Opus-4.6 max model, agent kicks subagent to do some work, and the subagent gets stuck in an infinite (appearing) thinking loop running for 30+ minutes. Need some sort of control on how big the thinking buffer is, or how much time it gets. Burned 1M tokens doing nothing then died because context was exceeded.
Hey, thanks for the report. This is a known issue. Subagents can get stuck in long thinking loops, especially in max mode, and can burn through the context window without producing useful output. We’ve seen similar reports across different models.
A couple of things that might help:
If you see a subagent stuck thinking for more than a few minutes, stopping it manually is the best move for now.
You can try running the same task without max mode to see if it happens less often.
Can you grab the Request ID from the chat where this happened? Open the top-right menu in the chat, then click Copy Request ID. That’ll help us look into what went wrong on the backend.
Let me know if you have the Request ID, or if this keeps happening.