Although much research has investigated oral correction feedback (OCF), little has addressed EFL teachers’ cognition and practices about OCF provision in blended learning environments. This qualitative case study aims to occupy this gap. This study explores four issues: teachers’ beliefs about giving OCF; their practices; convergences and divergences between their beliefs and practices; and the key factors influencing these relationships. Data were collected from six teachers teaching in a blended learning programme in a Vietnamese university. The data collection process began with semi-structured interviews, classroom observations followed by stimulated recall sessions, then focus group discussions, and narrative frames. All data were subjected to a process of grounded analysis. A key issue that the findings revealed was the appropriate balance between providing immediate feedback and delaying the feedback. Much research has focussed on the former (e.g., Brown, 2016; Li, 2010; Lyster & Saito, 2010; Lyster, Saito, & Sato, 2013; Mackey & Goo, 2007; Russell & Spada, 2006), and much less attention has been paid to delaying feedback (e.g., Li, 2018; Quin, 2014; Rolin-Ianziti, 2010). On the whole, there were more convergences than divergences between the teachers’ beliefs and practices. The teachers rarely provided immediate correction while students were performing in pairs or groups: much more often, they delayed their error treatment until the students had finished their performances in front of the whole class. This was because they believed that doing so would promote the students’ confidence and fluency and avoid demotivating them. Furthermore, it was found that they mainly corrected the errors in the whole-class setting as they believed that other students could learn from hearing the teachers and peers correct the other student’s error and then avoid committing the same mistakes. Although they would like the students to correct the errors, they could not transfer this into their practice. Such approaches to error correction perhaps led to limited improvement in the students’ fluency. The teachers’ cognition and practices were also viewed through the lens of principles of scaffolding and the framework of Cultural Historical Activity Theory (Engeström, 1987). Findings indicated the teachers did not fully apply van Lier’s (1996) six principles of successful scaffolding; therefore, impeding the students’ learning. In addition, findings revealed different levels of contradictions (Engeström, 1987). Primary contradictions occurred between teachers’ cognitions, emotions and practices due to cognitive and emotional dissonance. Secondary level contradictions took place between teachers’ cognitions, emotions, practices and learning outcome mainly due to: (1) the absence of expert knowledge within the academic community; (2) students’ low proficiency level, motivation and autonomy; (3) institutional constraints, such as large size class, time limits, and ineffective regulatory guidelines; and (4) a teacher-led convention. Quaternary contradictions emerged between the central activity (OCF) and its concurrent activity (online learning) because of little cross-referencing between face-to-face and online tasks and lack of online tools for teachers to facilitate students’ online learning. To improve students’ learning outcome, the study suggests expansive learning including a transformative change in the activity of OCF and the provision of conditions for boundary crossing to take place, such as relevant teacher professional development/learning.