Random &cool

Better multi-step coding reliability, fewer logic bugs and missing edge cases in longer code outputs. Stronger repo-level coherence, better at keeping imports, file boundaries, and refactor intent consistent across multiple files. More stable instruction-following, less format drift and fewer ignored constraints in long prompts. Improved tool and agent follow-through, fewer broken tool sequences and more complete multi-step task execution. Higher output stability, fewer abrupt truncations and fewer random mid-answer shifts in tone or direction.