claude-code - 💡(How to fix) Fix [Bug] Auto-compact triggers far below 1M window on Opus 4.7 [1M] — severe token burn regression on Max 5x (post-2.1.116) [2 comments, 2 participants]

Summary

On claude-opus-4-7[1m], auto-compact now triggers far below the 1M context window despite 1M being selected and shown in /model. The behavior changed ~2 days ago — prior to that, compact correctly triggered near the actual ~1M limit. No user-side configuration changes.

In addition to the premature compact, token consumption per turn has increased substantially: 5-hour Max plan windows are now exhausted within a fraction of previously-normal usage on identical coding workflows.

Environment

• Claude Code version (current): 2.1.119
• Onset: ~2 days ago (cannot pinpoint the exact version where the regression appeared, possibly 2.1.118 or 2.1.119, or a server-side flag rollout — see below)
• Model: claude-opus-4-7[1m] (1M context, confirmed in /model)
• Plan: Max 5x
• Settings (unchanged for months):
  • alwaysThinkingEnabled: true
  • effortLevel: "high"
  • ultrathink used in-session for hard tasks
• OS: Linux (Ubuntu)
• Install method: native (single-binary install)

Behavior

Before (~2+ days ago) — correct

• Auto-compact triggered close to 1M tokens (near the actual context limit)
• Token consumption per turn was reasonable
• 5-hour rate-limit window comfortably covered a normal coding day
• Long coding sessions sustainable end-to-end

Now — regression

• Auto-compact triggers much earlier than before — well below the 1M window. The exact percentage varies between sessions, but consistently far below where it used to fire.
• Severe token burn: 5-hour limits exhausted in a small fraction of normal time on identical workflows
• Constant compacts invalidate cache, re-inject CLAUDE.md / memory, slow iteration significantly
• Sessions are practically unusable

Ruled out

• CLAUDE.md size: 667 bytes (negligible)
• Memory files: ~12K total (negligible)
• MCP servers: 0 installed
• Plugins: 0 installed
• No settings change in months — these settings worked fine before

Note on root cause

Because the regression appeared abruptly without a user-side update event, this could be either:

• a CLI version regression (auto-update happened in the same window), or
• a server-side / GrowthBook flag rollout that silently changed the resolved compact threshold (see #46331 for prior precedent)

A downgrade to 2.1.116 has been performed as a workaround; behavior on that version will be reported back if the issue persists or resolves.

Workaround

Manual downgrade to 2.1.116 + auto-updater off:

curl -fL -o ~/.local/share/claude/versions/2.1.116 \
  https://downloads.claude.ai/claude-code-releases/2.1.116/linux-x64/claude
chmod +x ~/.local/share/claude/versions/2.1.116
ln -sfn ~/.local/share/claude/versions/2.1.116 ~/.local/bin/claude
echo 'export DISABLE_AUTOUPDATER=1' >> ~/.bashrc

Likely related issues

• #36014 — autocompact triggers at 17% on 1M context model (effectiveWindow likely capped at 200K) — matches this symptom
• #52981 — auto-compact triggers at ~8% context usage (84K/1M)
• #46331 — GrowthBook experiment tengu_amber_redwood silently reduces autocompact window
• #50803 — 1M context window not auto-applied on Max plan; --model flag drops [1m] suffix
• #52522 — adjacent: documents 2.1.117 threshold-detection change with subscriber-impact concerns
• #52153 — excessive token consumption per prompt on Opus 4.7 1M

Ask

1. Confirm whether the 1M window detection has regressed for Opus 4.7 on Max 5x accounts (or any plan) within the last few days, either via CLI release or server-side flag
2. Check whether a GrowthBook flag (e.g. tengu_amber_redwood or similar) is targeting accounts and overriding the resolved compact threshold below the model's actual window
3. Either fix the detection or expose the resolved compact threshold in /context output (currently it shows the model window but not the actual percent at which compact will fire), so users can verify what the CLI is actually using
4. Given the subscriber cost impact (5x plan exhausted on identical workflows), surface defaults/threshold changes in release notes with explicit "Impact on subscribers" callouts going forward