Claude Code — autoCompact Configuration

Control when Claude Code compresses its context window and how to reduce token waste from noisy tool output.

Why / When to Use

autoCompact fires at 95% context by default. For small model windows (e.g. 32K DeepSeek) this leaves almost no room for the response — adjust the threshold or disable entirely.

Core Concept / Commands

Adjust the threshold

Add to settings.json under the env block:

"env": {
  "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE": "70"
}

70% fires compaction at ~70% of the context window, giving the model breathing room before the limit.

Disable autoCompact entirely

claude config set -g autoCompactEnabled false

Or via environment variable:

export DISABLE_AUTO_COMPACT=true

Collapse blank lines to save tokens

Verbose tools (test runners, linters) often emit multiple blank lines. Each \n is a token. Collapse runs of 3+ newlines to 2:

import re
re.sub(r'\n{3,}', '\n\n', text)

Small per-command, but stacks across a full session.

Key Options / Variants

Variable	Default	Effect
CLAUDE_AUTOCOMPACT_PCT_OVERRIDE	95	% of context window at which compaction fires
DISABLE_AUTO_COMPACT	unset	Set to true to disable entirely
claude config set -g autoCompactEnabled false	—	Persists disable to ~/.claude.json

Gotchas

• At 95% default + 32K window → compaction fires at ~30K, leaving only ~2K for the response.
• 70% on 32K → fires at ~22K, giving ~10K headroom.
• CLAUDE_AUTOCOMPACT_PCT_OVERRIDE goes in the env block of settings.json, not as a top-level key.

Source

Conversation: “Analyzing GitHub issue resolution” — 2026-05-16