Copilot premium requests, against the clock

UsagePool resolves the user’s own gh executable and runs exactly one fixed read-only request: gh api --include copilot_internal/user --hostname github.com, with a minimal environment, bounded output, a deadline, cancellation and child cleanup.

--include prints the response headers ahead of the body. They are read for one narrow purpose: telling GitHub’s three different 403s apart — an exhausted request budget, an organization policy, and bad credentials — and reading the retry-after and x-ratelimit-reset values that set the retry delay.

The adapter hashes the returned login for account continuity and projects only the plan, the monthly reset, and the known Chat, Code Completions and Premium Requests quota snapshots. Limits marked unlimited or unavailable stay hidden rather than being drawn as full.

copilot_internal/user is a private GitHub endpoint. It is not a documented, versioned API, and it may change without notice. If it changes, the Copilot meters can go stale or disappear until the adapter is updated — that is a real risk and it is stated here rather than discovered later.

This is also why the Copilot adapter reads response headers: GitHub returns three materially different 403s through the same door, and a meter that cannot tell "budget exhausted" from "your organization disabled this" would be worse than no meter.