Cluster facts live in the core skill; procedures live in recipe skills¶

lab-hpc is the core skill and holds the foundational facts: which cluster is which, the safety rules, and the per-cluster reference pages under references/. lab-jupyter and lab-containers are task-recipe skills: one repeatable procedure each, with their own trigger vocabulary. A recipe builds on lab-hpc and never repeats a cluster fact — it points at the reference.

New content is routed by what it is, not by who needs it:

Why¶

A cluster fact copied into two skills goes stale in one of them, and nothing in the gate can tell which copy is right. Lint proves a file is well-formed; it cannot know that squeue --me does not exist on Slurm 18.08. The two clusters differ sharply — arc runs modern Slurm with cost and queue tiers, ircbc runs Slurm 18.08 with no internet on its compute nodes — so a fact that drifted is not a stale sentence, it is an agent issuing a wrong command against a real cluster.

The two kinds of content also change for different reasons and on different evidence. A fact is re-verified with read-only commands on a compute node and carries a date. A procedure changes when a step or a flag changes, and is tested by an eval case. Keeping them apart means each is edited against the evidence that can actually settle it.

What it costs¶

A recipe reads as incomplete on its own: an agent has to follow a pointer to lab-hpc before it can act. That is accepted because lab-hpc's description makes it load first and its step 0 runs before anything else — the pointer is followed in practice, not in theory.

The routing table is also a judgement call at the margin, and a one-liner that keeps growing eventually wants to become its own skill. Move it when it does; do not let the table's third row become where procedures accumulate.