zJaaal
Back to the map

Artemis

Nov 2026 – now in progress

Mine from the first commit: I designed it and built it, then used AI to steer it and move faster.

Why it exists

At dotCMS I kept running the same steps to bring up my development environment. So I wrote a small command engine: add steps, and it runs them in order. It started Docker, opened tabs and panes in Warp, and ran the commands that set everything up, all from one command, with nothing left to do by hand. That engine is what Artemis is today.

Then agents arrived. They like the command line, but nobody keeps a record of what they run, and they go rogue, by mistake or by skipping steps. Artemis keeps that record, stops anything destructive until it is confirmed, and behaves the same way every time.

Key decisions

  • Rules built in, not written down

    Every command gets --help, one JSON result, exit codes that mean something, a confirmation before anything destructive, a dry run and a line in the audit log. The framework enforces them, so they hold even for the commands an agent writes itself.

  • State lives in code, not in the context

    A persistent store keeps what commands need between runs, so an agent doesn't have to carry it in its context.

  • Worktrees managed by commands

    Git worktrees are created, tracked and cleaned up by code, not by an agent remembering which ones exist.

See it work

art db reset --json
{"ok":false,"command":"db reset","error":"art db reset changes or deletes data","hint":"Pass --yes to confirm. Pass --dry-run first to see what it would do."}

art db reset --dry-run --json
{"ok":true,"command":"db reset","result":null,"dryRun":true,"plan":["delete .data/db.json","run 5 migrations","load the seed"]}

art db reset --yes --json
{"ok":true,"command":"db reset","result":{"migrations":5,"seeded":true}}

art history
2026-11-14T03:29:45.981Z  db reset  exit 0
2026-11-14T03:29:45.912Z  db reset  dry run
2026-11-14T03:29:45.841Z  db reset  exit 1
A destructive command refuses, shows its plan on a dry run, runs once confirmed, and every attempt lands in the history.

The hard part

Artemis started as a way to save tokens: let agents write and run scripts for complex tasks instead of working step by step. My benchmarks said no. It burns about the same tokens as doing the job with plain Bash or a CLI.

The same benchmarks showed what it is good at: letting agents run commands safely, with an audit of everything they ran, and managing worktrees. So that is what it became.

Where it is now

For now I'm its only user. It's open source on GitHub and npm, and I'm sharing it with my colleagues next.