Slartibartfast: hands-off coding in the wee hours
by Maxwell Joslyn. (updated
I've spent the last few nights rolling myself a "software factory", a system for using an unattended coding agent to autonomously pick and work on issues from my GitHub repositories.
The system's name is Slartibartfast, after the hapless designer of fjords in Hitchhiker's Guide to the Galaxy. For short, I call it "Bart".
The primary Bart process is a good old-fashioned program, not an AI. It runs on the same server that hosts my 3 live websites (Zingor; Unsphere; and this blog).
On startup, Bart builds a list of all open issues from my GitHub repositories which satisfy the following criteria:
- no pull requests implement that issue
- previous Bart sessions have not failed to implement that issue twice
- the issue has been tagged, by me, with the label
bot-ok
With that list in hand, Bart proceeds one issue at a time. It starts a worker: an autonomous Claude Code session on a separate server, an ephemeral dev environment which exists only for that coding session.
Bart gives the worker its instructions to work on the chosen issue, and the worker has 60 minutes to attempt to implement a fix before Bart shuts it down. Bart also hands over verification instructions, and all that jazz. If the worker is successful, it files a pull request (PR) on that issue's repo. If unsuccessful, it comments on the issue with what progress it made, and where it got stuck.
Whether the worker succeeded or not, Bart records what happened, cleans up the dev environment, chooses another issue (possibly the same one) and kicks off another worker.
I'm still sanding off rough edges, but I just finished my first manual end-to-end run of a Slartibartfast session, and it resulted in a successful PR. I am excited, and I am going to go label some issues for Barting, turn on the scheduler, and -- hopefully -- wake up to some PRs!
