Cascade opened 14 files then undid my only good change
had Cascade refactor a Next.js route handler in Windsurf last night. it opened like 14 related files, rewrote half my zod schema, then hit undo on the one edit that actually fixed the 500.
the diff looked confident. the app still returned ZodError: Required at "userId" after.
anyone else seeing Cascade "fix" something then walk it back without saying so?
5 comments
Join the discussion
Log in to comment.
same class of bug with Warp for me. agent rewrites, silent undo, CI still red.
I put a hard timeout + require confirm on multi-file edits now. painful but cheaper than hunting ghost diffs at 1am.
yeah the confirm gate is the only thing that saved me after a Cascade pass ate our GraphQL input types.
i also cap it at 5 files before it has to spit a plan. anything over that and Zod/Prisma tend to drift together and the 500s get weird. still red CI sometimes, but at least i know which edit landed.
did the undo show up in the Cascade timeline or only in the editor history?
I started logging every agent edit to a fails.md with the file list — noticing undos that never get a chat message. curious if Windsurf surfaces that at all.
only in editor history for me. Cascade chat acted like the fix still existed.
my fails.md line for that night: "undid good Zod userId patch; no agent message". Windsurf did not surface it. if yours logged anything in the timeline i want to know — mine was silent.
the 14-file open is the part that gets me. once the editor tabs look like a moodboard, i already lost the plot of what "good" meant.
had Cascade do a similar schema pass last week — fixed the empty state, then silently reverted the spacing tokens i actually cared about. no chat line for the undo. just… gone.
do you keep a pre-agent snapshot of the files it touches, or is that too slow?