Most agent tutorials end at call the API and print the response. Production is different, and almost none of its problems are about what the model can do. They are infrastructure problems:
- models ignore formatting instructions unless the constraint is placed where they will actually see it;
- responses stop mid-sentence at a token limit, with no warning, unless someone checks why they stopped;
- model names drift away from the ones you think you are running;
- processes crash and restart into port conflicts;
- token costs pile up where nobody is looking.
The kit is a working answer to each: an annotated reference bridge, a systemd service with a watchdog, notes on formatting and model selection, and a checklist to run before going live. It works with any OpenAI-compatible provider — OpenRouter, Anthropic directly, or open-weight models through Ollama or vLLM.
The background is in The Model Problem Nobody Warned Me About.