To address the "scientific experience bottleneck" caused by fragmented toolchains and implicit domain conventions in scientific software, researchers introduced ScienceIDE, an infrastructure designed to transform complex scientific code repositories into programmable, agent-learnable environments. This framework enables autonomous agents to convert existing repositories into executable workspaces by utilizing expert-defined scientific cases and rigorous acceptance criteria. These environments facilitate automated task generation, execution, and specialized scientific verification, creating a structured foundation for supervised fine-tuning and reinforcement learning. Leveraging these verified interaction trajectories, the researchers developed the PhAI-IDE model family, comprising the 72B, 9B, and 4B parameter versions. Empirical results demonstrate that training on these scientific environments yields significant performance improvements in held-out scientific code repair tasks. Furthermore, the models exhibit positive transfer effects, showing enhanced capabilities across general-purpose benchmarks involving reasoning, code generation, and domain knowledge. By turning scientific software into a shared substrate for agent training, ScienceIDE provides a scalable method for developing scientific intelligence through integrated, executable workspaces.