Authelia runs as a managed native service on Wild Central, providing
two integration patterns for network services:
- Forward-auth via HAProxy for apps without native SSO (Lua auth-request
script intercepts requests, redirects unauthenticated users to login portal)
- OIDC provider for apps with native support (Gitea, Grafana, etc.)
Backend: new internal/authelia/ package with service manager, config
generation, file-based user management (argon2id), and OIDC client
management. API endpoints for status, config, users CRUD, and OIDC
clients CRUD.
HAProxy: config generator extended with lua-load, auth-request
directives, and Authelia backend. Auth directives scoped to session
cookie domain — only subdomains of the auth portal's parent are
eligible for forward-auth.
Frontend: Authentication page with enable/disable, user management,
OIDC client management (add/edit/delete), and protected domains
toggles. Advanced subsystem page for raw config view. Dashboard
service card and sidebar entries.
InstanceConfig was a ~50-field struct designed for Wild Cloud k8s instances,
but only 3 fields were ever read by dnsmasq (the sole consumer). The
reconciliation bridge constructed fake InstanceConfig objects from registered
domains just to satisfy this interface.
Replace with DNSEntry{Domain, IP} — a 2-field struct purpose-built for
dnsmasq. All domains get local=/ directives to prevent AAAA queries from
leaking to upstream DNS (Happy Eyeballs / RFC 8305 latency on LAN).
Removed dead code: InstanceConfig, NodeConfig, LoadCloudConfig,
SaveCloudConfig, DeepMerge, LoadMergedInstanceConfig, EnsureInstanceConfig,
instance path helpers, instanceLoadBalancerIP, GenerateInstanceConfig,
and all modular per-instance dnsmasq methods.
When haproxy -c fails, parse ALERT line numbers, map to service domains
via # service: comments in generated config, exclude broken services,
and regenerate. One retry — prevents a single bad registration from
blocking all config updates.
Linter-contributed improvements:
- EnsureCentralService() method: registers Central's own domain as a
service from config, with cleanup on domain change
- Tests for EnsureCentralService: registration, idempotency, domain
change cleanup, no-domain-is-noop
- Improved reconciliation tests with Central as a registered service
- Certbot handler cleanup
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Tests added (13 new):
- HAProxy ACL ordering: L7 before L4 wildcard, L4 exact before L4 wildcard
- HAProxy subdomains: false=exact only, true=wildcard+exact
- Reconciliation integration: HAProxy config from services, DNS config
from services, Central config domain injection, correct DNS target IPs
- dnsmasq: reach:internal has local=/, reach:public does not
- Services: idempotent registration, no duplicates in list
Bug fix:
- Run initial reconciliation on startup so Central's own domain (from
config) gets HAProxy + DNS routes even with zero registered services.
Previously, Central only got routes after the first service was
registered, causing 503 on fresh restart.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The reconciliation was building InstanceRoutes without setting the
Subdomains field, so all services were exact-match only — breaking
wildcard subdomain routing for *.cloud.payne.io and *.cloud2.payne.io.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Rewrite HAProxy config generation with correct ACL ordering:
1. L7 HTTP exact matches first (central, wild-cloud app)
2. L4 exact matches (payne.io, civilsociety.dev)
3. L4 wildcard matches (*.cloud.payne.io) — only when subdomains:true
4. Default → L7 termination
Key changes:
- InstanceRoute: removed ExtraDomains, added Subdomains bool
- GenerateOpts: removed CentralDomain (Central is now just another
HTTPRoute injected by reconciliation), renamed WildcardCert→CertsDir
- Central's domain comes from config, injected as HTTPRoute with
the actual running port (no more hardcoded 5055)
- Added API.SetPort() so reconciliation knows the running port
- subdomains:false → exact match only (no *.domain shadowing)
New tests: TestGenerate_WithSubdomains, TestGenerate_ACLOrdering
(verifies L7 exact matches come before L4 wildcards).
Fixes: BUG 1 (central routing to wrong port), BUG 2 (wild-cloud
getting traefik cert), BUG 3 (*.payne.io too broad).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Rewrite the service registration model per the architecture plan:
Model changes:
- Domain is the unique key (removed Name field)
- Removed ExtraDomains — each domain is its own registration
- Removed SourceNode, BackendStatic, ReachOff
- Added Subdomains bool for wildcard matching control
- Files stored as <domain_with_underscores>.yaml
- API routes: /services/{domain:.+} (regex for dots in path)
- Added DeregisterBySource for cleanup before re-registration
Reconciliation changes:
- No ExtraDomains iteration — each service IS one domain
- reach:internal → sets InternalDomain (generates local=/)
- reach:public → sets Domain (no local=/)
- tcp-passthrough → DNS points to backend IP
- http → DNS points to Central IP
All 22 tests pass (9 manager + 8 handler + 5 config/cert).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The auto-detect picks the default route interface (wlan0 at
192.168.8.152) but dnsmasq is bound to eth0 (192.168.8.151).
Now uses cloud.dnsmasq.ip from config as the Central IP for DNS
entries, falling back to auto-detect only if not configured.
Fixes DNS resolution failures when dnsmasq config points services
to the wrong interface IP.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Fix dnsmasq reconciliation to use the correct DNS target IP:
- tcp-passthrough (k8s instances): DNS → k8s LB IP (LAN direct)
- http/static (Central apps): DNS → Central IP (HAProxy terminates)
Also handle ExtraDomains properly:
- Internal domains (internal.cloud.payne.io) get local=/ entries
to prevent upstream DNS forwarding
- Extra domains get their own dnsmasq address= entries
- HAProxy already propagated ExtraDomains for SNI ACLs
Verified: cloud.payne.io → 192.168.8.240 (direct to k8s),
cloud2.payne.io → 192.168.8.80, central.payne.io → 192.168.8.151.
SNI routing through HAProxy works for both instances.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remove automatic wildcard cert assumption. Each service gets its own
cert by default. Wildcards are user-initiated via the Certificates page.
Backend:
- HAProxy L7 frontend uses cert directory (/etc/haproxy/certs/)
instead of single wildcard file — loads all PEMs, serves by SNI
- Cert status API shows every registered service individually with
coveredBy field when a wildcard cert covers the domain
- Reconciliation filters L7 routes to services that have a cert
- No auto-provisioning in reconciliation (just warnings)
Frontend:
- Certificates page shows per-service cert status
- "Provision" button for individual certs
- "Add Wildcard" form for opt-in wildcard provisioning
- Fixed CloudflareComponent type errors from cert API changes
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New dedicated Certificates page in Wild Central UI showing:
- All tracked certs (central, wildcard, per-service)
- Status (valid with days remaining, or missing)
- Per-domain Provision button
- Renew All button
Cert provisioning is now user-initiated only — reconciliation logs
warnings about missing certs but does NOT auto-provision. This
prevents unexpected wildcard cert requests on startup.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
HAProxy validation fails if the config references cert files that
don't exist (wildcard.pem for L7 routes, domain.pem for central TLS).
Now reconciliation checks for cert existence before including:
- L7 HTTP routes: only generated when wildcard cert exists
- Central domain TLS frontend: only generated when domain cert exists
This ensures HAProxy never enters a failure state due to missing
certs. Once certs are provisioned (via certbot), the next
reconciliation picks them up automatically.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Services registered via the service API have a domain but no internal
domain (that's a Wild Cloud instance concept). The dnsmasq generator
was producing broken entries like `local=//` and `address=//`.
Fix: skip internal domain entries when InternalDomain is empty, in
both config.go and config_modular.go. Also fix DNS entries to point
to Central's IP (where HAProxy listens) rather than the backend
address — all traffic flows through Central.
Added 3 tests covering: service-only, mixed services+instances.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When HTTP services are registered (where Central terminates TLS),
reconciliation now checks for and provisions TLS certificates:
- Derives gateway domain from central domain (e.g., payne.io)
- Checks for existing wildcard cert (*.payne.io)
- If no wildcard exists and Cloudflare token + operator email are
configured, provisions one via certbot DNS-01
- For services outside the gateway domain, provisions individual certs
- Services with TLS passthrough (k8s) are skipped (backend handles TLS)
- Graceful: skips silently if no Cloudflare token or email configured
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When a service is registered, updated, or deregistered, Central now
automatically regenerates HAProxy routes and dnsmasq DNS entries to
match the current set of registered services.
- L4 tcp-passthrough services (Wild Cloud) → HAProxy SNI routes
- L7 http services (Wild Works) → HAProxy Host-based reverse proxy
- All services with reach != off → dnsmasq DNS entries
- SSE events broadcast on config changes
- Reconciliation runs asynchronously (non-blocking)
This closes the loop: service registration is no longer inert — it
actually updates the networking stack.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Extract the Central networking functionality from wild-cloud/api into
a standalone service. Wild Central manages DNS (dnsmasq), gateway
(HAProxy), firewall (nftables), VPN (WireGuard), TLS (certbot),
security (CrowdSec), and DDNS — all the network appliance concerns.
All Cloud-specific code (instances, clusters, nodes, apps, backups,
operations, kubectl/talosctl tooling) has been removed. The API struct
and route registration contain only Central endpoints. Tests updated
to match the new API signature.
Builds and all tests pass.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
