1M context window silently removed for Max 5x accounts in v2.1.112

Summary

Upgrading from v2.1.97 to v2.1.112 silently downgrades the context window from 1M to 200K tokens on Max 5x accounts. No changelog entry, no user-facing notification, no setting to restore it.

Evidence

Same session, same account (Max 5x), same machine:

• v2.1.92 (Apr 8): System prompt says "Opus 4.6 (with 1M context). The exact model ID is claude-opus-4-6[1m]"
• v2.1.97 (Apr 14-16): Still claude-opus-4-6[1m]
• v2.1.112 (Apr 17): Model string is claude-opus-4-6 — no [1m] tag. /context shows 200k as the window.

What we found in the source

The 1M context gate flows through YX(), which checks for the context-1m-2025-08-07 experiment flag. In v2.1.112, this flag is either no longer being set for Max 5x accounts, or the gating logic changed.

The CLAUDE_CODE_AUTO_COMPACT_WINDOW env var exists and accepts up to 1M, but the implementation does Math.min(modelWindow, configured) — so if CC thinks the model window is 200K, the env var is capped at 200K regardless of what you set.

The /autocompact slash command appears in /context output but is not recognized as a command (Unknown command: /autocompact).

There is no user-facing toggle in the claude.ai settings UI to control this.

Impact

• Auto-compaction now triggers at ~167K (200K minus 33K buffer) instead of ~967K
• Long sessions that previously ran for hours now compact every 15-20 exchanges
• Each compaction is lossy — accumulated context drift degrades session coherence
• Users paying for Max 5x are getting 1/5th the context window they had yesterday

Expected behavior

Max 5x subscribers on Opus 4.6 should retain 1M context, or at minimum have a way to opt in. The CLAUDE_CODE_AUTO_COMPACT_WINDOW env var should respect the user's setting when the model supports it.

Environment

• CC version: 2.1.112
• Plan: Max 5x
• Model: claude-opus-4-6
• OS: Linux 6.8.0-107-generic