Towards behavior-aware code recommendation : a system architecture

Abstract

Many code-completion assistants use source-code and repository context as their main inputs. Live interaction signals, such as pauses, deletions, typing patterns, and movement between files, are less often treated as inputs to a sequential prediction model. This thesis proposes an architecture that combines real-time IDE interaction data with local code context to estimate whether a suggestion is likely to be accepted at an eligible trigger point. The study adopts a computational design science research (CDSR) orientation and applies the problem-identification, objective-definition, and artifact-design stages of the design science research methodology. The output is a design specification rather than a working implementation. It includes a Visual Studio Code extension for telemetry collection, a pandas-based sessionization and feature-engineering pipeline, mixed-input LSTM and Transformer models, and a local FastAPI inference service. The thesis also defines how a future prototype should be demonstrated and evaluated through predictive-performance, calibration, latency, ablation, and participant-based measures. The main contribution is a traceable architecture that links the research problem to design objectives, system components, and future evaluation evidence.