The freshness argument here cuts both ways though. Search_docs/read_doc always serving the current production docs kills staleness against the docs site, but the actual repo the agent is writing into is very often pinned to an older release of whatever SDK it's touching. I hit a version of this building a GraphQL skill for Hashnode's own API: the schema we described was correct for the live API, but an agent still confidently wrote a field that had been renamed since whatever training data or cached doc it half-remembered, and the fix wasn't fresher docs, it was reading the live schema at call time instead of trusting any static description, live or not. Does read_doc have a way to pin to whatever version is actually in package.json/requirements.txt, or does it always resolve to latest regardless of what the target repo is running? If it's always latest, you've traded "hallucinated from stale training data" for "hallucinated from a newer API than what's installed," which is a smaller bug but the same shape.