Context compaction too aggressive on 1M context window (Opus 4.6)

When using Claude Code with Opus 4.6 (1M context window), the automatic context compaction triggers too early and is too aggressive, losing ~90% of information at critical moments.

During a session reviewing technical documents (PDF analysis, email thread retrieval via IMAP/Graph API, iterative email drafting), the compaction kicked in and stripped out detailed analysis produced earlier. When referencing earlier output to finalize an email, the details were gone — resulting in a much weaker output.

Feature request

1. Configurable compaction threshold — let users on 1M context models control when it triggers
2. Compaction aggressiveness setting — allow users to choose how much is retained vs. compressed
3. Pre-compaction prompt — this used to exist in earlier versions, let users confirm before compacting
4. Model-aware defaults — 1M context should have much higher thresholds than smaller windows

Context

With a 1M token context window, there should be plenty of room for long sessions. Yet the compaction fires at the same threshold as smaller context models, making the 1M upgrade feel wasted. Users working on complex multi-step tasks (code review + email drafting + document analysis) are especially impacted, as the compacted summary loses the detailed intermediate work product.