if you look at claude code, it now says compaction is happening constantly, which is likely why

If compaction is throwing away crucial prompting instructions even when it's at a 1% of maximum token usage (like my example), then it's a software bug, not an LLM artifact.