Public orchestrator ops checklist
Keep marklovely/lovely-hub-provision in sync with workflow templates on home-dashboard main (especially after #761 lifecycle reporting or #765 sync tooling).
Priority order
1. Sync public orchestrator workflows (high)
marklovely/lovely-hub-provision must match templates in home-dashboard or signup/teardown/E2E jobs run stale YAML.
cd home-dashboard
node scripts/sync-all-public-orchestrator-workflows.mjs --repo marklovely/lovely-hub-provisionMinimum after public orchestrator: hub-provision-dispatch.yml (composite provision + billing teardown on ubuntu-latest).
After #761 / #763: sync hub-lifecycle-dispatch.yml (artifacts, R2 upload, lifecycle POST).
Requires gh auth login with push access to the public repo (agents often get 403 — run locally).
2. Public repo secrets (high)
node scripts/sync-hub-tfvars-to-github-secrets.mjs \
--repo marklovely/lovely-hub-provision \
--profile provision \
--from-terraform-state \
--skip-missingConfirm on the public repo:
| Secret | Notes |
|---|---|
HOME_DASHBOARD_REPO | marklovely/home-dashboard |
HOME_DASHBOARD_REF | Usually main |
HOME_DASHBOARD_DISPATCH_TOKEN | PAT Contents: Read on private repo (legacy name HOME_DASHBOARD_CHECKOUT_TOKEN still accepted by template) |
| Full provision set | Same names as platform-provision.md |
3. Deploy hub-jobs Worker (high)
Queue consumer must include wait-for-completion on provision/teardown and public repo routing.
- Redeploy
lovely-home-hub-jobsfrom this repo (platform admin Terraform / Worker deploy you use for hub-jobs). - Default in code: provision/teardown →
marklovely/lovely-hub-provisionunless Worker envUSE_LEGACY_VPS_HUB_PROVISION=true.
Optional explicit bindings (recommended in hub.tfvars):
github_provision_repo = "marklovely/lovely-hub-provision"
github_provision_workflow = "hub-provision-dispatch.yml"Then Platform admin Terraform → apply-platform-admin.
4. Hub-jobs PAT scopes (high)
PLATFORM_GITHUB_TOKEN on the Worker must actions:write on lovely-hub-provision (dispatch) and still reach private platform-site-registry.yml.
5. Smoke test (medium)
- Public repo → Hub provision dispatch →
job_action=provision, test site id,skip_platform_admin=true. - Confirm job runs on
ubuntu-latestand completes. - Optional: cancel a Stripe test subscription and confirm teardown dispatches and finishes (deprovision retry path from public orchestrator).
6. Stuck hubs after failed deprovision (medium)
For sites that stayed live with billing canceled before public orchestrator:
- Platform site billing deprovision (manual) with
use_github_hosted: trueif the VPS runner is busy, or - Public E2E sandbox cleanup dispatch / billing deprovision composite on
lovely-hub-provision.
Do not purge D1 while the site is still on origin/main — see e2e-sandbox-cleanup.md.
7. Private repo manual workflows (low)
Private Platform site provision / billing deprovision / deprovision support use_github_hosted on manual dispatch — use when self-hosted queue is long.
Registry record / drop stays on the private repo (self-hosted or ubuntu-latest detect jobs only).
Dependency updates (Dependabot)
home-dashboard uses .github/dependabot.yml for weekly npm (root, worker/, ops-handbook/), GitHub Actions, and Terraform updates. Review and merge Dependabot PRs on main; after workflow or action bumps, re-run sync public orchestrator workflows if templates changed. Enable Dependabot alerts under repo Settings → Code security if security-only PRs are not appearing (version updates are configured separately in the YAML above).
Kept quiet on purpose (#994). Every Dependabot PR dispatches hosted CI (test, Hub smoke, Hub visual), so a burst of PRs or rebases makes your own CI queue. The config therefore:
- runs Sunday 04:00 Europe/London, outside the UK working day;
- groups majors as well as minor/patch, so there are at most 2 PRs per ecosystem (
open-pull-requests-limit: 2); - sets
rebase-strategy: disabled, so merging tomaindoes not re-run CI on every open Dependabot PR. To bring one up to date, comment@dependabot rebaseon it when you're ready to merge it.@dependabot recreaterebuilds it from scratch.
If a grouped major PR fails, either fix forward on that branch or comment @dependabot ignore {dependency} major version to drop the one that's blocking it.
Supply-chain guard (#1165).
.npmrcin the repo root andworker/setsignore-scripts=true, sonpm cinever runs a dependency's install script, in CI, in the public orchestrator (it checks this repo out) or on a laptop.ops-handbookhas its ownpredev/prebuildhooks, soscripts/deploy-ops-handbook.shpasses--ignore-scriptson the install instead.- npm updates have
cooldown: default-days: 7: Dependabot only proposes a release once it has been public for a week. Hijacked versions are usually pulled within days. Security updates are not delayed. - In the public CD template, Cloudflare credentials are set on the deploy step only, not on the job that installs packages.
- If a new dependency really needs its install script, run it explicitly (
npm rebuild {pkg}) in the one step that needs it, without secrets in that step's env.
Big major groups get split by hand. When a grouped major PR bundles several unrelated upgrades and fails, close it and land them one per PR, each with a full Hub lifecycle dispatch regression run (checkout_ref = the PR branch). That's how #1034 was handled: ESLint 10 first (its recommended config adds preserve-caught-error and no-useless-assignment; the code was fixed rather than the rules turned off).
Vite 8 builds with Rolldown instead of Rollup (hub build ~4× faster). vite.config.js uses build.rolldownOptions, and the website QR rebuild probe checks rolldown (if its native binary is missing, the committed website/vendor/lovely-qr.js is used). npm 10 can crash with Cannot read properties of null (reading 'edgesOut') when changing Vite majors in the lockfile; regenerate it with npx npm@11 install (npm 10 npm ci reads the result fine).
Vitest 5 (root only; worker/ is still on Vitest 4. It used to wait for @cloudflare/vitest-pool-workers, but the worker tests run in plain Node and that package was unused, so it was removed in #1171 and the worker can move to Vitest 5 on its own). In the jsdom environment window.document is read-only, so tests must not assign global.document = new JSDOM(...).window.document; reset the environment's own document.body.innerHTML instead.