Dangerous commands die at the gate

An agent that can deploy is an agent that can delete. So every autonomous step in my fleet passes a policy gate first — small, boring, non-negotiable checks that fail loudly:

1. Secrets never travel. A guard script fails CI if .env is tracked, if contact data strays into git, if a dev-default JWT is still wired. The same check runs locally and in the pipeline — one rule, two places.

2. Boot fails loudly. Services start only with real secrets present (${VAR:?set me}). No silent dev defaults in production paths. A missing secret is a 6 a.m. error message, not a 6 p.m. incident.

3. Consoles stay private. Ops dashboards sit behind a token cookie gate with a 30-day device grant — readable by me anywhere, invisible to everyone else, de-indexed from search.

The pattern is always Demo → Dry-run → Live-with-approval. Autonomy with a leash. Start with the free check.