Setup

---
name: setup
description: Install an oh-my-mlip MLIP model environment and verify energy+force on GPU. Triggers on requests to "install", "set up", "setup", or "get working" any MLIP model (MACE, SevenNet, NequIP, Allegro, ORB, UMA, etc.) via oh-my-mlip. Also triggers when a user wants to run a model for the first time and the env is not yet materialized. Also triggers on GENERIC natural-language intent to use a machine-learning interatomic potential — "set up an MLIP", "install an ML potential / machine-learned force field / foundation interatomic potential" — even when no specific model and no "oh-my-mlip" is named; in that case list the registry roster (oh_my_mlip.list_models()) and confirm the model choice (MACE is the quickstart default). Also triggers on requests to install SEVERAL models or ALL models at once — "install everything", "set up all the MLIPs", "MACE and SevenNet and ORB". Also triggers on ROSTER questions — "which MLIPs can I install", "list the available models", "what does oh-my-mlip support" — answered as a pure registry read with no install. Works from ANY working directory: the skill never assumes the current directory is the oh-my-mlip repo. Do NOT trigger on purely informational MLIP discussion (papers, theory, definitions) with no install/run/roster intent.
argument-hint: "<model … | all | all except <model …>>  e.g. MACE-MPA-0 · MACE SevenNet ORB · all · all except UMA eSEN"
---

The hub is `$OH_MY_MLIP_HOME` (default `~/.oh-my-mlip`); if neither exists,
clone it there first (`$OH_MY_MLIP_HOME/AGENTS.md §9.0`). Read the AGENTS.md
and run the scripts of that clone, never the copy inside the plugin directory.

Defer entirely to `AGENTS.md §9` (installing model environments — single,
multiple, or all), plus `AGENTS.md §5` (gated models) and `AGENTS.md §8`
(error-class recovery policy) that §9 leans on. Read them now and follow
them verbatim; this skill carries no procedure of its own.

Success oracle: `scripts/setup_verify.py <variant> --json` passes with
`degraded: false`; it computes energy and forces and confirms GPU use.
`install.sh` exit 0, or energies printed by `run_examples/single_point.py`,
are never sufficient on their own.

A first build outlasts one tool call: start `install.sh` detached and poll
it as `AGENTS.md §9.4` step 1 describes; never wait on it inside a single call.

Deterministic entry scripts, called in the order `AGENTS.md §9` drives them:
`scripts/setup_survey.py` (state + disk-budget read; `--table <model>` for a
single target), `install.sh` (build / adopt-or-heal),
`scripts/setup_guardrail.py` (stop-condition verdict),
`scripts/setup_verify.py` (compute witness + local-ledger upsert),
`scripts/setup_sweep.py` (batch driver + JSONL report, for `all` or several
targets). Use these scripts instead of re-deriving any fact they compute.

Routing notes:
- Bootstrapping the repo from an arbitrary cwd (locating vs. cloning,
  `OH_MY_MLIP_HOME`, conda/mamba check, `env.sh`) —
  `AGENTS.md §9.0`.
- A pure roster/listing question installs nothing — `AGENTS.md §9.1`.
- Target resolution for one model, several, `all`, or `all except <...>`,
  and which of those skip the approval gate — `AGENTS.md §9.2`.
- The human approval checkpoint before an `all` / `all except` build starts
  (survey → render → approve); named models are approved by being named —
  `AGENTS.md §9.2`, `§9.3`.
- The per-target install → guardrail → verify loop and the batch sweep +
  recovery pass — `AGENTS.md §9.4`.
- Arch-pinned (NequIP/Allegro) first-run compilation — `AGENTS.md §9.5`
  and `§6`.
- Gated models (`gated: true` in `models.json`): never proceed without a
  license + token; never ask for, echo, or store the token value —
  `AGENTS.md §5` / `docs/hf_token.md`.

Gated targets (`gated: true` / `license_url` in `models.json`): before
installing them, follow `AGENTS.md §5` (a presence-only token check; the user
logs in themselves; never a token in the chat).

Whole-job requests (interview, scoped plan, approval, execution, evidence)
follow the shared recipe `recipes/setup.md` — read it after the sections
above; it says which helpers exist today and which are still planned.

`/oh-my-mlip:run <model>` and `/oh-my-mlip:catbench` are the next skills once
a model is installed and verified here.