CI runner strategy (private repo)
This repository is private. GitHub-hosted (ubuntu-latest) Actions minutes are billable, so workflows stay on the self-hosted lovely-home runner unless you deliberately pay for hosted capacity.
Current runner
| Label | Host | Role |
|---|---|---|
lovely-home | Primary VPS | CI, Terraform validate, platform auto-merge, provision/deploy, secret sync, releases |
Hub smoke runs here too: Playwright hits public https://demo.lovely-home.co.uk over HTTPS; OS browser deps should be installed once on the VPS (scripts/vps-install-playwright-deps.sh).
Public-repo hosted CI (PR checks, $0 Actions minutes)
ci.yml runs a thin dispatch on lovely-home, then hub-ci-dispatch.yml on marklovely/lovely-hub-provision runs lint/test/build and Playwright smoke on ubuntu-latest. PR statuses test and Hub smoke (demo) are posted via the Checks API. Setup: public-ci-dispatch.md.
Do not add Dispatch hosted CI as a required check — only test and Hub smoke (demo).
Public-repo hosted Hub Pages CD (manual, $0 heavy minutes)
cd-hub-pages.yml runs a thin dispatch on ubuntu-latest, then hub-cd-dispatch.yml on marklovely/lovely-hub-provision resolves sites, builds once, and deploys hubs in parallel on ubuntu-latest. Setup: public-cd-dispatch.md.
Deploy on merge on the public repo (#1227)
deploy-on-merge.yml plans on lovely-home, then hands the deploys to hub-deploy-dispatch.yml on marklovely/lovely-hub-provision without waiting, so merges no longer queue their deploys on the VPS. Repo variable DEPLOY_ON_MERGE_RUNNER=vps switches back. Setup: public-deploy-dispatch.md.
Queue relief (no paid GitHub-hosted on private repo)
- Fewer triggers — CI
paths-ignorefor docs/ and ops-handbook/ (handbook deploys separately), docs/social-only changes and hub registry ledger overlay files (platform/sites.yaml, wrangler env blocks, manifest, etc.) when those are the only paths touched; provision/deprovision push detect skips ledger commits (platform: record|drop … in site registry) onubuntu-latest; Terraform validate only whenterraform/**changes; ensure-hub-api-bindings only when its scripts change; auto-merge reconcile cron every 10 minutes (was every 1). - Concurrency — CI and Terraform validate cancel stale runs per ref (
cancel-in-progress: true). Provision/deploy/deprovision keepcancel-in-progress: falseso in-flight hub work is not dropped. - Remove dead weight — Dropping GitHub Pages deploy (#495) removes a heavy self-hosted job on every
mainpush.
GitHub CLI on the VPS
Platform PR auto-merge uses gh. The composite action .github/actions/setup-gh-cli installs gh into RUNNER_TOOL_CACHE when it is missing. Alternatively install gh on the VPS image once (package or tarball) so jobs skip the download step.
Measuring the queue (#847)
Grafana dashboard Lovely Home — CI runner queue charts this hourly (grafana-dashboards.md). For an ad-hoc check, run node scripts/report-runner-queue.mjs (uses your gh login, read-only). It reports how long jobs waited for a runner, as p50 / p95 / max, per pool (lovely-home, ubuntu-latest), for this repo and the public CI orchestrator. Narrow it with --since 2026-10-19, --runs 100 or --repo owner/name.
Trigger for adding capacity: a lovely-home p95 that stays over 10 minutes at signup peaks. On 19 Oct 2026, with an 8-PR Dependabot burst, lovely-home reached p95 8.9 min and max 9.9 min, while hosted runners stayed under 1 min. Dependabot is now quiet (#994), so no second runner is needed yet.
Cheapest next step if it trips: move the thin CI dispatch job (seconds long) to ubuntu-latest. It costs about 1 billable private-repo minute per PR or push, and it takes CI off the VPS queue entirely. Only after that, add a second VPS (below).
Real scale-out
- Do not add a second runner on the same small VPS — you only run two jobs at once on the same CPU/RAM.
- Do not move bulk CI to GitHub-hosted on a private repo unless you accept minute charges.
- Do add a second VPS with its own runner and label (e.g.
lovely-home-deploy) and route long provision/deploy/terraform jobs to that label while keeping lighter jobs onlovely-home— or split by workflow in YAML when the second host exists.
Public orchestrator (hub signup provisions)
To run heavy provision/teardown on GitHub-hosted runners (~2 min) instead of the VPS queue (~14 min), hub-jobs dispatches a public repo workflow that checks out home-dashboard and runs the same scripts on ubuntu-latest. See hub-provision-public-actions.md and hub-provision-public-repo-setup.md. Until GITHUB_PROVISION_REPO is set on the hub-jobs Worker, behavior is unchanged.
Related
- public-cd-dispatch.md — hosted Hub Pages CD
- public-deploy-dispatch.md — Deploy on merge on the public orchestrator
- Internal audit: Project store
internal/runner-ci-audit-2026-09-16.md - PR #496 — concurrency, path filters, cron,
setup-gh-cli