Deep-Context-Awareness-Based LLM Code Generation and Accurate-Defect-Repair Integrated Architecture

In response to the issues of context fragmentation and delayed defect repair in large language model (LLM) code generation, this paper proposes a deep context-aware generation-repair fusion architecture. Through a bidirectional collaborative mechanism, it achieves a paradigm shift towards "generation as correctness." This architecture innovatively constructs multi-granularity context encoding models, dynamically integrating code structure, developer intent, and project-level constraints. It combines a neural-symbolic collaboration framework to deeply couple LLM's generative capabilities with the reliability of formal verification. The —— generation module is implemented using graph attention networks to achieve cross-file semantic association, while the repair module accurately locates and fixes defects through probabilistically guided patch search and hierarchical verification strategies. The architecture supports multi-objective optimization and human feedback reinforcement learning (RLHF), balancing code quality, performance, and security requirements, and generates traceable decision chains to ensure ethical compliance. This study provides a new generation of solutions for automated software engineering that combines efficiency and credibility, and lays the technical foundation for future directions such as multimodal context expansion and quantized code analysis.