XirtaM investigates silent failure in autonomous agents — the cases where an agent appears to succeed while quietly dropping part of its task. Our BGI Sprint submission (Track 1, Holding the Thread) is a reproducible red-team harness for OmegaClaw and an accompanying taxonomy of thread-holding failures, organized by how and where coverage breaks down over long, interrupted runs. We ground each failure class in a concrete airworthiness-checking example, producing testable cases for OmegaClaw maintainers and BGI researchers. The underlying reasoning holds up under adversarial pressure; the contribution is naming precisely how it fails when it fails — a prerequisite for trustworthy agents.