The runner's Docker client negotiates API 1.53 but the daemon on the
NAS only supports up to 1.43. Pin the version for all docker commands
in the e2e job, including the new network connect step.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The act_runner job runs inside a Docker container. Docker Compose port
mappings bind to the Docker host, not the job container's localhost —
so localhost:5433 was always refused.
Fix: after compose starts, connect the job container (identified by
/etc/hostname) to the archive-net compose network. Then switch the
backend startup args to use service names db:5432 and minio:9000
instead of host-mapped ports.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The e2e job was calling plain `docker compose up` without the CI
override file, so it used the base compose bind-mount for MinIO
(./data/minio) which doesn't exist on the runner. The CI override
replaces bind mounts with ephemeral named volumes.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Both SvelteKit API proxy routes were hardcoding http://localhost:8080,
breaking typeahead search in Docker environments.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The userGroup hook was hardcoding http://localhost:8080 instead of
reading API_INTERNAL_URL from the environment. In Docker this caused
the /api/users/me fetch to fail silently, leaving event.locals.user
unset and triggering the handleAuth guard to redirect every page to
/login — including the login form action itself, creating an infinite
redirect loop.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
