Research Planner/User Prompt Refinement Agent

[moritz-baumgart]

Currently, a vague user prompt can lead to garbage results or to the agent using unsensible practices.
It sometimes even leads to complete failures, e.g., when no specific evaluation setup with train/val/test ratios is specified, the agent sometimes skips validation completely, which leads to failures in RecBole, because validation is required there.

We should add some kind of feedback loop before the current pipeline to enable the agent to elicit missing details from the user and to improve the starting user prompt.