A Real, Local Operator. Lights-out ops.
When your coding agent is down or rate-limited and the app breaks at 3am, you type your own words and arlo hands back the exact command to run — grounded in your system's own ground truth. It runs on an independent local path and never invents a command that isn't real.
The product is a skill your agent invokes — SKILL.md — not a command line.
You install it, set it up once while the frontier is up, then just talk to your agent:
/plugin marketplace add 3dl-dev/arlo
/plugin install arlo@arlo
/arlo:start # arlo learns this project's real commands
Then ask ("how do I restart the deriver?") and arlo answers with a real command, or says
"no confident match" rather than invent one. Lights-out, at a shell: arlo <your words>.
Full install: docs/INSTALL.md.
The arlo/ package is a small neutral core the skill calls — a subordinate library,
not the product. It exists to hold arlo's one invariant by construction (a command it
returns is always real ground truth); see arlo/README.md for the rule
on what earns code here versus what stays skill prose. Standard library only.
- Design, trust gradient, lineage:
docs/design.md - Decisions made and still open:
docs/decisions.md - How to think while building arlo (read the altitude invariant first):
CLAUDE.md
arlo is graded by running the skill, not by running its code: an agent executes
SKILL.md in a project it hasn't seen and you read what it does. There is deliberately no
grader, harness, or eval loop (that apparatus was deleted — we ship a skill, not code). What
a real session reveals is distilled into a generalized, project-agnostic expectation in
cases/, the growing record of what the skill must do. The hermetic checks in
tests/ grade only the core's mechanics (the structural guarantees); a green test suite is
not a graded product.
LGPL-2.1-or-later. Full text in COPYING. You may add arlo to any project,
proprietary included, and link it freely; changes to arlo itself stay open. Copyright
(C) 2026 the arlo authors.