#2492: Can't see the actual reply content, only Yellow thinking process content.

Found the issue. Root Cause: There are 1582 orphan reasoning messages in your session file. These messages have the role set to assistant with completely empty content fields, while lengthy thought texts are stored in the reasoning field. Take Message [1] as an example: role=assistantcontent: '' ← Empty reply contentreasoning: 'The user wants me to read a file at "D:\Documents\ 单词江湖 \Android 游戏...' ← Yellow thinking process content This is abnormal. In a standard session, assistant messages should contain valid reply content in the content field, with reasoning only serving as attached metadata for reference. These 1582 entries only retain reasoning data without formal reply texts. Cause Analysis: When the MiniMax provider performs inference, it stores thought processes into the reasoning field via the on_reasoning callback. Meanwhile, the verbal-style reasoning streamed through token output is rendered by the client as thinking card content. If the visible text in the token stream consists solely of thought processes with no substantial answers, the content field will remain blank. These 1582 orphan reasoning records are most likely duplicated reasoning events logged into the messages array during historical session replay, rather than genuine lost reply content. Why you see replies flickering and disappearing with full-screen yellow thinking content You send a query → MiniMax runs inference → thought processes are written into thinking cards, while formal replies go into the content field (which may also be empty or extremely short). The client streams and renders increasingly lengthy thinking card content. Upon receiving the done event, local S.messages gets overwritten by messages returned from the server. Part of the 1582 historical orphan reasoning records are incorrectly bound to the current conversation turn (caused by mispositioned attachment during finalizeThinkingCard execution or session replay). Abnormally excessive and lengthy thinking cards generated during renderMessages completely obscure the actual reply content.