Phase 1 — Environment

Goal: a checkout that can actually run a job.

Prerequisites (confirm before anything else)

• uv installed (uv --version). Deps are pinned through uv, including the exact Harbor git rev — do not pip install around it.
• Docker running (docker info). Every task runs in a local Docker sandbox; Harbor also builds/pulls task images. No Docker → nothing runs.
• Python ≥ 3.12 (python3 --version). uv sync provisions the interpreter, but a system Python below 3.12 will bite you in ad-hoc python3 -m ... calls.

Steps

1. uv sync — installs all deps incl. the pinned Harbor rev. Re-run after any pull that touches pyproject.toml / uv.lock.
2. Ensure .env exists: cp .env.example .env if missing. Leave the values empty for now — which env-var names you need isn't known until the model is configured (Phase 2), and the names are driven by the model YAML, not a fixed list. .env.example is documentation, not a required-var checklist.
3. Ensure the dataset(s) are downloaded (Phase 0, see 00-datasets.md). A fresh checkout has an empty datasets/; the run can't proceed without the subset the job targets.

Common failure modes

• uv: command not found → install uv first (https://docs.astral.sh/uv/), then re-run.
• Cannot connect to the Docker daemon → Docker isn't running / no permission. Start Docker; on Linux confirm the user is in the docker group.
• Harbor import / version errors after a pull → stale env; re-run uv sync.
• Everything below assumes uv run ... — running bare python3 uses the wrong interpreter/env. Always uv run ./scripts/run.sh ....

Guardrail

Do not edit anything outside configs/ and .env during setup. Dataset and task files are read-only inputs, downloaded from HuggingFace (Phase 0) rather than tracked in git.