docs-deploy.yml

name: Docs Deploy

# The ONE path by which this repository's docs site reaches production: the
# `deploy` job, Vercel project `objectstack`, Root Directory `apps/docs`.
#
# The Object UI docs site (www.objectui.org) was retired when its pages folded
# into this one. Its old URLs answer host-scoped permanent
# redirects from `apps/docs/redirects.mjs` (`objectuiRedirects`) once the
# objectui.org domains point at this project; nothing deploys it any more.
#
# ## Why the build runs here and not on Vercel
#
# Vercel bills Build CPU Minutes, and it bills them for deployments that do not
# end up building anything. Measured on 2026-10-12 against this project: ~600
# deployments a day, ~533 of them Git-integration PREVIEW deployments for
# `claude/*`, `gh-readonly-queue/*`, `board-archive*` and `changeset-release/*`
# branches. Each one cloned the monorepo, ran the old Ignored Build Step script
# for ~25 s and was then canceled — and was still billed: billed CPU-minutes
# divided by production-build wall-minutes came out at 12–15, more than the
# 8-vCPU build machine can produce on its own, so the canceled ones are in the
# bill. They were ~40% of this project's cost; the ~34 production builds a day
# that pushes to `main` triggered were most of the rest. Build CPU Minutes were
# 97% of the team's invoice.
#
# The maintainer's ruling, verbatim and untranslated:
# 「文档站应该只有主分支需要部署，而且应该用免费的」 — only `main` deploys the docs,
# and on free compute. This repository is public, so a standard
# `ubuntu-latest` runner costs nothing. So this job builds the site on that
# runner with `vercel build` and uploads the result with
# `vercel deploy --prebuilt`: Vercel stores and serves it and builds nothing.
# `apps/docs/vercel.json` turns the Git integration's deployments off
# (`git.deploymentEnabled: false`), so no branch push creates a Vercel
# deployment at all — billed, canceled or otherwise.
#
# ## Failure direction: a silent skip is the worst outcome
#
# A run that cannot deploy goes RED, never green-and-skipped. The site going
# quietly stale with no error anywhere is exactly what the old Ignored Build
# Step script was written around ("a wrong skip silently stops publishing
# documentation"), and moving the build here does not change which direction
# is the dangerous one. In particular a missing `VERCEL_TOKEN` secret (an
# objectstack-ai organization secret; a repository secret of the same name
# resolves identically) is an `::error::` and a failed run, never a notice:
# unlike an optional mirror, this job IS the publishing path, and without it
# nothing else publishes the docs.

on:
  schedule:
    # Every six hours, at :43 past 00, 06, 12 and 18 UTC. The maintainer's
    # decision: the docs refresh at most six hours behind `main`, and
    # "Run workflow" on this workflow's Actions page (workflow_dispatch below)
    # publishes `main` immediately when that is too slow. Builds cost nothing
    # here at any frequency; four a day (against ~87 commits a day on `main`)
    # keeps this job from competing with CI for runners and needs no
    # paths-filter list to keep in step with what the site is built from.
    #
    # The HOURS keep out of this repository's crowded scheduled windows, read
    # from `grep -rn 'cron:' .github/workflows/` on the day this landed: 01–05
    # UTC carry most of the nightlies (stale, CodeQL, validate-deps,
    # coverage, the release-coverage and required-set patrols, the smokes and
    # nightly tiers), 07 carries showcase-smoke, and the board
    # snapshot owns :07 of 02, 08, 14 and 20. Of the six six-hour sets, only
    # 0/6/12/18 and 1/7/13/19 have hours with nothing scheduled after :00; the
    # first is taken because its :00 neighbours are short (release.yml's
    # version-pr refresh, the ci/lint full runs), while 07:00 starts the
    # showcase smoke. Since #25771 those ci/lint runs are 4-hourly
    # (00/04/08/12/16/20 :00), so they now meet this set only at 00 and 12 and
    # 06/18 lose that neighbour; the choice still holds and nothing moved. The MINUTE is off the top of the hour for the reason every
    # patrol here gives: scheduled runs queue behind everyone else's :00 cron.
    #
    # This job spends almost nothing of the per-repository GITHUB_TOKEN budget
    # (the checkout, and no API reads), so unlike the board readers it is not
    # placed against that pool.
    #
    # ⚠️ GitHub delays scheduled runs under load and occasionally drops one. For
    # docs that costs one more interval of staleness, which a dispatch buys back.
    - cron: '43 0,6,12,18 * * *'
  workflow_dispatch: {}

# Read the tree, nothing else. Vercel is reached with VERCEL_TOKEN, not with
# any GitHub scope.
permissions:
  contents: read

# One production deploy at a time, and the newest one wins. Cancelling an
# in-flight run is safe HERE, unlike a backup: every run builds the tip of
# `main`, which contains everything an earlier run would have published, so a
# cancelled run loses nothing a later run does not deliver.
concurrency:
  group: docs-deploy-production
  cancel-in-progress: true

jobs:
  deploy:
    name: Build the docs site and deploy it to Vercel production
    runs-on: ubuntu-latest
    # A forced docs build (spec generators + `next build`) measures ~4 minutes
    # on this runner class in ci.yml's Build Docs job; the rest is install and
    # upload.
    timeout-minutes: 30
    env:
      # Which Vercel project receives the deployment. These are identifiers,
      # not credentials; with both set, the CLI needs no `.vercel/project.json`
      # in the repository and no interactive `vercel link`.
      VERCEL_ORG_ID: team_06nvlGKZCkwhzdXmRdxeCNnB
      VERCEL_PROJECT_ID: prj_KbPkuEVCj8viErlaC0wBUgMIYZWe
      # Pinned, so a CLI release cannot change how production is built between
      # two runs of an unchanged tree. Moving it is a reviewed edit here.
      VERCEL_CLI_VERSION: '60.1.3'
    steps:
      # The maintainer's ruling is "only main deploys". `schedule` always runs
      # on the default branch, but `workflow_dispatch` offers a branch picker,
      # and a dispatch from any other ref would publish that branch to
      # production. Refused loudly rather than skipped: a skip reads as a
      # successful run of the button somebody just pressed.
      - name: Refuse any ref but main
        if: github.ref != 'refs/heads/main'
        run: |
          echo "::error::Docs Deploy publishes to PRODUCTION and runs only from main; this run is on ${GITHUB_REF}. Re-run it with 'main' selected in the branch picker."
          exit 1

      # Checked before anything is built, so a missing credential costs seconds
      # rather than a full build, and it fails RED: see the header.
      - name: Require the Vercel token
        env:
          VERCEL_TOKEN: ${{ secrets.VERCEL_TOKEN }}
        run: |
          if [ -z "${VERCEL_TOKEN:-}" ]; then
            echo "::error::The secret VERCEL_TOKEN is not visible to this run, so the docs site cannot be deployed and production is NOT being updated. It is provided as an objectstack-ai ORGANIZATION secret: check that it still exists and that this repository is in its repository access list (org Settings -> Secrets and variables -> Actions). To replace it, create a token at vercel.com -> Account Settings -> Tokens scoped to the Object Stack team, and store it as VERCEL_TOKEN there (or as a repository secret of the same name). Then re-run this workflow."
            exit 1
          fi
          echo "VERCEL_TOKEN is set."

      - name: Checkout repository
        uses: actions/checkout@v7

      # The same key ci.yml's jobs use, so a scheduled run restores the store
      # CI already saved for this lockfile instead of downloading it again.
      # `vercel build` runs vercel.json's `installCommand` again; with the tree
      # already installed from the frozen lockfile that pass has nothing to do,
      # and a lockfile problem fails HERE, under its own name.
      - name: Setup workspace
        uses: ./.github/actions/setup-workspace

      # ## Every Vercel command runs from the REPOSITORY ROOT, not apps/docs
      #
      # The project's Root Directory setting (`apps/docs`) comes down with
      # `vercel pull`, and the CLI applies it relative to the directory it is
      # run from: `vercel build` builds `<cwd>/<rootDirectory>` and writes
      # `<cwd>/.vercel/output`, and `vercel deploy` checks that
      # `<cwd>/<rootDirectory>` exists before uploading `<cwd>/.vercel/output`
      # (read from vercel/vercel's packages/cli build and deploy commands). Run
      # from `apps/docs`, the same setting would point at `apps/docs/apps/docs`.
      # From the root, the build reads `apps/docs/vercel.json` and runs its
      # `installCommand` / `buildCommand` with `apps/docs` as the working
      # directory — the same command, in the same place, that ci.yml's Build
      # Docs gate reads out of that file and runs on every docs PR.
      - name: Pull the production project settings
        env:
          VERCEL_TOKEN: ${{ secrets.VERCEL_TOKEN }}
        run: npx --yes "vercel@${VERCEL_CLI_VERSION}" pull --yes --environment=production --token="$VERCEL_TOKEN"

      - name: Build the site (vercel build --prod)
        env:
          VERCEL_TOKEN: ${{ secrets.VERCEL_TOKEN }}
          # Build THIS tree, never replay one. A fresh runner has no local turbo
          # cache and this job passes no remote-cache credentials, so nothing
          # could be replayed today; forcing keeps that true if either ever
          # changes, for the reason ci.yml's Build Docs step gives: a replayed
          # `.next` would publish a stale site and report success.
          TURBO_FORCE: 'true'
        run: npx --yes "vercel@${VERCEL_CLI_VERSION}" build --prod --token="$VERCEL_TOKEN"

      # An exit code of 0 is not proof that the docs app was built — see
      # ci.yml's "Assert the docs app actually built" for the measured case of a
      # build command that matched no package and exited 0. Assert both the
      # Next.js build and the Build Output API directory the upload reads.
      - name: Assert the build produced a deployable output
        run: |
          set -euo pipefail
          if [ ! -f apps/docs/.next/BUILD_ID ]; then
            echo "::error file=apps/docs/vercel.json::vercel build exited 0 but produced no apps/docs/.next/BUILD_ID - the docs app was not built, so nothing is deployed."
            exit 1
          fi
          if [ ! -f .vercel/output/config.json ]; then
            echo "::error::vercel build exited 0 but wrote no .vercel/output/config.json at the repository root - there is no prebuilt output to deploy."
            exit 1
          fi
          echo "Built docs app: BUILD_ID=$(cat apps/docs/.next/BUILD_ID)"

      - name: Deploy the prebuilt output to production
        id: deploy
        env:
          VERCEL_TOKEN: ${{ secrets.VERCEL_TOKEN }}
        run: |
          set -euo pipefail
          # stdout is the deployment URL and nothing else; progress goes to
          # stderr. `set -e` fails the step on a non-zero exit before the URL
          # is used.
          url=$(npx --yes "vercel@${VERCEL_CLI_VERSION}" deploy --prebuilt --prod --token="$VERCEL_TOKEN")
          echo "url=${url}" >> "$GITHUB_OUTPUT"
          echo "Deployed: ${url}"

      - name: Publish the run to the summary
        if: always()
        run: |
          {
            echo "### Docs Deploy — ${{ job.status }}"
            echo
            echo "Commit: \`${{ github.sha }}\` (${{ github.event_name }})"
            echo
            echo "Deployment: ${{ steps.deploy.outputs.url || 'none (the run stopped before the deploy step finished)' }}"
          } >> "$GITHUB_STEP_SUMMARY"