Peptides are becoming a major therapeutic modality. Although peptides are integral to biological processes, designing therapeutic peptides de novo remains a challenging prospect. In this paper, we exploit the rich biological inductive bias of amino acids and introduce HelixFlow, a flow-matching model to design full-atom peptide structures. We incorporate a hotspot-specific sequence-conditioned SE(3)-equivariant flow matching module for full-atom helical structure generation and a novel pocket-flow module to generate the binding peptides given target receptors. HelixFlow presents substantial new architectural features over the previous HelixDiff family of models including an equivariant all heavy atom representation, a transformer-based model for flow prediction, and flexible-length generation. By one-shot generation without assembling and direct coordinate generation, HelixFlow could become a more powerful tool for realistic peptide design and open a door for more concise conditional generations on the atom level. As a proof of concept, we designed an acetylated D-peptide of Insulin-like peptide 5 (INSL5) that selectively activates the relaxin family peptide receptor 4 (RXFP4). Our designed D-INSL5 peptide shows good biological activity, comparable AKT phosphorylation levels and high resistance to protease degradation, underscoring the successful integration of deep learning and structure-based modeling and simulation for target-specific peptide design.