[guided]The point of this step is to make precise what “restricting $d$ to CR antiholomorphic directions” means. Start with an open set $U\subset M$ on which the bundle $T^{0,1}M$ has a smooth local frame, and write
\begin{align*}
\alpha_U:=\alpha|_U\in C^\infty(U,\Lambda_b^{0,q}M).
\end{align*}
A CR antiholomorphic $q$-form is only defined on $q$ arguments from $T^{0,1}M$. To apply the ordinary exterior derivative $d$, we first choose an ordinary complex-valued smooth $q$-form
\begin{align*}
\widetilde\alpha\in C^\infty(U,\Lambda^q(\mathbb C T^*M))
\end{align*}
whose value on $T^{0,1}M$ agrees with $\alpha_U$. This extension exists because, locally, a vector bundle complement to $T^{0,1}M$ in $\mathbb C TM$ can be chosen, and an alternating form on $T^{0,1}M$ can be extended by specifying its values on the complementary directions.
We then define the local value of $\bar\partial_b\alpha_U$ by
\begin{align*}
(\bar\partial_b\alpha_U)(\bar L_0,\dots,\bar L_q):=(d\widetilde\alpha)(\bar L_0,\dots,\bar L_q)
\end{align*}
for local sections $\bar L_0,\dots,\bar L_q\in C^\infty(U,T^{0,1}M)$. The only issue is whether this depends on the extension $\widetilde\alpha$.
Suppose
\begin{align*}
\widetilde\alpha_1,\widetilde\alpha_2\in C^\infty(U,\Lambda^q(\mathbb C T^*M))
\end{align*}
are two such extensions, and define their difference by $\beta:=\widetilde\alpha_1-\widetilde\alpha_2$. Then
\begin{align*}
\beta(\bar K_1,\dots,\bar K_q)=0
\end{align*}
for every collection of local sections $\bar K_1,\dots,\bar K_q$ of $T^{0,1}M$, because the two extensions agree on exactly those arguments.
Now evaluate $d\beta$ on $\bar L_0,\dots,\bar L_q$. The Koszul formula for the exterior derivative consists of terms in which one vector field differentiates the function obtained by inserting the remaining arguments into $\beta$, and terms in which $\beta$ is evaluated after replacing two arguments by their Lie bracket. The derivative terms vanish because they differentiate functions of the form
\begin{align*}
\beta(\bar L_0,\dots,\widehat{\bar L_i},\dots,\bar L_q),
\end{align*}
and these functions are identically zero. The bracket terms vanish because formal integrability gives
\begin{align*}
[\bar L_i,\bar L_j]\in C^\infty(U,T^{0,1}M)
\end{align*}
for every pair $i,j$, so $\beta$ is still being evaluated only on $T^{0,1}M$ arguments. Therefore
\begin{align*}
(d\beta)(\bar L_0,\dots,\bar L_q)=0.
\end{align*}
Since $d\beta=d\widetilde\alpha_1-d\widetilde\alpha_2$, this proves
\begin{align*}
(d\widetilde\alpha_1)(\bar L_0,\dots,\bar L_q)=(d\widetilde\alpha_2)(\bar L_0,\dots,\bar L_q).
\end{align*}
Thus the restricted exterior derivative is well-defined, and this is precisely the local operator $\bar\partial_b$.[/guided]