Building ai_resets: tracking AI usage limit resets across Claude, Codex and Gemini

I am Jeremy Koski, an independent Computer and Information Research Scientist in Honolulu. I develop with several AI coding tools in parallel, including six Codex accounts, Claude Code, Antigravity, and Cursor. I built ai_resets, a Python project backed by the PostgreSQL database ai, to track resets, plan changes, credits, promotions, and student offers. The most useful design decision was keeping three things separate: a provider's published rule, an account's observed state, and a prediction derived from those inputs. A reset is not one universal event. Rolling windows, fixed daily boundaries, billing cycles, weekly pools, and manual claims require different calculations and different evidence. The engine accepts an explicit now value so boundary behavior is reproducible. Time is stored with its context, including timezone aware timestamps. Claims are verified against evidence for the specific stored assertion. Account state is collected without crossing user boundaries. Uncertainty belongs in the interface beside the clock.