Articles, notes and updates on the agent, the tools, and the craft of building software with AI.
Two new commands. One lets you stop babysitting the agent. The other lets you stop leaving the agent to update it.
We spent the last few releases building extensions — LSP, SEO tools, workflow pipelines. Good features, all of them. But this release is different. Instead of adding new things, we made the tools you already use every minute of every session meaningfully better.
For twenty years, SEO meant one thing: make Google happy. Write good titles, add meta descriptions, build backlinks, and wait for the crawler. The rules were well understood, the tools were mature, and most developers had internalized the basics even if they didn't enjoy them.
This release has no flashy new commands. No new UI. Just two changes that address real friction we've been seeing in daily use: wasted money on cache misses, and wasted tokens on guessing where code lives.
We've all been there. You ask the agent to refactor a module, it gets three turns deep, and somewhere along the way things went sideways. The conversation branched into a dead end, and your files are in a state that's hard to untangle.
There's a quiet problem with CLI tools. You install them, they work, and then you forget about them. Updates ship with new features, performance improvements, and bug fixes, but you never find out because nothing tells you they exist.
There's a moment in every coding session where you want to try something different. You've been building a feature one way, but you're not sure it's the right approach. You want to explore an alternative without losing the work you've already done.
There's a familiar moment when you're working with a coding agent. You ask it to review a pull request, and it gives you back something polite and generic — "consider adding error handling here, maybe rename this variable, looks good overall." Useful in the way a horoscope is useful. You wanted a sharp opinion. You got a vibe.
There's a moment in every developer's workflow where you realize you've been solving the same problem a slightly different way each time. Code reviews follow a different checklist depending on who's doing them. Commit messages drift in shape across the team.