What problem does it solve? Debugging failing Kubernetes pods — especially GPU workloads — often turns into random log-diving that misses the real cause. This Skill enforces a fixed triage order (pod status, events, logs, node conditions, resources, GPU state, DRA claims) so root causes like CUDA driver mismatches, missing GPU allocatable, or unallocated ResourceClaims are found before you change code or resource limits. ## Core Features & Use Cases - Ordered triage workflow: Runs pod status, events (before logs), container logs, node conditions, and resource requests in a fixed sequence using kubectl. - GPU-specific diagnostics: Checks GPU allocatable counts, the nvidia-device-plugin DaemonSet, MIG profiles, and CUDA driver/runtime mismatches via a full scheduling checklist. - DRA troubleshooting: Walks ResourceClaim allocation, driver registration via resourceslices, CDI spec generation, and /dev/nvidia* device presence. - Use Case: A training pod is in CrashLoopBackOff with a CUDA error. The Skill checks events and logs, then flags the common pattern — a driver/runtime mismatch rather than an application bug — saving you from rewriting working code. ## Quick Start Ask the AI to run k8s-debug on the pod training-job-7f8 in namespace ml because it is stuck in CrashLoopBackOff.
