An agent built a feature from a plan. When using Agent mode to do a follow-up change the response made no sense and mentioned repos I’ve never heard of (I work by myself so I have no organization where there could be repos I don’t know about). The response includes functions not in my code, a repo name not in my code, tech that’s not in my code (database, external tools, etc.), , etc.
Steps to Reproduce
Cursor code view (not Agents window).
Use plan mode to create a plan (Composer 2.5 MAX fast). Tell it to build (Composer 2.5 MAX not fast).
After finished ask a question in Agent mode (Composer 2.5 MAX not fast).
Expected Behavior
I expect changes to my repo and the response to be about the changes we’ve made, the info in the agent chat, and the repo I’m in.
Hey, I get why you’re worried. When you see unfamiliar repos and functions in a reply, the first thought is usually a data leak. Good news, it’s not someone else’s data.
I checked your Request ID. The logs only show your account and a single session. There are no other users or other chats there, so isolation wasn’t broken.
What actually happened is what you noticed. Right before the issue, the chat context got summarized. In long chats, the model compresses early context, loses the link to your code, and starts filling in realistic but made-up repo names, files, functions, and stack traces. It looks scary, but it’s a model hallucination after summarization, not real data coming from anywhere.
What helps in practice:
Start a new chat for each task. Don’t drag one long chat across many steps, the longer the history, the higher the chance of drift after summarization.
Ground the model with clear @ mentions of files and folders, especially in follow-up questions after a lot of work.
This is a known quality issue with long context and it’s on our radar. If you catch a fresh case in a short chat with no summarization, send a new Request ID and I’ll take a separate look.
Thanks for the reply. I’ve found the summarization feature to actually work pretty well so far - I often continue after summarization and it stays on track. The only difference is I don’t usually use max mode. Not sure if it would work differently when using max mode or not, but I figured I would mention that in case it helps you guys track anything down for how to improve it.
Thanks for clarifying, that’s a useful detail. If summarization usually stays within bounds for you and the issue only shows up on MAX mode, that can really help narrow down where the model loses grounding after the context gets compressed.
If you catch another case like this on MAX, ideally in a fairly short chat so the summarization moment is visible, send over a fresh Request ID and I’ll take a separate look and add it to the notes. For now, for day to day work the advice is the same, start a new chat for each task, and use @-mentions for the files you need in follow-up questions.