fix(ui): discover the nginx resolver at run time instead of hard-coding Docker's

docker-compose.yml

@@ -137,16 +137,50 @@ services:
       init-perms:
         condition: service_completed_successfully
 
+  # Renders nginx/default.conf into a volume with the DNS resolver this runtime actually uses.
+  #
+  # It cannot be done in the UI image: that image's entrypoint is `nginx` directly — no
+  # /docker-entrypoint.d/, no /etc/nginx/templates, and it runs as nonroot 65532 — so the usual
+  # envsubst-on-a-template trick is unavailable. nginx itself has never read /etc/resolv.conf for its
+  # `resolver` directive, so the address has to be substituted before nginx starts.
+  #
+  # Asking the network rather than branching on the runtime: any container on this network is handed
+  # the correct nameserver by the runtime itself, so busybox reads its own /etc/resolv.conf and there
+  # is no docker-vs-podman case to keep in sync (verified: 127.0.0.11 on docker, the network gateway
+  # on podman).
+  #
+  # $$ not $: compose interpolates $VAR in `command:` before the shell sees it, and an interpolated-away
+  # variable yields an empty resolver and an nginx that will not load its config at all.
+  init-nginx-conf:
+    image: busybox:1.36
+    command:
+      - sh
+      - -c
+      - |
+        R="$$(awk '/^nameserver/{print $$2; exit}' /etc/resolv.conf)"
+        [ -n "$$R" ] || { echo "no nameserver in /etc/resolv.conf" >&2; exit 1; }
+        sed "s|__RESOLVER__|$$R|" /tpl/default.conf > /out/default.conf
+        chmod 0644 /out/default.conf
+        echo "nginx resolver = $$R"
+    volumes:
+      - ./nginx/default.conf:/tpl/default.conf:ro
+      - nginx_conf:/out
+
   duplo-ui:
     image: quay.io/duplocloud/duplo-ai-helpdesk-ui:${UI_TAG}
     restart: unless-stopped
     ports:
       - "${UI_PORT:-4200}:8080"
     volumes:
-      # Published image serves the SPA only; this conf proxies API prefixes to the studio (same-origin, no CORS).
-      - ./nginx/default.conf:/etc/nginx/conf.d/default.conf:ro
+      # Published image serves the SPA only; the rendered conf proxies API prefixes to the studio
+      # (same-origin, no CORS). The whole directory is replaced: the image ships only default.conf in
+      # there, and tests/test-ui-nginx.sh guards that assumption.
+      - nginx_conf:/etc/nginx/conf.d:ro
     depends_on:
-      - duplo-ai-studio
+      duplo-ai-studio:
+        condition: service_started
+      init-nginx-conf:
+        condition: service_completed_successfully
 
   xterm:
     image: quay.io/duplocloud/duplo-xterm:${XTERM_TAG:-main-d323e0b}
@@ -209,6 +243,7 @@ services:
       - DUPLO_ADMIN_TOKEN=${DUPLO_ADMIN_TOKEN:-}
 
 volumes:
+  nginx_conf:
   mongo_data:
   platform_data:
   extension_studio_data:

nginx/default.conf

@@ -36,9 +36,13 @@ server {
         # of the container, so recreating the studio (any compose change to it OR to mongo, which it
         # depends_on) strands this container proxying to a dead address — every API call 502s and the
         # login page renders with no auth methods, since it is those 502s that decide what to show.
-        # Docker's embedded DNS at 127.0.0.11 plus a variable in proxy_pass forces a fresh lookup.
+        # The resolver address differs per runtime — Docker's embedded DNS is 127.0.0.11, podman's
+        # aardvark-dns sits on the network gateway (10.89.0.x, varying by network) — so it is NOT
+        # hard-coded here. __RESOLVER__ is substituted at startup by the init-nginx-conf service, which
+        # reads the value out of its own /etc/resolv.conf on this same network. Hard-coding either
+        # address serves the SPA fine on the other runtime while every proxied API call fails.
         # $request_uri is required: a variable proxy_pass no longer forwards the URI implicitly.
-        resolver 127.0.0.11 valid=10s ipv6=off;
+        resolver __RESOLVER__ valid=10s ipv6=off;
         set $studio http://duplo-ai-studio:60021;
         proxy_pass $studio$request_uri;
         proxy_http_version 1.1;

tests/test-ui-nginx.sh

@@ -0,0 +1,57 @@
+#!/usr/bin/env bash
+# Checks that the UI's nginx config is runtime-agnostic — specifically that the DNS resolver it proxies
+# through is discovered at run time rather than hard-coded to one container runtime's address.
+#
+# Static checks only: nothing here starts the stack or needs docker/podman installed, so it runs the
+# same on a docker laptop, a podman box and CI.
+set -uo pipefail
+cd "$(dirname "$0")/.." || exit 1
+PASS=0; FAIL=0
+t()   { printf '  %s … ' "$1"; }
+ok()  { echo "ok"; PASS=$((PASS+1)); }
+bad() { echo "FAIL: $1"; FAIL=$((FAIL+1)); }
+
+CONF=nginx/default.conf
+COMPOSE=docker-compose.yml
+
+echo "UI nginx resolver is runtime-agnostic:"
+
+# 127.0.0.11 is Docker's embedded DNS and does not exist under podman, whose aardvark-dns sits on the
+# network gateway (10.89.0.1 on a default install, but it varies by network). Hard-coding either one
+# breaks the other runtime: the SPA still serves, every proxied API call fails, and the login page is
+# simply inert — the worst shape of broken.
+t "nginx conf does not hard-code a runtime-specific DNS address"
+if grep -qE 'resolver[[:space:]]+(127\.0\.0\.11|10\.[0-9]+\.[0-9]+\.[0-9]+)' "$CONF"; then
+  bad "resolver is pinned to one runtime's DNS: $(grep -E '^[[:space:]]*resolver' "$CONF" | tr -d ' ')"
+else ok; fi
+
+t "nginx conf carries a resolver placeholder for substitution"
+grep -qE '^[[:space:]]*resolver[[:space:]]+__RESOLVER__' "$CONF" && ok || bad "no __RESOLVER__ placeholder"
+
+t "compose defines an init service that renders the conf"
+grep -qE '^[[:space:]]*init-nginx-conf:' "$COMPOSE" && ok || bad "no init-nginx-conf service"
+
+# Extract ONLY the service definition, from its own key to the next top-level service. A plain
+# `grep -A<n>` also matches the depends_on reference in duplo-ui and runs off the end into the next
+# service, which made this file's first draft report xterm's ${XTERM_TAG} as an unescaped variable.
+svc_block() { awk '/^  init-nginx-conf:$/{f=1;next} f&&/^  [a-z]/{exit} f' "$COMPOSE"; }
+
+t "the init service discovers DNS from its own resolv.conf, not a literal"
+if svc_block | grep -q '/etc/resolv.conf'; then ok
+else bad "init does not read /etc/resolv.conf"; fi
+
+t "duplo-ui waits for the render to complete before starting"
+if grep -A22 'duplo-ui:' "$COMPOSE" | grep -A2 'init-nginx-conf:' | grep -q 'service_completed_successfully'; then ok
+else bad "duplo-ui does not depend on init-nginx-conf completing"; fi
+
+# Compose interpolates $VAR inside `command:` before the shell ever sees it, so a shell variable must be
+# written $$VAR. Getting this wrong yields an empty resolver and nginx fails to load the conf at all.
+t "shell variables in the init command are escaped for compose interpolation"
+BLOCK="$(svc_block)"
+if printf '%s' "$BLOCK" | grep -qE '(^|[^$])\$[A-Za-z{(]'; then
+  bad "unescaped \$ in the init command — compose will interpolate it away"
+else ok; fi
+
+echo
+echo "$PASS passed, $FAIL failed"
+[ "$FAIL" -eq 0 ]