kedge-cache

A deterministic, content-addressed cache for AST compaction — never for LLM responses.

Caching LLM output is a production hazard: cache invalidation is hard, and returning a stale/hallucinated answer because a prompt looked similar breaks things silently. Tree-sitter compaction, by contrast, is a pure function of the file's bytes: sha256(file_contents) is a perfect key. If a file hasn't changed, its skeleton hasn't either — return the cached result and skip the parser entirely.