Critical fix: Claude Code stores model IDs without the [1m] suffix in the JSONL. All Opus/Sonnet 4.5/4.6 sessions were being treated as 200K context instead of 1M, causing premature guard pruning and inflated context percentages (e.g., 81% instead of 16%).
Before: Context: 81% of 200.0K (claude-opus-4-6) — guard fires hard prune at 150K tokens
After: Context: 16% of 1.00M (claude-opus-4-6) — correct thresholds at 750K tokens
Haiku stays at 200K. Older claude-3 models stay at 200K. Pro plan users (200K) can override with COZEMPIC_CONTEXT_WINDOW=200000.