We've been working on a similar problem, but went the route of pushing this into the model and runtime layer instead of the orchestration layer. Less scaffolding, more baked into how the model reasons. Happy to share more if useful

Hiding details of reasoning and work traces to delay compaction as much as possible makes sense because in my experience compaction essentially borks our sessions

Good luck keeping your skills and prompts intact with a system wide compaction operation!

The holy grail for developers: never have to do marketing yourself again

> more to do with the deeply unfussy work of actual context engineering.

I agree. There is something to be said about autonomous systems producing valid outcomes within the current constraints of RL