The interesting tradeoff here is not just better retrieval, but whether the pre-compiled wiki makes behavior more predictable over long-running tasks. It would be useful to see comparisons against a plain markdown repo plus Claude/Codex: task completion, citation accuracy, stale-context failures, and the cost of keeping the wiki updated. Multi-model support and bring-your-own-key options also seem important for enterprises that don't want their company's context tied to a single provider.