First of all, it is a search to decision reduction: we are using an algorithm that distinguishes $A_{\mathbf{s},\chi}$ (the LWE distribution with secret $\mathsf{s}\in \mathbb{Z}_q^n$ and error distribution $\chi$) from $U$ (the uniform distribution on $\mathbb{Z}_q^n\times\mathbb{Z}_q$) in order to find the secret vector $\mathbf{s}=(s_1,\cdots,s_n)$. A decision to search reduction for LWE (like for most other problems) is trivial.
Let's see first how the reduction works when $q$ is prime [R, Lemma 4.2].
Prime $q$. Suppose we have a distinguisher $\mathsf{D}$ for LWE that works with overwhelming probability (i.e., $1-\mathsf{negl}(n)$) -- any distinguisher $\mathsf{D}'$ that works with probability that is bounded away from half by a non-negligible value can be turned into one that works with overwhelming probability using standard amplification techniques: i.e., carry out independent trials, take the majority and apply the Chernoff bound (see e.g., [AB, Lemma 7.9] to see how this is used for $\mathbf{BPP}$). We will use $\mathsf{D}$ to construct an algorithm $\mathsf{F}$ that finds $\mathbf{s}$ as follows. For $k\in\mathbb{Z}_q$ and $l\leftarrow\mathbb{Z}_q$ (where $\leftarrow$ denotes sampling uniformly at random), consider the transformation
$$(\mathbf{a},b)\mapsto(\mathbf{a}+(l,0,\cdots,0),b+lk).$$
If $s_1$ denotes the first entry of the secret vector $\mathbf{s}$ in the LWE distribution $A_{\mathbf{s},\chi}$, there are two cases:
$k=s_1$: here, the transformation maps $A_{\mathbf{s},\chi}$ to itself, as shown below
$$(\mathbf{a},b)=\left(\mathbf{a},\langle\mathbf{a},\mathbf{s}\rangle+e\right)\mapsto \left(\mathbf{a}+(l,0,\cdots,0),\langle\mathbf{a},\mathbf{s}\rangle+e+ls_1\right)=\left(\mathbf{a}',\langle\mathbf{a}',\mathbf{s}\rangle+e\right)=(\mathbf{a}',b'),$$
where $\mathbf{a}'$ denotes the vector $\mathbf{a}+(l,0,\cdots,0)$.
$k\neq s_1$: the transformation maps $A_{\mathbf{s},\chi}$ to the uniform distribution (as $l$ masks everything).
Now it is easy to recover $s_1$: $\mathsf{F}$ simply runs $\mathsf{D}$ over all $k\in\mathbb{Z}_q$ and outputs the $k$ for which $\mathsf{D}$ outputs $0$ (indicating "LWE"). Since the size of the modulus $q\leq\mathsf{poly}(n)$, this is still efficient. The remaining entries of $\mathbf{s}$ can be recovered by iterating the above procedure, shifting the $(l,0,\cdots,0)$ vector to the right in each iteration. $\square$
Note that the step in the proof which requires $q$ to be prime is the second case ($k\neq s_1$): if $q$ is composite then the map $b\mapsto b+lk\bmod{q}$ is not necessarily bijective and the distribution gets skewed away from uniform when $l$ is one of the factors of $q$. Extending this to the case when $q=p^e$ was sketched in [ACPS, Lemma 1]. The complete proof can be found in [MP, Theorem 3.1] (albeit phrased differently from below).
Prime power $q=p^e$. For simplicity, let's focus on $e=2$ as the argument can be easily extended for $e>2$. For $i=0,1,2$, consider the distributions
$$A_{\mathbf{s},\chi}^i:=\left\{(\mathbf{a},b+r\cdot p^{2-i}\bmod{q}):(\mathbf{a},b)\leftarrow A_{\mathbf{s},\chi},r\leftarrow\mathbb{Z}_q\right\}.$$
Note that $A_{\mathbf{s},\chi}^0$ is identical to $A_{\mathbf{s},\chi}$ as $r\cdot p^2=0 \bmod{p^2}$. On the other hand, $A_{\mathbf{s},\chi}^2$ is identical to $U$, the uniform distribution on $\mathbb{Z}_q^n\times\mathbb{Z}_q$ as $r$ masks everything.
Step $1$. The first step is to use the distinguisher $\mathsf{D}$, which distinguishes $A_{\mathbf{s},\chi}^0$ from $A_{\mathbf{s},\chi}^2$ to extract $\mathbf{s}'=\mathbf{s}\bmod{p}$. By the hybrid argument, it is guaranteed that $\mathsf{D}$ distinguishes $A_{\mathbf{s},\chi}^0$ from $A_{\mathbf{s},\chi}^1$ or $A_{\mathbf{s},\chi}^1$ from $A_{\mathbf{s},\chi}^2$ with overwhelming probability. We show that in either of the cases it is possible to extract $\mathbf{s}'=\mathbf{s}\bmod{p}$ using the ideas we saw for the case when $q$ is prime. Let's focus on the first case as the second case is analogous, and the only difference is in the transformation.
Case $1$. We are given a distinguisher $\mathsf{D}$ that distinguishes $A_{\mathbf{s},\chi}^0$ from $A_{\mathbf{s},\chi}^1$ with overwhelming probability. The idea is to use a slightly different transform from the previous case. To be precise, we use
$$(\mathbf{a},b)\mapsto(\mathbf{a}+(lp,0,\cdots,0),b+lkp),$$
where $l\leftarrow\mathbb{Z}_q$ (as before) and $k\in\mathbb{Z}_p$. If $s_1=s_1'+s_1''p$, there are two cases:
$k=s_1'$: the transformation maps $A_{\mathbf{s},\chi}=A_{\mathbf{s},\chi}^2$ to itself, as shown below
$$(\mathbf{a},b)=\left(\mathbf{a},\langle\mathbf{a},\mathbf{s}\rangle+e\right)\mapsto \left(\mathbf{a}+(lp,0,\cdots,0),\langle\mathbf{a},\mathbf{s}\rangle+e+lps_1'\right)=\left(\mathbf{a}',\langle\mathbf{a}',\mathbf{s}\rangle+e\right)=(\mathbf{a}',b').$$
Here, we use the fact that $(a_1+lp)s_1=a_1s_1+lp(s_1'+s_1''p)=(a_1s_1+s_1'lp)\bmod{p^2}$, where $a_1$ is the first component of $\mathbf{a}$.
$k\neq s_1'$: the transformation maps $A_{\mathbf{s},\chi}$ to $A_{\mathbf{s},\chi}^1$ as
$$(\mathbf{a},b)=\left(\mathbf{a},\langle\mathbf{a},\mathbf{s}\rangle+e\right)\mapsto \left(\mathbf{a}+(lp,0,\cdots,0),\langle\mathbf{a},\mathbf{s}\rangle+e+lpk\right)=\left(\mathbf{a}',\langle\mathbf{a}',\mathbf{s}\rangle+e+rp\right),$$
where $r=lk\bmod{q}$. Here, multiplying by $k$ preserves the distribution $l\leftarrow\mathbb{Z}_q$ since $k$ is coprime to $p$.
Now we recover $s_1'=s_1\bmod{p}$ as before: $\mathsf{F}$ simply runs $\mathsf{D}$ over all $k\in\mathbb{Z}_p$ and outputs the $k$ for which $\mathsf{D}$ outputs $0$ (indicating "LWE"). $\mathbf{s}'=\mathbf{s}\bmod{p}$ is obtained by the shifting trick.
Case $2$. Now, given a distinguisher $\mathsf{D}$ that distinguishes $A_{\mathbf{s},\chi}^1$ from $A_{\mathbf{s},\chi}^2$ with overwhelming probability, the only difference is in the transform, which is now
$$(\mathbf{a},b)\mapsto(\mathbf{a}+(l,0,\cdots,0),b+pr+lk),$$
for $l$, $k$ as above and $r\leftarrow\mathbb{Z}_q$. Note that when $k=s_1'$ the transformation maps $A_{\mathbf{s},\chi}$ to $A_{\mathbf{s},\chi}^1$ and when $k\neq s_1'$ it maps $A_{\mathbf{s},\chi}$ to $A_{\mathbf{s},\chi}^2=U.$
For general $e>2$, the hybrid distributions are defined as
$$A_{\mathbf{s},\chi}^i:=\left\{(\mathbf{a},b+r\cdot p^{e-i}\bmod{q}):(\mathbf{a},b)\leftarrow A_{\mathbf{s},\chi},r\leftarrow\mathbb{Z}_q\right\}$$
and the $i$-th transformation is
$$(\mathbf{a},b)\mapsto(\mathbf{a}+(l\cdot p^{e-j},0,\cdots,0),b+(pr+lk)\cdot p^{e-j}).$$
Step $2$. Given $\mathbf{s}'=\mathbf{s}\bmod{p}$, $\mathbf{s}$ can be recoved as described in [ACPS]. $\square$
[AB]: Arora and Barak, Computational Complexity: A Modern Approach
[ACPS]: Applebaum et al., Fast Cryptographic Primitives and Circular-Secure Encryption Based on Hard Learning Problems, Crypto 2009
[MP]: Micciancio and Peikert, Trapdoors for Lattices: Simpler, Tighter, Faster, Smaller, Eurocrypt 2012
[R]: Regev, On lattices, learning with errors, random linear codes, and cryptography, JACM 2009
