Replace the pending-approval error page with a dedicated screen (#802)

* Replace the pending-approval error page with a dedicated screen

A user waiting for an administrator to approve their account was shown
the generic /error page: an error icon, a raw response_message dump and
"if you believe this is an error, please contact your administrator".
Waiting for approval is an expected part of signing up, so it now gets a
screen of its own, following the onboarding layout of logo, centered card
and a stepper placing the user between "Account Created" and
"Join Account".

The owner to contact is read out of the refusal management sends, which
names a masked address, and the copy asks for the owner without naming
one when the address is absent.

Steps.Step gains an optional status and size, so a step can render as
complete, current or upcoming at a larger scale. Steps that pass neither
keep the look used by the instruction-list steppers.

* Detect the pending-approval screen by test id in e2e

The three places that recognise a pending user matched on "User Approval
Pending", the heading the old error page rendered. The screen that
replaced it has no heading, so the approval branch of the login races
could never resolve and every leg timed out instead: both fixtures treat
approval as a legitimate outcome, not a failure.

They now key off the data-testid on the screen's root, which does not
move when the copy does.

* Point the pending-approval footer at the docs and support pages

A mailbox asks the reader to compose something; the approval docs explain
what they are waiting for, and the support page routes them the way every
other NetBird surface does. The support address constant went with it.

* Render the pending-approval screen from the caller-identity response

Every call a pending user makes is refused, but only /users/current
names the owner who can approve them. Redirecting to /error on whichever
403 landed first therefore threw that name away about as often as it
kept it, and which screen appeared came down to a race.

Pending users are no longer routed to /error at all. The provider that
already waits on /users/current renders the screen from that response,
which also ends the wait it could never finish: the caller identity it
blocks on is exactly what such a user is refused.

* Decide the refused-user screen from the responses, not the first 403

Routing on whichever refused call landed first meant the screen depended
on a race, and once only /users/current could name the owner it meant
the name was usually thrown away. Management older than that reports a
pending user as merely blocked there, so it also meant a pending user
landed on the blocked screen, which is what the e2e run caught.

UserProfileProvider now decides from the responses it already waits on:
/users/current first, since it alone names the owner, then the user list,
which reports pending on every management version. Blocked keeps the
error page, decided after pending so the older wording cannot win.

* Wait for both refusal responses before choosing a screen

Concluding blocked the moment /users/current said so meant concluding it
while the user list was still in flight — and on management that reports
a pending user as merely blocked there, that list is the one response
that would have said otherwise. The screen it was meant to prevent was
reachable again, just through a narrower window.

The decision moves into resolveRefusedUser, which holds off while a
better-informed response is still coming: the named one before the list,
the list before blocked. Navigating to the error page becomes an effect
rather than a call made while rendering.

* Move the refused-user decision next to the provider that makes it

It sat in utils only because two callers needed it, but the second one —
the guard keeping these errors off the error boundary — needs nothing
more than to recognise them, which is a line of its own. So the decision
moves beside the provider that acts on it, and reads an error into a
RefusalReason rather than assembling a discriminated union.

The cost is that both files now spell out management's wording. The one
here is the tested reading; the one in @utils/api only has to be no
narrower than it.

* Assert the refusal kind the resolver actually returns

The result property is kind, not reason. toEqual types its argument
loosely, so the stale expectations type-checked and only failed when run.

.github/workflows/build_and_push.yml

@@ -98,6 +98,24 @@ jobs:
             type=ref,event=pr
             type=sha
 
+      - name: Docker meta (rootless)
+        id: meta_rootless
+        uses: docker/metadata-action@v5
+        with:
+          images: |
+            ${{ env.DOCKERHUB_IMAGE }}
+            ${{ env.GHCR_IMAGE }}
+          flavor: |
+            latest=false
+            suffix=-rootless,onlatest=true
+          tags: |
+            type=schedule
+            type=ref,event=branch
+            type=ref,event=tag
+            type=raw,value=latest,enable=${{ startsWith(github.ref, 'refs/tags/') }}
+            type=ref,event=pr
+            type=sha
+
       - name: Log in to Docker Hub
         if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
         uses: docker/login-action@v3
@@ -123,7 +141,18 @@ jobs:
           tags: ${{ steps.meta.outputs.tags }}
           labels: ${{ steps.meta.outputs.labels }}
 
+      - name: Build and push rootless Docker image
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          file: docker/Dockerfile.rootless
+          push: ${{ github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository }}
+          platforms: linux/amd64,linux/arm64,linux/arm
+          tags: ${{ steps.meta_rootless.outputs.tags }}
+          labels: ${{ steps.meta_rootless.outputs.labels }}
+
       - if: github.event_name != 'pull_request' || github.event.pull_request.head.repo.full_name == github.repository
         run: |
           echo '### Pushed tags' >> $GITHUB_STEP_SUMMARY
           echo '${{ steps.meta.outputs.tags }}' >> $GITHUB_STEP_SUMMARY
+          echo '${{ steps.meta_rootless.outputs.tags }}' >> $GITHUB_STEP_SUMMARY

build.sh

@@ -11,6 +11,7 @@
 #   ./build.sh                # build + export netbird-dashboard.tar.gz
 #   IMAGE_TAG=v1 ./build.sh   # override the image tag
 #   SKIP_BUILD=1 ./build.sh   # reuse an existing out/ (skip npm build)
+#   ROOTLESS=1 ./build.sh     # build the amd64-rootless image variant
 #
 set -euo pipefail
 
@@ -19,8 +20,26 @@ cd "$(dirname "$0")"
 
 IMAGE_NAME="${IMAGE_NAME:-netbird-dashboard}"
 IMAGE_TAG="${IMAGE_TAG:-amd64}"
+DOCKERFILE="docker/Dockerfile"
+CONTAINER_PORT=80
+OUTPUT_SUFFIX=""
+
+case "${ROOTLESS:-0}" in
+  0) ;;
+  1)
+    DOCKERFILE="docker/Dockerfile.rootless"
+    CONTAINER_PORT=8080
+    OUTPUT_SUFFIX="-rootless"
+    [[ "${IMAGE_TAG}" == *-rootless ]] || IMAGE_TAG="${IMAGE_TAG}-rootless"
+    ;;
+  *)
+    echo "ERROR: ROOTLESS must be 0 or 1." >&2
+    exit 1
+    ;;
+esac
+
 IMAGE="${IMAGE_NAME}:${IMAGE_TAG}"
-OUTPUT="${OUTPUT:-${IMAGE_NAME}.tar.gz}"
+OUTPUT="${OUTPUT:-${IMAGE_NAME}${OUTPUT_SUFFIX}.tar.gz}"
 
 # Build the static export (out/) unless told to reuse an existing one. Next 16
 # no longer runs ESLint during `next build`; types are checked by the build via
@@ -39,17 +58,10 @@ if [[ ! -d out ]]; then
   exit 1
 fi
 
-# The .dockerignore lives under docker/, but Docker reads it from the build
-# context root. Stage it for the duration of the build so node_modules/.next
-# don't get shipped into the build context.
-cleanup() { rm -f .dockerignore; }
-trap cleanup EXIT
-cp docker/.dockerignore .dockerignore
-
 echo "==> Building ${IMAGE} for linux/amd64"
 docker build \
   --platform linux/amd64 \
-  -f docker/Dockerfile \
+  -f "${DOCKERFILE}" \
   -t "${IMAGE}" \
   .
 
@@ -60,13 +72,13 @@ echo "==> Done: ${OUTPUT} ($(du -h "${OUTPUT}" | cut -f1))"
 echo
 echo "Copy and run on the server:"
 echo "  scp ${OUTPUT} user@your-server:/tmp/"
-echo "  docker load < /tmp/${OUTPUT}"
+echo "  docker load < /tmp/${OUTPUT##*/}"
 echo
-echo "Run (nginx serves on container port 80). Set the runtime env vars the"
+echo "Run (nginx serves on container port ${CONTAINER_PORT}). Set the runtime env vars the"
 echo "dashboard templates at startup — for the Agent Network onboarding test,"
 echo "flip NETBIRD_AGENT_NETWORK_ONLY=true:"
 echo "  docker run -d --name ${IMAGE_NAME} --restart unless-stopped \\"
-echo "    -p 8080:80 \\"
+echo "    -p 8080:${CONTAINER_PORT} \\"
 echo "    -e NETBIRD_MGMT_API_ENDPOINT=https://your-mgmt:443 \\"
 echo "    -e AUTH_AUTHORITY=https://your-idp/realms/netbird \\"
 echo "    -e AUTH_CLIENT_ID=netbird-dashboard \\"

config.json

@@ -24,5 +24,6 @@
   "hubspotSignupFormId": "$NETBIRD_HUBSPOT_SIGNUP_FORM_ID",
   "hubspotOnboardingFormId": "$NETBIRD_HUBSPOT_ONBOARDING_FORM_ID",
   "hubspotSurveyFormId": "$NETBIRD_HUBSPOT_SURVEY_FORM_ID",
-  "analyticsExcludedEmails": "$NETBIRD_ANALYTICS_EXCLUDED_EMAILS"
+  "analyticsExcludedEmails": "$NETBIRD_ANALYTICS_EXCLUDED_EMAILS",
+  "announcement": "$NETBIRD_ANNOUNCEMENT"
 }

docker/Dockerfile.rootless

@@ -0,0 +1,29 @@
+FROM alpine:3.24
+
+RUN apk add --no-cache bash ca-certificates gettext nginx tzdata && \
+    rm -rf /var/cache/apk/* && \
+    mkdir -p /tmp/nginx
+
+STOPSIGNAL SIGINT
+EXPOSE 8080
+
+WORKDIR /usr/share/nginx/html
+
+COPY docker/default.conf /etc/nginx/http.d/default.conf
+COPY docker/nginx.conf /etc/nginx/nginx.conf
+COPY docker/rootless.conf /etc/nginx/http.d/rootless.conf
+COPY docker/init_react_envs.sh /usr/local/init_react_envs.sh
+COPY out/ /usr/share/nginx/html/
+
+RUN sed -i \
+      -e 's/listen 80 default_server;/listen 8080 default_server;/' \
+      -e 's/listen \[::\]:80 default_server;/listen [::]:8080 default_server;/' \
+      /etc/nginx/http.d/default.conf && \
+    sed -i '/^user nginx;$/d' /etc/nginx/nginx.conf && \
+    chmod +x /usr/local/init_react_envs.sh && \
+    chgrp -R 0 /etc/nginx/http.d /usr/share/nginx/html /tmp/nginx && \
+    chmod -R g=u /etc/nginx/http.d /usr/share/nginx/html /tmp/nginx
+
+USER 10001:0
+ENTRYPOINT ["/usr/local/init_react_envs.sh"]
+CMD ["/usr/sbin/nginx", "-e", "/dev/stderr", "-g", "pid /tmp/nginx.pid;"]

docker/README.md

@@ -1,23 +1,59 @@
 # NetBird Dashboard
-NetBird Dashboard is NetBirds Management server UI. It allows users to signin, view setup keys and manage peers. This image is **not ready** for production use.
-## Tags
-```latest``` ```vX.X.X``` not available yet.
 
-```main``` built on every PR being merged to the repository
+NetBird Dashboard is NetBird's management server UI.
+
+## Tags
+
+Each published image tag also has a rootless variant with the `-rootless`
+suffix. For example, `main` is also published as `main-rootless`.
+
 ## How to use this image
-HTTP run:
+
+HTTP:
+
 ```shell
-docker run -d --rm -p 80:80 wiretrustee/dashboard:main
+docker run -d --rm -p 80:80 netbirdio/dashboard:main
 ```
-Using SSL certificate from Let's Encrypt®:
+
+Using an SSL certificate from Let's Encrypt:
+
 ```shell
 docker run -d --rm -p 80:80 -p 443:443 \
   -e LETSENCRYPT_DOMAIN=app.mydomain.com \
   -e LETSENCRYPT_EMAIL=hello@mydomain.com \
   netbirdio/dashboard:main
 ```
-> For SSL generation, you need to run this image in a server with proper public IP and a domain name pointing to it.
+
+For certificate generation, the server needs a public IP and a domain name
+pointing to it.
+
+### Rootless and OpenShift
+
+The rootless image listens on port `8080`, runs as UID `10001` by default, and
+supports OpenShift's arbitrary UID when the container retains GID `0`:
+
+```shell
+docker run -d --rm \
+  --user 12345:0 \
+  -p 8080:8080 \
+  -e AUTH_AUTHORITY=https://idp.example.com \
+  -e AUTH_CLIENT_ID=netbird-dashboard \
+  -e AUTH_AUDIENCE=netbird-dashboard \
+  -e AUTH_SUPPORTED_SCOPES='openid profile email' \
+  -e USE_AUTH0=false \
+  -e NETBIRD_MGMT_API_ENDPOINT=https://api.example.com \
+  netbirdio/dashboard:main-rootless
+```
+
+The rootless image does not include Certbot or expose HTTPS. Terminate TLS at
+the OpenShift Route, Ingress, or another reverse proxy.
+
 ## Environment variables
-* ```NGINX_SSL_PORT``` Changes the port that Nginx listens to. Defaults to ```443```
-* ```LETSENCRYPT_DOMAIN``` Enables Certbot`s client execution for the specified domain. Defaults to ```none```
-* ```LETSENCRYPT_EMAIL``` Email used in Certbot`s client execution to register the certificate request. Defaults to ```example@local```
+
+- `NGINX_SSL_PORT` changes the port that Nginx listens to. Defaults to `443`.
+- `LETSENCRYPT_DOMAIN` enables Certbot for the specified domain. Defaults to
+  `none`.
+- `LETSENCRYPT_EMAIL` is used to register the certificate request. Defaults to
+  `example@local`.
+
+These certificate variables apply only to the standard image.

docker/init_react_envs.sh

@@ -73,6 +73,8 @@ export NETBIRD_HUBSPOT_SIGNUP_FORM_ID=${NETBIRD_HUBSPOT_SIGNUP_FORM_ID}
 export NETBIRD_HUBSPOT_ONBOARDING_FORM_ID=${NETBIRD_HUBSPOT_ONBOARDING_FORM_ID}
 export NETBIRD_HUBSPOT_SURVEY_FORM_ID=${NETBIRD_HUBSPOT_SURVEY_FORM_ID}
 export NETBIRD_ANALYTICS_EXCLUDED_EMAILS=${NETBIRD_ANALYTICS_EXCLUDED_EMAILS}
+# A permanent banner for this deployment, shown to every user; empty for none.
+export NETBIRD_ANNOUNCEMENT=${NETBIRD_ANNOUNCEMENT}
 
 echo "NetBird latest version: ${NETBIRD_LATEST_VERSION}"
 
@@ -163,7 +165,7 @@ sed -i "s|add_header Content-Security-Policy \"[^\"]*\" always;|$CSP_HEADER|g" /
 }
 
 # replace ENVs in the config
-ENV_STR="\$\$USE_AUTH0 \$\$AUTH_AUDIENCE \$\$AUTH_AUTHORITY \$\$AUTH_CLIENT_ID \$\$AUTH_CLIENT_SECRET \$\$AUTH_SUPPORTED_SCOPES \$\$NETBIRD_MGMT_API_ENDPOINT \$\$NETBIRD_MGMT_GRPC_API_ENDPOINT \$\$NETBIRD_HOTJAR_TRACK_ID \$\$NETBIRD_GOOGLE_ANALYTICS_ID \$\$NETBIRD_GOOGLE_TAG_MANAGER_ID \$\$AUTH_REDIRECT_URI \$\$AUTH_SILENT_REDIRECT_URI \$\$NETBIRD_TOKEN_SOURCE \$\$NETBIRD_DRAG_QUERY_PARAMS \$\$NETBIRD_AUTH_SERVICE_URL \$\$NETBIRD_WASM_PATH \$\$NETBIRD_LICENSED \$\$NETBIRD_CLOUD \$\$NETBIRD_AGENT_NETWORK_ONLY \$\$NETBIRD_AGENT_NETWORK_ENABLED \$\$NETBIRD_HUBSPOT_PORTAL_ID \$\$NETBIRD_HUBSPOT_SIGNUP_FORM_ID \$\$NETBIRD_HUBSPOT_ONBOARDING_FORM_ID \$\$NETBIRD_HUBSPOT_SURVEY_FORM_ID \$\$NETBIRD_ANALYTICS_EXCLUDED_EMAILS"
+ENV_STR="\$\$USE_AUTH0 \$\$AUTH_AUDIENCE \$\$AUTH_AUTHORITY \$\$AUTH_CLIENT_ID \$\$AUTH_CLIENT_SECRET \$\$AUTH_SUPPORTED_SCOPES \$\$NETBIRD_MGMT_API_ENDPOINT \$\$NETBIRD_MGMT_GRPC_API_ENDPOINT \$\$NETBIRD_HOTJAR_TRACK_ID \$\$NETBIRD_GOOGLE_ANALYTICS_ID \$\$NETBIRD_GOOGLE_TAG_MANAGER_ID \$\$AUTH_REDIRECT_URI \$\$AUTH_SILENT_REDIRECT_URI \$\$NETBIRD_TOKEN_SOURCE \$\$NETBIRD_DRAG_QUERY_PARAMS \$\$NETBIRD_AUTH_SERVICE_URL \$\$NETBIRD_WASM_PATH \$\$NETBIRD_LICENSED \$\$NETBIRD_CLOUD \$\$NETBIRD_AGENT_NETWORK_ONLY \$\$NETBIRD_AGENT_NETWORK_ENABLED \$\$NETBIRD_HUBSPOT_PORTAL_ID \$\$NETBIRD_HUBSPOT_SIGNUP_FORM_ID \$\$NETBIRD_HUBSPOT_ONBOARDING_FORM_ID \$\$NETBIRD_HUBSPOT_SURVEY_FORM_ID \$\$NETBIRD_ANALYTICS_EXCLUDED_EMAILS \$\$NETBIRD_ANNOUNCEMENT"
 
 OIDC_TRUSTED_DOMAINS="/usr/share/nginx/html/OidcTrustedDomains.js"
 envsubst "$ENV_STR" < "$OIDC_TRUSTED_DOMAINS".tmpl > "$OIDC_TRUSTED_DOMAINS"
@@ -173,6 +175,12 @@ for f in $(grep -R -l AUTH_SUPPORTED_SCOPES /usr/share/nginx/html); do
     rm "$f".copy
 done
 
+# When a command is provided, initialization happens before the server starts.
+# The rootless image uses this path so nginx never needs to be reloaded.
+if (( $# > 0 )); then
+    exec "$@"
+fi
+
 # Reload nginx so the patched CSP header takes effect.
 # supervisord starts nginx (priority 100) before this script (priority 201) and
 # never reloads it, so without this nginx keeps serving the static default.conf

docker/rootless.conf

@@ -0,0 +1,7 @@
+# nginx includes this file inside the http context.
+# All runtime state stays in a directory writable by the arbitrary UID used by OpenShift.
+client_body_temp_path /tmp/nginx/client_temp;
+proxy_temp_path /tmp/nginx/proxy_temp;
+fastcgi_temp_path /tmp/nginx/fastcgi_temp;
+uwsgi_temp_path /tmp/nginx/uwsgi_temp;
+scgi_temp_path /tmp/nginx/scgi_temp;

e2e/environment/.gitignore

@@ -1,6 +1,7 @@
 # Ignore zitadel environment
 .env
 Caddyfile
+agentgateway-stub.Caddyfile
 management.json
 turnserver.conf
 zitadel.env

e2e/environment/clean-test-env.sh

@@ -19,7 +19,7 @@ check_docker_compose() {
 DOCKER_COMPOSE_COMMAND=$(check_docker_compose)
 
 $DOCKER_COMPOSE_COMMAND down --volumes
-rm -f docker-compose.yml Caddyfile zitadel.env .env dashboard.env machinekey/zitadel-admin-sa.token turnserver.conf management.json proxy.env proxy-no-ports.env
+rm -f docker-compose.yml Caddyfile agentgateway-stub.Caddyfile zitadel.env .env dashboard.env machinekey/zitadel-admin-sa.token turnserver.conf management.json proxy.env proxy-no-ports.env
 rm -rf proxy-certs proxy-certs-no-ports
 rm -f ../../.test-config.json ../playwright.env.json
 rm -f ../fixtures/auth/owner.json ../fixtures/auth/user.json

e2e/environment/create-test-env.sh

@@ -532,7 +532,7 @@ initEnvironment() {
     echo "Generated files already exist, if you want to reinitialize the environment, please remove them first."
     echo "You can use the following commands:"
     echo "  $DOCKER_COMPOSE_COMMAND down --volumes # to remove all containers and volumes"
-    echo "  rm -f docker-compose.yml Caddyfile zitadel.env dashboard.env machinekey/zitadel-admin-sa.token turnserver.conf management.json proxy.env proxy-no-ports.env && rm -rf proxy-certs proxy-certs-no-ports"
+    echo "  rm -f docker-compose.yml Caddyfile agentgateway-stub.Caddyfile zitadel.env dashboard.env machinekey/zitadel-admin-sa.token turnserver.conf management.json proxy.env proxy-no-ports.env && rm -rf proxy-certs proxy-certs-no-ports"
     echo "Be aware that this will remove all data from the database, and you will have to reconfigure the dashboard."
     exit 1
   fi
@@ -540,6 +540,7 @@ initEnvironment() {
   echo Rendering initial files...
   renderDockerCompose > docker-compose.yml
   renderCaddyfile > Caddyfile
+  renderAgentGatewayStub > agentgateway-stub.Caddyfile
   renderZitadelEnv > zitadel.env
   echo "" > turnserver.conf
   echo "" > management.json
@@ -612,6 +613,30 @@ initEnvironment() {
   echo "Password: $ZITADEL_ADMIN_PASSWORD" | tee -a .env
 }
 
+# Saving an Agent Network provider makes management call the upstream and
+# refuses the save unless it answers a model listing, so a provider spec needs
+# a reachable OpenAI-compatible endpoint. agentgateway is self-hosted and has
+# no public one; this stub stands in for it, inside the compose network where
+# management can reach it and without depending on a vendor being up.
+renderAgentGatewayStub() {
+  cat <<'EOF'
+{
+	admin off
+	auto_https off
+}
+
+:8088 {
+	handle /v1/models {
+		header Content-Type application/json
+		respond `{"object":"list","data":[{"id":"agentgateway-e2e-model","object":"model","owned_by":"netbird-e2e"}]}` 200
+	}
+	handle {
+		respond 404
+	}
+}
+EOF
+}
+
 renderCaddyfile() {
   cat <<EOF
 {
@@ -818,6 +843,15 @@ services:
     volumes:
       - netbird_caddy_data:/data
       - ./Caddyfile:/etc/caddy/Caddyfile
+  # Stand-in upstream for the Agent Network provider specs: answers the model
+  # listing management probes before it accepts a provider save. Reachable
+  # only from inside the compose network, at http://agentgateway-stub:8088.
+  agentgateway-stub:
+    image: caddy
+    restart: unless-stopped
+    networks: [ netbird ]
+    volumes:
+      - ./agentgateway-stub.Caddyfile:/etc/caddy/Caddyfile
   # Management
   management:
     image: ghcr.io/netbirdio/management-cloud:${MANAGEMENT_IMAGE_TAG}
@@ -892,6 +926,16 @@ services:
       # GeoLite2 download removes a startup stall of up to 2 minutes
       # when pkgs.netbird.io is slow.
       - NB_PROXY_DISABLE_GEOLOCATION=true
+      # Declare the private capability so this cluster can host the agent
+      # network endpoint. That endpoint is a private service — reachable only
+      # from connected peers — and management refuses to bootstrap it onto a
+      # cluster without a private-capable proxy; the connect-provider wizard
+      # reads the same flag (supports_private) to pick its cluster. Without
+      # it no cluster qualifies and the wizard stops at the provider tab,
+      # which strands every spec that walks through it. The no-ports proxy
+      # below stays centralised, so the pick has to skip past a cluster that
+      # does not qualify.
+      - NB_PROXY_PRIVATE=true
     volumes:
       - ./proxy-certs:/certs:ro
     command: [

e2e/helpers/auth.ts

@@ -33,7 +33,7 @@ export async function loginToApp(
   // Use locators that match either outcome — Playwright auto-waits.
   const appReady = page.getByTestId("left-navigation-item").first();
   const setupModal = page.getByTestId("setup-netbird-modal");
-  const approvalPending = page.getByText("User Approval Pending");
+  const approvalPending = page.getByTestId("pending-approval");
   const onboarding = page.getByText("Add new device to your network");
   const selectAccount = page.getByText("Select account");
   const loginInput = page.locator("input[id=loginName]");

e2e/helpers/control-center.ts

@@ -24,6 +24,31 @@ export async function openControlCenter(page: Page, tab?: FlowView) {
   await dismissBlockingOverlays(page);
 }
 
+/**
+ * Opens the groups view already showing `groupName`.
+ *
+ * The view picks its own group on first render (getFirstGroup: the highest
+ * peers_count that is a policy source), and that ranking runs over every group
+ * in the account — including ones a spec on the other worker seeded moments
+ * ago. Deleting the spec's own fixtures first is not enough to pin it, so
+ * select the group by name and let the view rebuild around it.
+ */
+export async function openGroupView(page: Page, groupName: string) {
+  await openControlCenter(page, "groups");
+  const selector = canvasNode(page, "select-group-node");
+  await expect(selector).toBeVisible();
+  // Unforced on purpose: the opening fitView animates the canvas, and the
+  // actionability check is what waits for the node to stop moving.
+  await selector.click();
+  await page.getByTestId("select-dropdown-search").fill(groupName);
+  await page
+    .locator('[role="option"]')
+    .filter({ hasText: groupName })
+    .first()
+    .click({ force: true });
+  await expect(selector).toContainText(groupName);
+}
+
 export async function switchFlowView(page: Page, view: FlowView) {
   await dismissBlockingOverlays(page);
   await page.getByTestId(`cc-flow-${view}`).click({ force: true });

e2e/tests/agent-network-agentgateway-provider.spec.ts

@@ -59,7 +59,12 @@ test.describe
         .getByTestId("agent-network-provider-option-agentgateway")
         .click({ force: true });
 
-      const upstreamURL = "https://agentgateway.e2e.example";
+      // Saving a provider makes management call the upstream, and it refuses
+      // the save with a 422 unless the upstream answers a model listing — so a
+      // made-up host cannot be used here. agentgateway is self-hosted and has
+      // no public endpoint to point at, so create-test-env.sh runs a stub that
+      // answers that probe inside the compose network.
+      const upstreamURL = "http://agentgateway-stub:8088";
       await page
         .getByTestId("agent-network-provider-upstream-url")
         .fill(upstreamURL);

e2e/tests/agent-network-bootstrap-cluster.spec.ts

@@ -0,0 +1,206 @@
+/**
+ * Agent Network bootstrap-cluster spec.
+ *
+ * The account's agent network endpoint is assigned once, from the proxy
+ * cluster the connect-provider wizard pins it to, and it is immutable after
+ * that. Not every live cluster can host it: the endpoint is a private service,
+ * reachable only from connected peers and authorised by their tunnel identity,
+ * which only a cluster with private capabilities can serve. Management
+ * reports that per cluster as supports_private and refuses a
+ * bootstrap onto a cluster without it.
+ *
+ * This spec pins the wizard's half of that contract: it picks a
+ * private-capable cluster even when a centralised one is listed first, and it
+ * blocks setup — with a message naming what is missing — when no listed
+ * cluster qualifies. Both the settings and domains routes are mocked, so the
+ * assertions do not depend on which clusters the test backend happens to have
+ * connected, and no provider is created against it.
+ */
+import { type Browser, expect, type Page, test } from "@playwright/test";
+import { loginToApp } from "../helpers/auth";
+
+const AGENT_NETWORK_CONFIG_KEY = "netbird-test-agent-network";
+
+// Listed first, and unusable: a centralised proxy cluster. Picking it is the
+// bug this spec guards against.
+const CENTRAL_CLUSTER = "central.proxy.netbird.io";
+// Listed second, and the only valid pick.
+const PRIVATE_CLUSTER = "private.proxy.netbird.io";
+
+const UNBOOTSTRAPPED_DEFAULTS = {
+  endpoint: "",
+  proxy_address: "",
+  dedicated: false,
+  enable_log_collection: true,
+  enable_prompt_collection: false,
+  redact_pii: false,
+  access_log_retention_days: 30,
+};
+
+const BOOTSTRAPPED_SETTINGS = {
+  ...UNBOOTSTRAPPED_DEFAULTS,
+  endpoint: `violet.${PRIVATE_CLUSTER}`,
+  proxy_address: PRIVATE_CLUSTER,
+  created_at: "2026-01-01T00:00:00Z",
+  updated_at: "2026-01-01T00:00:00Z",
+};
+
+/** A free (cluster-backed) domain as /reverse-proxies/domains reports it. */
+function freeDomain(name: string, supportsPrivate: boolean) {
+  return {
+    id: name,
+    domain: name,
+    validated: true,
+    type: "free",
+    supports_private: supportsPrivate,
+    supports_custom_ports: true,
+    require_subdomain: true,
+  };
+}
+
+async function json(route: import("@playwright/test").Route, body: unknown) {
+  return route.fulfill({
+    status: 200,
+    contentType: "application/json",
+    body: JSON.stringify(body),
+  });
+}
+
+/**
+ * Opens the providers page for an account that has not been bootstrapped,
+ * with the given clusters on the domains endpoint. Returns the page plus a
+ * promise-friendly getter for the bootstrap POST body, which is captured and
+ * answered locally so the real backend is never mutated.
+ */
+async function openWizard(
+  browser: Browser,
+  domains: ReturnType<typeof freeDomain>[],
+): Promise<{ page: Page; close: () => Promise<void> }> {
+  const context = await browser.newContext({
+    storageState: "e2e/fixtures/auth/owner.json",
+  });
+  try {
+    await context.addInitScript(
+      ([key, value]) => {
+        try {
+          window.localStorage.setItem(key as string, value as string);
+        } catch (e) {}
+      },
+      [AGENT_NETWORK_CONFIG_KEY, "enabled"],
+    );
+    const page = await context.newPage();
+
+    await page.route("**/api/agent-network/settings", (route) => {
+      const method = route.request().method();
+      if (method === "GET") return json(route, UNBOOTSTRAPPED_DEFAULTS);
+      if (method === "POST") return json(route, BOOTSTRAPPED_SETTINGS);
+      return route.continue();
+    });
+    await page.route("**/api/reverse-proxies/domains", (route) =>
+      route.request().method() === "GET"
+        ? json(route, domains)
+        : route.continue(),
+    );
+    // Keep the create hermetic: the wizard only reaches this once the
+    // bootstrap it does first has succeeded.
+    await page.route("**/api/agent-network/providers", (route) =>
+      route.request().method() === "POST"
+        ? json(route, { id: "prov-e2e", name: "e2e-provider" })
+        : route.continue(),
+    );
+
+    await loginToApp(page, "owner");
+    await page.goto("/agent-network/providers");
+    await page.keyboard.press("Escape");
+    return { page, close: () => context.close() };
+  } catch (e) {
+    await context.close();
+    throw e;
+  }
+}
+
+/** Fills the provider tab with the minimum a create needs. */
+async function fillProviderTab(page: Page) {
+  await page
+    .getByPlaceholder("https://api.openai.com")
+    .fill("https://api.openai.com");
+  await page.getByPlaceholder("sk-...").first().fill("sk-e2e-bootstrap-key");
+}
+
+test.describe("Agent Network bootstrap cluster @agent-network", () => {
+  test("bootstraps onto a private-capable cluster, skipping a centralised one listed first", async ({
+    browser,
+  }) => {
+    const { page, close } = await openWizard(browser, [
+      freeDomain(CENTRAL_CLUSTER, false),
+      freeDomain(PRIVATE_CLUSTER, true),
+    ]);
+    try {
+      await page
+        .getByRole("button", { name: "Connect Provider" })
+        .first()
+        .click({ force: true });
+
+      // The unusable cluster is listed, but it is not a blocker: a usable one
+      // exists, so the wizard proceeds without the warning.
+      await expect(
+        page.getByTestId("agent-network-no-cluster-callout"),
+      ).toHaveCount(0);
+
+      await fillProviderTab(page);
+
+      const bootstrap = page.waitForResponse(
+        (resp) =>
+          resp.url().includes("/agent-network/settings") &&
+          resp.request().method() === "POST",
+        { timeout: 30_000 },
+      );
+
+      await page
+        .getByRole("button", { name: "Continue" })
+        .click({ force: true });
+      await page
+        .getByRole("button", { name: "Connect Provider" })
+        .last()
+        .click({ force: true });
+
+      const payload = (await bootstrap).request().postDataJSON();
+      expect(payload.proxy_address).toBe(PRIVATE_CLUSTER);
+    } finally {
+      await close();
+    }
+  });
+
+  test("blocks setup when no listed cluster has private capabilities", async ({
+    browser,
+  }) => {
+    const { page, close } = await openWizard(browser, [
+      freeDomain(CENTRAL_CLUSTER, false),
+    ]);
+    try {
+      await page
+        .getByRole("button", { name: "Connect Provider" })
+        .first()
+        .click({ force: true });
+
+      const callout = page.getByTestId("agent-network-no-cluster-callout");
+      await expect(callout).toBeVisible({ timeout: 10_000 });
+      await expect(callout).toContainText("private capabilities");
+      // The fix lives in the agent network's own Clusters section, not the
+      // generic Reverse Proxy page, so that is where the callout sends users.
+      await expect(callout.getByRole("link")).toHaveAttribute(
+        "href",
+        "/agent-network/configuration?tab=clusters",
+      );
+
+      // With nothing to pin to there is no bootstrap to make, so the wizard
+      // cannot advance — rather than sending a request management refuses.
+      await fillProviderTab(page);
+      await expect(
+        page.getByRole("button", { name: "Continue" }),
+      ).toBeDisabled();
+    } finally {
+      await close();
+    }
+  });
+});

e2e/tests/control-center-draft-matrix.spec.ts

@@ -364,13 +364,17 @@ test.describe.serial("Control Center Draft Matrix @control-center", () => {
   }) => {
     const group = await place(page, "group", 0.6, 0.5);
     await expectChangeCount(page, 1);
-    // React Flow ignores delete keys while an input is focused, and the panel's
-    // search takes focus on click, so blur it before pressing Backspace.
     await group.click();
     await expect(group).toHaveClass(/selected/);
-    await page.evaluate(
-      () => (document.activeElement as HTMLElement | null)?.blur(),
-    );
+    // React Flow ignores delete keys while an input is focused, and the panel
+    // the click opens takes focus into its search — from an effect, about
+    // 100ms later. Blurring before that lands leaves the input to claim focus
+    // afterwards and swallow the Backspace, so wait for the focus, then drop
+    // it.
+    const panelSearch = page.locator("#cc-group-panel input").first();
+    await expect(panelSearch).toBeFocused();
+    await panelSearch.blur();
+    await expect(panelSearch).not.toBeFocused();
     await page.keyboard.press("Backspace");
     await expect(canvasNode(page, "group-new-")).toHaveCount(0);
     await expectChangeCount(page, 0);

e2e/tests/control-center-draft-remove.spec.ts

@@ -19,6 +19,7 @@ import {
   dismissBlockingOverlays,
   enterDraft,
   openControlCenter,
+  openGroupView,
   resetDraftState,
   reviewButton,
 } from "../helpers/control-center";
@@ -37,7 +38,6 @@ test.describe
     await deleteGroupsByPrefix(page, PREFIX);
   });
 
-  // Clean-slated so the draft group view's auto-select is deterministic.
   async function seedPolicyGroupView(page: Page) {
     await deletePoliciesBySubstring(page, PREFIX);
     await deleteGroupsByPrefix(page, PREFIX);
@@ -49,7 +49,9 @@ test.describe
       src.id,
       dst.id,
     );
-    await openControlCenter(page, "groups");
+    // Selected by name: the view's own pick ranks every group in the account,
+    // so a peer-bearing group from another spec on the other worker wins it.
+    await openGroupView(page, src.name);
     await enterDraft(page);
     const policyNode = canvasNode(page, `policy-${policy.id}`);
     await expect(policyNode).toBeVisible({ timeout: 15_000 });

e2e/tests/control-center-live.spec.ts

@@ -18,6 +18,7 @@ import {
   clickContextMenuItem,
   dismissBlockingOverlays,
   openControlCenter,
+  openGroupView,
   resetDraftState,
   switchFlowView,
 } from "../helpers/control-center";
@@ -47,8 +48,6 @@ test.describe.serial("Control Center Live Mode @control-center", () => {
   });
 
   async function seedPolicyAndOpenGroupView(page: Page, enabled = true) {
-    // The group view auto-selects a group that HAS a policy, so stale policies
-    // would make the rendered group non-deterministic.
     await deletePoliciesBySubstring(page, PREFIX);
     await deleteGroupsByPrefix(page, PREFIX);
 
@@ -64,15 +63,10 @@ test.describe.serial("Control Center Live Mode @control-center", () => {
     );
 
     const policyNode = canvasNode(page, `policy-${policy.id}`);
-    // A stale SWR read or a build/fit race can fail to place the node.
-    for (let attempt = 0; attempt < 3; attempt++) {
-      await openControlCenter(page, "groups");
-      const shown = await policyNode
-        .waitFor({ state: "visible", timeout: 10_000 })
-        .then(() => true)
-        .catch(() => false);
-      if (shown) break;
-    }
+    // Selecting src by name rather than trusting the view's own pick: the
+    // auto-select ranks every group in the account, so a peer-bearing group
+    // another spec seeded on the other worker outranks this one.
+    await openGroupView(page, src.name);
     await expect(policyNode).toBeVisible({ timeout: 10_000 });
     return { src, dst, policy, policyNode };
   }
@@ -512,7 +506,6 @@ test.describe.serial("Control Center Live Mode @control-center", () => {
   test("Should save group resource membership in live mode (correct payload)", async ({
     dashboardAsOwner: page,
   }) => {
-    // Clean slate so the group view deterministically lands on our group.
     await deletePoliciesBySubstring(page, PREFIX);
     await deleteGroupsByPrefix(page, PREFIX);
 
@@ -529,7 +522,7 @@ test.describe.serial("Control Center Live Mode @control-center", () => {
       [],
     );
 
-    await openControlCenter(page, "groups");
+    await openGroupView(page, src.name);
     const groupNode = canvasNode(page, `group-${dst.id}`);
     await expect(groupNode).toBeVisible({ timeout: 15_000 });
 
@@ -610,7 +603,7 @@ test.describe.serial("Control Center Live Mode @control-center", () => {
     );
     await createPolicy(page, generateRandomName(PREFIX), src.id, d2.id);
 
-    await openControlCenter(page, "groups");
+    await openGroupView(page, src.name);
     const policyNode = canvasNode(page, `policy-${p1.id}`);
     await expect(policyNode).toBeVisible({ timeout: 15_000 });
     await expect(

e2e/tests/login.spec.ts

@@ -102,7 +102,7 @@ async function loginAndSave(
   const skipButton = page.locator("button[name=skip]");
   const appNav = page.getByTestId("left-navigation-item").first();
   const modal = page.getByTestId("setup-netbird-modal");
-  const approval = page.getByText("User Approval Pending");
+  const approval = page.getByTestId("pending-approval");
 
   let after_login: "2fa" | "app" | "modal" | "approval";
   try {

e2e/tests/team-users-approval-and-billing.spec.ts

@@ -20,7 +20,7 @@ test.describe.serial("User Approval & Billing Admin @team", () => {
     });
     const page = await context.newPage();
     await loginToApp(page, "user");
-    await expect(page.getByText("User Approval Pending")).toBeVisible();
+    await expect(page.getByTestId("pending-approval")).toBeVisible();
     await context.close();
   });