1M context window silently removed in CC v2.1.112 — affects all cache-fix users

[cnighswonger]

CC v2.1.112 silently downgrades the context window from 1M to 200K tokens. We confirmed this on a Max 5x account that had 1M on v2.1.97 — same session, same machine, no user action.

What happened

The [1m] tag that CC appends to the model string (e.g. claude-opus-4-6[1m]) is gone in v2.1.112. This tag is controlled by a server-side experiment flag (context-1m-2025-08-07) gated through a function called YX() in the CC source. Either the flag was removed from accounts, or the gating logic changed.

The CLAUDE_CODE_AUTO_COMPACT_WINDOW env var exists but cannot override this — the code does Math.min(modelWindow, configured), so if CC thinks the window is 200K, you are capped at 200K.

There is no toggle in the claude.ai settings UI. The /autocompact slash command shows in /context output but returns Unknown command.

Why this matters for cache-fix users

At 200K, auto-compaction triggers at ~167K tokens. Long sessions that previously ran for hours now compact every 15-20 exchanges. Each compaction cycle:

1. Is lossy — context coherence degrades over multiple cycles
2. Can bust the prompt cache prefix (compaction rewrites the conversation summary)
3. Negates much of the value of sticky markers and historical cache preservation

The interceptor's cache-stability fixes are most valuable in long sessions with large context. Cutting the window to 1/5th undermines the use case.

Filed upstream

anthropics/claude-code#50083

If you are on Max 5x (or any plan that previously had 1M), check your /context output after updating to v2.1.112. If you see 200k where you used to see 1M, you are affected.

cc @fgrosswig @deafsquad @ArkNill @wadabum @thepiper18 @JEONG-JIWOO

[cnighswonger]

Workaround (v2.1.112 only)

We found a bypass in the CC source that forces the 1M window regardless of the server-side experiment flag:

export DISABLE_COMPACT=1
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=1000000

Set these before launching CC. Confirmed working — /context shows 120.6k/1m tokens (12%).

Tradeoffs

• Auto-compaction is completely disabled. You must manually /compact or you will hit the context wall and the request will fail.
• We are running a cron monitor that checks context usage every 7 minutes and warns at 70% / 85% of 1M.
• CLAUDE_CODE_MAX_CONTEXT_TOKENS only exists in v2.1.112+. It does not exist in v2.1.97 or earlier.
• CLAUDE_CODE_AUTO_COMPACT_WINDOW does NOT work as an alternative — the code caps it with Math.min(modelWindow, configured), so if the server flag says 200K, you are stuck at 200K.

Context from upstream duplicates

The bot flagged three prior reports of the same regression:

• [BUG] 1M context window regressed to 200K on Max plan (~March 26-27) anthropics/claude-code#41082 — March 26-27 regression on Max plan
• [Bug] Regression: Claude Opus context window downgraded from 1M to 200K on Max plan anthropics/claude-code#44403 — Max 20x, mentions GrowthBook flag tengu_hawthorn_window returning 200K server-side
• [BUG] Server-side regression: tengu_hawthorn_window reset to 200K for Max 20x accounts — 1M context lost anthropics/claude-code#47549 — April 13, tengu_hawthorn_window reset again

This is a recurring pattern — Anthropic keeps toggling the flag. None of those issues have an official response yet.