this isn't a marketing story. nothing here is dressed up — it's reconstructed from actual session transcripts and actual commit history. every date is accurate, every bug was real, and every fix (or deliberately-left-open bug) really happened. if you ask how a methodology is born, the honest answer is: usually not from a plan. it's born from a forgotten file, a deletion nobody noticed, and the nerve to ask "did i actually verify that number, or did i just count it?"
this is the story of what happened while trying to test pheronagent — a macos agent running a local, 9-billion-parameter model on my own machine, sending not a single byte to the cloud. months of sessions, crash logs noticed at odd hours, and one question i kept asking myself over and over: is this actually working, or does it just feel like it is? and how, at some point, it stopped being one project's test file and became a discipline worth sharing.
a month of silence
it didn't start with a decision to "write tests." in early may, buried inside an entirely unrelated infrastructure commit — something about an xpc daemon, some automation scripts — sat a 34-line file, tucked in like an afterthought: six simple scenarios, a handful of prompts, a few expected tool ids. no engine behind it, no plan, not even a thought about how it would ever get run. it read like someone jotting down a couple of sanity checks in a notebook while building something completely different.
