Michael Vierhauser

Michael Vierhauser

Ass.-Prof
Department of Computer Science
Quality Engineering

Living Architecture: IDE-Integrated Recovery and Live Evolution of Design Intent During AI-Assisted Development


Modern software is increasingly written with the help of AI coding assistants, where developers express intent informally through prompts, issues, and lightweight specifications rather than explicit requirements or architecture documents. In these "vibe-coding" workflows, architecturally significant decisions are often made implicitly and at speed — and the rationale behind them is rarely captured. As the codebase grows commit by commit, this architectural intent becomes buried in the implementation, leaving developers and AI assistants alike with little visibility into why the system is built the way it is.

Recent research has shown that this latent intent can be recovered directly from code using LLM-assisted analysis to surface a system's Architecturally Significant Concerns, the Design Decisions that address them, and the typed relationships connecting them. So far, however, this model is produced through an offline, batch-style pipeline and inspected in a standalone viewer. To be genuinely useful during everyday development, it needs to live where developers work — inside the IDE — and stay current as the code changes, rather than being regenerated after the fact.

The research focus should include:

  • IDE Integration (VS Code): Designing and building a VS Code extension that surfaces Architecturally Significant Concerns, Design Decisions, and their relationships directly in the editor, and links them back to the relevant source locations.
  • Live, Incremental Evolution: Triggering and integrating incremental model updates on development events — commits, pull requests, or save actions — so the model reflects the current state of the code rather than a past snapshot.
  • Change Awareness for AI-Assisted Development: Presenting what changed architecturally after an AI-assisted edit — newly introduced, strengthened, weakened, or removed concerns and decisions — so developers can review the architectural consequences of generated code.
  • Usability and Developer Experience: Investigating how this information should be presented (inline annotations, side panels, diff-style change views) so that it informs without overwhelming, and evaluating it with developers.

Thesis Goal The goal of this Master's thesis is twofold. First, to review existing work on architecture recovery, design-rationale recovery, and IDE-integrated developer tooling, with a particular focus on keeping recovered architectural knowledge synchronized with evolving code. Second, to design, implement, and evaluate a VS Code extension that integrates the Living Twin Peaks Model into the development workflow and updates it incrementally as the system evolves through AI-assisted development.