true, or can we weaken it? (I know this is supposed to be true for, for example, Hilbert and Siegel modular forms, but I think the reasoning above applies in those cases.)