Long chats and compression
Automatic context compaction, Compress history, Continue in a new chat and chat size warnings.
Models can only read a limited amount of text. CX handles long chats in three ways.
Automatic context compaction#
Settings → AI models → Default models:
- Enable automatic context compaction – summarise older messages when the chat approaches the model's context limit.
- Compaction threshold – a Percentage of the context or a Token count (e.g. 8k).
- Keep recent tool calls – how many recent tool calls keep their raw input/output (0 = compact everything).
- Compression target – size of the summary in thousands of tokens (blank = 1% of context).
- Compress Model – which model writes the summary.
Compress history (manual)#
Add to task → Compress History. Choose Summary size, Keep recent messages and optional Additional Instructions ("Specify what to preserve"). Compaction only changes the summary sent to the model; your original messages stay.
Continue in a new chat#
When a chat gets long, CX offers This chat is getting long → Continue in a new chat: a new chat starts with a short summary and your last messages. The old chat stays as it is and the new one shows "Continued from previous chat · View original chat".
Size warning#
Extremely large chats show "This conversation data is too large" with a recommendation to start a new chat, to avoid lag or data loss.