mirror of
https://github.com/hanzoai/ai.git
synced 2026-11-20 14:13:23 +00:00
Two quadratics braided into one stall. The token budget re-tokenized the whole accumulated section for every line appended — the count is now a running sum, one encode per line, with one shared encoder instead of a per-line encoding resolve. And section order came from probing the text for each breadcrumb key, which is a composed name the text does not contain, so every probe scanned the whole document to find nothing; the walk that builds the sections now records the order it walked. A 111KB changelog: 55s to split before, 0.03s after. Co-authored-by: Hanzo Dev <[email protected]>