Defesa – Tese de Nicole da Costa Davila

Aluno(a): Nicole da Costa Davila
Orientador(a): Erika Fernandes Cota
Coorientador(a): Igor Scaliante Wiese

Título: Toward Author-Guided Review: A Theory and Agentic Architecture for Reflective Code Review
Linha de Pesquisa: Engenharia de Software

Data: 16/06/2026
Hora: 09:00
Local: Esta banca ocorrerá de forma remota. Acesso público disponibilizado pelo link https://mconf.ufrgs.br/webconf/00108898.

Banca Examinadora:
-Karina Kohl Silveira (UFRGS)
-Cleidson Ronald Botelho de Souza (UFPA)
-Gustavo Henrique Lima Pinto (UFPA)

Presidente da Banca: Erika Fernandes Cota

Resumo: Modern code review (MCR) is a widely adopted practice in software development that enhances code quality, promotes knowledge sharing, and fosters team collaboration. Despite a large body of research on tool support for reviewers, limited attention has been given to how authors can be systematically supported in reasoning about their own code changes before submission. As a result, the pre-review stage remains largely implicit and unsupported: contextual and domain-specific knowledge underlying design decisions is rarely externalized, requiring reviewers to reconstruct author intent through clarification during the review itself. This thesis investigates how to support code authors in reasoning about their own code changes before peer review, particularly by making contextual and domain-specific knowledge explicit. To characterize the problem and its context, we conducted three complementary empirical studies: a fine-grained analysis of coding issues found and fixed during code review in TypeScript projects; an industry case study on the adoption of AI-based programming assistants, revealing that their effectiveness decreases substantially when contextual reasoning is required; and a gray literature review on generative AI in code review, showing that current approaches predominantly support reviewers after submission, with limited attention to the author’s pre-submission reasoning. Grounded in these findings, we propose the author-guided pre-review theory, which conceptualizes pre-review as a structured, guided reasoning process defined by three core constructs: Contextual Review Knowledge, Reviewer Likely Concerns, and Guided Self-Reflection. To instantiate the theory, we design an agentic architecture comprising Contextualizer Agents, which analyze heterogeneous knowledge sources to identify potential concerns; an Orchestrator Agent, which generates reflective prompts grounded in those concerns; and an Interaction Agent, which engages the author in Guided Self-Reflection while extracting and preserving the Contextual Review Knowledge externalized during the session. The theory and architecture are evaluated analytically through a stress test of their foundational premises and design decisions.

Palavras-Chave: modern code review, author support, agentic architecture, generative artificial intelligence