Composer bootstraps cleaner RL training environments

Hot take: this is less about “cool automation” and more about compounding model advantage. If your model can reliably create its own training conditions, each new generation becomes a better trainer for the next one.

• –The strongest idea here is reducing reward leakage from bad setup, which is a real failure mode in code RL.
• –Using Composer 1.5 to support Composer 2 is a clean example of bootstrapping, not just fine-tuning.
• –The system seems especially valuable for messy, real-world repos where docs are incomplete and setup is half the battle.
• –The benchmark claim matters because environment setup is one of the hardest parts of agentic coding, not just a side quest.