Devtools must be open source

There was a stretch of years when most engineers, even good ones, had basically nothing they'd built purely for themselves. They spent all day writing software for other people using software other people had written, tweaking configs and plugins at the margins, but rarely going further. The math never worked: writing your own tools ate time you didn't have, and coming back a year later to fix something you'd forgotten was miserable. That calculation has quietly flipped.

The shift comes from agents that can do two things at once: hack together a personal version of some piece of software, and then keep it in sync with upstream automatically, night after night, without you touching it. Once that second part is handled, the whole equation changes. Getting started is cheap and staying current is cheap too. Build the workflow into an open-source agent as a simple skill — some text instructions — and you don't even need to write the prompts yourself each time. That's exactly what got built into Shelley, an agent where typing something like "make the UI high-contrast" is enough to personalize it on the spot.

A smaller side project called meat.dev shows how far this goes. The idea behind it: after months of reviewing AI-generated code, most of what used to matter in a diff — nil-checks, error handling, import blocks — turned out to be things models now handle more reliably than humans ever did, so it made sense to build a tool that strips that noise out of diffs before a human looks at them. Wiring that into Shelley's diff view, so commits get pre-processed the moment they're created instead of waiting on a slow terminal command, took a single prompt. The only real hiccup was the agent's choice to slap an emoji on the toggle button. Trying to bolt the same background-processing behavior into something like a VS Code extension or vimdiff would have been a much uglier project, because those tools weren't built with that kind of extension point in mind.

That gap is the real story here. Big, complicated software like Vim grew config files and plugin systems because it never made sense for one person to learn the whole codebase just to make line numbers print by default — better to build a shared extension mechanism and spread the cost across everyone. Agents erase that math. For software used by one person, in narrow, well-understood conditions, a capable agent can often just add the feature directly, and "does it seem to work" replaces the review process a shared plugin system would have demanded. Which raises an obvious question for teams too: why configure a sprawling, generic task manager or CRM around your workflow when you can assemble only the parts you actually need.

None of it works, though, without the source. The same personalization trick that turns Shelley into a build-your-own agent applies just as well to other open-source agents, and could be applied, with more effort, to Codex. Claude Code is a different story — it's closed, so there's no source to hand an agent and no way to personalize it beyond whatever hooks its makers decided to expose. Old-fashioned customization still works there, as long as what you want happens to fit inside those hooks.