route runner build container clones via public URL with CF Access

The Gitea Actions runners on Legion need build containers to be able to clone repos. They run with network: host so they can't resolve gitea.git.svc.cluster.local — they have to use the public URL. The public URL sits behind Cloudflare Access, which is why the previous naive switch (#3) had to be reverted. This change keeps the runner daemon registered against the in-cluster URL (no CF Access on the polling loop) but rewrites http://gitea.git.svc.cluster.local:3000/ to https://git.neuralplatform.ai/ inside the build container, with the CF Access service-token headers injected via git extraHeader. The redirect script is sourced before every job step via BASH_ENV; the CF Access credentials reach the build container through act_runner's container.env, which we now populate from the new gitea-runner-cf-access Vault path that PR #5's Terraform writes. Known limitation documented in the init script: actions/checkout's per-job auth header is keyed to the in-cluster URL and gets dropped after the insteadOf rewrite. Public repos work; private repos that need that token will need a follow-up.
2026-05-04 16:19:35 -05:00
6 changed files with 196 additions and 91 deletions
@@ -0,0 +1,165 @@
+---
+# Gitea CI runner — general-purpose (legion)
+# Uses host Docker socket for container management and docker build/push.
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: gitea-runner
+  namespace: ci
+  labels:
+    app: gitea-runner
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      app: gitea-runner
+  template:
+    metadata:
+      labels:
+        app: gitea-runner
+      annotations:
+        config-version: "2026-04-27-containerd-sock"
+    spec:
+      securityContext:
+        runAsNonRoot: false  # act_runner needs root for container management
+      initContainers:
+        - name: register
+          image: registry.neuralplatform.ai/ci-base:latest
+          workingDir: /data
+          command: ["/bin/sh", "-c"]
+          args:
+            - |
+              act_runner register \
+                --instance "$GITEA_INSTANCE_URL" \
+                --token "$GITEA_RUNNER_REGISTRATION_TOKEN" \
+                --name legion \
+                --labels "self-hosted:docker://registry.neuralplatform.ai/ci-base:latest,ubuntu-latest:docker://registry.neuralplatform.ai/ci-base:latest,ubuntu-24.04:docker://registry.neuralplatform.ai/ci-base:latest,linux,x64" \
+                --no-interactive
+              cat > /data/config.yaml << 'EOF'
+              runner:
+                capacity: 2
+                timeout: 3h
+              container:
+                network: host
+                docker_host: "unix:///run/k3s/containerd/containerd.sock"
+                force_pull: false
+                valid_volumes: []
+                default_image: "registry.neuralplatform.ai/ci-base:latest"
+                extra_hosts:
+                  - "gitea.git.svc.cluster.local:10.43.1.53"
+              EOF
+          envFrom:
+            - secretRef:
+                name: gitea-runner-secret
+          volumeMounts:
+            - name: data
+              mountPath: /data
+      containers:
+        - name: runner
+          image: registry.neuralplatform.ai/ci-base:latest
+          workingDir: /data
+          command: ["act_runner", "daemon", "--config", "/data/config.yaml"]
+          envFrom:
+            - secretRef:
+                name: gitea-runner-secret
+          volumeMounts:
+            - name: data
+              mountPath: /data
+            - name: docker-sock
+              mountPath: /var/run/docker.sock
+          resources:
+            requests:
+              memory: 512Mi
+              cpu: 250m
+            limits:
+              memory: 4Gi
+              cpu: "4"
+      volumes:
+        - name: data
+          emptyDir: {}
+        - name: docker-sock
+          hostPath:
+            path: /run/k3s/containerd/containerd.sock
+            type: Socket
+---
+# Neuron Technologies CI runner
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: neuron-technologies-runner
+  namespace: ci
+  labels:
+    app: neuron-technologies-runner
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      app: neuron-technologies-runner
+  template:
+    metadata:
+      labels:
+        app: neuron-technologies-runner
+      annotations:
+        config-version: "2026-04-27-containerd-sock"
+    spec:
+      securityContext:
+        runAsNonRoot: false
+      initContainers:
+        - name: register
+          image: registry.neuralplatform.ai/ci-base:latest
+          workingDir: /data
+          command: ["/bin/sh", "-c"]
+          args:
+            - |
+              act_runner register \
+                --instance "$GITEA_INSTANCE_URL" \
+                --token "$GITEA_RUNNER_REGISTRATION_TOKEN" \
+                --name neuron-technologies \
+                --labels "self-hosted:docker://registry.neuralplatform.ai/ci-base:latest,ubuntu-latest:docker://registry.neuralplatform.ai/ci-base:latest,ubuntu-24.04:docker://registry.neuralplatform.ai/ci-base:latest,linux,x64" \
+                --no-interactive
+              cat > /data/config.yaml << 'EOF'
+              runner:
+                capacity: 2
+                timeout: 3h
+              container:
+                network: host
+                docker_host: "unix:///run/k3s/containerd/containerd.sock"
+                force_pull: false
+                valid_volumes: []
+                default_image: "registry.neuralplatform.ai/ci-base:latest"
+                extra_hosts:
+                  - "gitea.git.svc.cluster.local:10.43.1.53"
+              EOF
+          envFrom:
+            - secretRef:
+                name: neuron-technologies-runner-secret
+          volumeMounts:
+            - name: data
+              mountPath: /data
+      containers:
+        - name: runner
+          image: registry.neuralplatform.ai/ci-base:latest
+          workingDir: /data
+          command: ["act_runner", "daemon", "--config", "/data/config.yaml"]
+          envFrom:
+            - secretRef:
+                name: neuron-technologies-runner-secret
+          volumeMounts:
+            - name: data
+              mountPath: /data
+            - name: docker-sock
+              mountPath: /var/run/docker.sock
+          resources:
+            requests:
+              memory: 512Mi
+              cpu: 250m
+            limits:
+              memory: 4Gi
+              cpu: "4"
+      volumes:
+        - name: data
+          emptyDir: {}
+        - name: docker-sock
+          hostPath:
+            path: /run/k3s/containerd/containerd.sock
+            type: Socket
@@ -23,16 +23,3 @@ resource "cloudflare_record" "np_web_stage" {
  proxied = true
  ttl     = 1
 }
-
-# vault.neuralplatform.ai — GCE Raft HA Vault cluster via GCP Global HTTPS LB.
-# DNS-only (not proxied) — GCP managed TLS cert terminates at the LB.
-# Vault nodes listen on plain HTTP 8200 internally; LB does TLS.
-# IP: terraform output vault_lb_ip from servers/gcp workspace = 34.54.164.21
-resource "cloudflare_record" "np_vault" {
-  zone_id = local.zone_neuralplatform_ai
-  name    = "vault"
-  type    = "A"
-  content = "34.54.164.21"
-  proxied = false
-  ttl     = 60
-}
@@ -1,42 +0,0 @@
-# Cloudflare Zero Trust Access — git.neuralplatform.ai (Gitea)
-#
-# The Gitea Access application itself is currently managed in the Cloudflare
-# dashboard, NOT in Terraform. This file only manages the *service token* the
-# Gitea Actions runners use to authenticate through CF Access while still
-# keeping the human Google-OAuth gate for browser users.
-#
-# Why not import the application here?
-#   - Importing the existing dashboard app risks drifting the human-auth
-#     policy (Google IdP, allowed emails) which is settled and working.
-#   - Service tokens can be added to a dashboard-managed app without
-#     importing the app itself; the token resource lives at the account
-#     level and is referenced from a policy.
-#   - We pay only the cost we need to. If we later want all Access apps
-#     in TF we can do a focused import pass.
-#
-# After `terraform apply` produces the token id/secret, Will must:
-#   1. Run `vault kv put secret/gitea-runner-cf-access ...` (see outputs).
-#   2. In the Cloudflare dashboard, edit the existing "Gitea" Access
-#      application's policies and add a new policy:
-#         Action:  Service Auth (decision = non_identity)
-#         Include: Service Token = "gitea-runner"
-#      This grants the service token bypass through CF Access on
-#      git.neuralplatform.ai without changing the human-auth flow.
-
-resource "cloudflare_zero_trust_access_service_token" "gitea_runner" {
-  account_id = var.cloudflare_account_id
-  name       = "gitea-runner"
-  # Default duration is "8760h" (1 year). Rotate via re-apply when needed.
-  duration = "forever"
-}
-
-output "gitea_runner_cf_access_client_id" {
-  description = "CF Access service token client ID for the Gitea Actions runner. Store in Vault at secret/gitea-runner-cf-access."
-  value       = cloudflare_zero_trust_access_service_token.gitea_runner.client_id
-}
-
-output "gitea_runner_cf_access_client_secret" {
-  description = "CF Access service token client secret. Store in Vault at secret/gitea-runner-cf-access. Only emitted at creation time."
-  value       = cloudflare_zero_trust_access_service_token.gitea_runner.client_secret
-  sensitive   = true
-}
@@ -82,12 +82,7 @@ spec:
    apiVersion: apps/v1
    kind: Deployment
    name: neuron-marketing
-  # minReplicas=1 to match the file's own convention (see header comment).
-  # Kubernetes only allows minReplicas=0 when at least one Object or External
-  # metric is configured (queue depth, custom signal, etc.); with only a
-  # Resource (CPU) metric, scale-to-zero is rejected and the whole HPA is
-  # invalid — which was blocking neuron-prod's Argo CD sync.
-  minReplicas: 1
+  minReplicas: 0
  maxReplicas: 8
  metrics:
    - type: Resource
@@ -117,32 +117,6 @@ spec:
            matchLabels:
              kubernetes.io/metadata.name: neuron-prod
 ---
-# ── dharma: accept from Traefik (kube-system) and neuron-prod namespace ──────
-# The dharma pod was healthy and the IngressRoute was correct, but cross-
-# namespace ingress from kube-system (Traefik) was denied by default-deny-all,
-# so every external request landed at Traefik and bounced back as 502. This
-# allow rule mirrors `allow-mcp-ingress` and brings dharma into line with the
-# other neuron-prod services.
-apiVersion: networking.k8s.io/v1
-kind: NetworkPolicy
-metadata:
-  name: allow-dharma-ingress
-  namespace: neuron-prod
-spec:
-  podSelector:
-    matchLabels:
-      app: dharma
-  policyTypes:
-    - Ingress
-  ingress:
-    - from:
-        - namespaceSelector:
-            matchLabels:
-              kubernetes.io/metadata.name: kube-system
-        - namespaceSelector:
-            matchLabels:
-              kubernetes.io/metadata.name: neuron-prod
---
 # ── Egress: all prod pods may reach platform (postgres/redis), vault,
 #            monitoring (alloy OTLP), kube-dns, and the internet (external APIs) ─
 apiVersion: networking.k8s.io/v1
@@ -92,11 +92,37 @@ resource "cloudflare_zero_trust_tunnel_cloudflared_config" "legion" {
      }
    }

-    # vault.neuralplatform.ai — moved to GCP Global HTTPS LB (34.54.164.21)
-    # DNS is now a direct A record (not proxied) in dns-neuralplatform.tf
+    ingress_rule {
+      hostname = "vault.neuralplatform.ai"
+      service  = "https://traefik.kube-system.svc:443"
+      origin_request {
+        no_tls_verify = true
+      }
+    }

-    # watch.nook.family, jellyfin.nook.family, bazarr.nook.family — removed
-    # This infrastructure is focused on Neuron; nook.family media stack retired
+    ingress_rule {
+      hostname = "watch.nook.family"
+      service  = "https://traefik.kube-system.svc:443"
+      origin_request {
+        no_tls_verify = true
+      }
+    }
+
+    ingress_rule {
+      hostname = "jellyfin.nook.family"
+      service  = "https://traefik.kube-system.svc:443"
+      origin_request {
+        no_tls_verify = true
+      }
+    }
+
+    ingress_rule {
+      hostname = "bazarr.nook.family"
+      service  = "https://traefik.kube-system.svc:443"
+      origin_request {
+        no_tls_verify = true
+      }
+    }


    # fornax.neuralplatform.ai — Fornax torrent coordinator (qBittorrent API proxy)