In the previous log, I intentionally stopped after setting up the infrastructure.

The VPS was secured.
Moltbot was running in Docker.
Costs were controlled.
Nothing was exposed publicly.

The next step sounded trivial:

“Just run one prompt.”

It wasn’t.

The goal this time

I wanted to verify one thing only:

That a real prompt could travel end-to-end through the system:

• CLI / TUI
• Gateway
• Agent
• Model provider
• And back again with a response

No channels.
No automation.
One sentence in, one sentence out.

Where things actually broke

Nothing failed catastrophically. Instead, I ran into a series of structural issues:

• The gateway was running in a dev profile
• The CLI initially used the main profile
• Tokens existed — but in the wrong place
• The dev agent had no access to the Anthropic API key
• Model defaults existed — but were overridden at agent level

The system was doing exactly what it was designed to do.

My mental model just wasn’t complete yet.

The key realization

In Moltbot, auth is scoped per agent workspace.

Having a provider configured globally is not enough. Each agent must explicitly have access to the credentials it needs.

Until that was fixed, prompts didn’t “fail” — they just hung, waiting for something that could never succeed.

Once the auth store was correctly wired to the dev agent, everything changed.

The moment it worked

Using the terminal UI, I sent the smallest possible test:

“Say only ‘ok’ and tell me which model you are using.”

The reply came back:

“ok — I’m using anthropic/claude-haiku-4-5.”

That single line confirmed everything:

• The gateway was reachable
• The agent was executing
• Auth was valid
• The correct model was selected
• The cost profile was exactly what I intended

End to end. No assumptions left.

Final state

After verification, I stopped the container again.

Nothing is running continuously. Nothing is exposed. Nothing is spending money.

But now I know — for real — that the system works.

What this step was really about

This wasn’t about “using AI”.

It was about understanding:

• how agent systems fail
• where state actually lives
• why defaults are rarely enough
• and why one controlled test is worth more than ten features

The infrastructure wasn’t complete until it answered once.

What’s next

The next step isn’t technical.

It’s conceptual:

• Should this agent get an identity?
• Should it talk to anyone but me?
• Should it exist at all?

Those are product questions — not infrastructure ones.

This log closes the setup phase.