Automatic Compaction Triggers at 50% of Context Window #39798

Description

Automatic Compaction Triggers at 50% of Context Window, when a single prompt instructs model to read a few long .md markdown files, upon reaching approximately half of the context window, automatic compaction is triggered. Reproducible with same instructions in new session. This problem only seems to have appeared after recent updates.
(context window of provider model is 1m but triggers compaction at around 500k, but config buffer is set ass only 10k)
(provider is nvidia nim deepseek v4 pro routed behind a proxy)

Plugins

oh-my-openagent

OpenCode version

1.18.8

Steps to reproduce

1. Install opencode 1.18.8
2. Prepare a few long markdown files
3. Instruct model to read all of them in full without truncation +/- other instructions
4. Compaction triggers around 50% of context window usage

Screenshot and/or share link

No response

Operating System

debian

Terminal

No response