While code summaries are critical for software comprehension and maintenance, evaluating the quality of automatically generated summaries remains a significant challenge. Conventional metrics rely heavily on rigid lexical or semantic similarities. Conversely, emerging LLM-as-a-judge methods improve flexibility but often exhibit systematic stylistic biases and fail to reflect developers’ practical needs. To address this, we propose MultiDRECS, a multi-dimensional, requirement-aligned framework for automated code summarization evaluation. By employing a collaborative multi-agent architecture, MultiDRECS successfully mitigates the redundancy biases typical of single-LLM judges. Within this design, it performs statement-level scoring across three developer-centric criteria—\emph{adequacy}, \emph{conciseness}, and \emph{faithfulness}—integrated with an information value ranking mechanism that assigns priority-based weights to different summary types. Consequently, MultiDRECS achieves strong agreement with human judgments while requiring only lightweight backbone models. To rigorously validate such evaluation capabilities, we further introduce DiscrimEval, a novel benchmark comprising real-world methods with rich contexts and systematically synthesized variants (\emph{stylistic} and \emph{error-injected}) designed to stress-test metric discriminability. Experiments on DiscrimEval using Qwen2.5-Coder-7B-Instruct demonstrate that MultiDRECS outperforms traditional metrics and larger LLM judges (e.g., Qwen-Plus and DeepSeek-R1) in evaluation accuracy. Crucially, it achieves this at a lower deployment cost and facilitates practical adoption in automated software engineering workflows.