Protostar Docs Deploy

Protostar docs site (docs.voidprojects.ai)

Operational reference for how the Protostar docs site is built, hosted, and recovered.

Architecture: public content + private deploy

Two repos under voidprojectssoftware, split so the public repo carries zero hosting detail:

• protostar-docs (PUBLIC): Docusaurus content hub. Holds the site shell, docs.manifest.json, and scripts/fetch-docs.mjs. Its ONLY workflow is a PR/push build-check that compiles the site (no secrets, no deploy). Never add a deploy job, secrets, or pull_request_target here.
• protostar-docs-infra (PRIVATE): the deploy + IaC. Holds the deploy and terraform workflows, Terraform, and the Azure/OIDC config. It checks out the public repo, builds, and ships.

Docs are authored next to code in each product repo (a docs/ folder) and lifted into the hub at build time.

How content is lifted (fetch-docs.mjs)

• docs.manifest.json is the single source of truth for which product repos contribute docs. Each entry: id, label, repo, ref, docsPath (usually docs), routeBasePath, navbarPosition.
• scripts/fetch-docs.mjs --remote blobless+sparse-clones each product's docsPath at the pinned ref into external/<id>/docs. Local sibling checkouts win when present (for authoring) unless --remote/CI/DOCS_SOURCE=remote.
• Most common failure: a manifested product repo has no docs/ folder yet → Error: <repo>@<ref> has no 'docs' folder yet and the script exits 1. This breaks BOTH the public build-check AND the private deploy (they run the same script). Fix = add the docs/ folder to the product repo (or remove/repoint the manifest entry), then re-run.
• Adding a product to the hub: add an entry to docs.manifest.json in protostar-docs. Routing, sidebar, and "edit this page" wire up automatically. The product repo must have the docsPath folder at the pinned ref.

How it deploys (deploy workflow, private repo)

• Triggers: workflow_dispatch, repository_dispatch (docs-content-updated), and a daily schedule at cron: '0 7 * * *' (UTC).
• Checks out the public repo, npm ci, node scripts/fetch-docs.mjs --remote, npm run build:no-fetch, then deploys ./build to Azure Static Web Apps with @azure/static-web-apps-cli deploy --env production.
• Auth is OIDC (azure/login); the SWA deployment token is fetched at runtime via az staticwebapp secrets list. No stored Azure or SWA secrets.

Azure / hosting facts

• App: swa-protostar-docs (Azure Static Web App, Free tier) in resource group rg-protostar-docs, region East US 2.
• SWA default host: orange-river-0d06b480f.7.azurestaticapps.net (always serves the latest deploy; test here to isolate "is it built?" from "is the custom domain bound?").
• DNS: managed manually in Cloudflare (one docs CNAME), NOT in Terraform — Cloudflare has no OIDC, so rather than store a long-lived token the CNAME is added by hand. The whole pipeline stays credential-free.
• Custom domain docs.voidprojects.ai is bound on the SWA via Terraform (azurerm_static_web_app_custom_domain, validation_type = "cname-delegation"), gated by the repo variable ENABLE_CUSTOM_DOMAIN.

Sibling site: apex root voidprojects.ai

A SECOND, independent Azure SWA serves the root/landing site at the apex, separate from the docs app:

• App: swa-voidprojects-site in resource group rg-voidprojects-site, region East US 2.
• SWA default host: proud-hill-01ebd2b0f.7.azurestaticapps.net.
• Custom domain voidprojects.ai was bound manually (Azure portal + az), NOT via Terraform (as of 2026-06-11). If/when Terraformed, mirror the docs terraform/site pattern but with validation_type = "dns-txt-token".

Apex domains differ from the docs subdomain

An apex (root) name cannot be a CNAME per the DNS spec, so two things change vs. docs.:

1. Validation = TXT token, not cname-delegation. az staticwebapp hostname set ... --validation-method dns-txt-token returns a token; read it with ... hostname show ... --query validationToken -o tsv. Put it in a Cloudflare TXT record on @.
2. Traffic record = CNAME on @ relying on Cloudflare CNAME flattening. Cloudflare synthesizes A records at the apex from the SWA default host. Target = proud-hill-01ebd2b0f.7.azurestaticapps.net, DNS only (grey cloud) during validation (same grey-cloud rule as gotcha 1). The flattened A will be whatever Azure front-end IP Cloudflare resolves the host to (Azure SWA fronts on several anycast IPs, so the apex A and a direct dig of the host can differ; both are valid).

• www.voidprojects.ai is NOT configured yet. To add it: a second SWA custom domain (CNAME www → the host, grey cloud) or a Cloudflare redirect rule www → https://voidprojects.ai/$1.
• "Apex looks broken locally right after setup" is almost always DNS, not Azure. The bind, cert, and Cloudflare flattening can all be correct (verify with curl --resolve voidprojects.ai:443:<azure-ip> → 200) while the LAN resolver still serves a stale negative answer. Blake's workstation resolves through the homelab Technitium server, which caches NODATA for ~30 min; flush it (see the technitium-api skill's negative-cache gotcha). Triage tell: Resolve-DnsName voidprojects.ai -Server 8.8.8.8 works but the local resolver does not.

Terraform layout (private repo)

• terraform/bootstrap — run ONCE, LOCALLY, by an owner. Creates the remote-state storage account, the rg-protostar-docs RG, the GitHub OIDC Entra app + federated creds, and the role assignments. Uses LOCAL gitignored state (chicken-and-egg: it creates the backend everything else uses) and Azure-CLI auth (az login), NOT OIDC. CI never touches bootstrap.
• terraform/site — the SWA + (gated) custom domain. Uses the azurerm backend and OIDC auth (use_oidc = true, ARM_* env vars set by the workflow). Runs in CI via the terraform workflow: plan on PRs, apply on push to main + workflow_dispatch.
• The CI OIDC identity has Contributor scoped to only rg-protostar-docs + Storage Blob Data Contributor on the state account + (see gotcha 2) Reader at subscription scope.

Gotchas (both bit us bringing the custom domain up)

1. cname-delegation needs the Cloudflare CNAME on DNS-only (grey cloud)

cname-delegation validation reads public DNS and must see the CNAME pointing at the SWA default host. If the Cloudflare record is proxied (orange cloud) it returns Cloudflare anycast IPs and hides the CNAME, so Azure can't validate → 400 BadRequest: CNAME Record is invalid (ExtendedCode 51021), and the proxied edge serves Azure's default 404 for the unrecognized Host.

• Fix: set the docs CNAME to DNS only (grey cloud), target = orange-river-0d06b480f.7.azurestaticapps.net. You can re-enable the proxy AFTER validation if desired.
• Check proxy status: nslookup docs.voidprojects.ai 8.8.8.8 — if it shows the CNAME / an Azure IP it's grey-cloud; Cloudflare anycast (104.21.x, 172.67.x, 2606:4700::) means it's still proxied.

2. The RG-scoped CI identity can't poll SWA custom-domain async ops

Creating/deleting a SWA custom domain is an async Azure operation. The create runs inside the RG (Contributor covers it), but Terraform then polls the operation RESULT at a subscription/location scope OUTSIDE the RG: Microsoft.Web/locations/<region>/operationResults/<id>. Contributor-on-RG can't read it → 403 AuthorizationFailed ... 'Microsoft.Web/locations/operationResults/read'. Azure still CREATES the domain, but the apply errors before recording it in state → permanent drift and already exists - needs to be imported on every later apply.

• Fix: grant the CI service principal Reader at subscription scope (in terraform/bootstrap). Least-privilege alternative: a custom role with Microsoft.Web/locations/operationResults/read + operations/read. Read-only either way.

Runbooks

Site is 404 / not available at docs.voidprojects.ai

1. Is content built? curl -I https://orange-river-0d06b480f.7.azurestaticapps.net/. 200 = built; the issue is the custom domain. 404 = the deploy never produced an artifact → check the deploy workflow (almost always the fetch-docs "no docs folder" failure).
2. Custom-domain 404 with Server: cloudflare: check proxy status (gotcha 1) and that the binding exists: az staticwebapp hostname list --name swa-protostar-docs --resource-group rg-protostar-docs. Empty [] = not bound.
3. errorMessage/status on the hostname entry tells you where validation is (Validating → wait; cname-delegation + managed cert takes a few minutes, longer right after a delete due to ~15-min propagation).

Build / deploy action failing

• <repo>@main has no 'docs' folder yet → add the docs/ folder to that product repo (or fix the manifest), then re-run. Same root cause hits both build-check and deploy.

Recreate a drifted custom domain cleanly (no import)

Done when the resource exists in Azure but not in TF state (gotcha 2). Requires an owner Azure login locally:

1. Land + apply the Reader-role RBAC fix: terraform -chdir=terraform/bootstrap apply (LOCAL state, az login auth — confirm the plan adds ONLY the role assignment, nothing else).
2. Delete the drifted domain so TF can create fresh: az staticwebapp hostname delete --name swa-protostar-docs --resource-group rg-protostar-docs --hostname docs.voidprojects.ai --yes (the site 404s until step 3; deletion edge-propagation can take ~15 min).
3. Re-run the site apply: gh workflow run terraform --repo voidprojectssoftware/protostar-docs-infra. With Reader in place and the resource absent, the create polls successfully (~6-7 min for validation + cert) and records state. Verify: curl -I https://docs.voidprojects.ai/ → 200, and openssl s_client -connect docs.voidprojects.ai:443 -servername docs.voidprojects.ai </dev/null | openssl x509 -noout -subject → CN=docs.voidprojects.ai.

Alternative to teardown: a config-driven import {} block adopts the existing domain without downtime, but it doesn't fix the RBAC gap (future destroy/recreate 403s again), and the import plan may want a replace.

Bootstrap local-state gotcha

terraform/bootstrap state is LOCAL and gitignored — it lives only on the owner's machine that first ran it. Before terraform -chdir=terraform/bootstrap apply, confirm terraform/bootstrap/terraform.tfstate exists locally; without it Terraform tries to recreate the RG/storage/Entra app and fails on already exists. Also note the infra repo's default branch was renamed master→main; a stale local clone tracking origin/master shows a false "diverged" — realign with git reset --hard origin/main (the old local commit is a superseded Cloudflare-provider design).

Non-secret identifiers (handy for ops; none are credentials)

• Subscription: 4d53a45e-234e-451f-b733-df7d7aac9e0b ("Main")
• OIDC app client id: 2d115dbe-6091-4c8e-9b22-40dad052aa1e; SP object id: 2edc0699-ca13-47aa-8664-5a73d8cc5f98
• Tenant: 65c0a7b8-1f48-4b79-8bd2-d13415a790e8
• State backend: RG rg-tfstate, storage stvptfstate, container tfstate
• Repo var that gates the custom domain: ENABLE_CUSTOM_DOMAIN (true after the CNAME exists)