If you haven't heard of gastown it is my sincere pleasure to be the one to fix that. If you're like me and you have way more ideas than time or ability to type them, gastown is an absolute game changer. I haven't felt this jazzed about programming in decades, like, I'm using antiquated slang like "jazzed" without embarrassment.
Well. Not embarrassment about that.
I awoke to a polite note from a colleague saying I had apparently pushed a bunch of random branches to the upstream Mesa repo and are we sure I wasn't hacked. No, I wasn't, those branches were definitely things I was working on, but I had been working on them locally, nothing should have even been pushed to my personal gitlab repo. I was using gastown to do that work so obviously we start looking there...
Gastown is built on beads and beads is built on git. Every atom of work within a gastown has a bead, which means updates to those beads are absolutely critical, if they don't happen then the town chugs to a halt. Gastown is also built on claude, or whatever, but claude is what I was using. Not to pick on claude, here, just using it as a generic brand name, claude is too polite to be a robust tool, sometimes. It'll lose some critical bit of context and want to stop and ask for directions. For a coding assistant that's a great way to be; for a code factory it's less awesome. It is slightly comic to read how much of gastown is just different ways of exhorting a lilypond of claudes to please do their work, please.
In gastown you wrap an upstream project in a rig, and I'm an old so I have all my git clones of everything already in ~/git, so I had set up my rigs to use those as the local storage so I didn't have to wait for things to clone again. The town mayor or one of his underlings dispatches work to ephemeral claude instances by writing the orders down in a bead, the work happens in the worker instance's git worktree instance of the rig. I would use gastown for the automagic git worktree management alone, forget the rest of the automation.
But here's where the projectile weapon starts pointing towards the positive gradient of the ol' G field. One way to try to help when claude loses context is to put important orientation information into CLAUDE.md, and helpfully, the /init command will build that for you by inspecting the current project. So I ran that at the top level of my gastown so it, gastown, wouldn't have to keep parsing "gt --help" output just to rediscover how to update a bead, hopefully. In doing so, claude lifted that directive about bead updates being seriously no-kidding mandatory up to the top level.
From that point on, claude instances would interpret that directive to apply to the project in the rig! So now, work isn't done until the work-in-progress branch is pushed. Until it is pushed to origin. Which, too bad that you cloned it from https, I'm going to discover the ssh URL to your personal mesa repo from the local repo's git config and change the URL in the rig to use ssh, because I was told to resolve the push failure or else.
So, cautionary tale, right? Maybe determinism in your tools still has value. Maybe plaintext English isn't the best idea for a configuration language. Maybe agent prompts need to be extra careful about context. Maybe careful sandbox construction would mitigate that kind of escape. Maybe an open source agent would be more trustworthy in terms of configurable tool usage since you would actually be able to see and control the boundary instead of just trusting that there is a boundary at all.
