Deterministic folding for LLM agents: continuity without LLM compaction

I just open-sourced Context Warp Drive , a continuity engine for LLM agents. Repo: github.com/dogtorjonah/context-warp-drive Right now, the industry has two bad ways of dealing with long agent horizons: Just ride the 1M-2M context window. Use an LLM to summarize older messages ("compaction"). LLM summaries are inconsistent, they burn an extra model round-trip, they quietly drop the exact identifiers your agent needs (UUIDs, paths, hashes), and worst of all, they constantly rewrite the prefix—which t