There are several shortcomings in your definition. You need to first absorb the argument to \ref and only after it check whether a c follows.
Also the conditional should end before calling the “gobbling” macro.
Finally, \ref gets redefined by hyperref at begin document, so your own redefinition should be delayed.
Another small issue is the usage of \let with the \ref command that hyperref makes robust; here I use \NewCommandCopy that's been available for more than a year; if your TeX distribution is older, use \LetLtxMacro from the letltxmacro package.
\documentclass[10pt]{article}
\usepackage{amsthm}
\usepackage{hyperref}
\usepackage{cleveref}
\newtheorem{theorem}{Theorem}
\makeatletter
\AtBeginDocument{%
\NewCommandCopy{\old@ref}{\ref}%
\def\ref#1{\def\temp@ref{#1}\futurelet\@nexttoken\new@ref}%
}
\def\new@ref{%
\if c\noexpand\@nexttoken
\expandafter\@firstoftwo
\else
\expandafter\@secondoftwo
\fi
{\cref@eatc{\temp@ref}}%
{\old@ref{\temp@ref}}%
}
\def\cref@eatc#1c{\cref{#1}}
\makeatother
\begin{document}
\begin{theorem}[sdfs] \label{sdf} asfa \end{theorem}
\cref{sdf} = theorem 1
\ref{sdf}c = theorem 1
\ref{sdf} c = 1 c
\end{document}
Why \noexpand? Because you don't want to expand the following token which might be a macro or ~.
You may appreciate an expl3 version of the same, which is easier to implement, but not conceptually different.
\documentclass[10pt]{article}
\usepackage{amsthm}
\usepackage{hyperref}
\usepackage{cleveref}
\newtheorem{theorem}{Theorem}
\ExplSyntaxOn
\AtBeginDocument
{
\NewCommandCopy{\__refc_saved_ref:n}{\ref}
\cs_set_eq:NN \ref \__refc_new_ref:n
}
\tl_new:N \l__refc_arg_tl
\cs_new_protected:Nn \__refc_new_ref:n
{
\tl_set:Nn \l__refc_arg_tl { #1 }
\peek_charcode_remove:NTF c
{
\exp_args:NV \cref \l__refc_arg_tl
}
{
\exp_args:NV \__refc_saved_ref:n \l__refc_arg_tl
}
}
\ExplSyntaxOff
\begin{document}
\begin{theorem}[sdfs] \label{sdf} asfa \end{theorem}
\cref{sdf} = theorem 1
\ref{sdf}c = theorem 1
\ref{sdf} c = 1 c
\end{document}
