Start

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 — undocumented Makefile targets, every real verb of a dispatcher, --help subcommands even under a nonzero exit — per §2's rules. Cards are generated, never authored, so they cannot drift from the tool. Stamp the cache with the git SHA / mtime it was harvested at.

3. Write the lights-out fallback. Generate .arlo/LIGHTS-OUT.md (a dated last-resort snapshot) and any project runbook from the harvested cards.

4. Wire the command. Put arlo on the PATH (the bundled bin/arlo launcher resolves arlo's home and walks up from any cwd to find .arlo/). After this, arlo <your words> answers from a shell with no agent and no frontier.

5. Verify against real ground truth. Run two or three real intents through arlo <words> and confirm each answer is a real command from the harvested surface, labeled with its rung, and that a nonsense intent yields "no confident match" rather than a guess.

After setup

Cards are a regenerated projection of live ground truth, never stored truth — the ground shifts (a flag renamed, a script added, a verb removed) and arlo itself improves. So this is not one-and-done: run /arlo:update periodically and after the project's commands change, to re-harvest and surface drift. Uninstall cleanly with /arlo:remove.