Slurm

A read-only MCP server exposing Slurm scheduler state to agents, with the allowlist enforced in code and progressive disclosure of the tool surface.

Production HPC and GPU cluster runbooks packaged as loadable Agent Skills, each one naming the wrong diagnosis it exists to prevent.

A public incident-diagnosis benchmark for HPC schedulers — ten reproducible failure scenarios with measured ground truth and degenerate baselines.

Multi-agent diagnosis over an explicit cluster dependency graph — five ablatable configurations, specified before results existed.

Drain a node for a persistently faulty GPU — and never for a transient one.

Slurm on Kubernetes from nothing in one command — and the auth-key rotation nobody wants to test in production.

Correlate InfiniBand and RoCE counters with the Slurm job that owns them, so a slow collective can be traced to the fabric.

Find wasted GPU allocations on a Slurm cluster and escalate them through alert, drain, and cancel.

Test Slurm priority and backfill policy against a job trace before it reaches a live controller.

Three versions, two config renames that break on restart, and one cgroup change that quietly breaks every job-attribution tool you own.

I built a simulator to tune Slurm priority weights. It kept telling me the weights were not the problem.