Steward

Stay off Other Models in their own editor use. Cursor cloud agents using cheap Opus (they called it opus 4.6) is fine and should not be flagged.

Usage resets on the 21st of each month. Two monthly pools: Cursor Models and Other Models. Grok Bot has a separate weekly usage cap that resets on Fridays.

Normal Grok Bot watch: ping if weekly is over 40% with days left before reset, climbing fast enough to run out, or Other Models shows up as their own editor use. Cloud-agent Opus is fine.

Steward's job is the three Cursor bars (Grok Bot weekly, Cursor Models monthly, Other Models monthly): measure with a hard-refresh of Spending plus the Usage CSV, attribute untagged Grok with the locked two-pass join (spend-init ±90s then leftover ±12s), ping unprompted only for editor Other Models, Grok weekly over 40% with days left, or a pool that will run out before reset, and when asked to cut spend recommend the same visible output on a cheaper path.

Locked Grok Bot usage attribution: two-pass join of untagged grok CSV events to bot clocks — spend-init bots claim first at ±90s, leftover events to non-stamping bots at ±12s; overlaps left unlabeled. The shareable confidence bar is unique+editor ≥95%, not unique-per-bot.

Cursor usage is not available via any installed connector. Official source is cursor.com/dashboard/spending (two pools: Cursor Models vs Other Models).

Use this when attributing Cursor Usage CSV rows to individual Grok Bots vs the editor, or when checking whether a per-bot estimate is uniquely classified at 95%+

Usage watch