Overflow to disk

Red-team lab (Module 6, L6.6): a tool result over the threshold never enters the prompt. A 40,000-token test log inlined pushes the window from 0.600 to 0.970 and forces compaction within 4 steps; spilled, the prompt grows by a 12-token locator and a ranged read returns the 11 FAIL lines. Then attack the store: a naive path-string scheme loses to a stale summary, a check-then-use race and three injected locator strings (3 of 3 land); the hardened store of harness-minted ids, mint-time sha256 and verify-on-read refuses all three while legitimate reads still work. In-memory dicts only; no real files.

This Harness Engineering lesson is part of a paid plan. Sign in with a plan that includes this course to read it.