Recasts are an important type of implicit negative feedback that has attracted much attention in both L1 research and SLA research. The utility of recasts in face-to-face interaction has been empirically established, and the contingency of recasts is argued to be the key. However, the efficacy of recasts in computer-mediated communication (CMC), text-based online chatting in particular, remains questionable due to the possible violation of this contingency factor in the "split negotiation routines" commonly observed in CMC discourse. This study used a repeated-measure design to examine the potential impact of the contingency of recasts on noticing as well as some contextual factors that might mediate the contingency effect on the noticing of recasts. In this study, 17 ESL learners were invited to chat with one researcher on two dyadic communication tasks, one preceded by prewriting and the other without. Think aloud protocols and stimulated recalls were used to measure the noticing of recasts. It was found that participants noticed contingent recasts significantly more often than noncontingent recasts. Furthermore, working memory and prewriting were found to mediate the contingency effect, and learner proficiency level was found neither pertinent to the noticing of recasts nor mediative of the contingency effect.