Station

A Station is the runtime template. It pairs one AgentDefinition with a Kubernetes PodTemplateSpec and a few run-management knobs. It is where “the recipe” meets “the cluster”.

flowchart LR
    AD["AgentDefinition"] --> ST["Station<br/>runtime"]:::hi --> AG["Agent"] --> JB["Job"]
    classDef hi fill:#2f5fd8,color:#fff,stroke:#16224f;
    class ST hi

What it carries

• agentDefRef: the name of the AgentDefinition this Station runs.
• template: a standard Pod template. The controller wires one container (named agent) with the rendered prompt and the injected runtime; the rest of the template (base image, volumes, node selectors, resource limits) is yours.
• deadlineMinutes: wall-clock limit per run, default 30. Becomes the Job’s activeDeadlineSeconds (× 60).
• successfulRunsHistoryLimit / failedRunsHistoryLimit: how many terminal Agents to keep per phase before the controller prunes the oldest. Both default to 3.

Like an AgentDefinition, a Station has no status. It is a template, not a run.

The injected runtime

The Station’s container image does not need the agent toolchain. The controller adds an init container that copies the runtime, the agent CLI, and the supervisor into a shared emptyDir, then overrides the main container to run the supervisor. The only requirement is a glibc-based image (for example debian:bookworm-slim), because the injected runtime is glibc-linked. See Agent runtime for details.

The full field reference is in Station CRD.