compliance/missing-human-gate
Severity: warning Category: compliance Fixable: no
What it catches
A binding rule that permits a high-impact action — deploy, release, send, charge, revoke — without any human-in-the-loop cue ("approval", "review", "sign-off", "oncall", "ask").
Why it matters
Destructive-but-necessary actions are the riskiest part of any agent spec. Even if the agent behaves correctly 99% of the time, the 1% that doesn't will cost you a customer, a payment, or a database. An explicit human gate in the spec gives you a contractual checkpoint: the agent must stop, not just should.
Related: tools/destructive-no-confirm catches explicit tool references; this rule catches the broader class of sensitive verbs in directive prose.
Example — bad
- You must deploy the release candidate every Monday.
Example — good
- You must deploy the release candidate every Monday, after getting approval from the oncall engineer.