[BUG] Server-side regression: tengu_hawthorn_window reset to 200K for Max 20x accounts — 1M context lost

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

The tengu_hawthorn_window feature flag was reset from 1,000,000 to 200,000 server-side approximately 1 hour ago (around 18:30-19:00 UTC, April 13, 2026). This is a new regression — not the same as #47019 (Pro→Max upgrade caching) or #44403 (earlier regression).

Key evidence this is a new, separate issue:

• 3 separate Max 20x ($200/mo) accounts all affected simultaneously
• All accounts have been on Max 20x for weeks and were working at 1M until today
• No local version change triggered it — binary was locked with chflags uchg at v2.1.87, confirmed no auto-update occurred
• Logging out and back in does not fix it — server continues to push 200000
• Upgrading to v2.1.104 does not fix it

Diagnostics

{
  "tengu_hawthorn_window": 200000,
  "hasExtraUsageEnabled": false,
  "cachedExtraUsageDisabledReason": "org_level_disabled",
  "crystal_beam": { "budgetTokens": 0 }
}

What I've Tried

• [x] Updated from v2.1.87 → v2.1.104
• [x] Logged out and back in (/login)
• [x] Switched between 3 separate Max 20x accounts — all show 200K
• [x] Set CLAUDE_CODE_AUTO_COMPACT_WINDOW=1000000 env var (doesn't override API-level context limit)

Impact

All 3 accounts lost 1M context simultaneously, causing premature autocompaction at 200K. This points to a server-side feature flag deployment that incorrectly set tengu_hawthorn_window to 200000 for Max plan users.

Environment

• Claude Code v2.1.104 (also reproduced on v2.1.87)
• macOS (darwin, arm64)
• Native install
• Max 20x plan ($200/mo) × 3 accounts