Claude Code fills entire disk via /tmp/claude-{uid} temp files (771 GB crash loop)

Bug Report

Claude Code version: v2.1.101
OS: macOS Sequoia (Darwin 25.4.0), Mac Mini M2 Pro 32GB
Filesystem: APFS, 926 GB volume

Problem

Claude Code wrote ~771 GB of temporary files to /private/tmp/claude-501, filling a 1TB drive to 100% (101 MB free). This caused Claude Code itself to crash with "Configuration Error: invalid JSON" because it couldn't write to ~/.claude.json, which then triggered a restart loop creating even more temp files.

Timeline

1. ~/.claude.json was corrupted (truncated mid-write, likely during an SDK session refresh)
2. Claude Code detects corrupted config → crashes
3. On restart, writes temp files to /private/tmp/claude-501
4. Crashes again → restarts → writes more temp files
5. Disk fills to 100% → config writes fail even harder → infinite loop
6. All other services on the machine were affected (tee: No space left on device, Telegram bot 409 conflicts, etc.)

Evidence from logs

Claude configuration file at /Users/mini/.claude.json is corrupted: JSON Parse error: Unable to parse JSON string
The corrupted file has been backed up to: /Users/mini/.claude/backups/.claude.json.corrupted.1775866517894
A backup file exists at: /Users/mini/.claude/backups/.claude.json.backup.1775865061735

This message repeated 50+ times in stderr, each iteration creating more temp data.

tee: /Users/mini/.openclaw/logs/preflight.log: No space left on device
tee: stdout: No space left on device

Impact

• 771 GB of temp files written to /private/tmp/claude-501
• Disk filled to 100%, all services on the machine affected
• Required manual intervention: sudo rm -rf /private/tmp/claude-501 + reboot
• ~45 minutes of downtime to diagnose and recover

Expected Behavior

1. Claude Code should not write hundreds of GB to /tmp — there should be a size limit or periodic cleanup for temp files
2. A corrupted ~/.claude.json should not cause an infinite restart loop — after N failures, Claude Code should stop retrying or recreate the config from the backup it already detects
3. The backup detection message shows Claude Code knows a backup exists but doesn't auto-restore from it

Suggested Fixes

• Add a max size cap for /tmp/claude-{uid} (e.g., 5 GB) with automatic cleanup of old temp files
• On config corruption: auto-restore from backup after first detection instead of crash-looping
• Add a circuit breaker: if Claude Code crashes N times within M minutes, stop restarting and log a clear error

Workaround

sudo rm -rf /private/tmp/claude-501
cp ~/.claude/backups/.claude.json.backup.* ~/.claude.json
# reboot to clear any stuck processes