Compaction updates made the models dumber #25746

Description

With the latest changes done by #23870, the model compaction step has made frequent compactions much worse than before: the new prompt(s) will not carry over enough detail.

I've had a lot of new sessions begin with wrong assumptions, missing context, or just straight up not them not getting important info or preferences carried over.

The hard limit to 2k tool output also seems to cut off a lot of detail that then gets lost, and the AI gets confused about what to do next.

Please allow much better configuration of:

• the limits (i would be fine with a 16k response when compacting) -> Opus sometimes took like 4-5 mins to compact the session, but it was a good summary of everything so far
• The internal prompt (currently, the layout seems to be hardcoded).

Plugins

none

OpenCode version

1.14.30

Steps to reproduce

1. Use a model like Opus 4.6 via GitHub Copilot, so the compaction kicks in at around 128k context
2. Have it compact.

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response