How-to guides

Task-oriented recipes for getting a specific job done in ConfPlanner. Each guide assumes you already know what you want and shows the steps to achieve it. If you are starting from zero, work the tutorials first; to look up a flag, role name, or annotation, jump to the reference; for the why behind a design, see the explanation.

Run and operate

• Run ConfPlanner against a real model — set a provider key in .env, start the shell, and invoke a goal.
• Run a deterministic demo with mock mode — the mock Spring profile for key-free, token-free, repeatable runs.
• Enable Zipkin tracing — bring up the collector, run with the observability profiles, and read the trace at port 9411 (plus the no-Docker fallback).
• Run the eval lane — the @Tag("eval") lane: sample schedule quality with an LLM-as-judge (./mvnw -Peval test), separate from the deterministic keyless gate.

Change the agent

• Add a new typed action — drop an @Action into the pipeline and let GOAP route through it (the Lab 2 shape).
• Add a guardrail — an invariant the goal pre-requires, a precondition, and a budget (the Lab 3 shape).
• Route models by role — withLlmByRole plus application.yml, and keep MODEL_ROUTING.md in step.
• Route a step to a local model — keep a step's data in the building: bind a role to a local (Ollama) model under a Spring profile, key-free (the Lab 6 regulated-environment lever).

Drive it through the harness

• Drive a change with the build-time harness — Track C: orchestrator → java-implementer → two-stage review.