skills/generate-runbook/references/safety-gates.md

# Safety gates

## Gate levels

### Gate 0 — Read-only

Proceed when the target is clear and inspection does not expose secrets or private content.

### Gate 1 — Reversible change

Require a verified target, current health check, recovery path, post-change verification, and bounded scope.

### Gate 2 — Production or external change

Require explicit authorization, named operator, blast-radius assessment, stop conditions, rollback owner, communication plan, and live monitoring.

### Gate 3 — Destructive or irreversible change

Require explicit approval immediately before execution, independently verified targets, recoverable backup when possible, dry run, precise exclusions, and a second-person check when the organization supports it.

## Mandatory stops

Preserve and reuse an existing authorization when its action, target, scope and
limits still apply. A resumed task does not renew spent authority. Reapproval is
needed for changed scope, exhausted/expired/revoked grants, or explicit action-time
gates such as Gate 3. Finish authorized preparation before requesting approval;
continue independent authorized work while a gated action waits.

Stop the affected action and dependent work immediately for the conditions below.
For an active privacy/security incident or outage, prioritize containment and
preservation; otherwise continue independent authorized preparation where safe.

Stop conditions:

- unexpected access to another user’s data
- data loss or corruption
- credential or secret exposure
- production outage outside the stated tolerance
- a target that differs from the approved scope
- recovery that requires an unapproved destructive action
- verification signals that are unavailable or contradictory
- an uncertain external-action outcome that has not been reconciled before retry

Do not convert a failed verification into an improvised repair outside the
authorized recovery scope. Contain and preserve evidence; follow already
authorized recovery or obtain a new decision when the scope changes.

## Command rules

- Resolve paths and identifiers with read-only checks.
- Avoid broad globs, unresolved variables, and recursive deletion.
- Never target a home directory, repository root, filesystem root, or shared production resource ambiguously.
- Prefer reversible operations and non-interactive commands.
- Redact secret values from commands and evidence.