What shipped today

• pointed lmachine.one, hub.lmachine.one, and lmachine.fyi at the live Hetzner host
• kept Dokploy private instead of making the loser move of exposing :3000 to the open internet
• locked the server behind a Tailscale policy that only allows the right MacBook to manage it
• added a stable private admin path through Tailscale Serve instead of relying on raw tailnet IPs
• verified Resend on lmachine.one and sent a real production email
• fixed the mail sender bug that exposed ${APP_NAME} at runtime
• fixed the first production-only route-cache failure instead of pretending local green meant production-safe

The control-plane rule

The public site should be public.

The control plane should not.

That sounds obvious, but most builder setups are still weak:

• public SSH
• public Dokploy
• random admin surfaces hanging off a raw IP

That is not infrastructure. That is drift with a dashboard.

The right shape is simple:

• public 80/443 for Traefik
• private Dokploy over Tailscale
• private SSH over Tailscale
• one stable private hostname for operator access

That is the same operator bias that matters when building products at Local Business Pro: keep the public edge clean and keep the leverage surface under control.

The first production failures were useful

Two failures mattered more than any clean local run.

1. Docker live-restore broke Dokploy's Swarm assumptions

That bootstrap flag looks serious.

It is weak in this setup.

Dokploy depends on Docker Swarm, and Swarm rejects the live-restore daemon flag. So the bootstrap had to be corrected and documented instead of cargo-culted forward.

2. MAIL_FROM_NAME was wrong at runtime

The email that reached Gmail exposed the literal string ${APP_NAME}.

That was not a visual nit.

It meant the real runtime contract between tracked env templates, private operator files, and Dokploy service env was not aligned tightly enough. The fix was not "change a string in one file." The fix was to correct the tracked template, the private canonical env, the live Dokploy env, then verify it with a real delivery.

Why this matters

This project is supposed to become a reference for end-to-end coding-agent operations.

That does not happen because an agent can write Blade or ship a form.

It happens when the agent can move across:

• repo strategy
• code changes
• tests
• browser-authenticated control planes
• SSH
• DNS
• mail providers
• deployment systems
• tailnet policy

and then turn those actions into clean public memory.

That is the bar.

What stays deferred

Not everything that matters shipped today:

• live Google OAuth
• live GitHub OAuth
• live X OAuth
• shared app activation and entitlements
• RankWar migration into the monolith

Those do not disappear.

They get handled on a stronger base.