opus 4.6 randomly "not thinking" in claude code desktop, found out why
Heat trend
The percentage is based on available heat signal, not comment count or independent people.
Users of Claude Code desktop app noticed Opus 4.6 sometimes appeared to "not think" before answering, even for complex questions, while other times it would "think" for simple ones. Investigation into session JSONL files revealed that thinking blocks were always present, but often had empty text, despite the signature field and output token counts indicating actual processing. This suggests that the "thinking" text was simply not being streamed to the app, potentially leading to user complaints about adaptive thinking performance.
kept asking actually hard questions in claude code (desktop app) the past couple days and half the time it would answer instantly, no thinking block at all. then a throwaway "test" session thinks on literally one word. felt completely random and not tied to question difficulty at all. anyone else had extended thinking triggering way less lately, or just inconsistently?
so i dug into the session jsonl files. the thinking blocks are there on every single turn, even the "didnt think" ones. they just have empty text. the signature field (the encrypted reasoning) is huge and the output token counts show real thinking happened. the text just never gets streamed to the app
turns out the desktop app updated and now launches the CLI with --thinking-display omitted, its right there in ps aux. older builds showed summarized thinking by default. so the model is still doing extended thinking every turn, the app just throws the text away. and the mode can apparently flip mid session, which is why it felt so random
so if opus 4.6 "stopped thinking" for you after a desktop update, it probably didnt. wondering how many "adaptive thinking got worse" complaints are actually just this