AI Coding Quota: Two Ratios That Predict the Wall, and a Preheating Trick
Developers burning through AI coding credits mid-session lose context along with time. These rules turn an unpredictable wall into a schedule you can plan around, and the symlink trick preserves session history across accounts, so switching providers doesn't mean re-explaining everything.
AI coding quotas come in three metering styles: short windows that reset a few hours after the first message, weekly or monthly caps, and daily per-request buckets. Providers expose only a usage percentage and a reset time, so the method compares how much time has elapsed against how much quota is used. If usage extrapolates beyond 110%, a shutdown is coming; if a quarter of time remains and a quarter of quota sits unused, that quota will be wasted. Extrapolation uses recent speed, not the window average, because a burst of activity can exhaust a 5-hour window in under an hour.
When a hit is coming, the cheapest move is switching to a smaller model for lightweight tasks like commit messages, keeping the session and context intact. After a hit, the escalation order is wait, switch accounts, or switch agents, with waiting most cost-effective when reset is under an hour away. Preheating sends a minimal request earlier in the day to shift the reset to lunchtime, and sharing session history across multiple subscriptions works by symlinking the history directory while keeping credentials separate, with the caveat that settings.json must be copied rather than symlinked. These tactics are packaged in the open-source workbench Antumi.
The method of predicting quota exhaustion with explicit thresholds is more reliable than asking the model itself, which costs quota and gives inconsistent answers.
The distinction between a rolling window and a sliding window clarifies a common misconception; the start is fixed at the first message, giving the user control over when the window begins.
Preheating exploits the service's own semantics by scheduling a minimal request, turning a fixed constraint into a flexible resource.
The symlink workaround shows that CLI tools are not designed for multi-account management, so users have to patch around assumptions about single-tenant config directories.