on vercel

push mode, for hosts that can't run the runner


vercel can't run the runner, so it lives on another machine with a clone and pushes each change to github, which vercel builds like any push

on the runner

computah.config.mjs
import { gateway, vercel } from "@tetraslam13/computah/server";

export default {
  app: { install: "pnpm install --frozen-lockfile", build: "pnpm build" },
  identity: gateway(process.env.COMPUTAH_SECRET),
  trust: { owners: ["you@example.com"], members: "everyone" },
  release: { push: { branch: "main", host: vercel({ projectId: "prj_..." }) } }, // needs VERCEL_TOKEN
};

no start cause vercel serves the app. gateway means "believe the app's server about who's signed in when it sends the shared secret", and you run it with computah serve somewhere the app's server can reach

in the app

one route passes the box's requests to the runner and says who's signed in:

app/api/computah/[...path]/route.ts
import { createProxy } from "@tetraslam13/computah/next";

export const { GET, POST } = createProxy({
  runner: process.env.COMPUTAH_RUNNER!,
  secret: process.env.COMPUTAH_SECRET!,
  identify: async (request) => {
    const user = await yourAuth(request);
    return user && { id: user.id, name: user.email };
  },
});

yourAuth is however your app knows who's signed in (null for nobody), and the box points at the route:

tsx
<Computah endpoint="/api/computah" />

it only passes on changes from your own pages carrying the box's header, and it only uses standard Request and Response so any framework with those works

what happens to a change

owner or trustedpushed to the branch vercel serves, and the box waits for the deploy and reloads
memberpushed to their own branch, computah/u/ plus their name, and the box links vercel's preview as "open your version"
the buildchecked on the runner before pushing (check: false skips that), and if vercel's fails anyway the end of its log goes to the agent like any failed build
someone pushed meanwhileit merges their work and pushes again, once, and a conflict means the change doesn't go out
commit authorvercel refuses authors it can't match to the team, so commits use the runner machine's git email (make it someone on your team) and only the name is whoever asked
blocked deployreported in the box, not handed to the agent, cause code can't fix it

what you lose

it takes as long as vercel takes, and there's no live preview, so nobody gets to watch it happen :(

another host

it's one function, deployed({ sha, branch, production }), that waits for the build and returns where it's served or throws with the build log for the agent to fix. more on adapters

how much it's been tested

i've run the vercel adapter against a real project with a scripted agent standing in: a production deploy, a member preview and a failed build that got fixed. the app-side route (createProxy) has only run in tests so far