菜单
Work on a Slurm cluster through the private cluster reference and the harness job tools instead of re-probing the cluster every time. Use for sbatch/srun submissions, queue or partition questions, node and GPU availability, job monitoring, and post-job log review. Skip on machines without Slurm.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
基于 SOC 职业分类
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/eightmm/oh-my-setting --skill oms-slurm命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
Shared state, graph-guided work, typed runtime projections, recovery, peer consultation and review, isolated delegation, patch landing, bounded autonomy, experiments, Draft PRs, and cross-provider handoff.
Resolve implementation-shaping choices for new projects, repo onboarding, unresolved drafts, or broad features. Inspect first and ask only about material gaps; clear bounded changes do not require an interview.
Register datasets and splits in a manifest and check group leakage and split drift before training: declared group keys (not row identity) decide train/eval separation, so the same protein, patient, or source never sits on both sides.
| name | oms-slurm |
| description | Work on a Slurm cluster through the private cluster reference and the harness job tools instead of re-probing the cluster every time. Use for sbatch/srun submissions, queue or partition questions, node and GPU availability, job monitoring, and post-job log review. Skip on machines without Slurm. |
| compatibility | Requires a Slurm cluster (sinfo on PATH) and an oh-my-setting install for the oms job tools. |
Static cluster facts live in the private reference, not in repeated probes.
local/slurm.md under the oh-my-setting
install root (oms status reports the root). It records partitions,
node/GPU inventory, and — from sacctmgr show assoc/show qos — your
account associations and submission limits (GrpTRES, MaxTRES per job,
MaxJobs, MaxSubmit, MaxWall). Check those limits there before sizing a
job, instead of discovering them from a rejected submission. Validate or
refresh with oms snapshot --cluster --check (regenerate without
--check when the cluster changed). Never commit it.squeue for the
current queue, sacct for finished jobs. Do not re-run sinfo topology
scans when the reference already answers the question.PROJECT.md before submitting.$SLURM_SUBMIT_DIR, fail fast (set -euo pipefail),
and write job-specific logs.oms run-ledger (pre-flight gate, one
JSONL row per run) so parallel sessions see them.oms job-digest <log> distills a job log to agent-sized context instead of
pasting raw logs.oms run-reconcile reconciles finished Slurm jobs against the run ledger.squeue -j <id> on a sensible interval, then digest
the log and report; do not tail interactively.