Budget and calibration

Budget

Gru tracks per card and in total: tokens (what the harness reports; otherwise bytes/4, marked ~), wall-clock and replans. Paid tokens are tokens on paid models, Gru itself included; free and local models are tracked but not counted against the cap.

• Defaults: a one-line warning at 300k paid tokens (spent by slot, what is left, what is running), stop-and-ask at 400k (raise it, shrink reviews to the risky hunks, shift work to cheaper slots, or stop). Set other numbers in gru.yml (team default) or in the preview; with a cap Gru warns at 75% and stops at 100%. A wall-clock cap behaves the same.
• Cost in the preview: an estimate per card from its size and roles. Tokens always; dollars when prices is set in local.yml. Estimates start from rough defaults and sharpen with calibration.
• On a warning: no more bumps without asking, no optional should-fix work.
• During long unattended runs Gru sends a push notification (if available and allowed) when it is blocked on you or a threshold is hit.

Quality over savings: if a card has no real checks, Gru says so and asks rather than shipping green-but-wrong.

Calibration

When a card closes, Gru records its size, the swe slot and model, replans, findings and tokens in local.yml, and keeps stats per size and slot (cards, first-pass rate, average replans, bumps). It then suggests, never applies on its own:

• "S cards on weak keep needing replans in this repo: move S to mid?"
• "M cards on mid pass first try: try weak for small M?"

Accepted suggestions update size_slots. Because local.yml lives with the project's logs, calibration carries across Gru sessions on your machine.