grid5000-skill

A portable AI-agent skill for the Grid'5000 HPC testbed — the knowledge an LLM coding agent needs to reserve nodes, submit jobs, and run GPU/ML workloads on Grid'5000 (including the Inria "abaca" / production queue) without getting it wrong.

Grid'5000 uses the OAR scheduler (oarsub) — less widespread than SLURM, so agents rarely know it well — together with a three-tier access → frontend → compute-node workflow and per-site, un-backed-up storage. Out of the box, assistants tend to reach for the wrong commands or compute on the frontend. This skill encodes the correct commands, the gotchas, and a ready-to-use end-to-end ML workflow.

What's in it

A single SKILL.md — a self-contained knowledge document with YAML frontmatter (name + description) following the Agent Skills convention. It covers:

• OAR resource model and oarsub recipes (interactive, passive/batch, submission scripts, property filters, advance reservations)
• GPU reservations and CUDA on the default environment (no kadeploy needed for normal ML)
• The abaca / production queue: walltime tiers, cluster/GPU inventory, access requirements
• Storage: per-site homes, group NFS volumes, node-local scratch, and how to move data in/out
• Python envs: module, conda, and uv on nodes; outbound internet
• kadeploy (when you actually need custom OS / root)
• Project isolation for shared logins, and a list of common gotchas

Concrete examples use public information from the Grid'5000 wiki (site names, queue names, cluster/GPU models). No personal account details are included — replace <login>, <group>, <site>, and storage hosts with your own.

Install

The skill is harness-agnostic prose; drop it wherever your agent loads skills from.

Claude Code (user-level, available in every project):

mkdir -p ~/.claude/skills/grid5000
cp SKILL.md ~/.claude/skills/grid5000/SKILL.md

Codex CLI:

mkdir -p ~/.codex/skills/grid5000
cp SKILL.md ~/.codex/skills/grid5000/SKILL.md

Other agents / manual use: paste SKILL.md into your system prompt, project instructions, or RAG knowledge base. The frontmatter description is written to help an agent decide when the skill is relevant; the body is the actual guidance.

Scope & accuracy

Grid'5000's hardware inventory and per-user quotas change over time. The skill deliberately tells the agent to verify live (oarnodes, quota -s, the site Hardware wiki page) rather than trust hardcoded values. Cluster names (grat, musa, …) and GPU models are illustrative snapshots, not guarantees.

Contributing

Corrections and additional recipes welcome — especially for sites/clusters not covered here. Keep SKILL.md harness-agnostic and free of personal account details (see CLAUDE.md).

License

MIT.