I’ve been busy and couldn’t reply immediately, but this bug has been causing significant issues for me all week—it’s still unresolved and has cost me a lot of time and credits over the weekend. I suspect the gravity of the problem isn’t fully appreciated because it involves permanently lost work.
I consistently encounter this freeze with composer agent mode AND normal mode, and I finally managed to capture a session of it failing completely. I left the recording running because this issue occurs about every 20 minutes. Once a thread freezes, it’s lost for good and all progress is gone.
I’m a developer, so I’ve recorded a detailed video that may be the best repro you’ll get. In the video, you’ll see:
- Initial Frozen Thread– I start off with an already-frozen thread to show how it fully hangs. In that case, I had to restart to get past it.
- Regular Workflow – I continue working in Composer Mode (no Agent Mode here) for about 20 minutes. During that time, everything functions but occasionally struggles—typical whack-a-mole issues while coding with LLM tools… but composer isn’t doing great here either, numerous apply mistakes.
- New Freeze– At the end, I get another freeze that kills the thread entirely, causing me to lose my work again.
It’s very frustrating and undermines productivity and creativity. The user experience becomes especially poor if you’re left waiting indefinitely for a response that never comes. Although it was reported over a month ago and was said to be fixed, the issue seems to have reemerged almost immediately. For reference, it doesn’t appear specific to Anthropic—4o and o1-mini also exhibit this behavior. From my perspective, it might be tied to the “shadow indexing” aspect of the client.
Please escalate or prioritize this bug—it’s a critical issue. It might be related to my particular development toolset, but I’m not doing anything out of the ordinary: just using Jest test tools and ESLint. I’ve also opened DevTools at the end of the recording, which might shed more light on what’s happening.
Thank you for looking into this. I’m happy to share any additional details or testing insights you need to help diagnose and solve the problem… wasted quite a bit of time tho.