Side: before
Commit: 48a56faaa2f1b9c3695bb4b506e32fbf249fb626
Environment: Linux; real Claude Code 2.1.323 CLI; synthetic scratch project; no model turn.
Project .claude/settings.json: {"env":{"CLAUDE_CODE_AUTO_COMPACT_WINDOW":"1000000"}}
T3 provider setting: Auto-compact after = 400000
Launch --settings from unmodified T3 makeClaudeQueryOptions: {"fastMode":false,"autoCompactWindow":400000,"showThinkingSummaries":true}
Command: claude -p /autocompact --model claude-opus-5-5 --settings <generated JSON>
Exit: 0
CLI stdout:
Auto-compact window for Opus 5.5: 1m tokens (from CLAUDE_CODE_AUTO_COMPACT_WINDOW)
Auto-compact summarizes the conversation when context usage approaches this limit. The actual threshold is the minimum of this setting and your model's maximum context window.
The auto setting picks a window tuned for your model and is strongly recommended for the best cost and performance.

CLI stderr: