for_code

Generate one repo-grounded oral-grade deep-learning idea together with a paper-level method, a full experiment program, and the first formal batch.

Implement the active formal batch from .codex/active_idea.md, disclose the batch with request_user_input, and launch the formal run in screen.

Review one completed formal batch, then either improve the idea through multi-round clarification, abandon it, or finish it.