Developing a Void Platform

The framework, CLI, and platform live in one repository. You can change the platform, test it locally, and deploy a runtime built from your fork into your own Cloudflare account.

If you want to run the released platform without changing its implementation, start with Install a Void Platform. Use Platform Administration for managing an installed platform.

Setting Up the Repository

Clone your fork and install the workspace dependencies with Vite+:

git clone https://github.com/your-org/void.git
cd void
vp install
vpr install:void-dev
void-dev --help

Use the Node.js version recorded in .node-version. The workspace uses public npm packages; a GitHub Packages token is not required.

install:void-dev builds the CLI, shared packages, and platform runtime, then makes this checkout's built CLI globally available as void-dev. Public packages export their built files, so run vp run build:core after later source changes to refresh the alias. The installer refuses to replace an unrelated global command. Remove only this checkout's alias with vpr uninstall:void-dev. All commands in this guide run from the repository root.

Finding the Implementation

Directory	Purpose
packages/void	Framework, application runtime, and CLI
packages/platform	Installer contracts, packaged Workers, and platform migrations
packages/deploy-core, packages/deploy-cloudflare	Shared deployment contracts and Cloudflare upload code
platform/packages/api	Users, projects, deployments, provisioning, and administrator API/UI
platform/packages/dispatch	Application routing and static assets
platform/packages/proxy	AI, remote bindings, and revalidation
platform/packages/tail	Runtime log ingestion
platform/packages/dashboard	Dashboard source and local UI components in ui/

The @voidcloud/* names identify workspace implementation packages. Their private: true flags prevent publishing those packages to npm. Deployable core Workers are bundled into the public @void/platform package.

The core installer creates the API, dispatch, proxy, and tail Workers. The dashboard and managed GitHub build services are available in the source tree but are not included in that installation. Adding an optional service requires its infrastructure, bindings, authentication, and capability configuration together.

Running the API Locally

Start a local PostgreSQL server and make psql and pg_isready available on your path. Then create the development database, apply its migrations, and seed an administrator:

vp run --filter @voidcloud/api setup --admin-email dev@example.com
vp run --filter @voidcloud/api dev --local --enable-containers=false --host localhost --port 8787

The command disables the optional build containers, so basic API and admin work does not require Docker. To develop managed builds, install Docker and run the API with containers enabled.

Open http://localhost:8787/admin/. Setup writes local development values to the API and dashboard .dev.vars files, including the development authentication bypass and local database connection. Those files are ignored by Git. Production installations get separate credentials through the installer.

The API's development bypass lets you work on the browser admin UI without setting up OAuth. Operator CLI sessions still require administrator authentication; they do not use the browser bypass.

Running the Dashboard Locally

The dashboard is a separate source app. After API setup, start it in another terminal:

vp run --filter @voidcloud/dashboard dev

Its local .dev.vars should point to the API you started:

API_URL=http://localhost:8787
SITE_DOMAIN=apps.example.com

Open the Vite URL and choose Dev login (local API). That button appears for a localhost API and uses the local development sign-in endpoint.

You can also point the dashboard at an API that you operate. If Cloudflare Access protects that API, install cloudflared and opt into the dashboard's local token helper with matching origins:

API_URL=https://platform.example.com
CF_ACCESS_APP_URL=https://platform.example.com
SITE_DOMAIN=apps.example.com

The helper refreshes an Access session before the dev server starts. Human dashboard requests require the user's Access session as well as their Void login; a service token does not represent that user. Service-token pairs are for scoped machine operations. This is dashboard development configuration; Access credentials are separate from platform login credentials.

For a deployed dashboard, bind its API service to your platform API, configure DASHBOARD_URL on both the dashboard and API, and include that exact dashboard origin in the platform's Access protection application. The dashboard passes the browser's company identity to the API using that service binding. Its login page shows the platform's currently enabled methods, and Account supports adding an additional login identity.

Testing Changes

Run tests for the area you changed while developing:

vp test run platform/packages/api/test/integration/operator-auth.test.ts
vp run check

Before preparing a release, build the packages and run the complete checks:

vp run build:all
vp run check
vp lint
vp run lint:platform
vp fmt --check
vp test run
vp run build:docs

Platform integration tests use their local test database by default. To test against PostgreSQL, set VOID_TEST_DATABASE_URL to a disposable database: these tests clear tables between cases.

To exercise a deployed application, deploy a disposable copy of playground/kitchen-sink and pass its URL as SMOKE_URL to the smoke test described in platform/scripts/kitchen-sink-smoke-test.md. That test creates and removes application data.

Deploying Your Runtime

Build the runtime. From a Git checkout, the build automatically records the current commit in the runtime manifest. An uncommitted checkout is recorded with a -dirty suffix:

vp run build:core

Build systems may set VOID_PLATFORM_SOURCE_REVISION when they need to override automatic detection with another immutable revision. A source archive without Git metadata builds normally but leaves the revision unrecorded.

The runtime is written to packages/platform/dist/runtime. Use the built CLI to preview a new installation from those files:

void-dev platform install \
  --name my-team \
  --application-domain example.app \
  --zone example.app \
  --runtime packages/platform/dist/runtime \
  --plan

Before applying the plan, complete the first-install walkthrough. Run the command without --plan to install. For testing before your domain is ready, replace --application-domain and --zone with --workers-dev; PostgreSQL and the runtime/GitHub/R2 credentials are still needed. Use void-dev in place of void and keep --runtime packages/platform/dist/runtime on install and resume commands. For an interrupted installation, keep that runtime build available until it completes.

Later, run void-dev platform domain set example.app. Domain setup uses the existing installed runtime and needs no --runtime, rebuild, or app redeploy. It keeps the platform API URL and original workers.dev app URLs available.

For an existing installation, preview an upgrade using its connection ID:

void-dev platform upgrade <installation-id> \
  --runtime packages/platform/dist/runtime --plan

Then run it without --plan to apply the upgrade. Source-built runtimes go through the same artifact, database, ownership, and health checks as released runtimes. The CLI preserves a disabled installation's state and records the source revision in its installation checkpoint.

Optional GitHub Webhook Ingress for Access-Protected APIs

The core installer does not deploy the dashboard, GitHub App, build Containers, or webhook ingress. If a source-built installation adds those optional services and Cloudflare Access protects its API hostname, GitHub cannot deliver directly to /webhooks/github: GitHub does not present your Access credentials. Do not add an Everyone or bypass policy to the API application.

The API source package includes an optional, path-isolated Worker for this case. It accepts only POST /github, validates GitHub's signature over the raw body, and forwards one authenticated internal operation over an API service binding. The API independently verifies both that internal proof and GitHub's signature before running the normal webhook handler. Installations without perimeter protection can continue using the API's direct /webhooks/github endpoint.

To deploy the optional ingress:

1. Deploy the source API/build runtime containing the internal operation and its build bindings, then finish the separate GitHub App/build-service configuration. A core install or upgrade alone does not add managed builds. Enable the platform's managed-build capability only when that infrastructure is ready.

2. Edit platform/packages/api/wrangler.github-webhook-ingress.jsonc. Give the ingress a name unique to the installation and set its API service binding to the exact installed API Worker name. Keep its public hostname separate from every human/API hostname covered by Access.

3. Deploy it from the repository root:

   vp run --filter @voidcloud/api deploy:github-webhook-ingress --env production

   Use --env staging for the staging entries in the same config.

4. In the Cloudflare dashboard, add encrypted Worker secrets. Set the GitHub App's existing GITHUB_WEBHOOK_SECRET on both the API and ingress Workers. Generate a separate high-entropy value, such as openssl rand -base64 32, and set it as GITHUB_WEBHOOK_INGRESS_SECRET on both Workers. Do not reuse a platform management, JWT, Access, or GitHub webhook credential for that value.

5. In the GitHub App settings, keep Content type set to application/json, keep the same webhook secret, and change Webhook URL to the isolated ingress URL ending in /github. Use GitHub's test delivery and confirm a 2xx response before relying on push builds.

The ingress has no login, dashboard, project, operator, proxy, or arbitrary forwarding route. It does not make the GitHub integration part of the core installer, provision build executors, create a GitHub App, configure Cloudflare Access, or manage either Worker's secrets. Its body limit is 25 MiB, based on GitHub's documented 25 MB webhook payload cap; malformed or larger deliveries are rejected before event processing.

Deploying Source Builds from CI

Build @void/platform from your checkout and pass its runtime directory to install, upgrade, repair, enable, or rollback with --runtime. Custom runtimes get the same integrity, migration, health, and rollback checks as packaged releases.

Void records the runtime's manifest digest, whether it was packaged or custom, and its source revision. A build from a Git checkout automatically records HEAD, or <HEAD>-dirty when the checkout has uncommitted files. Build systems can set VOID_PLATFORM_SOURCE_REVISION to override automatic detection.

A fresh CI runner can discover the installation each time. Set CLOUDFLARE_API_TOKEN and VOID_PLATFORM_DATABASE_URL through its protected environment, then:

CI commands and recovery inputs

vp run build:core

export VOID_PLATFORM_REGISTRY_DIR="$RUNNER_TEMP/void-platform-registry"
export VOID_PLATFORM_RECOVERY_KEY="$(openssl rand -base64 32)"
node packages/void/dist/cli/cli.mjs platform discover --account "$CLOUDFLARE_ACCOUNT_ID" \
  --installation "$VOID_PLATFORM_INSTALLATION"
node packages/void/dist/cli/cli.mjs platform upgrade "$VOID_PLATFORM_INSTALLATION" \
  --runtime packages/platform/dist/runtime --plan
node packages/void/dist/cli/cli.mjs platform upgrade "$VOID_PLATFORM_INSTALLATION" \
  --runtime packages/platform/dist/runtime --yes

Omit the installation selector only if the account has one discoverable installation. Run one deployment per installation at a time, and let it finish before starting the next. Cancelling during migrations or Worker rollout can leave an installation waiting for recovery.

Keep the management token, database URL, JWT signing secret, email signing secret for email-enabled installations, and complete project-encryption keyring in a protected CI environment. Upgrades inherit deployed Worker secrets. The original values are needed when recreating a missing Worker; configuration credentials are also needed when explicitly rotating them.

Changing the Platform Schema

The schema lives in platform/packages/api/src/schema.ts. Generate a migration after changing it:

vp run --filter @voidcloud/api db:generate

Review the generated SQL and migration journal before committing them. Released migrations are append-only. An upgrade applies new migrations while the previous Workers may still serve requests, so schema changes must remain compatible with that code.

platform/packages/api/drizzle/compatibility.json records which earlier runtimes have been verified against the new schema. Add an edge only after testing that compatibility. The runtime build checks that the migration files and recorded schema history agree. See packages/platform/README.md for the manifest contract.

CI in a Fork

The uncredentialed test workflows use GitHub-hosted runners in forks. They build and test the workspace without requiring Void Cloud accounts, tokens, or deployment access.

If your fork has a released platform version, set the repository variable VOID_PLATFORM_PRODUCTION_REF to its immutable 40-character commit SHA. Platform pull requests then create that version's database, seed representative user, project, and encrypted-secret records, and apply the proposed migrations. The check verifies existing data and runs the previous runtime against the upgraded schema. Mutable branch or tag names are rejected.

Without an explicit SHA, required compatibility checks use the latest successful GitHub deployment to Prod (or VOID_PLATFORM_PRODUCTION_ENV). They require read access to deployment history and stop if no completed deployment is found. Advancing the production branch does not change the selected baseline. Fork workflows that call platform CI must grant deployments: read alongside contents: read.

The upstream repository also contains workflows for deploying Void Cloud and publishing the official npm packages. Forks should configure their own release workflow around the built CLI and platform upgrade --runtime; the source-build CI example shows the required inputs. Keep production credentials in protected environments restricted to the appropriate release refs.

Public releases use matching versions for the CLI, scaffolder, adapters, and packaged platform. The publish workflow rejects mismatched versions and previously unpublished void versions that npm cannot reuse. Stable versions use latest; prereleases use their named channel, such as beta or rc. Numeric prereleases use next. The scaffolder installs its exact matching CLI version, so create-void@beta cannot silently select the stable CLI. Packaged platform runtimes record the release commit as their source revision.

The npm release job uses trusted publishing from GitHub-hosted runners, with id-token: write and no stored npm publishing token. Configure a trusted publisher for each public package using this repository, publish.yml, and the Release environment, allowing direct npm publish. A brand-new package needs an initial authenticated publication before its trusted publisher can be configured; subsequent releases use OIDC.

The managed build fallback CLI also derives its version from packages/void/package.json; there are no separate version pins to update. Build its image from the repository root with docker build --file platform/packages/api/container/Dockerfile .. The root .dockerignore limits that build context to the agent, Dockerfile, and public SDK manifest.

Publishing requires both SDK CI and platform CI, including the platform unit and API integration suites. Release tags also run the Windows SDK checks; a passing SDK-only build cannot publish a changed control plane.

Retrying a Release

To retry a failed release without moving an existing tag, add +retry.N to a new Git tag, with N starting at 1. Keep the package versions unchanged:

Git tag	Package version	npm channel
v0.21.0	0.21.0	latest
v0.21.0+retry.1	0.21.0	latest
v0.21.0-beta.1+retry.2	0.21.0-beta.1	beta

For example, when the packages are at 0.21.0, commit the release fix and tag that commit:

git tag -a 'v0.21.0+retry.1' -m 'Retry 0.21.0 publication.'
git push origin 'refs/tags/v0.21.0+retry.1'

The retry suffix belongs only in the Git tag, not in package.json. A -1 suffix is a distinct prerelease version, not a retry. Tag and package versions are checked before dependency installation and the full CI jobs; retries still run the normal release checks.

Retries publish only package versions that are still missing from npm. They cannot replace an already-published version. If an earlier attempt partially published the release and you changed its package contents, bump the version instead of combining different contents under the same version.

For implementation history, use the design archive at platform/meta/design-docs/README.md. Its proposals explain earlier decisions; the source and current guides define the supported behavior.