Context window

The context window is the maximum amount of text, measured in tokens, that a model can take in on a single model provider request. The system prompt, every message in the session, every file the agent read and every command result it saw all have to fit inside it together. The size is a property of the model: current frontier models offer somewhere between roughly 200k and a million tokens, and the number moves with each release.

The window fills faster than people expect. A coding session appends constantly: your prompts are small, but one cat of a large file or a noisy test run can consume tens of thousands of tokens in a single step. When the window is nearly full, the harness has to do something: compact the history, drop old turns, or refuse to continue. Long before that point, quality usually slips, because the model is weighing more material than it can hold in focus (see attention degradation).

Treat the window as a budget rather than a limit. Load what the task needs and no more, prefer targeted reads over whole-repo dumps, and open a fresh session for a new task instead of stacking it on the old one. A bigger window is not a licence to be careless about what goes in; it just moves the cliff further out.