Auto-compaction triggers at ~6% context usage on Opus 4.6 (1M context)

Bug Description

Auto-compaction is triggering far too early on claude-opus-4-6[1m] (1M context window). Conversation context is being compressed when usage is well under 100K tokens — nowhere near the 950K threshold that CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=95 should enforce.

Environment

• Claude Code version: latest (auto-updates enabled)
• Model: claude-opus-4-6[1m] (1M context)
• Platform: macOS (Darwin 25.4.0, Apple Silicon)
• Terminal: iTerm2
• Settings:

``json
{
"env": {
"CLAUDE_AUTOCOMPACT_PCT_OVERRIDE": "95"
},
"model": "claude-opus-4-6[1m]"
}
``

Steps to Reproduce

1. Start a session with claude-opus-4-6[1m] model
2. Run a moderate conversation (PR reviews with 5 background agents returning results, some bash commands)
3. Observe that context compression kicks in despite low usage

Evidence

Running /context after the conversation showed:

63.8k/1m tokens (6%)
Messages: 44.1k tokens (4.4%)
Free space: 915.3k (91.5%)
Autocompact buffer: 21k tokens (2.1%)

Yet earlier messages in the conversation were already compressed/summarized, losing the full PR review output and agent results.

Expected Behavior

With CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=95 and a 1M context window, auto-compaction should not trigger until ~950K tokens are consumed.

Actual Behavior

Compaction triggered at what appears to be well under 100K tokens of actual conversation content.

Hypothesis

The autocompact threshold may be calculated against a default context window size (e.g., 200K) rather than the actual 1M window of the claude-opus-4-6[1m] model. 95% of 200K = 190K, which could explain premature triggering if agent results temporarily inflate the token count.

Workaround

Setting DISABLE_AUTO_COMPACT=1 in env and using manual /compact as needed.