Context Window Advisor

Overview

Codex logs every model call's input context, cache split, output size, and every compaction event. That is enough to measure the parameters of the Agent Task Cost cost model — and once those are known, the context window that a task consumes least at follows from the model rather than from guesswork.

This skill reads the user's local session history, computes the least wasteful window for three task sizes, compares it against the window they are actually running, and offers to change the setting.

Everything the user sees is stated as a multiple of what the task would consume at its best window. That reads the same on a subscription and on the API, because metered consumption is weighted the way the price list is — usage and bill are one quantity.