Auto-compact triggers at ~420K tokens despite Opus having 1M context window — threshold should be configurable

Problem

Claude Code's auto-compact triggers at approximately 420K tokens, even when using Claude Opus (which has a 1M token context window via the API). This means only ~42% of the purchased context window is actually usable before compaction is forced.

The bottom-right status bar shows "0% until auto-compact" at ~419K tokens — far from the 1M limit.

Expected behavior

The auto-compact threshold should either:

1. Default to a value much closer to the model's actual context window limit, or
2. Be user-configurable via settings.json (e.g., autoCompactThreshold: 800000)

Why this matters

• Users paying for Opus specifically for its large context window are not getting what they paid for
• The threshold is not documented anywhere — users discover it by hitting the wall
• There is no knob in settings.json to adjust this behavior (confirmed by reviewing the full settings schema)

Environment

• Claude Code CLI
• Model: Claude Opus (1M context window)
• Observed compact trigger: ~419K–420K tokens
• Platform: Linux

Requested change

Expose autoCompactThreshold (or equivalent) as a configurable field in settings.json, so users can set it closer to the model's actual limit.