A build-tool swap that took 70 days and 166 files, and the build was the easy part

The load-bearing sentence in the whole account is that Create React App was never a bundler in the first place, it was an API ^[16]. Over five years, DreamNet's source had absorbed webpack's semantics: `process.env` reads in browser code, `%PUBLIC_URL%`, `import { ReactComponent as Icon }`, JSX living in `.js` files, implicit Node globals, and automatic Babel transpilation of every CommonJS dependency the team had ever installed ^[17]. None of those are React features and all of them were holding something up, the author writes ^[17]. So the first task was not the config file but a survey of every place the source assumed only webpack could satisfy, and that list came out roughly ten times the size of the config ^[18].

The `constants.js` episode is the mechanism in miniature. It was a data module that had quietly grown JSX inside it, nav configs and ingredient lists and session-OS maps each carrying an `icon: <AacIngredientIcon />` ^[15]. What exposed it was esbuild's refusal to parse JSX in a `.js` file, which Babel had been doing without complaint ^[13]. The fix was to split the JSX-bearing exports into `constant.jsx` and leave `constants.js` as data, a change the author says was worth making on its own merits ^[15]. Note the direction of causation. The new toolchain did not create that mess; it declined to keep concealing it.

Then there is the shape of the review. Twenty-eight or more diff revisions attracted 19 review comments ^[2], which is fewer than one comment per revision ^[22], spread across 70 days at roughly one revision every two and a half days ^[23]. That is not a merge request stuck in an argument. That is one engineer repeatedly finding the next thing the old build had been supplying for free. The 166 changed files also exceed the roughly 130 source files in `src/scene` alone ^[24], so the blast radius reached well past the screens.

Two things temper the read. This is a single first-person write-up, submitted to DEV's Summer Bug Smash powered by Sentry ^[21], and the payoff is stated as intent rather than measurement: faster cold start, near-instant HMR, better production builds ^[19]. The old feedback loop is described in units of checking Slack and, on a cold start, making coffee ^[20]. Anyone approving this work on those grounds is buying 70 days of attention against a number nobody in the account has yet published.

The transferable part is the discipline around the deployment contract. Vite defaults to `dist` and port 5173, but the GitLab CI jobs, Docker image and deploy scripts expected `build/` on port 3000, so both were pinned in config ^[7]. The author's rule: absorb the churn in configuration, not in infrastructure, because changing the bundler and the deployment contract together produces a broken pipeline you cannot attribute to either ^[8].