Skip to content

Where knowledge lives

Labmill's know-how comes from two places: the generic skills and the site profiles. This page states the rules for which knowledge goes where.

The rules

  1. Site-specific information lives only in sites/<domain>/. Only the public sites/generic/ fallback is in this repository. Profiles for real institutions are not part of the Labmill client; they are kept and shipped apart from it. Site-specific means anything that is true of one institution or one of its clusters, and not of research computing in general:
    • Slurm configuration: partitions, queues, QOS, accounts, limits, and how they are configured.
    • Software: where it can be found, what is already installed (modules, versions, images, drivers, CUDA), and container registries.
    • Access: how to log in, bastion and login hosts, transfer hosts, 2FA, proxies, and how to get an account.
    • Storage: data storage paths and their specific configuration (quotas, purge rules, scratch areas).
    • Hardware and support: GPU and node inventories, documentation and help-desk URLs.
    • Implicit facts: a number, model name or anecdote that only one site has, even without its name. Examples are a node's core count, the site's CUDA floor, or "it converged on GPU X but not Y" where X and Y are one site's GPUs.

  2. All generic knowledge lives in the skills (skills/).

    • Slurm manuals and best practices.
    • General research computing: virtual environments, conda, Jupyter notebooks, containers, POSIX file systems, storage tiers, SSH, shell hygiene, data transfer and credentials.
    • Research computing support: how to get help and how to write a good ticket.

    A site profile holds only the site's values, and points to the generic skill for everything else. 3. No site-specific information may ever appear in the skills. That covers skills/ in full, its provenance notes included. It also applies to Labmill's runtime code and embedded data (system prompt, tool descriptions, catalogs, defaults, UI text), to the public sites/README.md, and to all documentation. These read site values from Site.yaml / Cluster.yaml at run time, or use neutral examples (example.edu, bigcluster, /data/mygroup). 4. The README and the docs stay anonymous. README.md, CLAUDE.md and everything under docs/ name no institution, host, cluster, person or path of a real site. 5. Skills hold all the logic. A Claude Code user who installs the generic skills plus their site skill gets everything Labmill knows. Every change to domain logic also goes into the skills, with the shell commands a Claude Code user would run.

How it stays enforced

  • The leak test (internal/site/leak_test.go, TestGenericSkillsHaveNoSiteFacts) checks the embedded generic skills against every term derived from the site profiles that are present, plus curated forbidden_terms for implicit facts.
  • The drift test (internal/app/skillmap.go) ties each typed tool's logic to a skill section.
  • Every feature's acceptance includes "skills updated". The reviewer checks both directions:
    • no site facts in skills, code or docs;
    • no generic teaching in a site profile.