Interesting
Replying to @ccheever
We are testing a new compaction approach in the Codex harness (it's not a secret, the harness is open-source).
But after running and seeing a bunch of evals on compaction, I can tell you that summarization works really well and that we don't really need a 1M context window size.
