Context compaction triggers too frequently on claude.ai web (Opus 4.6)

Description

Context compaction on claude.ai web (not Claude Code) triggers excessively frequently, making long-running project sessions unusable.

Environment

• Platform: claude.ai web
• Model: Claude Opus 4.6
• Subscription: Max plan

Behavior

Context compaction triggers every 3-5 exchanges in conversations with pre-loaded project files (~1000 lines of markdown). Previously, the same setup worked fine throughout August-September 2026 without frequent compaction. The issue started around early October 2026.

Impact

• Long-running project sessions become unusable
• Critical context (file paths, prior decisions, ongoing analysis) gets lost after each compaction
• Users need to constantly re-explain context that was already established
• Pre-loaded reference documents get compressed away within minutes of starting the conversation

Expected Behavior

• Compaction should not trigger every 3-5 exchanges when using a 1M token context window
• Users should have control over compaction frequency or threshold
• Pre-loaded project context should be preserved longer, especially on Max plan with 1M context

Notes

• Related to #42590 (locked) and #32691
• This issue specifically affects claude.ai web, not just Claude Code
• The compact-2026-09-04 beta update appears to have changed compaction behavior significantly
• No user-facing controls exist on claude.ai web to adjust compaction settings