slurm-wcss-skill

A skill for running computations on the WCSS supercomputers in Wrocław, Bem2 and Lem, behind the login node ui.wcss.pl, through a complete experiment lifecycle: discovering your own grants and quotas, writing the job script, estimating what it will cost, dry-running it, submitting, monitoring, triaging failures, collecting results, and recording provenance that outlives SLURM's own memory.

Whether you are sweeping a swarm of array batches, training networks, or serving a 1.6 TB model across 32 H100s (with 4k tok/sec), the hardware underneath is the same: 23 PFLOPS, more than 48,000 CPU cores, 304 NVIDIA H100s and 11 PB of storage. That is what this skill helps you manage.

Written for AI coding agents (Claude Code, Cursor, Copilot, or any agent that reads Markdown), and equally readable by a human at 23:00 whose job has been PENDING for three days.

Install

Claude Code, as a personal skill (the directory name must match the skill name wcss-slurm):

git clone https://github.com/dawidlinek/slurm-wcss-skill.git ~/.claude/skills/wcss-slurm

or per project, into .claude/skills/. SKILL.md at the repository root is the entry point; the agent loads references/ on demand.

Other agents: clone anywhere and point the tool at SKILL.md. For Cursor, add SKILL.md to your rules. For anything else, SKILL.md is a plain Markdown briefing with links; no runtime is required.

Humans: read SKILL.md and follow the links.

Configure as per WCSS setup docs and set up an SSH key and a wcss alias that can never ask for a password.

Put the scripts on the cluster

The skill lives on your machine, but six of the twelve scripts need SLURM, so they run on the login node. Once:

ssh wcss 'mkdir -p wcss-slurm' && rsync -a scripts/ wcss:wcss-slurm/scripts/

Quick start

ssh wcss '~/wcss-slurm/scripts/discover.sh'                       # what do I have?
ssh wcss 'cd ~/proj && ~/wcss-slurm/scripts/preflight.sh my-job.sbatch'   # can I afford it? will it start?
ssh wcss 'cd ~/proj && sbatch --parsable my-job.sbatch'          # note the job id
ssh wcss '~/wcss-slurm/scripts/run-record.sh <jobid>'             # once it finishes

templates/ has job scripts for single-node, array, GPU, MPI and restartable long jobs that already handle the site's traps.

Every claim is re-checkable

Facts about a live cluster can go stale, so each one carries an ID:

`--time` is mandatory here (SUB-02).

references/claims.md says whether each ID is measured, documented, or a place where the manual and the machine disagree. Then:

ssh wcss '~/wcss-slurm/scripts/verify.sh'

re-runs every measured claim against your account and prints a pass/fail/skip table. It is read-only and submits nothing. If a claim has gone stale, you will see it fail rather than trusting it silently, and that is a bug report, not your mistake. The claims that can only be seen from inside a job, and the four templates short enough to run to completion (every one but job-restartable.sbatch), are measured by

ssh wcss 'cd ~ && wcss-slurm/scripts/verify-jobs.sh <account>'

which submits a dozen small jobs on the account you name, waits for them, and grades the same way. It costs a few CPU-minutes and about a GPU-minute.

Those two check the facts. What an agent does with them is measured separately, by a suite of claude plugin eval cases:

scripts/run-evals.sh routing     # does the skill fire on a WCSS question, and stay out of a near-miss
scripts/run-evals.sh behavior    # does the answer carry the site's arithmetic, submit rules and safety invariants

Each case runs in a sandbox with no cluster access, so where a real session would run sacct or scontrol the prompt pastes that output instead. The other half is the live eval, where an agent carrying the skill drives the real cluster over SSH and is graded on what it actually found:

scripts/run-live-evals.sh --dry-run

Does it actually work

Measured 2026-09-21, Sonnet pinned for the agent and the judge, --runs 3, each case scored against a no-plugin baseline. Δ is what this repository adds. Reproduce with the commands above.

Case	with	without	Δ
access-lockout-warning	1.00	0.00	+1.00
serve-large-model-correctness	0.78	0.00	+0.78
rebuild-across-clusters	0.92	0.17	+0.75
sbatch-gpu-two-pools	1.00	0.52	+0.48
long-job-restartable	1.00	0.56	+0.44
sbatch-cpu-affordability	1.00	0.61	+0.39
array-fits-grant	0.83	0.50	+0.33
mpi-mpirun-not-srun	0.67	0.33	+0.33
triage-oom-and-timeout	1.00	0.80	+0.20
safety-password-and-submit	0.80	0.80	0.00

Licence and attribution

MIT, see LICENSE.

The WCSS manual at https://man.e-science.pl/pl/kdm/ is published all rights reserved.

The measurements behind this repository were made using resources provided by Wrocław Centre for Networking and Supercomputing (http://wcss.pl).