gpu-custom

/gpu-custom — runtime GPU operations in custom mode

When ROCKIE_GPU_MODE=custom, the agent doesn't drive GPU provisioning through rockie's router. Instead, it follows the user's own flow as documented in .claude/gpu-custom.md (populated once by /gpu-custom-setup).

This skill is the agent's gateway to that flow.

Pre-flight checks

# Are we in custom mode?
case "${ROCKIE_GPU_MODE:-router}" in
  custom) ;;  # ok, proceed
  router)
    echo "[gpu-custom] ROCKIE_GPU_MODE=router — use rockie-gpu / /gpu-spend instead"
    exit 0 ;;
  *) echo "[gpu-custom] ROCKIE_GPU_MODE=${ROCKIE_GPU_MODE} — not custom; nothing to do"; exit 0 ;;
esac

# Is the setup file present?
test -s .claude/gpu-custom.md || {
  echo "[gpu-custom] .claude/gpu-custom.md missing or empty"
  echo "  → run /gpu-custom-setup first to onboard your custom GPU flow"
  exit 0
}

If both pass, read .claude/gpu-custom.md and route based on the user's intent.

Routing user intent → setup file section

User intent	Read this section of gpu-custom.md
"spin up a GPU" / "provision" / "start training"	§2 Provision (then §3 Connect)
"connect to my pod" / "ssh in"	§3 Connect
"what's running" / "list pods" / "status"	§4 Monitor → Status
"what's it costing" / "spend" / "burn rate"	§4 Monitor → Cost
"tear down" / "terminate" / "stop the pod" / "we're done"	§5 Terminate (verbatim — don't improvise)
"preemption" / "got kicked off"	§6 Optional → Preemption (if present)
"checkpoint" / "save state"	§6 Optional → Persistent storage (if present)

The hard rule about §5 Terminate

The terminate command captured during onboarding was reviewed by the user for correctness. Run it verbatim. Do NOT:

• Substitute "similar" commands you think might work
• Add --dry-run or --no-op flags the user didn't include
• Skip the command because you're worried about side effects (the user wrote it; they own it)
• Run a different terminate command for "their other pod" you noticed — only act on the pod the user explicitly told you to terminate