This answer (https://physics.stackexchange.com/q/348673) provides good intuition for why Legendre transformation induces 1-particle irreducible graphs: It mainly tries to convey the idea that the condition for tree level functional expansion for the effective action is equivalent to a Legendre transformation, and due to the uniqueness for decompositions of connected graphs into trees with 1PI vertices, Legendre transformation should leads to one-particle irreducible graphs.
Is there a similar argument for 2PI effective actions? Or a derivation of 2PI skeleton expansion to all orders?
For the moment, I would like to consider the Dyson equation as an extra consistency condition between 1PI and 2PI, such that the functional derivative give rise to the 1PI self-energy: \begin{equation} \frac{\delta \Gamma_2[\phi,G]}{\delta G} = \Sigma_{\text{1PI}}, \end{equation} which means that $\Gamma[\phi,G]$ must only contain two-particle irreducible graphs (Because functional derivative w.r.t. $G$ can be represented graphically as cutting one line.). The Dyson equation looks like following: \begin{equation} \frac{\delta[S[\phi]+\imath\text{Tr}\ln(G^{-1})+\imath\text{Tr}G_0^{-1}G+ \Gamma_2[\phi,G]]}{\delta G}=0=-\imath G^{-1}+\imath G_0^{-1}+\Sigma_{\text{1PI}}. \end{equation} The expression on the left hand side come from the double Legendre transformation.