RepoTruth packs an entire repository into one Beam call, finds where the docs contradict the code, and shows a measured token thrift meter. Judges: paste a public GitHub URL below for a live audit.
Beam’s long context reads README + docs + source together — the gap no thin chat demo covers.
We keep only true positives we can re-read in the code: file + line evidence, then publish to the Hub dataset.
On tinyclick, reasoning_effort=low used 246 reasoning tokens vs 572 at medium for the same 3 findings (−57%).
Beam-1.0-501B-A23B · rate limit ~5/IP/hour · ~150/day · zip ≤2.5MB · ≤120 files · ≤280k packed chars.
encode/httpcore — 5 verified · prompt 56,742 · reasoning 4,292 · 36.5s · effort=medium
proxy.request(...); httpcore.Proxy has no request (use ConnectionPool).Request(...).headers already has Host; constructor leaves headers empty.httpcore.backends.asyncio.AsyncioBackend — module does not exist (use AnyIOBackend).Also verified: itsdangerous dumps bytes-vs-str example, requests “7 methods” omitting options(), starlette “only requires anyio” vs conditional typing_extensions. Full JSON in the Hub dataset.