github.com/Zhanyl-tech/cluster-ops-skills · Python · MIT
Eleven runbooks for on-premises Slurm operations, written as loadable Agent Skills.
The expensive mistakes on a cluster are not missing information — they are wrong confident diagnoses. Recommending a priority-weight change that measurement shows does nothing. Reporting that a storage stall halted scheduling when jobs were completing throughout. So every skill carries a mandatory What not to conclude section naming the plausible wrong answer, and an Escalate when section, both enforced by a validator that fails the build.
Nine of the eleven map onto a fault family that slurm-rca-bench already reproduces — storage, accounting, controller, GPU, fabric, scheduler config — so each encodes a failure that has been measured rather than imagined. The validator also fails any skill quoting a percentage without a link to the repo that measured it.
The odd one out is evidence-is-gone, about deciding an incident is not
diagnosable and abstaining. It exists because the benchmark awards full credit
for abstention on its deliberately undiagnosable scenarios and zero for naming
even the most plausible cause. A confident wrong diagnosis sends someone to
replace healthy hardware and stops the search.
Whether loading these improves an agent’s accuracy is unmeasured, and the harness for answering that is the benchmark above.