Add --max-context flag to cap context window usage

Problem

With the recent upgrade of Opus 4.6 to 1M context, my API quota burns ~5x faster than before. I was working comfortably at 200K and have no need for 1M in most sessions. There's currently no way to limit the context window size Claude Code uses.

Proposed Solution

Add a --max-context <tokens> CLI flag (and/or a settings option) that caps the effective context window. For example:

claude --max-context 200000

This would:

• Trigger auto-compaction at the same relative thresholds (64-75%) but relative to the cap, not the model's max
• Prevent the context from growing beyond the specified limit
• Allow users to control their token spend per session

A settings-level default would also be useful:

{
  "maxContextTokens": 200000
}

Use Case

• Quota management: Users on usage-based billing who don't need the full context window
• Consistent behavior: Researchers benchmarking across sessions want reproducible context sizes
• Cost control: Especially relevant as context windows keep growing (200K → 1M → ?)

Alternatives Considered

• Frequent /compact: Manual, error-prone, doesn't prevent the context from growing in between
• Shorter sessions: Loses continuity
• Hooks: PreToolUse hooks don't have access to current context token count, so can't build this as a workaround

Environment

• Claude Code on Windows (MSYS2/Git Bash)
• Opus 4.6 with 1M context
• Previously worked well with 200K limit