What problem does it solve? Before running real training jobs on a remote server, you need to know whether the training stack is actually ready: which conda environments exist, which verl or veomni versions are installed, and whether a minimal one-step training run can pass with a decreasing loss. This Skill automates that readiness probe so you avoid wasting time on misconfigured environments. ## Core Features & Use Cases - Environment Inspection: Parses conda env list and pip list output to inventory available training environments and installed packages. - Version Detection: Detects verl or veomni framework versions on the remote server. - One-Step Smoke Test: Runs a minimal 1-step training case and verifies that the step completes and the loss decreases. - Health Table Output: Produces a train stack health summary for readiness review before data collection. - Use Case: Before launching a distributed RLHF experiment on server nvidia-01, run the stack probe to confirm the conda env, verl version, and minimal runner behavior are all healthy. ## Quick Start Run the stack probe against my remote server nvidia-01 using the config at ./config/config.yaml and tell me whether the training stack is healthy.
