This series covered one evening: from a question about SonarQube to a shared quality gate on five repositories. Part 1 ordered the checks by speed. Part 2 described the setup. Part 3 went through what broke, and part 4 through how changes were proven.

The numbers for that evening: 40 merged pull requests across 5 repositories, 7 releases of quality-gate, three libraries and a site migrated to the same flow.

An agent did most of the typing. Below is how the agent and I split the rest of the work, and why that split made the speed safe.

Plans first, with break points

Nothing ran before I’d read a plan. For anything with more than one step, the agent proposed the plan, I corrected it, and we agreed on break points: “we wait until 1, 2 and 3 are validated.” At each break point, the agent stopped and showed what had happened before going on.

Break points cost a few minutes each. They’re also where I can catch a wrong assumption before it turns into a series of pull requests.

The decisions that stayed mine

The agent researched, compared and recommended. It didn’t decide:

  • bun or pnpm. It looked for adoption data, found none that was reliable, and laid out the context. I chose pnpm.
  • The duplication budget. The duplication gate failed this site at 0.23 %, because jscpd’s --exit-code fails on any clone and the configured 5 % threshold was never applied. Whether 5 % is an acceptable budget, and whether Markdown code samples count as code, is a judgment call. I made it.
  • No squash. After the incident in part 3, I settled it: “overall, I don’t want squash” (translated: we talked in French).
  • The license of this site.

None of these are hard problems. They depend on what I want, and the agent can’t know that.

Stop before what you can’t undo

The agent stopped and asked before every outward-facing or irreversible step: making a repository public, publishing a package, deleting anything. Agents merge into canary when the checks are green and I asked for it. They never merge into main. Every release and every production deploy that evening went through me.

Some rules exist mostly to protect other people. stable-error exported the same function twice, once named and once as a default export. The dead-code gate flagged the duplicate. Removing either export would have broken the projects that import it. It got an @alias instead. Never break users, even when a gate complains.

The agent’s mistakes, out loud

The agent made mistakes, and it had to state each one plainly:

  • It reported a run as green before it was. The check was still running. That one became the rule from part 4: only report a check you saw pass.
  • It invented a “rule” from a migration note it had misread, and presented it as a convention.
  • It wrote a ticket that used a step name as a required check. Required checks match job names, not steps, so that check would have waited forever.

Each one belongs in the instructions or the checklist, so the next agent doesn’t repeat it. Same pattern as the incidents in part 3: a symptom, a cause, and a rule stored in the repository.

What wasn’t tested

Every pull request ends with what wasn’t tested. The evening gets the same treatment.

Since then, the site’s first production release went through the new flow: release pull request, v0.1.1, changeset consumed, tag and GitHub release, canary brought back up to date, production deployed with pnpm.

Still untested:

  • The first npm release of a library through the shared workflow. It’s waiting for the next real changeset.
  • A Dependabot chore(deps) pull request getting its automatic changeset.

Takeaways

  • Review the plan before anything runs, and agree on break points.
  • Keep the decisions that depend on what you want: tools, budgets, policies, licenses.
  • The agent stops before irreversible or outward-facing steps. Releasing stays human.
  • Never break users, even to satisfy a gate.
  • Turn every agent mistake into a rule in the repository. A lecture in the chat dies with the session.
  • Say what you didn’t test. It’s the most useful line in a pull request.

The agent made the evening fast. The checks from part 1, the shared setup from part 2 and the decisions above are why I’m still happy with what it shipped.

Share this post

← All posts