feat: make "only the app publishes a port" a binding rule

A published database port puts the data on the public internet, and Docker's
own iptables rules bypass ufw — the server looks firewalled while the port is
open. Previously this was only a comment in compose.yaml, which an agent adding
a db service later is free to ignore.

Anchored at every point where a service could be added or shipped:
- rules/security.md: the rule, the reasoning, and the audit commands
- rules/backend.md + architecture SKILL: constraint when picking the data layer
- agents/backend-dev.md: non-negotiable for any service it adds
- agents/qa-engineer.md: a published port is a Critical finding
- skills/deploy: blocking pre-flight audit
- compose.yaml: db example now uses expose:, with depends_on healthcheck
- docs/DEPLOYMENT.md + README: SSH tunnel instead of publishing for debugging
This commit is contained in:
Florian Wilde
2026-08-09 01:12:06 +02:00
parent e30d02d41a
commit 75e99f75c6
9 changed files with 161 additions and 7 deletions
+4
View File
@@ -36,6 +36,10 @@ The non-negotiables:
as a fallback
- Every new env var lands in `.env.example` in the same change
- Whatever you add must run inside the Docker Compose stack
- **Any service you add to `compose.yaml` gets no `ports:` entry.** Only `app`
publishes a port. A database, cache or queue is reached by service name over the
compose network (`db:5432`) and must stay unreachable from the host. Publishing it
exposes the data to the internet and bypasses the server's firewall.
## Tests
For each route handler, write a Vitest test next to it
+4
View File
@@ -33,6 +33,10 @@ The feature spec path and the list of previously deployed features to regression
fallback rate-limited? A missing or weakened passkey flow is a **High** finding.
- Injection through every user-controlled input (XSS, SQL)
- Secrets or internal detail leaking in API responses, logs or the network tab
- **Published ports:** run `docker compose config | grep -B8 "published:"`. Only the
`app` service may publish, and only on `127.0.0.1`. A database, cache or queue with
a published port — or anything bound to `0.0.0.0` — is a **Critical** finding: it
exposes the service to the internet and bypasses the host firewall.
## Tests you write
- **Unit (Vitest), co-located in `src/`:** hooks with real logic, pure transforms,
+4
View File
@@ -18,6 +18,10 @@ decision exists, do not install an ORM or write schema code.
Whatever is chosen must be self-hostable inside the Docker Compose stack —
the deployment target is a plain server, not a managed platform.
When you add it to `compose.yaml`, it gets **no published port**. The app reaches it
by service name over the compose network. See the port-exposure rule in
`.claude/rules/security.md`.
## Route Handlers
- Validate every request body with a Zod schema before touching it. Reuse the
+48
View File
@@ -59,6 +59,54 @@ Requirements whenever auth is present:
- Never interpolate user input into SQL, shell commands or file paths
- Escape or avoid `dangerouslySetInnerHTML`
## Port exposure (mandatory)
**Only the app service publishes a port. Nothing else. Ever.**
Databases, caches, queues, search indexes, admin UIs and migration containers get
**no `ports:` entry at all**. Services reach each other by service name over the
Compose network — `postgres://user:pass@db:5432/app` works without publishing
anything. Use `expose:` if you want the port documented; it changes nothing about
reachability from outside.
```yaml
db:
image: postgres:17-alpine
expose:
- "5432" # documentation only — not reachable from the host
# ports: # NEVER. This would put the database on the public internet.
```
Why this is not a style preference: `ports: "5432:5432"` on a public server means
the only thing between an attacker and the whole dataset is the database password.
Worse, Docker writes its own nat/DOCKER-USER iptables rules that **bypass ufw** on a
typical setup — the server looks firewalled while the port is wide open. This is one
of the most common ways self-hosted stacks get breached.
The app service itself binds to loopback (`${APP_BIND:-127.0.0.1}`) because a reverse
proxy on the host terminates TLS and forwards to it. Binding `0.0.0.0` is only
correct when there is no proxy, and it means serving plain HTTP — which also breaks
passkeys.
If a port genuinely must be reachable for local debugging, bind it explicitly to
loopback and never bare:
```yaml
ports:
- "127.0.0.1:5432:5432" # ok for an SSH-tunnelled session
- "5432:5432" # WRONG — this is 0.0.0.0, the whole internet
```
Verify what is actually published before deploying:
```bash
docker compose config | grep -A3 "ports:" # what the merged config declares
docker compose ps --format "{{.Service}}\t{{.Ports}}" # what is really bound
```
Anything other than the app service showing a `0.0.0.0:` binding is a Critical
finding — treat it as such in `/qa` and block the deploy.
## Transport and Headers
Security headers are configured in `next.config.ts`: `X-Frame-Options: DENY`,
+4 -2
View File
@@ -41,8 +41,10 @@ This template deliberately ships **no** database or ORM. Decide per feature:
> es eine Datenbank im Compose-Stack.
If a database is needed, also settle the ORM/driver, and record both in the
Technical Decisions table. Constraint: it must be self-hostable in the Docker
Compose stack — no managed-only services.
Technical Decisions table. Two constraints: it must be self-hostable in the Docker
Compose stack (no managed-only services), and it runs **without a published port** —
only the app is reachable from outside, everything else talks over the internal
compose network.
### Authentication
+12
View File
@@ -36,6 +36,18 @@ Then check:
- [ ] Pending migrations are committed
- [ ] Everything committed and pushed to the remote the server pulls from
**Port audit — blocking.** Only the app service may publish a port, and only on
loopback:
```bash
docker compose config | grep -B8 "published:"
```
Any service other than `app` with a published port, or any binding to `0.0.0.0`,
stops the deploy. A published database port puts the data on the public internet and
bypasses ufw, because Docker writes its own iptables rules. Fix `compose.yaml`
before continuing.
## 3. First deployment only
Walk the user through this once. It happens on the server, over SSH.
+32
View File
@@ -435,6 +435,37 @@ docker compose up -d --build # Erststart
./scripts/update.sh # neuen Stand ausrollen
```
### Nur der App-Port geht nach außen
Im Stack veröffentlicht **ausschließlich der App-Service** einen Port, und auch den
nur auf Loopback (`127.0.0.1`). Jeder weitere Service — Datenbank, Cache, Queue,
Admin-UI — bekommt **keinen `ports:`-Eintrag**. Die Container erreichen sich über
den Compose-Netzwerknamen, dafür muss nichts veröffentlicht werden:
```
DATABASE_URL=postgres://user:pass@db:5432/app
```
> `ports: "5432:5432"` bindet auf `0.0.0.0` — die Datenbank steht damit im offenen
> Internet, geschützt nur durch ihr Passwort. Und Docker schreibt eigene
> iptables-Regeln, die **ufw umgehen**: Der Server sieht abgeschottet aus, während
> der Port offen ist. Das ist einer der häufigsten Wege, wie selbst gehostete Stacks
> kompromittiert werden.
Die Regel ist an vier Stellen verankert, damit sie beim Vibe-Coden nicht verloren
geht: `rules/security.md` (lädt automatisch beim Bearbeiten von `compose.yaml`), der
Backend-Sub-Agent, der QA-Sub-Agent (Verstoß = **Critical**) und der Pre-Flight von
`/deploy`, der den Deploy abbricht.
Prüfen, was wirklich gebunden ist:
```bash
docker compose config | grep -B8 "published:"
docker compose ps --format "{{.Service}}\t{{.Ports}}"
```
Zugriff zum Debuggen holst du dir per SSH-Tunnel, nicht per veröffentlichtem Port.
### Was `update.sh` macht
```mermaid
@@ -555,6 +586,7 @@ Vier Dinge, die nicht offensichtlich sind und je einmal wehtun:
| Fall | Was passiert |
|------|--------------|
| Ein DB-Service bekommt `ports: "5432:5432"` | Die Datenbank steht im offenen Internet — und Docker umgeht dabei ufw, der Server *wirkt* nur abgeschottet |
| `NEXT_PUBLIC_APP_URL` weicht von der Domain ab | Alle Passkeys scheitern, ohne verwertbare Fehlermeldung |
| Port 3000 ist von einem anderen Projekt belegt | Playwright testet still gegen den fremden Server — `E2E_PORT` setzen |
| Jemand legt eine `tailwind.config.ts` an | Wird in v4 ignoriert; Tokens gehören in `globals.css` |
+20 -2
View File
@@ -32,8 +32,13 @@ services:
max-file: "3"
# A database is deliberately not part of this template — /architecture picks one
# per project. When you add one, declare it here as a second service, put it on
# an internal network without a published port, and add its volume below.
# per project.
#
# RULE FOR EVERY SERVICE ADDED BELOW: only `app` gets a `ports:` entry. Databases,
# caches and queues get none — the app reaches them by service name over this
# compose network (`db:5432`). Publishing 5432 puts the database on the public
# internet, and Docker's iptables rules bypass ufw, so the server only *looks*
# firewalled. See .claude/rules/security.md.
#
# db:
# image: postgres:17-alpine
@@ -41,8 +46,21 @@ services:
# environment:
# POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:?set POSTGRES_PASSWORD in .env}
# POSTGRES_DB: ${POSTGRES_DB:-app}
# expose:
# - "5432" # documentation only — NOT reachable from the host
# volumes:
# - db-data:/var/lib/postgresql/data
# healthcheck:
# test: ["CMD-SHELL", "pg_isready -U postgres"]
# interval: 10s
# timeout: 5s
# retries: 5
#
# Then make the app wait for it:
# app:
# depends_on:
# db:
# condition: service_healthy
#
# volumes:
# db-data:
+33 -3
View File
@@ -82,14 +82,44 @@ to `HEALTH_TIMEOUT` seconds (default 120). If the container never turns healthy
prints the last 50 log lines, resets to the previous commit, rebuilds and restarts
that — so a bad deploy self-heals instead of leaving the site down.
## Port exposure
**Only the app service publishes a port, and only on loopback.** Everything else —
database, cache, queue, admin UI — gets no `ports:` entry. Services reach each other
by name over the Compose network, which needs nothing published:
```
DATABASE_URL=postgres://user:pass@db:5432/app
```
`ports: "5432:5432"` binds `0.0.0.0`, which means the database is on the public
internet. And Docker writes its own nat/DOCKER-USER iptables rules, so **ufw does not
block it** — the server looks firewalled while the port is open. This is a routine
way self-hosted stacks get breached.
Need access for debugging? Do not publish the port. Use an SSH tunnel:
```bash
ssh -L 5432:localhost:5432 user@server \
-t 'docker compose -f /srv/app/compose.yaml exec db psql -U postgres'
```
Audit what is really bound before and after every deploy:
```bash
docker compose config | grep -B8 "published:" # declared
docker compose ps --format "{{.Service}}\t{{.Ports}}" # actual
ss -tlnp | grep -v 127.0.0.1 # what the host exposes
```
## Database
The template ships none by design. When `/architecture` picks one:
1. Add it as a second service in `compose.yaml` — no published port, reachable only
over the internal Compose network
1. Add it as a second service in `compose.yaml` — **no `ports:` entry**, reachable
only over the internal Compose network. Use `expose:` if you want it documented.
2. Add a named volume for its data
3. Add `DATABASE_URL` to `.env.example`
3. Add `DATABASE_URL` to `.env.example` (pointing at the service name, e.g. `db:5432`)
4. Run migrations **before** the new container serves traffic. Either add a
short-lived `migrate` service that the app depends on, or add the migration
command to `scripts/update.sh` between build and restart.