Training Composer for longer horizons

Introduced self-summarization as a trained behavior for Composer, a specialized model for agentic coding. This technique allows the model to learn to summarize its own context when it approaches a fixed token-length trigger, enabling it to handle tasks requiring hundreds of actions by preserving critical information across extended trajectories. This approach is integrated into training, where self-summaries are rewarded, leading to more token-efficient compaction (around 1,000 tokens per summary compared to thousands for baseline methods) and a 50% reduction in error from compaction. The post details the self-summarization process, its token-efficient nature, and demonstrates its effectiveness on challenging tasks like 'make-doom-for-mips' from Terminal-Bench 2.0, where Composer successfully solved the problem over 170 turns by intelligently summarizing over 100,000 tokens.