As large language model (LLM) agents mature, their deployment in the legal domain has grown rapidly. Legal Judgment Prediction (LJP) requires multi‑step reasoning and the retrieval of external legal evidence. Existing legal search agents typically construct queries directly from the user's initial request, overlooking latent legal issues and salient details, which limits retrieval quality. To address this, we introduce LegalEvolver, a query‑optimization framework tailored for LJP. The framework works as follows: 1) treat high‑quality optimized queries as seeds and heuristically generate search trajectories; 2) abstract the experience of multi‑round retrieval into rubrics and use these rubrics to guide reinforcement learning (RL) during self‑evolution. In each iteration the model first refines its queries according to the current rubric, then updates its parameters via rubric‑guided RL, thereby surpassing its initial performance ceiling. Extensive experiments on public benchmarks show that optimized queries markedly improve recall and final judgment accuracy; compared with a baseline that applies supervised fine‑tuning (SFT) followed by RL, rubric‑guided training yields superior results; and dynamically evolving rubrics outperform fixed ones, confirming the benefit of self‑evolution.\n\nReview: LegalEvolver combines query generation with rubric‑driven self‑evolution, offering a practical route to boost retrieval quality in LJP systems. Its iterative reasoning and evidence‑gathering capabilities demonstrate strong adaptability and robustness.