Coding agents are rapidly moving from single‑file generation to repository‑scale, long‑horizon engineering, yet assessing their ability to author interactive extended reality (XR) applications remains an open problem. Unlike traditional software, XR apps run inside dynamic 3D physical spaces where runtime robustness depends jointly on continuous user motion, spatial input, and environmental sensing. To address this gap we introduce GenXR‑Bench, the first end‑to‑end benchmark for evaluating coding agents in XR development.\ \ GenXR‑Bench comprises 112 expert‑crafted tasks covering 24 feature implementations, 21 bug‑fixes, and 67 full‑application generation scenarios. Held‑out tests combine scripted and adaptive interactions with both programmatic and visual assertions to verify core functionality and runtime behavior.\ \ When run on consumer XR hardware, the suite achieves 96.5% accuracy against blinded expert judgments on 124 candidate applications. Evaluating ten models on GenXR‑Bench, the best configuration attains only 50.9% Pass@1. Failure analysis reveals recurring misunderstandings of asynchronous sensor lifecycles, physical grounding, and embodied interaction semantics.\ \ GenXR‑Bench thus provides a reproducible testbed for assessing coding agents under the interactive and physically situated conditions of spatial computing.\ \ Review: The benchmark exposes substantial gaps in current coding agents for XR, highlighting sensor lifecycle handling, physical constraint reasoning, and interaction semantics as key targets for future model improvements.