From 67263cb75c5410813dbf3a0aae46f393fca803d4 Mon Sep 17 00:00:00 2001
From: Sergei Poljanski <me@asxp.io>
Date: Thu, 2 Jul 2026 18:32:53 +0400
Subject: [PATCH] ci: gate build on tests, deploy waits for container health

---
 .forgejo/workflows/deploy.yaml | 53 +++++++++++++++++++++++++++++++---
 .forgejo/workflows/test.yaml   |  6 ++--
 CLAUDE.md                      |  4 ++-
 3 files changed, 56 insertions(+), 7 deletions(-)

diff --git a/.forgejo/workflows/deploy.yaml b/.forgejo/workflows/deploy.yaml
index 1cc2673..2309836 100644
--- a/.forgejo/workflows/deploy.yaml
+++ b/.forgejo/workflows/deploy.yaml
@@ -6,8 +6,44 @@ on:
       - main
 
 jobs:
+  # Same job as in test.yaml (which covers non-main branches); duplicated here
+  # so a red suite blocks the build+deploy — Forgejo has no cross-workflow needs.
+  test:
+    runs-on: nix-latest
+    services:
+      postgres:
+        image: postgres:17-alpine
+        env:
+          POSTGRES_USER: postgres
+          POSTGRES_HOST_AUTH_METHOD: trust
+          POSTGRES_DB: asxpio_test
+    steps:
+      - name: Prepare container for actions
+        run: |
+          echo "experimental-features = nix-command flakes" >> /etc/nix/nix.conf
+          # node for JS actions (checkout); git-minimal is already in the image
+          nix-env -iA nixpkgs.nodejs_22
+
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Wait for Postgres
+        run: |
+          for i in $(seq 1 30); do
+            nix develop -c pg_isready -h postgres -U postgres && exit 0
+            sleep 2
+          done
+          echo "Postgres service never became ready" >&2
+          exit 1
+
+      - name: Run test suite
+        env:
+          TEST_DATABASE_URL: postgres://postgres@postgres:5432/asxpio_test
+        run: nix develop -c bundle exec rake test
+
   build:
     runs-on: arch-latest
+    needs: test
     container:
       image: gcr.io/kaniko-project/executor:debug
     outputs:
@@ -115,7 +151,16 @@ jobs:
             docker pull $NEW_IMAGE
             docker compose up -d
 
-            sleep 5
-            docker ps | grep asxpio || exit 1
-
-            echo "Deployment successful!"
+            # Wait for the image HEALTHCHECK (GET /healthz) to report healthy;
+            # `docker ps | grep` passed even while the app crash-looped.
+            for i in $(seq 1 18); do
+              status=$(docker inspect --format '{{.State.Health.Status}}' asxpio 2>/dev/null || echo missing)
+              if [ "$status" = "healthy" ]; then
+                echo "Deployment successful!"
+                exit 0
+              fi
+              sleep 5
+            done
+            echo "asxpio never became healthy (last status: $status)" >&2
+            docker logs --tail 50 asxpio
+            exit 1
diff --git a/.forgejo/workflows/test.yaml b/.forgejo/workflows/test.yaml
index 36d6631..19c46df 100644
--- a/.forgejo/workflows/test.yaml
+++ b/.forgejo/workflows/test.yaml
@@ -1,9 +1,11 @@
 name: Tests
 
+# main is covered by the test job inside deploy.yaml (which gates the build);
+# this workflow covers every other branch.
 on:
   push:
-    branches:
-      - '**'
+    branches-ignore:
+      - main
 
 jobs:
   test:
diff --git a/CLAUDE.md b/CLAUDE.md
index dd39811..4eb5e66 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -84,6 +84,7 @@ Public:
 
 - `GET /i/:uuid` — landing page: client name, number, total, status badge, download button.
 - `GET /i/:uuid/pdf` — 302 to a 5-minute MinIO presigned URL (Content-Disposition: attachment).
+- `GET /healthz` — liveness (plus `SELECT 1` when invoicing is configured); used by the Docker HEALTHCHECK and the deploy pipeline's post-deploy wait.
 
 ### Storage model
 
@@ -138,10 +139,11 @@ The hedgehog is the user's pre-IE personal mark; once the IE has its own logo, t
 - **Run tests:** `nix develop -c bin/test` — full suite (minitest + rack-test) against an ephemeral Postgres it provisions and tears down itself. `bundle exec rake test` runs without a DB, skipping DB-backed tests. Mail uses `Mail::TestMailer`; S3 uses aws-sdk stubbed responses; tests never touch real services, and `.env` is deliberately not loaded when `RACK_ENV=test`.
 - **Local image build (sanity check):** `docker compose build`
 - **Production deploy:** automatic via `.forgejo/workflows/deploy.yaml` on push to `main`. Pipeline:
+  0. Test suite runs on the `nix-latest` runner label (postgres service container); a red suite blocks the build. Non-main branches get the same job from `test.yaml`.
   1. Kaniko builds the image and pushes it to the Forgejo registry.
   2. SSH copies `docker-compose.yml` to the deploy directory on the prod host.
   3. SSH writes `.env` (mode 600) from Forgejo secrets. The full set of invoicing envs (`DATABASE_URL`, `S3_*`, `ADMIN_*`) is built here; the per-app `ASXPIO_DB_PASSWORD` / `ASXPIO_S3_*` values must match the storage repo's Forgejo secrets, since storage's init containers provision the corresponding role + MinIO user.
-  4. SSH `sed`-substitutes the image tag in `docker-compose.yml` and runs `docker compose up -d`.
+  4. SSH `sed`-substitutes the image tag in `docker-compose.yml`, runs `docker compose up -d`, and waits (up to 90s) for the image `HEALTHCHECK` (`GET /healthz`) to report healthy — a crash-looping container fails the deploy.
 
 ## Where the secrets live