Compact earlier or never: Claude Code's /autocompact, autoCompactWindow, and CLAUDE_AUTOCOMPACT_PCT_OVERRIDE

In a long session you have probably noticed the conversation get summarized without warning, and a recent instruction lose its edge. The opposite need is just as common: compact earlier, before the context fills, to avoid the performance drop that comes with a full window.

In short: the size of the auto-compact window can be set in three places, /autocompact, the --autocompact flag, and CLAUDE_CODE_AUTO_COMPACT_WINDOW; the percentage of that window at which compaction fires can be lowered with CLAUDE_AUTOCOMPACT_PCT_OVERRIDE; and autoCompactEnabled: false turns it off.

The default point

Without a window set, Claude Code compacts when the conversation reaches the model's context limit, with these exceptions.

Session	Compaction boundary
Native 1M models on the Anthropic API (Sonnet 5, Fable models, Opus 4.7 and later)	before the window fills, about 967K tokens by default
Models running at 200K (Sonnet 4.6 / Opus 4.6 without extended context, Opus 4.8 / Opus 5 on Bedrock and similar)	the 200K boundary
1M models with CLAUDE_CODE_DISABLE_1M_CONTEXT=1	the 200K boundary
Cloud sessions	as the conversation approaches the limit
Model IDs Claude Code does not recognize, such as an LLM gateway alias	the window Claude Code assumes for that ID

Changing the window size

The window, meaning how full the context gets before compaction, can be set in three places.

Method	Scope	Notes
/autocompact 500k	this session and later ones	saved to user settings as autoCompactWindow; if a higher settings scope such as managed settings sets the key, the value is saved but the session keeps that scope's window
claude --autocompact 500k	one launch	overrides the saved setting without changing it, and is not preempted by managed settings
CLAUDE_CODE_AUTO_COMPACT_WINDOW=500000	while set	beats the command, the flag, and the setting; /autocompact reports the override

The command and the flag accept 100K to 1M tokens in any of these forms:

• a plain token count such as 200000
• a k or M suffix such as 500k or 1M
• a bare number from 100 to 1000 meaning thousands, so 200 is 200,000

The environment variable accepts only the plain count. Whichever you use, Claude Code caps the value at the model's context window. /autocompact auto or claude --autocompact auto returns to the model-tuned default.

{
  "autoCompactWindow": 500000
}

Compacting earlier, or not at all

CLAUDE_AUTOCOMPACT_PCT_OVERRIDE changes not the window but the percentage of it at which compaction fires. Give it 1 to 100; 50 compacts when the window is half full. It only lowers the threshold, so values above the default are ignored. It applies to the main conversation and subagents, and only in sessions that compact before the model's limit.

{
  "env": {
    "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE": "60"
  }
}

To disable auto-compact entirely, set autoCompactEnabled to false.

{
  "autoCompactEnabled": false
}

What survives a compaction and what does not is covered in Managing context in Claude Code.

Summary

• By default compaction runs when the conversation reaches the model's limit: about 967K for native 1M models, 200K for models running at 200K
• Change the window with /autocompact (saved), --autocompact (one launch), or CLAUDE_CODE_AUTO_COMPACT_WINDOW (highest priority), from 100K to 1M
• CLAUDE_AUTOCOMPACT_PCT_OVERRIDE sets the percentage of the window and only lowers it
• autoCompactEnabled: false disables auto-compact; manual /compact still works
• If disabled, compact manually with a focus, or /clear between tasks