[felixevers]

Full-stack e2e behind the nginx proxy + a real backend sort crash it caught

Follow-up to #319: the mock-based e2e can't see routing or database-level problems, so this adds a second e2e tier that runs against the real stack — the docker images built for the current commit, wired together exactly like production (nginx proxy in front of web, backend and Keycloak).

Proxied full-stack e2e

• docker-compose.e2e.yml — production-like stack (proxy → web / backend / Keycloak / Postgres / Redis / InfluxDB) with parameterized image references (BACKEND_IMAGE / WEB_IMAGE / PROXY_IMAGE) so CI runs this commit's images; ephemeral volumes, restart: on-failure on the backend so migrations retry until Postgres accepts connections.
• tests/e2e/proxy-fullstack.spec.ts (gated behind E2E_PROXY_TARGET=1, skipped in the mock-based runs):
  • proxy routing: / → web, /keycloak/... → Keycloak (issuer under the subpath), /graphql → backend (JSON, never the frontend catch-all)
  • authentication through the proxied Keycloak: password grant for API seeding and the real browser login flow (redirect → login form → /auth/callback → app)
  • seeds patients + tasks over the proxied GraphQL API, then asserts server-side filtering and sorting from the real query engine: name asc/desc, birthdate, NOT_CONTAINS, update-date sort, and sorting tasks by patient asc/desc (issue Sorting of my tasks list based on patients not possible #213)
  • UI: logs in, searches, and sorts the patient list through the whole stack
• e2e-proxy CI job (tests.yml) — builds the backend/web/proxy images for the commit with Buildx (cache-from/to: type=gha, same scopes now written by the Docker Build workflow, so layers are shared with the images built earlier), starts the compose stack, waits for readiness through the proxy, runs the spec, and uploads the report + stack logs on failure. Added to the required CI gate.
• E2E_TESTING.md documents how to run this locally.

Real bug the proxied tests caught (issue #213)

Sorting my tasks by patient crashed on PostgreSQL:

(sqlalchemy) for SELECT DISTINCT, ORDER BY expressions must appear in select list

The tasks/patients base queries use SELECT DISTINCT (location-scope joins), and every sort that orders by a joined or correlated expression — tasks by patient or assignee, patients by update date, ward/room/bed, clinic or position — produced an ORDER BY outside the select list and 500ed. This is invisible to the mock-based suite and to SQLite-backed tests; it reproduced immediately against the real Postgres backend through the proxy.

Fix: apply_unified_query now rewraps DISTINCT statements into a plain select over deduplicated ids before applying sorts (and clears stale join memos so sort adapters re-join the rewrapped statement). This also covers the pre-existing case where a property filter (needs_distinct) was combined with a position/clinic sort.

Verification

• All 5 proxied tests pass locally against the compose stack built from this commit (they fail on main with the DISTINCT crash when sorting tasks by patient).
• Backend: ruff clean; unit tests show no new failures.
• The spec is skipped without E2E_PROXY_TARGET=1, so the existing mock-based e2e job is unaffected.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WCSmjBuS6NbcB6oRmKvQgT

---

Generated by Claude Code