<p>In large-scale multiple hypothesis testing, controlling the false discovery exceedance (FDX) is emerging as a compelling alternative to the widely employed false discovery rate (FDR) when the false discovery proportion (FDP) exhibits high variability. Existing methods for FDX control have mainly focused on nondirectional discovery within a two-group model framework. However, this approach can significantly compromise the reliability of FDX control when directional decisions are required, especially in cases when the signal-to-noise ratio is low. Furthermore, the theoretical optimality of FDX control has not been thoroughly explored. In this paper, we introduce an empirical Bayes approach tailored for directional FDX control within a three-group model. We demonstrate that an oracle decision rule, which ranks and thresholds a directional version of the local false discovery rate (lfdr), achieves optimality by maximizing power while adhering to the directional FDX constraint. We also propose a data-driven procedure to mimic the oracle rule in practical applications and establish its asymptotic optimality. Through extensive simulation studies and a real-data application, we show the superior performance of our method in directional FDX control.</p>