gpu-spend

/gpu-spend — Rockie-GPU spend snapshot

The single source of truth for "what is the agent costing me right now?" rockie-gpu spent reconciles live state, sums project spend, and prints both the LLM-readable JSON and a human view — all under Rockie's own deidentified billing surface, with no supplier names and no supplier billing URLs leaking through.

When to invoke

• User asks anything about cost, spend, rate, burn, budget, or pods running.
• Before any decision to spin up a new GPU (read state before adding load).
• After terminating a pod, to confirm the bleed has stopped.
• Periodically during long autonomous runs (the budget-reconcile.sh hook fires this on every UserPromptSubmit, but the agent should also check on entering the Codify phase).

Do NOT invoke if the user asked a non-cost question and you'd just be showing off. The hook keeps the budget honest in the background.

What the skill does

Runs:

rockie-gpu spent --json

Output shape (LLM-ergonomic, deidentified — Rockie-priced, no supplier names):

{
  "project": "...",
  "compute_per_hr": 0.0,
  "storage_per_hr": 0.011,
  "running_pods": 0,
  "idle_volume_gb": 80,
  "grand_total_per_hr": 0.011,
  "grand_cumulative_usd": 2.20
}

Read the JSON. Synthesize a 2–3 line summary for the user that:

• Names the top spend driver (compute or storage).
• Calls out idle storage if running_pods=0 but idle_volume_gb > 0 (user is paying for nothing — recommend tearing the pod down).
• Compares grand_cumulative_usd to budget ceiling if known (rockie-gpu spent --json reports the ceiling when one is configured).

Composition

• rockie-gpu spent reconciles implicitly, so numbers are always fresh (within seconds of the call). No separate reconcile step.
• rockie-gpu audit-open-spend --demo-safe is the line-item audit when the user wants the full breakdown of where the dollars went — it is the only drill-down surface. There is no per-supplier CLI to reach for; rockie-gpu sees every dollar.