startlisted

# arlo:start — set arlo up for this project Run this **once per project, while the frontier is up**, so arlo can answer when it is down. Setup harvests this project's own ground truth, resolves the model that will run the skill, and leaves an `arlo` command on the PATH. Nothing here is arlo-specific to invent — every command arlo will later hand back is harvested from *this project's* real sources. `REFERENCE.md` (bundled beside this file) is the full arlo document — the trust gradient (rungs 0–6), the complete harvest rules (§2), and the lights-out use surface (§3–§4). Read it for depth; this skill is the setup procedure that stands on it. ## Steps 1. **Resolve the model tier (REFERENCE.md §1).** The model that runs arlo is the operator's choice along a resource gradient — the agent's own model, a small local one, or none. Record the choice in `.arlo/config.json`. If a local model was chosen, provision it now with the bundled `provision.sh` (it runs on an independent path, so it works when the frontier is gone). A trivial tier (use the harness model, or none) just records the choice. 2. **Harvest the ground truth (REFERENCE.md §2).** Point a small card spec at *this project's* real sources — its scripts, Makefiles, verb-dispatched CLIs, and the `--help` of the infra commands actually run here — and generate `.arlo/cards.json`: python3 -m arlo.cards <spec.json> --root . --out .arlo/cards.json Card the **canonical surface completely** — undoc