Cursor 的 SSH 重新連線曾將待處理的「Undo Create Diff」操作重播至使用中的工作區
Cursor's SSH reconnect replayed a backlog of undo-diffs onto a live workspace
Cursor 在 2026 年 11 月 6 日表示已推出修正,避免 SSH 重新連線或重新載入後還原待處理 Agent 變更時改寫或刪除磁碟上的檔案。同一故障模式在討論串中涉及三名用戶;其中一次有 100-200 個檔案遭改寫,部分檔案沒有 Local History 可供還原。
A Cursor user working over Remote-SSH into a Linux GCP VM had their SSH connection drop and auto-reconnect mid-session. Six seconds after the client reconnected, files across two repositories on that remote started rewriting themselves.
What the source says
Per the original forum report, the reconnect caused cursor-server to replay a backlog of pending "Undo Create Diff" operations against the live workspace: 56 history writes landed within two seconds of the reconnect, followed by a mass file-modification-time rewrite about six seconds later. New gitignored files in a sandbox folder came back empty, and tracked files — in that repository and in a second, unrelated repository on the same VM — were rewritten too.
In the original report, git-tracked content still matched HEAD, and the gitignored files were eventually restored from Local History after a second reconnect synced the workspace back up — a full recovery, but a manual one. Two other users hit the identical bug in the same thread over the following weeks. One, on August 4, had 100-200 files rewritten in roughly seven seconds and reported many had no Local History snapshot to fall back on — an unrecoverable loss. A third, on August 7, recovered via an agent chat session instead.
Cursor staff (Mohit, then Colin) confirmed the behavior was unintended, requested logs, and tracked it for over three months. On November 6, 2026, Colin reported a shipped fix: restoring pending Agent changes after a reconnect or reload no longer rewrites or deletes files on disk, and a dropped connection can no longer leave behind bad checkpoint data. The original reporter confirmed no recurrence since.
What it doesn't establish
The thread documents three users hitting the same failure mode, not a count of how many were affected in total, and there's no vendor postmortem beyond staff replies in the forum — just acknowledgment, tracking, and a fix. It's also not clear from the thread exactly which checkpoint-restore code path caused the 100-200-file case specifically, versus the more contained original report.
Why it matters anyway
This wasn't one person's fluke setup. The same trigger — a dropped-and-restored SSH connection to cursor-server — produced the same effect for at least three separate users over several weeks: a backlog of pending undo state gets replayed against files that were never part of that backlog, sometimes with no local snapshot available to undo it. If you were running Cursor's Remote-SSH against a Linux VM before November 2026 and have ever had a connection drop mid-session, it's worth checking whether anything in your workspace shows the same reconnect-timed rewrite pattern Cursor fixed — the fix means new sessions should be safe, not that any damage already done was found.
Full scoring breakdown and source: https://www.stupidllm.com/incident/STUPID-2026-0173/
來源:DEV Community · dev.to