skypilot-agent-skill

A Claude Code agent skill (ddps-sky-run) by DDPS Lab for running lab experiments on SkyPilot.

Describe an experiment in natural language — the skill asks only for what is missing, generates a SkyPilot YAML from lab-validated templates, shows the instance/cost table, and launches only after you approve (it never passes -y).

Install

pip install skypilot
aws configure        # your own cloud credentials; sky runs locally and reads them
sky check aws        # must show "enabled"
claude plugin marketplace add ddps-lab/skypilot-agent-skill
claude plugin install ddps-sky-run@skypilot-agent-skill --scope project

--scope project activates the plugin ONLY for the current project (recorded in the project's .claude/settings.json, so repo collaborators get it too). Use --scope local for project-only AND personal (not committed), or omit the flag / choose User scope for all your projects.

(Or interactively: run /plugin inside Claude Code, open the Discover tab, and install ddps-sky-run from the skypilot-agent-skill marketplace — the scope picker appears at install time.)

Then run claude in your own project and describe the experiment, e.g.:

"OLMoE 모델을 EP로 서빙 테스트하고 싶어. GPU는 L4 1장짜리 노드 2개."

The skill classifies the workload (online serving / offline inference / training), interviews you for the missing fields, and walks you to the cost-approval prompt.

Templates

Template	Layout	Status
online-ep-2node.yaml	2 nodes x L4:1 — TP=1 x DP=2, EP=2 (experts split across nodes)	VERIFIED (2026-07-22, vllm 0.25.1, spot)
online-ep-tp.yaml	1 node x L4:4 — TP=2 x DP=2, EP=4	pending (needs 48-vCPU spot quota)
online-ep-dp.yaml	1 node x L4:4 — TP=1 x DP=4, EP=4	pending (same quota)
offline-inference-gpu.yaml	L4:1 managed job — latency percentiles to S3, auto-teardown	runnable
offline-inference-neuron.yaml	Inferentia + custom AMI skeleton	placeholders (AMI id/region/command)

Repository layout

.claude-plugin/marketplace.json      marketplace catalog -> ./agent
agent/
  .claude-plugin/plugin.json         plugin manifest
  skills/ddps-sky-run/
    SKILL.md                         trigger + behavior rules (the skill itself)
    references/
      gotchas.md                     12 field-tested pitfall rules
      upstream.md                    routing to official SkyPilot reference docs
      templates/                     the 5 templates above
EXPERIMENT.md                        validation runbook (how templates get VERIFIED)

How it stays trustworthy

• Templates ship only after passing their EXPERIMENT.md phase. A template that has not been run for real does not get published (skeletons are marked as such).
• Every real failure becomes a rule. The 12 entries in references/gotchas.md were each hit in an actual run (secrets semantics, mount modes, quota shapes, KV-cache preallocation, ...) and are applied to every generated config.
• Out-of-template requests route to the official SkyPilot reference docs via references/upstream.md; lab rules take precedence on conflicts, and such configs are flagged as not yet lab-validated.
• Cost gate is non-negotiable. The skill never bypasses SkyPilot's "Considered resources" Y/n prompt.