I'm also one of the people working on this project.

Oversized result artifacts are session-namespaced files cached on disc. Coordinator runs are given a lookup manifest and the internal scratchpad tool can search, grep, etc. to find only the pieces they need with inner loops. This enables a sort of model-driven targeted context engineering.

You can explore the implementation here: https://github.com/mezmo/aura/tree/main/crates/aura/src/scra...