[Feature] Auto-Compact Mode Unnecessarily Slow

Environment

• Platform (select one):
• [X] Anthropic API
• [ ] AWS Bedrock
• [ ] Google Vertex AI
• [ ] Other: <!-- specify -->

Bug Description

Claude Code's auto-compact functionality slows down my workflow.

Anthropic already mentions publicly that it can use parallelized agents, yet when it comes to compacting your conversation, it happens only once you hit the limit. It's really annoying when you're just trying to get the agent to finish what it's busy with. Why does the compact have to happen right at the end? Why doesn't it just happen all the time while it's busy, and then it can immediately switch to the more compact version when it reaches the end?

Also, often times after auto-compacting it forgets the last thing that I told it and it just goes in a different direction.

Also, can the compacting not be done by a smaller model like Haiku to save my credits?