The Boring Compaction Winner Broke in Production. The Fix Is More Boring.
Friday's production context compactor, which won a benchmark for AI context compaction, failed in about 30% of attempts over 60 days, with failures costing up to 11.7 million input tokens and 77 minut…