[BUG] Autocompact buffer (232.9k) exceeds 200k context window (116.5%) on Max plan — instant compaction loop after upgrading from 2.1.74

Your diagnosis looks correct — the buffer is calculated for 1M but applied against 200k.

A few additional things to try beyond the 2.1.74 downgrade:

1. Check for stale config

The 1M context feature may have persisted a setting that 2.1.74 ignores:

# Check if there's a contextWindow or autocompact config
cat ~/.claude/settings.json | grep -i "context\|compact\|buffer"
cat ~/.claude/projects/*/settings.json | grep -i "context\|compact\|buffer" 2>/dev/null

If you see anything context-related, try removing it and restarting.

2. Try a fresh session

If settings look clean, the buffer calculation might be cached in the session state:

# Start a completely fresh session (no resume)
claude --no-resume

Then immediately run /status to see if the buffer is still miscalculated.

3. Model-specific test

Since the 1M context is tied to Opus, try starting with a different model to confirm it's model-specific:

claude --model sonnet

If Sonnet works normally on 2.1.76, it confirms the buffer calc is leaking from the 1M Opus path into sessions that should use 200k.

4. The CLAUDE_CODE_DISABLE_1M_CONTEXT=1 failure

The fact that this env var didn't fix it suggests the buffer is calculated before the flag is checked, or the flag only affects context window display but not the autocompact threshold calculation. Worth mentioning this specifically in the bug report — it narrows down where in the code the miscalculation happens.

Thanks for your reply @yurukusa but unfortunately nothing works:

1. No context, compact, or buffer settings found in any config file. The only matches are the statusLine display command and a plugin name (context7). No persisted 1M or autocompact configuration exists.

1. claude --no-resume doesn't exist (unknown option), and starting a fresh session on 2.1.76 still produces the same 232.9k/200k autocompact buffer (116.5%) immediately.

1. Same issue with Sonnet

I manage to resolve this issue, no further action required from Anthropic

Root Cause Found

After 3 days of debugging, the issue was caused by two environment variables that Claude Code itself added to my ~/.zshrc:

export CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=90
export CLAUDE_CODE_AUTO_COMPACT_WINDOW=90

These were injected by Claude Code during a session (I did not add them manually). After the v2.1.76 upgrade, these values conflict with the new 1M context window calculations, causing the autocompact buffer (232.9k) to exceed the 200k window — triggering an infinite compaction loop on every session start.

Fix

Remove both lines from ~/.zshrc (or wherever your shell config lives), quit Terminal completely, and reopen.

Recommendation to Anthropic

1. Claude Code should never write environment variables into user shell configs without explicit user consent and clear warnings
2. These undocumented env vars should be validated against the actual context window size on startup
3. If invalid values are detected, Claude Code should warn the user instead of entering an infinite loop

I install claude on Windows PC and another Macbook I have and only then I manage to figure out that this is local issue and found it eventually.

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.