docs.yml

name: Docs

# Publishes https://interlock.sourceshift.ai/: the landing page at the root
# and the manual, from the Markdown under docs/, at /docs/.
#
# The site is built by Zensical -- the successor to Material for MkDocs -- and
# deployed to Cloudflare Pages as a directory of static files. There is no
# Node build step and no framework: Zensical is a Python package, so this job
# only needs an interpreter.
#
# Zensical requires Python >= 3.11, which is why this job pins 3.13 rather than
# the 3.9 floor the SDK itself supports.
#
# Pinned to an exact Zensical version on purpose. It is 0.0.x and its output
# is the deployed artefact, so an unreviewed upgrade should be a commit someone
# made, not something that happens on a Tuesday because CI resolved `latest`.
#
# Two builds, in a fixed order. zensical.landing.toml roots a one-page build
# at / and owns site/; zensical.toml then writes the manual into site/docs/
# beside it (the landing config explains why one build cannot do both). The
# deploy step publishes site/, so the deployed directory and a local build
# have the same shape.

on:
  push:
    branches: [main]
    paths:
      - "docs/**"
      - "landing/**"
      - "overrides/**"
      - "zensical.toml"
      - "zensical.landing.toml"
      - "scripts/check_site.py"
      - ".github/workflows/docs.yml"
  pull_request:
    paths:
      - "docs/**"
      - "landing/**"
      - "overrides/**"
      - "zensical.toml"
      - "zensical.landing.toml"
      - "scripts/check_site.py"
      - ".github/workflows/docs.yml"
  workflow_dispatch:

permissions:
  contents: read

concurrency:
  group: docs-${{ github.ref }}
  cancel-in-progress: true

jobs:
  build:
    name: build
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v5

      - uses: actions/setup-python@v6
        with:
          python-version: "3.13"

      - name: Install Zensical
        run: pip install zensical==0.0.70

      # --strict aborts the build on any warning rather than printing it and
      # carrying on. Verified locally: a broken internal link makes this exit 1.
      # The docs are read on GitHub *and* on the site, so a link that resolves
      # in only one of the two places is exactly the drift this catches.
      - name: Build the landing page
        run: zensical build -f zensical.landing.toml --clean --strict

      - name: Build the manual
        run: zensical build --clean --strict

      # --strict already fails a broken internal link during the build; this
      # keeps the check visible as its own named step (and catches a relative
      # file link that Zensical would otherwise resolve differently from
      # GitHub). It walks the source Markdown, so a page outside the nav is
      # checked too.
      - name: docs-link-check
        run: |
          python3 - <<'PY'
          from pathlib import Path
          import re

          FENCE = re.compile(r"^ {0,3}(`{3,}|~{3,})(.*)$")
          CODE_SPAN = re.compile(r"(?<!`)(`+)(?!`)([^\n]*?)(?<!`)\1(?!`)")

          def strip_code(text):
              """Drop fenced blocks and inline code spans before parsing links.

              Brackets inside code are not links: `TOOLS[name](**safe_args)`
              is Python, not Markdown. Matching the raw text reports it as a
              broken target, so remove code first (a Markdown parser would do
              the same thing, but is not worth another dependency here).
              """
              kept = []
              fence = None
              for line in text.splitlines():
                  match = FENCE.match(line)
                  if fence is None:
                      if match:
                          fence = match.group(1)
                          continue
                      kept.append(line)
                      continue
                  if (match and match.group(1)[0] == fence[0]
                          and len(match.group(1)) >= len(fence)
                          and not match.group(2).strip()):
                      fence = None
              return CODE_SPAN.sub("", "\n".join(kept))

          missing = []
          for page in sorted(Path("docs").rglob("*.md")):
              text = strip_code(page.read_text(encoding="utf-8"))
              for target in re.findall(r"\]\(([^)\s]+)", text):
                  if target.startswith(("http://", "https://", "mailto:",
                                        "#")):
                      continue
                  path = target.split("#", 1)[0].split("?", 1)[0]
                  if not path:
                      continue
                  if not (page.parent / path).resolve().exists():
                      missing.append("{}: {}".format(page, target))
          if missing:
              raise SystemExit(
                  "docs-link-check: unresolved links\n" + "\n".join(missing))
          print("docs-link-check: every internal link resolves")
          PY

      # What neither build can check for itself: that the landing config still
      # mirrors the manual's, that the landing's links into the manual resolve
      # (they are served URLs, because the landing build cannot see the
      # manual's files), and that the tree has the shape the deploy relies on.
      # The landing build owns the root of site/ -- its index, its 404.html,
      # and landing/_redirects copied in, the only place Pages reads one. The
      # manual build owns site/docs/, whose own 404.html Pages picks for
      # anything missing under /docs/, since it serves the closest 404.html
      # walking up from the requested path.
      - name: Check the site
        run: python3 scripts/check_site.py

      - name: Upload the built site
        uses: actions/upload-artifact@v4
        with:
          name: site
          path: site
          retention-days: 7

  deploy:
    name: deploy
    needs: build
    # Only main publishes. A pull request builds (above) so a broken link or a
    # bad config fails the PR, but it must not touch the live site.
    if: github.event_name != 'pull_request'
    runs-on: ubuntu-latest
    environment:
      name: docs
      url: https://interlock.sourceshift.ai/
    steps:
      - uses: actions/download-artifact@v5
        with:
          name: site
          path: site

      - name: Deploy to Cloudflare Pages
        uses: cloudflare/wrangler-action@v3
        with:
          apiToken: ${{ secrets.CLOUDFLARE_API_TOKEN }}
          accountId: ${{ secrets.CLOUDFLARE_ACCOUNT_ID }}
          command: pages deploy site --project-name=interlock-docs --branch=main

      # The deploy is live within seconds, but the custom domain can lag the
      # pages.dev URL by a little, hence the retries. This checks the two
      # halves of the layout: the landing at the root, the manual at /docs/,
      # and a pre-move URL still finding its way there.
      - name: Smoke the live site
        run: |
          for attempt in 1 2 3 4 5 6; do
            code=$(curl -s -o /dev/null -w '%{http_code}' https://interlock.sourceshift.ai/docs/)
            [ "$code" = "200" ] && break
            echo "attempt $attempt: /docs/ returned $code, retrying"
            sleep 10
          done
          test "$code" = "200"
          test "$(curl -s -o /dev/null -w '%{http_code}' https://interlock.sourceshift.ai/)" = "200"
          test "$(curl -s -o /dev/null -w '%{http_code}' https://interlock.sourceshift.ai/docs/getting-started/)" = "200"
          test "$(curl -s -o /dev/null -w '%{redirect_url}' https://interlock.sourceshift.ai/getting-started/)" = "https://interlock.sourceshift.ai/docs/getting-started/"