Skip to content
Sign up

flagon-io/detect

public
0

Tuesday, October 13, 2026

  1. A deployment is a site or a Worker whenever it can be, and a container only as the fallback. With Build as on Auto, a Wrangler config or a framework's Cloudflare adapter is a Workers project; a framework's site, a build script's or an index.html is that site even beside a Dockerfile, and the log says "Built as a Vite site; Dockerfile is not used (a container is only the fallback; choose **Build as**: Container to use it)", where a Dockerfile that ran a server used to win; a Node.js server with no Dockerfile (Express, Koa, Hono, Polka, Restify, hapi, node:http) runs as a Worker; any other server g1t recognises, a Node.js server with its own Dockerfile, a framework set up to need a server, or a Dockerfile with nothing else to serve is a container.

    A Node.js server runs as a Worker through httpServerHandler from cloudflare:node with nodejs_compat (developers.cloudflare.com/workers/runtime-apis/nodejs/http: from the compatibility date 2025-09-01 nodejs_compat turns on the HTTP server modules, and g1t's date is 2026-09-26). deploy/node_worker.rs finds the file that starts it (what its start script runs with node, tsx or ts-node, else main, else server.js and the like), and the build installs and builds it, writes a prelude that keeps every server the app creates or listens on and lets listen take any port and address, and an entry that imports the app and answers with httpServerHandler, or with the app's default export's fetch when it has one (a Hono app, whose Response @hono/node-server replaces is turned back into the runtime's). wrangler deploy --dry-run bundles it, and wrangler dev starts it once with the project's variables and asks it for /: a module that cannot load, a top-level await that never settles, or a server that never listens fails there, not in production. A folder served with app.use(express.static("public")) is uploaded as the Worker's static files. What cannot run as a Worker is found before anything is built, and the app is a container from the Dockerfile g1t writes, the log saying "Could not run as a Worker: <reason>; deploying as a container.": a native addon (better-sqlite3, bcrypt, sharp and others), Fastify, NestJS, Sails, remix-serve or react-router-serve, a WebSocket server (socket.io, express-ws, WebSocketServer, an upgrade handler), child_process, cluster or worker_threads, files read or written while it runs, files served from disk under a path, or a start script g1t cannot read. A Worker that does not bundle or start falls back the same way. Build as Workers runs a Node.js server as one whatever is found, and Build as Container (shown as Container now, still `docker` as stored) runs its container. g1t-runner detect and g1t/detect print entry for such a Worker.

    syntaqx committed

    18c93d9

Monday, October 12, 2026

  1. g1t.page runs Docker deployments, behind a switch that is off until Cloudflare's container APIs are checked against g1t's account (`CONTAINERS` in the deployments service). Each container app is a Worker of g1t's own in the same Workers for Platforms namespace, so the dispatcher, custom domains, previews and take-downs reach it by name as any app: its `G1tApp` Durable Object holds a Container application (default scheduling policy, one instance, the project's **Container size**, `lite` unless chosen), starts it on the first request with the project's secrets and variables and `PORT` as its environment, waits up to 90 seconds for it to answer on its port, proxies to it, and stops it five minutes after the last request; it charges each minute awake ahead and the bytes it sends, and reports them with a token of its own to `POST /deployments/apps/<app>/meter`, which the API forwards. A build's new `registry` step gets a registry login g1t mints push-only for 15 minutes, with its app's own repository named; g1t runs only an image by digest in that repository, checked with the registry, and never hands a build a login that can pull. Container time is metered at Cloudflare's Containers price plus the margin through new price book meters (vCPU charged as busy while awake, and egress at the most a GB costs anywhere, so g1t never sells below cost), noted for the spend limit as it runs and charged with the month's Deployments line; a paused workspace's container apps stop and start no more. Settings, Deployments gets Build as (Auto, Static site, Workers, Docker), Dockerfile and Container size with each size's price an hour, and says what the last build found with what it was built with ("Static site · Astro", "Container · Dockerfile"), now that the finish report's `framework` is stored. Deploy sandboxes get Docker, `BUILD_AS` and `DOCKERFILE`, and reach `registry.cloudflare.com`; `start_deployment_build` answers `build_as` and `dockerfile` for `g1t/deploy`, which takes them as inputs, as `g1t/detect` does. The Deployments guide has Docker deployments, their prices and how they work with the evidence for it, the limitations page says what is left, and Deploy g1t to Cloudflare lists each live check to make before switching it on.

    Claudeagent and Claudeagent committed

    a6a6f43
  2. g1t.page builds know the framework a project is built with, from its package's dependencies and config files: Vite (and React, Vue, Svelte, Solid or Preact on it), Astro, Next.js with `output: 'export'`, SvelteKit with the static adapter, Nuxt by `nuxi generate`, React Router and Remix in SPA mode, Gatsby, Docusaurus, Eleventy, Angular, Create React App, Hugo, Jekyll and mdBook are each built with their own command and served from the folder they write, read from their config where it moves it. A Cloudflare adapter's Worker deploys as a Workers project with no Wrangler config of the project's own: Astro's, which now writes its config for wrangler in `.wrangler/deploy/config.json` (as `@cloudflare/vite-plugin` does, and that is read too), SvelteKit's `.svelte-kit/cloudflare/_worker.js`, and OpenNext's with a `wrangler.jsonc`; a site's `_worker.js`, `_routes.json` and what `.assetsignore` lists are never uploaded, and a written config's empty binding lists are no longer warned about. A framework set up to need a server (Next.js without a static export, Astro's server output, SvelteKit's auto or node adapter, a server-rendered React Router or Remix app) fails before anything is installed, saying what to change, in place of "found nothing to serve". The project's Build command and Output directory still win. The build's log names the framework, the finish report carries it as `framework` (which the deployments service ignores for now), and `g1t-runner detect` and the `g1t/detect` action output it. What deploys has a table of every framework, and the detect action's README lists them.

    Claudeagent and Claudeagent committed

    a8686a3
  3. g1t has actions of its own: g1t/deploy builds the checkout a job is in and puts it on g1t.page, opening the build with the job's token, masking the build's token and the project's secrets, and running g1t-runner deploy, with the app's url, the deployment's id and its kind as outputs; g1t/detect says what such a build would find (kind, install, build and output-dir) without building anything. Both use the g1t-runner on the machine, or download the newest release from g1t.sh/downloads/runner and check it against its SHA256SUMS. They live under actions/, and the Actions release workflow pushes each folder to its own repository in the g1t workspace and moves its v1 tag whenever main changes them, with the G1T_ACTIONS_PUBLISH_TOKEN secret its header describes.

    Claudeagent and Claudeagent committed

    7d7bef0