#AI Agents

A rename touches more than headings and copy. Fourteen localStorage keys, two cookie names, and a desktop bundle ID that locks in the moment one build ships all had to move together, or not at all. Here’s how the Subpath rename got through that without losing anyone’s saved work.

A Sunday backlog on coco that should have eaten the whole day took about an hour. Not because the AI wrote my code, but because steering files and specs meant it already knew the project instead of me re-explaining it every session. Here is what is in mine, and what I left out.

Every guide blames a fancy zsh prompt when your AI coding agent cannot read the terminal. So I deleted mine, then deleted bash too, and the commands still were not running. Here is where the trail actually led.

The regex-based structural diff extractors landed and worked – then hit a ceiling regex was never going to clear. Here’s how coco’s diff pipeline actually moved onto tree-sitter: sharper structural context for commit messages, a layered regex fallback, lazy-loaded grammars – plus the bonus the same parser bought us, syntax-highlighted diffs in the Workstation.

A weekend alone. Four projects. 222 commits in a single day. An honest look at what agentic coding is doing to my attention, what wisdom still costs, and how I am trying to stay balanced at light speed.

A small filesystem tweak that turned my open source projects into a single workspace my agents can actually reason about. Nested git repos, a centralized orchestrator, and a documentation loop that finally stays in sync.

A client needed help managing a few VPSes, so I wrote a small bash script. Three rewrites and 21 releases later, it’s Strut: an open-source CLI for deploying Docker stacks anywhere there’s SSH, built agent-first for the way I actually code now.