@Youssofal_ yeah open models have their quirks when it comes to thinking and tools - unless harness+provider specifically addresses those failure modes on a model by model basis you will have an unreliable experience
@ishaansehgal ye our whole arch has been based on this for a bit now: displayed UI, context window, etc are all just different projections of a session’s events
Take another look at that cache hit pricing - that’s $0.003625, or 0.83% of uncached input price. Compare this to the closed model standard of 10%. There’s no longer any excuses for providers to have poor cache management.
blessed to be working on a coding agent where the testing cycle includes vibe coding random silly things. I don't know what this is but Kimi K2.6 said it wanted to make it