Yeah, that downstream effect is a fair point. With the 1M default in place, the auto-compaction threshold is calculated based on 1M. So if a model only supports 200 to 272K, the upstream API hits its real limit before our compaction kicks in, and the chat stalls. The Auto-Mode round trip is basically the only user-side workaround right now.
It’s the same root cause as the original report. Without a way for custom OpenAI-compatible models to declare their actual context size, both the indicator and the compaction trigger are wrong. The best place to add this info is the same feature request thread: Unlock Full Context Window with Own API Keys. Calling out the auto-compaction failure helps with prioritization, since it turns this from a nice-to-have into a workflow blocker for sub-agents.