Risk Escalation

Agents can inspect risky areas and propose a plan.

They should not make risky changes without explicit human approval.

Stop Before Touching

• auth
• security
• payments
• billing
• Stripe
• production data
• data deletion
• migrations
• secrets
• environment variables
• public API compatibility
• licensing
• telemetry/privacy behaviour
• destructive commands
• production configuration
• major dependency upgrades

Production Data Work

Require:

• dry-run mode where feasible
• idempotent upserts
• idempotency key
• validation before write
• run logs
• failure report
• rollback or recovery notes
• no deletes unless explicitly approved

Good Escalation

Risk level: high

This touches production billing and a database migration.

I can inspect and draft a plan, but I need explicit approval before editing these files:

- `db/migrations/...`
- `src/billing/...`

Recommended plan:
1. add migration in a feature branch
2. add tests for existing billing states
3. run migration locally against a copy
4. prepare rollback SQL