More information: Now failed even when changing from ASK to AGENT mode, auto model, entered in infinite loop. Switching model to Sonnet 4.5 Thinking worked.
Can you try to reproduce it with opus 4.6 and opus 4.6 Max (both with thinking)? I’ve recently observed that behavior with those specific 2 models, not with the rest.
this sounds like the same issue someone reported in this thread. the “Auto” model selection seems to pick models that don’t always respect the ask/plan mode boundaries.
not sure if this is a fix, but one thing that’s worked for me when a model loops is just starting a fresh chat. once it gets into that loop state, the conversation history seems to make it worse. also worth trying: when you’re in ask mode, start your prompt with “don’t edit any files” just to reinforce it. annoying but it helps until they fix whatever’s going on with auto mode.
interesting that switching to Sonnet 4.5 Thinking worked. might be worth reporting which model Auto selected when it looped, if you can see it in the usage log.
Hey, thanks for the report. This is a known issue with Composer 1 in Ask mode. The model sometimes doesn’t realize it’s in read-only mode and tries to edit files.
I can see you already found a workaround by switching to Sonnet 4.5 Thinking. That’s exactly what we recommend in cases like this.
The team is aware of the issue. For now, the best option is to use a different model (Sonnet 4.5 works well) or explicitly remind it in your prompt that the Agent is in Ask mode and shouldn’t edit files.
Hey @RodriGas, could you share the Request ID from one of the sessions where you reproduced this? We need to check which exact Composer version handled the request.
How to get the Request ID: open the chat menu (three dots in the top right) and select “Copy Request ID”.
If you can reproduce the bug again, copy the Request ID right after the loop starts. This will help the engineers narrow it down.
I’ve been searching in my chat history, but I’ve been unable to find the failing request. Anyway I’m pretty sure it began to fail right after the update that came with composer 1.5. In fact, I’ve disabled composer 1.5, couldn’t test in depth because I’m at project deadline, but the few requests I did, for me, I think performed bad. As I’m using auto, can’t guarantee 100% it’s composer 1.5 the problem, but since I’ve disabled it, I see similar reply performance like before the updtae with composer 1.5. I will give it another try in the future, when I close the urgent project.
Of course, if the problem appears, will copy request ID and send it to you.