gpu-custom

name

gpu-custom

description

Runtime skill for users in ROCKIE_GPU_MODE=custom — invoked when the user (or agent) needs to do anything GPU-related (provision, connect, check status, check cost, terminate) in a project where Rockie's GPU router is bypassed in favor of the user's own setup. Reads `.claude/gpu-custom.md` (populated by /gpu-custom-setup) for the user's documented flow and follows it. Replaces the deidentified `rockie-gpu` surface and /gpu-spend in custom mode — those route to gpu.py which exits gracefully when ROCKIE_GPU_MODE is not 'router'. If `.claude/gpu-custom.md` doesn't exist, redirect to /gpu-custom-setup first.

/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

If the user's command needs an instance ID / job ID / pod ID substituted in, ask the user what to substitute. Don't guess.

When the file is incomplete

If the user asks for an operation that isn't covered in .claude/gpu-custom.md (e.g. they ask about cost but §4 doesn't have a cost subsection), respond honestly:

Your .claude/gpu-custom.md doesn't document a cost-tracking flow. Want me to add it now? (You'd paste the command or URL you check.)

Don't make up a flow. Don't fall back to rockie's router-mode commands like gpu.py cost — those will exit with a "custom mode bypassed" message anyway.

Budget integration (optional)

If the user has a way to surface live $/hr (URL, CLI), and it's documented in §4 Monitor → Cost, you can periodically query it and write to rockie's budget.py so the dollars ceiling stays honest:

# In a wrapper the user manually invokes (NOT auto-fired by hooks
# in custom mode — we don't presume to know their billing cadence):
python3 .claude/scripts/budget.py add dollars <amount>

But don't auto-wire this without explicit user opt-in — custom mode exists because the user wants to drive their own tooling.

Composition with other harness pieces

• autopilot loop still runs in custom mode. The training launcher inside autopilot.conf points at the user's own command (e.g. LAUNCHER_CMD="ssh worker bash /home/user/train.sh"); the ZCM supervisor still detects exit/anomaly/crash; the budget hook still gates on token counts.
• /post-run-review still works (independent of GPU layer).
• /propose-harness-change still works.
• The dollars budget dimension is what's affected. In custom mode it's only as good as the user's manual reporting (or the optional budget.py wrapper above).