SysClaw Ops (Server)

Server-side operations for the cross-agent communication system. This skill covers how SysClaw processes incoming requests, manages notifications, and communicates with agents.

Required Credentials

Minimal DB privileges required:

-- agent_requests table
GRANT SELECT, UPDATE (verdict, security_assessment, resolved_at, resolved_by, escalated, escalation_reason)
  ON agent_requests TO <sysclaw_role>;

-- issues table
GRANT SELECT, UPDATE (status, resolved_at, resolved_by) ON issues TO <sysclaw_role>;

-- notifications table
GRANT SELECT, INSERT, UPDATE (read) ON notifications TO <sysclaw_role>;

Installation

This skill is designed for the SysClaw operator (an OpenClaw agent running on the server that manages infrastructure). It is not a standalone daemon.

Components

1. This SKILL.md — Instructions for the SysClaw agent on processing requests
2. OpenClaw cron job — sysclaw-notification-check (see below)
3. Telegram integration — Via OpenClaw's session delivery to the human operator

Cron Job Setup

The sysclaw-notification-check is an OpenClaw cron job, not a system binary. It runs as an isolated agent session within OpenClaw:

{
  "name": "sysclaw-notification-check",
  "schedule": { "kind": "cron", "expr": "*/15 * * * *", "tz": "Europe/Copenhagen" },
  "sessionTarget": "isolated",
  "payload": {
    "kind": "agentTurn",
    "message": "Check for new agent notifications and process requests. Use sysclaw-ops skill for workflow.",
    "timeoutSeconds": 120
  },
  "delivery": { "mode": "announce" }
}

Created via OpenClaw cron API. The agent session has access to SSH and DB tools — no separate binary needed.

Telegram Integration

Escalated requests are flagged in the database by the cron job. SysClaw checks for escalated requests during its normal session workflow and notifies the human operator on Telegram using OpenClaw's messaging. The cron job itself does not send Telegram messages.

Database Tables

agent_requests — Requests from agents

issues — Issues reported by agents

notifications — Agent ↔ SysClaw communication

worklog — SysClaw action log

Full schema reference: See references/db-schema.md for detailed column constraints, indexes, role permissions, and status flows.

Processing Workflow

Check for new notifications

SELECT * FROM notifications WHERE recipient = 'sysclaw' AND read = FALSE ORDER BY created_at ASC;

Process a pending request

Agents submit requests because they can't do it themselves. When SysClaw approves, SysClaw executes.

Note: These queries show the pattern. Use parameterized queries or your agent's DB tooling — never interpolate values directly.

-- 1. Read the request
SELECT * FROM agent_requests WHERE id = <request_id>;

-- 2. Assess and update verdict
UPDATE agent_requests
SET verdict = 'approved',
    security_assessment = 'Low risk: standard package install on managed VM',
    resolved_at = NOW(),
    resolved_by = 'sysclaw'
WHERE id = <request_id>;

-- 3. Log to worklog
INSERT INTO worklog (request_id, action, target, command, status)
VALUES (<request_id>, '<action>', '<target>', '<what SysClaw will do>', 'in_progress');

-- 4. Execute the action (SSH to target, run command, verify)

-- 5. Update worklog with result
UPDATE worklog SET status = 'completed', result = '<outcome>', completed_at = NOW() WHERE id = <worklog_id>;

-- 6. Write response notification back to agent
INSERT INTO notifications (recipient, sender, related_request, message, urgency)
VALUES ('jobhunter', 'sysclaw', <request_id>, 'Request #5 DONE: cron installed and verified on MB-OpenClaw-01', 'normal');

-- 7. Mark sysclaw's notification as read
UPDATE notifications SET read = TRUE WHERE id = <notification_id>;

Execution by Request Type

Escalate to the human operator

For urgent requests or high-risk actions:

1. Set verdict = 'escalated', escalated = TRUE, escalation_reason = '<reason>'
2. The request is now flagged for human review
3. SysClaw checks for escalated requests during its next session and notifies the human operator on Telegram
4. The human operator reviews and decides
5. SysClaw updates the verdict and notifies the requesting agent

Note: The cron job does NOT send Telegram messages directly. It flags requests as escalated. SysClaw handles Telegram notification during its normal session workflow.

4. SysClaw updates verdict and notifies the requesting agent

Resolve an issue

UPDATE issues SET status = 'resolved', resolved_at = NOW(), resolved_by = 'sysclaw' WHERE id = <issue_id>;

Decision Guidelines

Auto-approve (low risk):

• Software installs on managed VMs (standard packages)
• Info/check requests (disk, service status)
• Config changes to agent's own workspace

escalate to the human operator (high risk):

• Access requests (SSH keys, DB credentials, new users)
• Firewall or network changes
• Production service restarts
• Anything modifying shared infrastructure
• Requests marked urgency = 'urgent'

The human operator and infrastructure owner. All high-risk decisions require the human operator's approval.

Deny:

• Requests without clear justification
• Dangerous operations (rm -rf, chmod 777)
• Conflicts with security policy

Cron Job Behavior

The sysclaw-notification-check OpenClaw cron job runs every 15 minutes:

1. Checks notifications WHERE recipient = 'sysclaw' AND read = FALSE
2. For each notification, reads the related request
3. Assesses risk per Decision Guidelines above
4. Low risk → approves, executes the action, logs to worklog, notifies agent with result
5. High risk → flags as verdict='escalated' for human review
6. Marks processed notifications as read

This is not a system cron job or binary. It runs within the OpenClaw framework as an isolated agent session with access to SSH and DB tools.

SysClaw Escalation Handling

SysClaw checks for escalated requests during its heartbeat and regular sessions:

1. Query: SELECT * FROM agent_requests WHERE verdict = 'escalated' AND resolved_at IS NULL
2. For each escalated request, notify the human operator on Telegram with request details
3. Wait for human operator's decision
4. Update verdict, write security_assessment, set resolved_at/resolved_by
5. Write response notification back to the requesting agent

This is configured in SysClaw's HEARTBEAT.md to run on every heartbeat cycle.

Useful Queries

-- Pending requests
SELECT id, requesting_agent, request_type, action, target, urgency, created_at
FROM agent_requests WHERE verdict = 'pending' ORDER BY created_at;

-- Recent activity
SELECT * FROM agent_requests ORDER BY created_at DESC LIMIT 10;

-- Open issues
SELECT id, source, severity, title, status FROM issues WHERE status = 'open';

-- Unread notifications by recipient
SELECT recipient, COUNT(*) FROM notifications WHERE read = FALSE GROUP BY recipient;

Security Considerations

• Principle of least privilege: DB role should only have UPDATE on specific columns, not full table access
• Audit trail: All verdict changes are logged via resolved_at, resolved_by, and security_assessment
• Human oversight: High-risk requests always escalate to the human operator — SysClaw never auto-approposes access/infra changes
• No secrets in skill: Credentials are provided via OpenClaw environment, not stored in this skill