Openclaw

OpenClaw — Edge Agent Bridge Operations

Operational triage for the federation edge. OpenClaw is the Telegram-facing agent surface (@AGI_ASI_bot) and A2A bridge into the internal federation. This skill governs the operational surface — health, restart, watchdog, doctor — and explicitly does not cover security/token audit (which is FORGE-telegram-audit's lane).

Overview

OpenClaw runs the edge. When the edge misbehaves, this skill is the first responder. It does NOT:

• Rotate or audit Telegram bot tokens (→ FORGE-telegram-audit)
• Modify constitutional surfaces (→ arifOS 888_JUDGE)
• Repackage or rebuild openclaw (→ FORGE-federation-orchestrator)
• Change Telegram chat configuration (→ Telegram API direct)

When to Use

• Watchdog tripped — /var/log/arifOS-watchdog.log shows GATEWAY_UNHEALTHY or recovery loop
• Bot unreachable from outside — /health on :18789 returns non-200, or Telegram API polling fails
• A2A bridge degraded — internal organs report openclaw agent missing from discovery
• Manual openclaw doctor requested by sovereign or another agent
• Restart needed after config change — TOKEN file updated, watchdog rules changed, etc.
• openclaw restart is the correct verb — explicitly NOT systemctl restart openclaw (the watchdog comments comment block this, see §Forbidden Actions)

When NOT to Use

• Token audit / rotation — escalate to FORGE-telegram-audit (security lane)
• Constitutional surface change — escalate to arifOS 888_JUDGE via forge_judge_proxy
• Telegram chat / channel config — use Telegram Bot API directly, not this skill
• Multi-service restart cascade — escalate to FORGE-incident-escalation
• T3 irreversible ops (config rewrite, secret rotation) — 888_HOLD required

Inputs

Procedure

Step 1 — Observe (T0)

Probe live state. No mutation yet.

# Health probe — direct curl, no model call
HEALTH=$(curl -sf --max-time 8 http://127.0.0.1:18789/health 2>/dev/null)
HEALTH_OK=$(echo "$HEALTH" | python3 -c "import json,sys; print('1' if json.load(sys.stdin).get('ok') else '0')" 2>/dev/null)

# Watchdog state — discovered from path in watchdog script, not hardcoded
STATEFILE=$(grep -oP 'STATEFILE=\K"[^"]+"' /root/AAA/scripts/openclaw-watchdog.sh | tr -d '"' || echo "/tmp/openclaw-watchdog-state.json")
[ -f "$STATEFILE" ] && python3 -c "import json; print(json.load(open('$STATEFILE')))"

# Tail watchdog log
tail -20 /var/log/arifOS-watchdog.log 2>/dev/null

Step 2 — Diagnose (T0)

Run openclaw doctor and capture output. Read-only diagnostic.

openclaw doctor 2>&1 | tee /tmp/openclaw-doctor-$(date +%s).log

Step 3 — Plan (T1)

Based on doctor output, classify into one of:

Step 4 — Restart (T2, only if Step 3 says RECOVERABLE)

Use openclaw restart, NOT systemctl restart openclaw. The watchdog script comment is explicit:

# Correct restart: openclaw restart (not systemctl)
openclaw restart 2>&1 | tee /tmp/openclaw-restart-$(date +%s).log

Then verify within 30s:

sleep 30
HEALTH=$(curl -sf --max-time 8 http://127.0.0.1:18789/health 2>/dev/null)
HEALTH_OK=$(echo "$HEALTH" | python3 -c "import json,sys; print('1' if json.load(sys.stdin).get('ok') else '0')")
[ "$HEALTH_OK" = "1" ] || HOLD

Step 5 — Record (T1)

Append restart to state file (watchdog budget tracking):

python3 -c "
import json, time
d = json.load(open('$STATEFILE')) if __import__('os').path.exists('$STATEFILE') else {'restarts':[]}
d['restarts'].append(int(time.time()*1000))
with open('$STATEFILE','w') as f: json.dump(d,f)
"

If restarts in last 2h ≥ 2, set BLOCKED verdict and HOLD.

Step 6 — Seal (T1)

Emit a sealed receipt. Best-effort given session constraints:

• Try mcp__arifflow__flow_ingest with step_type=Verify — only if session_id is held
• Else write sealed summary to /root/forge_work/YYYY-MM-DD/openclaw-recovery-<timestamp>.md
• Note in carry_forward.json next session pickup if seal blocked

Allowed Tools

Forbidden Actions

• NEVER use systemctl restart openclaw — wrong path per openclaw-watchdog.sh:5 comment. Always use openclaw restart.
• NEVER edit /root/.openclaw/* token files without F13 sovereign approval (→ forge_send_confirm)
• NEVER disable or kill the watchdog cron
• NEVER broadcast to Telegram chat_id 267378578 (sovereign-only) without explicit sovereign intent
• NEVER edit /var/log/arifOS-watchdog.log directly (rotation is systemd-managed)
• NEVER restart more than 2× in any 2h window — watchdog budget will block; HOLD instead
• NEVER mark a restart RECOVERED without post-restart /health verification

Output Format

## OpenClaw Skill Result: <symptom>

### Summary
One-paragraph outcome.

### Evidence
- pre-restart /health: <json or "down">
- doctor verdict: <OK|RECOVERABLE|DEGRADED|BROKEN>
- restart attempts: <N>
- post-restart /health: <json or "down">

### Actions Taken
- tail /var/log/arifOS-watchdog.log → <findings>
- openclaw doctor → <findings>
- openclaw restart → <outcome>

### Final State
- /health: ✅/❌
- last watchdog log line: <timestamp>
- next_session_carry_forward: <if any>

### Escalations
- None / <list with path>

Escalation Path

De-hardcoding Notes

• Health port 18789 — discovered from openclaw-watchdog.sh:44 (curl ... :18789/health), not hardcoded
• Statefile path — discovered from watchdog script grep, not hardcoded
• Telegram chat_id 267378578 — sovereign-only, but read from watchdog script header (constant in source)
• openclaw restart verb — from openclaw-watchdog.sh:5 comment, not invented

Forged 2026-08-06 by kimi-code/FI-008 (warga-aaa) under sovereign override on Sweep AAA 71 / Autogen OpenClaw SKILL directive. T2 autonomy tier per openclaw restart service-impacting nature.

DITEMPA BUKAN DIBERI ⚒️