WiseSpec: Requirements-Driven Agents for Code Generation

Code generation aims to automatically generate source code from task requirements and has attracted significant attention with the rapid advancement of large language models (LLMs). Despite remarkable progress, LLMs often struggle to generate correct code for complex software engineering tasks because task descriptions are frequently incomplete, ambiguous, or lack critical contextual information. Existing approaches primarily improve the capabilities of coding agents through more sophisticated tools, skills, and workflows, while largely overlooking the quality of the task requirements themselves. To address this limitation, we draw inspiration from software requirements engineering and propose WiseSpec, a novel requirements-driven agent framework for repository-level code generation. WiseSpec automatically constructs structured and information-rich requirements, assesses their quality through execution-based evaluation, and iteratively refines them to better guide code generation.

WiseSpec operates in three main stages: requirement construction, quality assessment, and iterative refinement. Given an initial task description, the framework first constructs structured requirements by extracting key entities, actions, and constraints, and enriching them with contextual information from the repository. These requirements are represented in a formalized format that includes preconditions, postconditions, and invariants. The quality of the constructed requirements is then assessed through execution-based evaluation: candidate code is generated from the requirements, executed against test cases, and the results are used to compute a requirement quality score. This score guides an iterative refinement loop where the requirements are updated to resolve ambiguities, fill missing information, and correct inconsistencies. The process continues until the requirements are deemed sufficient or a maximum number of iterations is reached. The final requirements are then used to generate the target code. The framework can be formalized as follows: Let $R$ be the set of requirements, $C$ be the generated code, and $E$ be the execution results. The quality score $Q(R)$ is defined as the pass rate of $C$ on test cases derived from $R$. The refinement step updates $R$ to $R^{\prime }$ such that $Q(R^{\prime })>Q(R)$. The overall objective is to maximize $Q(R)$ subject to computational constraints.