RFD 1059 details: the steps, why one script, what it does not cover

What the one step does

In order, stopping at the first failure:

  1. npm install, vitest run, npm run build. The JS side.
  2. mix deps.get, mix compile --warnings-as-errors, mix test against an ephemeral CockroachDB node the script starts and tears down itself (mix weftspun.crdb). The Elixir side.
  3. podman build (or docker build, whichever is on PATH) for both RFD 1058 images. No push, build only.

Each step must pass before the next runs. A JS test failure never reaches the Elixir suite. A compile warning fails the build the same as a test failure does, per --warnings-as-errors.

Why one script, not three CI steps

A step written only in YAML runs nowhere but GitHub’s machine. This repository’s one broken CI step proves what that costs. Nobody ran it locally, so nobody noticed it never worked. A shell script runs on a laptop and on a runner identically, through bash scripts/ci.sh, so the gap this RFD closes cannot reopen the same way.

CONTAINER_ENGINE picks podman first, falling back to docker, because this project develops against Podman (RFD 1058), while GitHub’s runners ship Docker. The Dockerfiles are engine-agnostic. Neither one assumes Podman.

Retracted for now: the GitHub workflow that called the script

.github/workflows/main.yml once installed the toolchains and called scripts/ci.sh, and nothing in it duplicated a step the script already ran. That file is deleted, on purpose.

scripts/ci.sh still exists, still runs the same one command, and a developer still runs it before committing. The browser client’s own test suite carries many pre-existing failures across several files, unrelated to any one commit, and every push turned red for that reason alone. RFD 1057 tracks restoring the workflow once that suite is fixed.

The decision this retracts is Fowler’s, not this repository’s, and it still holds. One command builds and self-tests the system. What is missing is the machine that runs it, not the command.

What this does not cover

Playwright (test:e2e) and the smoke scripts under test:anim-smoke / test:appearance-* stay outside scripts/ci.sh. They need a browser, or a running dev server this RFD does not stand up.

RFD 1058’s deploy-weftspun-quadlet.sh stays outside it too. Continuous integration is not continuous deployment, and RFD 1058’s open firewall question means that script cannot pass on this host yet regardless.