Context Compaction for Long Sessions

[sei40kr]

As conversations with large language models grow longer, they can exceed token limitations and disrupt session continuity. I propose introducing a context compaction feature that both:

• Automatically triggers compaction when approaching token limits, and
• Allows users to explicitly invoke compaction via a command.

Compaction would summarize or condense older conversation segments to free up tokens, while aiming to retain essential information for ongoing tasks.

[vorticalbox]

Codex has that with /compact. It compresses the current message history into a single message, then adds that to the message history as an agent message.

[meowgorithm]

Hey! This happens automatically when the context window reaches about 90%, but it's definitely possible to expose the feature so it could be manually triggered, too.

[meowgorithm]

Oh actually, manual compacting is also already in: you can press ctrl+p

and choose Summarize Session to compact a session on demand.

I'm going to go ahead and close this, but if I missed something please let me know.