I seeded a private repo from this one and burned a monthly Actions quota in forty-eight hours. The pipeline had not changed; the repository had. The multiplier that did it is the same thing that makes the gate fast — twelve parallel jobs, each rounded up to a full minute, green in three and a half minutes of wall clock and billed for twelve.
An agent can infer almost everything about a codebase from the codebase. What it cannot infer is the inventory — which components exist, what they are called, and which ones were deliberately refused. Here that lives on a design board committed beside the code, and the routing file points the agent at it before it writes a line.
Recruiters, leads, engineers, the people who use the thing and the person accountable for it all arrive at a portfolio through the same link, and the changelog is the one page that answers each of them without a line of code. What v1.3.0 shows, and why the release page itself is the point.
Six days before choosing an error tracker, I wrote a research doc recommending the other one. On the seam that made reversing it cheap, the free tier that changed the maths, and the adblocker that ate my first production error.
A dependency bot opened sixteen pull requests against this repo. Three merged themselves unattended. What makes that defensible is not the pipeline that said yes, but the three refusals that had to be written down first.
The autonomous orchestrator that lives in .sandcastle/ plans, implements, reviews and merges GitHub issues while I am away from the keyboard. Why it belongs in a portfolio repo, what "AFK" actually buys, and how every other tool in the monorepo lines up behind the same quality gate.
How to write Gherkin features that act as an executable contract instead of becoming a layer of indirection nobody reads.
A design-system major bump compiled cleanly, passed every unit test, and shipped a primary button at 1.06:1 contrast. On the bug class your type checker cannot see, and the test layer that caught it.
A branch protection rule blocked two merges to main in one afternoon. Both times it caught the assistant, never the autonomous agent — and the same session removed the rule for two seconds and put it back.
Why traditional snapshot testing breaks down when part of the code is written by an agent, and what strategy replaces it.