Chat
When summarizing the chat, give the user’s prompt as it is, not compressed.
Example: if you give GPT-5.4-High a task to work according to a certain algorithm through subagents, then if it comes to context summarization, after that it will work independently, and not through subagents.
You can turn on the MAX mode and use it with a much larger context window, but…
Hi @Artemonim , thanks for the request! I apologize, I wasn’t able to replicate your exact issue - even after chat compaction, the model still delegated to subagents as expected. Could you please make sure your Cursor is up to date, as a first troubleshooting step?
Additionally, it also sounds like you’re describing a feature request for an unlimited token context window. Summarization works to compact the chat context - it inherently cannot keep the exact prompt as it is written due to constraints of the context window. However, it should continue to delegate to subagents even after compaction.
If you’d like, I can take a closer look of your requests if you disable privacy mode and replicate the symptoms you’re describing.
Max mode is also a good alternative as it will take longer for the context window to fill up and trigger compaction.
The chat is divided into the user prompt, reasoning, output, tool calls, MCP calls, and subagent calls.
My request is that the “user prompts” block should remain unchanged after chat summarization, but everything else should be compressed by summarization.
I see. Thanks for clarifying. I don’t believe this is possible now, but definitely interesting to explore and think about.