Agent drops connection/fails while writing plan files

[Samuel_Talenti (Sam Talenti)]

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

When creating plans the agent will fail when writing the plan with a generic network error. Sending follow-up messages the agent has no context of the previous chat, and there is no plan, resulting in significant wasted tokens

Steps to Reproduce

Have reproduced with Opus 4.7 extra high & Opus 4.6 high.

1. In plan mode ask to create a sub-plan of some other plan documents (not sure if directly related), in my tests I’ve had at least one follow-up question/answer before it starts to create the plan
2. Observe, will spin on creating once, then go into a thought where it mentions needing to follow proper JSON, attempt another edit, at which point the network failure arises

Expected Behavior

It produces a plan file, or at least fails gracefully allowing further attempts while retaining context.

Screenshots / Screen Recordings

Operating System

MacOS

Version Information

Version: 3.4.20 (Universal)
VSCode Version: 1.105.1
Commit: 0cf8b06883f54e26bb4f0fb8647c9500ccb43310
Date: 2026-05-15T02:26:10.351Z
Layout: glass
Build Type: Stable
Release Track: Default
Electron: 39.8.1
Chromium: 142.0.7444.265
Node.js: 22.22.1
V8: 14.2.231.22-electron.0
OS: Darwin arm64 24.6.0

For AI issues: which model did you use?

Opus 4.7 extra high, Opus 4.6 high

For AI issues: add Request ID with privacy disabled

request ID: cfaf775d-0957-4779-a796-0da5752dbd53

Does this stop you from using Cursor

Sometimes - I can sometimes use Cursor

[Samuel_Talenti (Sam Talenti)]

Couldn’t attach two screenshots, but here is another showing the thought regarding JSON before the second attempt fails. Additionally I’ve reproduced this at least 3 times

[mohitjain (Mohit)]

This is on our end. Thinking models like Opus 4.7/4.6 at high effort can hit a server-side timeout during extended thinking, especially with larger conversation contexts. The plan file never gets written, which causes the context loss you’re seeing.

While we work on a fix:

• Auto mode can route to models less prone to this

• Claude 4.6 Sonnet is faster and less likely to be affected

• Stop + resend if the agent stalls before the error appears

Thanks for the detailed report and request ID.

[Samuel_Talenti (Sam Talenti)]

Appreciate the confirmation Mohit, just as a follow-on note the “create attempt” → “thinking mentioning JSON” → “create attempt times out” pattern is 3/3 on my attempts. While I’m sure the timeout is the final cause of failure, I’m suspicious there’s something with writing the plan that’s causing the overall execution time to balloon into timeout territory. But I’ll leave further investigation to y’all