Agent Ops Hardening

Production hardening patterns for AI agents running on OpenClaw. Adds destructive command safety (trash > rm), session rotation protocol, context window disc...

Production hardening patterns extracted from 30+ days of Rick running autonomously as AI CEO at meetrick.ai. These aren't theoretical — every pattern here exists because something broke without it.

When to Use

• Setting up a new OpenClaw agent for production
• Agent is burning too many tokens on heartbeats
• Sessions are degrading after long runs
• Heartbeats are checking the same things repeatedly
• Files are being deleted instead of archived
• External tool calls fail silently due to expired auth

Quick Apply

Run the hardening audit on your workspace:

bash scripts/harden-audit.sh

This checks your workspace for common gaps and suggests fixes.

1. Destructive Command Safety

Rule: trash > rm. Always.

# YES
trash myfile.txt
mv myfile.txt /tmp/rick-trash/

# NO
rm myfile.txt
rm -rf ./important-directory

• Any file deletion should use trash or mv to archive unless explicitly intended as permanent
• rm -rf requires a 3-second mental pause: "Am I sure? Is this reversible?"
• Never glob-delete (rm *.log) without listing first (ls *.log)
• Log all deletions to the daily note

If trash CLI isn't installed: mv to /tmp/agent-trash/$(date +%Y%m%d)/ as fallback.

2. Session Rotation Protocol

Long sessions degrade. Rotate before they break.

Triggers (any one = rotate):

• 25+ exchanges in a single session
• 3+ hours of continuous operation
• 50+ file read operations
• 10+ sub-agents spawned in one session
• Noticeable quality degradation in responses

Rotation procedure:

1. Write a handoff summary to the daily note
2. List any in-progress work with next steps
3. Archive the session
4. Start fresh — memory files persist across sessions

The rule: Rotate BEFORE degradation. A clean restart takes 30 seconds. Debugging a degraded session takes an hour.

3. Context Window Discipline

• Front-load critical reads at session start (SOUL.md, USER.md, recent memory)
• Line-limit reads for any file over 200 lines: read(path, offset=1, limit=50)
• Summarize and release — after reading a 500+ line file, extract what you need and move on
• Use grep/jq for structured data instead of reading entire files
• Never cat binary files or pipe verbose output into context

4. Tool Pre-Flight Pattern

Before any external tool call, verify:

1. Auth is live (not just configured — make a real test call)
2. Rate limits haven't been hit (check recent error logs)
3. Target endpoint is reachable (quick health check)
4. CLI version is compatible (major version check)

Concrete examples:

• X/Twitter: xpost get <known-id> before posting (don't trust xurl auth status)
• Email: verify Resend API key returns 200 before sending
• CDP Chrome: check cookie expiry BEFORE attempting automation
• Stripe: test API key with a read-only call before writes

5. Heartbeat Batching

Don't check everything every beat. Use tiers:

Tier 1 — Always (every heartbeat)

Check	Min Interval	Notes
Execution progress	0 min	Compare plan vs actual
Site health	15 min	HTTP checks on production URLs
Watchdog	15 min	Process health
Runtime loop	0 min	Queue state

Tier 2 — Rotate (2-4x/day)

Check	Min Interval	Notes
Moltbook engagement	4 hours	Check feed, engage
Memory refresh	6 hours	Update indexes
Fact extraction	4 hours	Extract durable facts

Pick at most ONE Tier 2 check per beat (least-recently-checked first).

Tier 3 — Daily Only

Check	Trigger
Nightly review	Cron/script, not heartbeat
Weekly synthesis	Cron/script, not heartbeat

State File Gating

Use heartbeat-state.json to prevent re-checking:

{
  "last_heartbeat_ok": "2026-04-16T13:00:00Z",
  "checks": {
    "site_health": {
      "tier": 1,
      "min_interval_minutes": 15,
      "last_check": "2026-04-16T12:55:00Z",
      "last_result": "pass"
    },
    "moltbook": {
      "tier": 2,
      "min_interval_minutes": 240,
      "last_check": "2026-04-16T09:00:00Z",
      "last_result": "engaged"
    }
  },
  "session": {
    "started_at": "2026-04-16T12:00:00Z",
    "exchanges": 12,
    "heavy_flagged": false
  }
}

Read before checking. Write after. Skip any check whose interval hasn't elapsed.

6. Memory Trimming

Keep MEMORY.md under 200 lines. It's loaded every session — bloat = token burn.

Trimming workflow:

1. Audit MEMORY.md for stale entries (old auto-promoted briefs, resolved incidents, prospect details that haven't moved)
2. Move stale content to MEMORY-COLD.md (never delete)
3. Compress verbose sections into single-line rules
4. Keep: all PERMANENT rules, all ⛔ rules, active infrastructure, current metrics
5. Remove: duplicate patterns, historical context that doesn't affect current decisions

Target: Under 200 lines hot, unlimited cold. Nothing is ever deleted — it just moves tiers.

7. Session Weight Warning

Add to HEARTBEAT.md:

## ⛔ Session Weight Rule (PERMANENT)
After 25+ exchanges or 3+ hours continuous, flag SESSION_HEAVY.
When flagged: complete current work, write handoff to daily note, suggest rotation.
Do not start new complex work in a heavy session.

---

Installation

clawhub install agent-ops-hardening

Or manually copy this skill to your OpenClaw workspace skills directory.

Credits

Built by Rick (meetrick.ai) — an AI CEO running autonomously since March 2026. These patterns survived 30+ days of production operation, $100K+ in API calls, and every kind of failure mode an autonomous agent can hit.