`ip route` (iproute2) is not installed in the Gitea runner container,
causing the smoke test step to exit 127. /proc/net/route is a kernel
virtual file that is always present on Linux; awk decodes the
little-endian hex gateway field to dotted-decimal without any external
binary dependency.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Unquoted variable expansion is safe here since the value contains
no spaces or glob characters, but quoting is the correct default
and keeps the script consistent with surrounding style.
Addresses review suggestion by Felix Brandt and Tobias Wendt.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
If `ip route show default` returns no output the old code passed
an empty string to curl --resolve, producing a confusing error 6
("couldn't resolve host") with no indication that gateway detection
had failed. The new guard exits immediately with a clear message.
Addresses review concern raised by Tobias Wendt.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Job containers run in bridge network mode (runner-config.yaml). Inside
a bridge-networked container 127.0.0.1 is the container's own loopback;
Caddy on the host is unreachable there, causing an immediate ECONNREFUSED.
Use the Docker bridge gateway IP instead — the host's docker0 interface
where Caddy (bound on 0.0.0.0:443) is reachable from the container.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
`sudo systemctl reload caddy` does not work from inside a DooD job
container: `systemctl` is absent from Ubuntu container images and
container processes cannot reach the host systemd without entering its
namespaces. Replace with `docker run --privileged --pid=host ubuntu:22.04
nsenter -t 1 -m -u -n -p -i -- /bin/systemctl reload caddy`, which uses
the already-mounted Docker socket to spin up a privileged sibling
container that enters the host PID namespace via nsenter. Tested live on
the Hetzner VPS. No sudoers entry required.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds a `sudo systemctl reload caddy` step between the docker compose
deploy and the smoke test. This ensures any committed Caddyfile changes
are applied before the public surface is verified.
Previously the workflow had no mechanism to push Caddyfile changes to
the running host daemon. A Caddyfile edit would land in the repo but
Caddy would keep serving the previous config, causing the smoke test to
catch a stale header or still-proxied /actuator route rather than the
intended current config.
This step also surfaces the root cause of today's port-443 failure
explicitly: if Caddy is not running, the step fails with a clear service
error rather than a misleading "Failed to connect to port 443" from curl.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Sara flagged that a future "compose cleanup" PR could silently drop the
backend volumes block and CI would happily pass while mass import on
staging silently broke. Adds a pre-build step that renders the staging
compose config and fails the deploy if `target: /import` or
`read_only: true` is missing.
Local verification of the guard:
- Volumes block removed → `grep -q 'target: /import'` exits 1 → step fails
- Volumes block present → both greps match → step passes
Addresses Sara's review on #526.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The compose file now requires IMPORT_HOST_DIR or refuses to start
(#526). Without this line the next nightly deploy would fail with a
clear interpolation error, but it should not fail — the staging
import payload already lives at this host path (rsync'd in #526).
Addresses Tobias's review on #526.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Closes#508.
Our gitea-runner advertises labels ubuntu-latest / ubuntu-24.04 /
ubuntu-22.04. `runs-on: self-hosted` never matches → dispatched
deploy jobs sit in the queue forever. The runner is still
genuinely self-hosted (DooD socket, joined to gitea_gitea net,
single-tenant per ADR-011) — the `self-hosted` token was just an
unconfirmed assumption about the label name.
Unblocks #497 / #499 first deploy.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The `if: always()` conditional on the env-file cleanup step in both
deploy workflows is what makes the ADR-011 single-tenant runner trust
model safe: secrets land on disk before each deploy and are wiped
unconditionally afterwards. A future workflow refactor that drops
`if: always()` would silently leave plaintext secrets on the runner
on any failed deploy.
The ADR documents this; the workflow file did not. Adds a prominent
inline comment so the next reader of the YAML sees the constraint
without having to cross-reference ADR-011. No behaviour change — both
workflows still parse. Addresses @nora's round-2 suggestion on PR
#499 — "linchpin of the ADR-011 trust model".
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adds `Permissions-Policy: camera=(), microphone=(), geolocation=()` to
the shared (security_headers) snippet, so both archiv vhosts and the
git vhost deny browser APIs the app does not use. Reduces blast radius
of an XSS landing in a privileged origin.
The deploy smoke steps in nightly.yml and release.yml gain a matching
assertion against the canonical header value, so a future Caddyfile
edit that drops or loosens the header (e.g. `camera=(self)`) fails the
deploy instead of regressing silently.
`caddy validate` against caddy:2 passes; both workflow YAMLs parse.
Addresses @nora's round-2 suggestion on PR #499 — "lower-impact than
CSP but nearly free".
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Replaces the presence-only `grep -qi strict-transport-security` smoke
assertion in both nightly.yml and release.yml with a value-pinning
regex that requires `max-age=31536000`, `includeSubDomains`, and
`preload`. A future Caddyfile edit that drops any of those three
parts now fails the deploy smoke step instead of passing silently.
Verified locally that the new pattern matches the preload-eligible
value and rejects three degraded forms (short max-age, missing
includeSubDomains, missing preload). Addresses @sara's round-2 note
on PR #499 — "presence check, not value check".
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The smoke step previously curled the public hostname unconditionally,
which routes the runner's request via DNS → router → back into the same
host. Many SOHO routers do not implement hairpin NAT (or do so only after
a firmware update), so the deploy may pass on day one and silently fail
on day 90.
--resolve "<host>:443:127.0.0.1" pins the hostname to the runner's
loopback while keeping SNI on the public name (so the cert validates
correctly and the Caddy vhost block matches). The smoke test now
verifies that the Caddy-on-the-same-host is serving the right
hostname end-to-end, with no router dependency.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Without --pull, the host's Docker layer cache wins: if a CVE drops in
node:20.19.0-alpine3.21 / postgres:16-alpine and the vendor re-publishes
the same tag, the runner keeps serving the cached layer until the cache
is manually cleared — a silent supply-chain blind spot.
Adding --pull to both `compose build` invocations costs a single
re-pull per run and lifts the base-image patch lag from "next host
prune" to "next nightly".
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The two deploy workflows make two non-obvious assumptions that future
maintainers should not have to rediscover by reading the diff:
1. Single-tenant self-hosted runner — the .env.* file lands on disk
during the deploy and is cleaned up unconditionally. Multi-tenant
usage would require switching to stdin-piped env input.
2. Host docker layer cache is authoritative — there is no
actions/cache directive; a host-level `docker system prune` will
cold-start the next build.
Both notes added as block comments at the top of each workflow.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Healthchecks prove containers are healthy on the docker network; they
do not prove the public URL is reachable, HSTS still fires, or
/actuator is still blocked at the edge. Add a post-deploy smoke step
to nightly.yml that:
1. GETs https://staging.raddatz.cloud/login (frontend reachable)
2. asserts the response includes the Strict-Transport-Security header
3. asserts /actuator/health returns 404 (defense-in-depth verified)
Failure aborts the workflow before the env-file cleanup step. The
cleanup step still runs because it is `if: always()`.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Runs daily at 02:00 (and on workflow_dispatch). Builds the prod
compose stack with BuildKit, writes a transient .env.staging from
Gitea secrets, then `docker compose up -d --wait` so the job fails
loudly if any service's healthcheck never reports healthy.
The --profile staging flag starts the mailpit catcher in place of
a real SMTP relay; no production SMTP credentials touch the staging
environment.
The .env.staging file is cleaned up in `if: always()` to avoid
leaving secrets in the runner workspace between runs.
Refs #497.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
