Publishing docs.hanzo.ai

One lane. A push to main builds the static export and publishes it to the Sites plane, and the bytes it uploads are what the site serves a second later. There is no image to tag and no pin to move.

push main
  -> git.hanzo.ai/hanzoai/docs          a pull mirror of github.com/hanzo-docs/docs
  -> .hanzo/workflows/deploy.yml        the forge reads this directory natively
  -> pnpm build --filter=docs           NEXT_EXPORT=1 -> apps/docs/out
  -> hanzoai/ci .github/actions/site    POST /v1/projects/docs-hanzo-ai/deployments
  -> s3://hanzo-sites/hanzo/docs-hanzo-ai   the prefix the docs.hanzo.ai route serves

The route is an IngressRoute + Middleware pair in hanzoai/universe (charts/app/values/hanzo/static-sites.yaml): Host(docs.hanzo.ai) -> docs-static -> root: s3://hanzo-sites/hanzo/docs-hanzo-ai. No pod, no replicas, no image. A site is files and a route.

The project is named for its host, as hanzo-ai is hanzo.ai's. docs is a reserved label on the Sites plane (cloud apps/sites/reserved.go), so no project can be created with it: the publish step asks for the project first, and that ask answered 400 "slug is a reserved subdomain" the first time a run got that far.

Every workflow lives in .hanzo/workflows/ and runs on git.hanzo.ai, which reads that directory first. GitHub reads only .github/workflows/, so none of them can queue there (Actions is also disabled on hanzo-docs/docs).

Fire it

Push to GitHub main, then pull the forge and dispatch. Nothing does either on its own today, measured 2026-09-27: the forge's copy of main sat at da72bf0527 for a day while GitHub moved (its update_mirrors cron logs every ten minutes and moves nothing), the repo carries no Actions unit so a push queues no run, and every recent deploy run is a workflow_dispatch.

curl -X POST -H "Authorization: token $FORGE_TOKEN" \
  https://git.hanzo.ai/v1/repos/hanzoai/docs/mirror-sync
curl -X POST -H "Authorization: token $FORGE_TOKEN" \
  https://git.hanzo.ai/v1/repos/hanzoai/docs/actions/workflows/deploy.yml/dispatches \
  -d '{"ref":"refs/heads/main"}'

FORGE_TOKEN is minted on the forge pod, as the git user: gitd admin user generate-access-token --username <you> --token-name <name> --scopes write:repository --raw. The forge also takes your Hanzo IAM login in its place, curl -u "<you>:$(hanzo auth token)", which carries write:repository and was answered 200 on mirror-sync and 204 on the dispatch (2026-09-28). KMS deploy/FORGE_TOKEN is read:repository only, and mirror-sync refuses it with 403. Run status: GET /v1/repos/hanzoai/docs/actions/runs/<id> with either credential, or action_run in the forge's Postgres (status 1 ok, 2 failed, 6 running).

One publish runs at a time (concurrency: deploy-docs, no cancel): a newer push waits for the running one, and the newest waiting commit is the one that publishes. Two at once would each delete the other's files on completion, because completing a deployment reconciles the prefix against the manifest it sends.

What the job proves before it publishes

The ingest key resolves. deploy/PUBLISHABLE_KEY (env prod) comes from KMS — the org's Actions secrets carry only KMS_CLIENT_ID/KMS_CLIENT_SECRET — and is POSTed to /v1/event before the build. A key that is present but names a deleted project answers 403, and a site built on it loses every pageview while looking perfect.
The export is a site. scripts/check-export.sh refuses an export without index.html, docs/index.html, 50+ pages, a rendered nav, and every path in apps/docs/export.require — the sections that vanish without breaking anything else (docs/studio/ is a submodule, docs/openapi/ and docs/mcp-tools/ are generated).
The key is in the bytes. A build-time env that never reached the bundle is invisible everywhere except the warehouse, so the export is grepped for it.
Every link resolves. build:post walks the export, follows every anchor through the serving ladder (/docs -> docs.html -> docs/index.html) and prints what points nowhere. It reports and does not refuse: the site is already built, and one rotted link is not a reason to stop shipping 2,900 pages.

Verify

curl -sI https://docs.hanzo.ai/ | grep -i last-modified   # the publish time
curl -s -o /dev/null -w '%{http_code}\n' https://docs.hanzo.ai/docs/quickstart
curl -s -o /dev/null -w '%{http_code}\n' https://docs.hanzo.ai/no-such-page  # 404

The publish itself reads back what it wrote (bin/site fetches the live URL and compares the bytes), so a deployment that completes has already proved the object store serves it.

The image lane is cold

ghcr.io/hanzoai/docs, the root Dockerfile, hanzo.yml's images: block and the docs app in hanzoai/universe are the previous way: an image wrapping the same export, served by hanzoai/static in a pod. The pod still runs and nothing routes to it — the docs.hanzo.ai router names the Sites middleware. Do not build or pin it to publish; retiring it is a universe change.

Sibling sites in this repo

apps/cloud publishes Sites project hanzo-cloud through the same action (.hanzo/workflows/deploy-cloud.yml). The remaining deploy-*-docs.yml workflows still wrangler pages deploy their app; each is the only deploy of its host, so they stay until that host has a Sites project of its own. The path for each is the one above: give the app output: 'export', publish a slug, then move its DNS.

5.5 KiB Raw Permalink Blame History