On the new version, the model’s reply often stops after only a few lines, even though it clearly hasn’t finished.
There’s another issue: sometimes a message doesn’t actually send. It looks like it went through, but then the text comes back into the input box. I’m wondering if this might be related to optimistic updates.
Steps to Reproduce
On the new version, the model’s reply often stops after only a few lines, even though it clearly hasn’t finished.
There’s another issue: sometimes a message doesn’t actually send. It looks like it went through, but then the text comes back into the input box. I’m wondering if this might be related to optimistic updates.
When I used the cursor APP, I tried to use the latest Claude fable5 model to build the software architecture for my envisioned new project, but both conversations ended automatically after thinking for a few seconds and ended a single round of conversation.
Hey, thanks for the report, and thanks to @SilasMeng too. There are two different symptoms here, so let’s cover both.
Response cuts off after a few lines, or the chat ends right after thinking. This is a known class of bugs. Most often it’s caused by an unstable or proxied connection that drops a long-lived stream mid-way. Here’s what to check:
Cursor Settings > Network > Run Diagnostics and send the results. If any *.cursor.sh domain resolves to 198.18.x.x, your traffic is being intercepted by a VPN or proxy like Clash, Surge, or V2Ray in fake-IP mode. If so, disable it or exclude cursor.sh from proxying.
Settings > Network > Disable HTTP/2 HTTP Compatibility Mode. This often helps on corporate networks and with proxies.
Text returns to the input box after sending. This is a known bug in the recovery logic. If the connection fails mid-stream, the text can get written back into the input field. We’ve already filed an issue, but there’s no fix timeline yet. The model response isn’t lost, this is just a recovery-path UI bug.
So we can pull server-side logs for your cutoffs, please send a Request ID with Privacy Mode turned off from a specific interrupted response. You can get it from chat menu ⋯ > Copy Request ID. The IDs you sent so far don’t show up in logs, likely because Privacy Mode was on, or because the request failed before billing.
Maybe another agent’s configuration was causing the issue. I removed the AGENT.md file from the working directory, and everything started working again.
