Engagement in Code Review: Emotional, Behavioral, and Cognitive Dimensions in Peer vs. LLM Interactions

Data The repository contains all anonymized study artifacts across two phases. For Phase I, it includes the source code submitted by 20 software engineers, the peer-authored code reviews (filenamed Re_P<author>_<reviewer>.pdf), and the matched ChatGPT-4o reviews for the same submissions, along with the Phase I interview guide and full interview transcripts. Phase II adds the follow-up interview guide and anonymized transcripts, plus a prompt-engineering report detailing how we aligned LLM feedback to participant preferences (structure, concision, tone, why-based rationales, and scope). The repository also provides the member-checking questionnaire and responses, a data-saturation monitoring spreadsheet, and the combined codebook (Phases I & II) with First Cycle codes and Pattern Codes per RQ (RQ2-RQ4, RQ1 is synthesized from these). A README file document is also supplied. All materials are shared in line with informed consent and ethics approvals. Methods The study follows a two-phase qualitative design examining code review as a socio-technical process with human and LLM reviewers. Phase I elicited peer-to-peer and ChatGPT-4o reviews for participant code and used semi-structured interviews to identify affective responses, behavioral engagement, and perceptions of LLM input. Phase II conducted follow-up interviews and a prompt-engineering alignment to tailor LLM feedback to stated preferences. This alignment allowed us to compare analysis of content characteristics and adoption. Analysis proceeded via iterative First Cycle coding to Pattern Codes, and structured monitoring of data saturation. Member checking was used to validate our results. Results We propose a high-level loop from emotional engagement (self-regulation) to behavioral engagement (social calibration, negotiation, sense-making) to resolution and implementation, with learning and team-norm codification as frequent outcomes. Engineers deploy reframing, dialogic regulation, avoidance, and occasional defensiveness to move from affect to resolution; these strategies are scaffolded by motivation for code quality, accountability (team and individual), and a growth mindset. Resolution pathways differ by locus: solo/internal sense-making (fast but potentially idiosyncratic), dyadic negotiation (moderate pace), and team escalation (slower but standardizing); partial adoption is common and recurrent issues often lead to guideline codification. LLM feedback when aligned with engineers' preferences on structure, concision, tone, why-based justifications, and fit-for-purpose scope reduces cognitive effort and increases stated likelihood of adoption, whereas verbosity, generic tone, missing context, or over-scoped feedback increase cognitive friction. Across both phases, participants positioned LLMs as augmenters of human review rather than social substitutes.