2026 OpenClaw in Practice: Tiered Model Cache & Cleanup Windows

Start with a service account (not a shared interactive login) so permissions and LaunchAgents stay stable across SSH sessions. Map two physical ideas even if you only have one disk today: a hot APFS volume or fast folder on internal NVMe for latency-sensitive shards, and a warm tier on external USB4/TB or a second volume for bulk archives. Document every environment variable your OpenClaw build reads—common patterns include a home-level config dir plus vendor caches for Hugging Face, Ollama, or llama.cpp; align names with your installation troubleshooting guide so teammates reproduce the same tree.

1. Pin Node 22+ and OpenClaw semver; record them in an internal runbook.
2. Create ~/Library/Logs/OpenClawCache/ with mode 755 for the service user.
3. Export HF_HOME, OLLAMA_MODELS, or equivalent in launchd EnvironmentVariables—never commit secrets.
4. Verify diskutil apfs list and mount paths survive reboot (avoid transient /Volumes/Untitled names).

Pick one canonical path your jobs always use (for example ~/.openclaw/cache/models). On disk, store blobs under tier-specific roots such as /Volumes/FastSSD/oc-models and /Volumes/BulkDisk/oc-models-archive. Replace the canonical directory with a symlink to the hot tier so CI scripts never hard-code mount letters.

# Example: canonical path points at fast tier (stop OpenClaw jobs first)
mkdir -p "/Volumes/FastSSD/oc-models"
rm -rf "$HOME/.openclaw/cache/models"   # backup if it was a real directory
ln -s "/Volumes/FastSSD/oc-models" "$HOME/.openclaw/cache/models"
readlink -f "$HOME/.openclaw/cache/models"

Promote models from warm to hot with rsync -a --partial --inplace or atomic mv within the same volume. Demote cold weights with the reverse path after recording checksums.

Never delete blindly while OpenClaw holds mmap handles. Prefer a cleanup window when CI is quiet—typical remote Mac pools use 02:00–04:00 local time. Use a user LaunchAgent with StartCalendarInterval (hour + minute) instead of cron on macOS unless you already standardize on /usr/libexec/atrun-style jobs; LaunchAgents inherit the graphical session user and are easier to audit per tenant.

For log rotation, append dated files under ~/Library/Logs/OpenClawCache/ and add a /etc/newsyslog.d/ stanza or a weekly gzip pass in the same plist. Keep at least seven days of cleanup stdout to explain sudden cache misses.

# Sketch: plist ProgramArguments call this after idle check
#!/bin/bash
set -euo pipefail
# if pgrep -f openclaw-gateway >/dev/null; then exit 0; fi  # optional guard
find "$HOME/.openclaw/cache/tmp" -type f -mtime +3 -delete
find "$HOME/.openclaw/cache/tmp" -type d -empty -delete

Pair quota ideas with simple df parsing in your patrol script. A practical 2026 matrix for shared Apple Silicon runners:

If your host supports per-volume quotas, set soft limits slightly below marketing capacity so APFS snapshots do not surprise you. Always check df -h and df -i; inode exhaustion mimics full disks but will not be fixed by deleting a single 200 GB file.

Wrap downloads with capped exponential backoff (for example 5s, 15s, 45s) and jitter so a fleet of Macs does not stampede a mirror. When a job fails:

• ENOSPC / “No space left” — confirm symlink target volume, not only $HOME.
• Stale file handle — remount external disk; verify APFS encryption unlock order at boot.
• HTTP 429 / 5xx — treat as registry policy; lower concurrency and respect Retry-After.
• Permission denied — compare umask vs service user; re-run openclaw doctor after macOS updates.

Cross-check gateway stability with failure recovery and retry patterns so automation does not fight a flapping daemon.

• Canonical symlink resolves on a cold boot before any job runs.
• Patrol script prints human-readable utilization and exits non-zero on >90% fast tier.
• Cleanup dry-run log archived with ticket ID; destructive mode only after two green dry-runs.
• CI references the canonical path only—no duplicate hard-coded mount URLs.
• Rollback doc explains how to swap symlink back to warm tier within five minutes.