菜单
Plans and launches an hmfast experiment (training, inference, or scan) on the cluster with safe resource limits. Use when the user wants to run something that costs nontrivial compute, especially on GPU.
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。
基于 SOC 职业分类
Reviews proposed changes to hmfast for correctness, JAX compatibility, API layering, test coverage, and hardware-cost surprises. Use when the user asks for a review, a sanity check on a diff, or before merging a feature branch.
Guides changes to hmfast Cosmology, emulator loading, and cosmological parameter sets. Use when editing cosmology.py, emulator_load.py, download.py, or adding new emulator_set keys.
Diagnoses a failed or stuck SLURM/HPC job for hmfast (OOM, timeout, NaNs, CUDA errors, missing outputs). Use when the user asks "why did my job fail" or shows a SLURM log.
Guides edits to hmfast's differentiable JAX halo model (HaloModel, mass function, bias, concentration, profiles). Use when changing src/hmfast/halos/, power-spectrum methods, or halo-model consistency flags.
Guides adding or modifying hmfast tracers and halo profiles. Use when editing src/hmfast/tracers/ or src/hmfast/halos/profiles/, or integrating new observables with HaloModel.
Drafts or revises a section of an astrophysics or cosmology manuscript based on hmfast results. Use when the user asks for a draft, an edit pass, or a publishability review of a section, including its figures.
| name | run-experiment |
| description | Plans and launches an hmfast experiment (training, inference, or scan) on the cluster with safe resource limits. Use when the user wants to run something that costs nontrivial compute, especially on GPU. |
Follow .claude/rules/cluster-usage.md and, for autonomous sessions, .claude/rules/autonomous-operation.md.
nvidia-smi check on the target node.8 bytes (x64 default). Confirm it fits with headroom for JAX's default 75% preallocation. Show the arithmetic.N, few steps) and extrapolate. If the extrapolation is unclear, ask the user before launching the full run.--time, --mem, --gres=gpu:1 (when GPU is needed), and --output=logs/%x-%j.out.source /scratch/scratch-lxu/venv/cmbagent_env/bin/activate
/scratch/scratch-lxu/... (working) or /rds/rds-lxu/... (persistent), never inside the repo.tail -f logs/<jobname>-<jobid>.out.run.json manifest with: git hash, full config, seed, MaxRSS, wall time, output paths.If during the run you see OOM, repeated NaNs, or runtime exceeding the estimate by more than 2x, stop and report before resubmitting. Do not blindly retry with the same parameters.