Every coding-agent scaffold compacts: when the context fills, the run is summarised and continues from the summary. On three long-horizon benchmarks that costs almost nothing — the objective is fixed and the half-built work sits on disk, so most of what a summary drops can be read back. The exception is the part that exists nowhere but the context. The first summary already dropped 0.92 of 3 of a benchmark’s hard rules, and across 44 real sessions the instructions about how to work and what not to touch were dropped at 35.9% and 36.8%, against 6.2% for descriptions of the feature itself.