runpod-ops

name	runpod-ops
description	RunPod endpoint operations runbook — checking worker state, forcing scale-down, reading the spend query, and cold-start expectations. Use before and after cloud inference sessions. Forward-looking — RunPod tooling lives outside this repo.

RunPod operations runbook

Reference for managing RunPod serverless/pod endpoints used for cloud inference and training. RunPod itself is external to this repository; this skill captures the operating procedure so sessions do not leave workers running or misread spend.

Status: forward-looking. Wire the concrete commands to whatever RunPod CLI / API wrapper the integration roadmap lands (make agent-usage becomes the spend entry point). Until then, treat the steps below as the checklist and use the RunPod console/API directly.

Sandboxed execution: agents have no access to the host, its make targets, or the RunPod console. The human executes each checklist step and reports the result; the agent tracks the checklist and records the reported worker state and spend.

Before a cloud session

• Confirm the target endpoint exists and its image tag is pinned (no :latest).
• Check current worker state: how many workers are warm vs cold, and max-worker cap.
• Note expected cold start: first request after scale-to-zero pays image pull + model load; do not attribute that latency to inference.

During

• Keep an eye on active worker count; a stuck request can hold a worker (and spend) open.
• Route measurement work through the bench protocol (see the benchmarking skill) so cloud numbers are comparable to local ones and cold-start runs are discarded.

After a cloud session (required)

1. Force scale-down to zero workers so idle GPUs stop billing.
2. Verify no workers remain warm.
3. Read the spend query (via make agent-usage once wired) and record it against the task.

The skeptic and the primary agent both run this checklist: primary before starting cloud work, skeptic when verifying a "completed" cloud task actually released its resources.