It seems to run for local users. ChatGPT suggests, and I agree, that the difference might be set source setvars.sh which we usually do locally but for historical reasons usually don’t do in the pipelines and try to get MKL anyway.
It was due to FEAST using MKL functions (marked as deprecated since around 2018) that were removed from Intel oneAPI in the 2026 release. The issue should have been seen by all developers with default build options using the most recent oneAPI. I reached out to the FEAST developers, and they already provided a fix which we will include in CFS soon: fix FEAST for MKL2026 (!396) · Merge requests · openCFS / CFS · GitLab