human-escalation

name	human-escalation
description	Determines when the agent should stop investigating and escalate to a human SRE. Used when confidence is low, domain is unknown, or destructive actions are needed.

Human Escalation Skill

When to Escalate

1. Unknown domain: Alert doesn't match any known skill
2. Database operations: Agent has no safe DB tools
3. Security incidents: Potential breach, unauthorized access
4. Data corruption: Risk of data loss
5. Multi-region failures: Blast radius too large
6. Confidence < 0.5 after 3 loops: Agent is going in circles
7. Contradictory evidence: Findings from analysts conflict

How to Escalate

1. Write current findings to plans/{id}/report.md with status: ESCALATED
2. Include what was investigated and what remains unclear
3. Send Slack message with noop_require_human tool
4. Do NOT attempt remediation

Escalation Message Format

Include:

• What was investigated (tools run, findings)
• What the top hypothesis is (even if uncertain)
• What specific information a human should look at
• Suggested next steps for the human