From de7053644b325060736136cec61c984cc092301f Mon Sep 17 00:00:00 2001
From: Marcel <marcel@familienarchiv>
Date: Sun, 17 May 2026 19:15:51 +0200
Subject: [PATCH] =?UTF-8?q?docs(adr):=20ADR-020=20=E2=80=94=20stateful=20a?=
 =?UTF-8?q?uth=20via=20Spring=20Session=20JDBC?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 ...0-stateful-auth-via-spring-session-jdbc.md | 94 +++++++++++++++++++
 1 file changed, 94 insertions(+)
 create mode 100644 docs/adr/020-stateful-auth-via-spring-session-jdbc.md

diff --git a/docs/adr/020-stateful-auth-via-spring-session-jdbc.md b/docs/adr/020-stateful-auth-via-spring-session-jdbc.md
new file mode 100644
index 00000000..78b62397
--- /dev/null
+++ b/docs/adr/020-stateful-auth-via-spring-session-jdbc.md
@@ -0,0 +1,94 @@
+# ADR-020 — Stateful Authentication via Spring Session JDBC
+
+**Date:** 2026-05-17
+**Status:** Accepted
+**Issue:** #523
+
+---
+
+## Context
+
+PR #521 (closing #520) introduced `AuthTokenCookieFilter` to unblock a production deploy.
+The filter promotes an `auth_token` cookie — which contains the full HTTP Basic credential
+(`Basic <base64(email:password)>`) — to an `Authorization` header so browser-direct `/api/*`
+calls authenticate correctly behind Caddy.
+
+This model has three concrete problems:
+
+1. **Cookie = credential.** A stolen `auth_token` cookie leaks the user's password in
+   base64-encoded plaintext. No decode step is needed; the cookie value is directly usable
+   as a credential forever.
+2. **No server-side revocation.** Logout deletes the local cookie but the credential
+   remains valid until the 24 h `Max-Age` elapses. An attacker who copied the cookie before
+   logout retains access.
+3. **No audit signal.** There is no server-side record of login or logout events. Observability
+   and compliance tooling cannot reconstruct "who was logged in when".
+
+Additionally, Nora flagged that `url.protocol === 'https:'` in `login/+page.server.ts` is
+incorrect behind Caddy: SvelteKit sees `http`, so `Secure=false` was set on the credential
+cookie in production, transmitting it in cleartext from Caddy to the browser on any network
+path without TLS.
+
+---
+
+## Decision
+
+Replace the `auth_token` / `AuthTokenCookieFilter` model with **Spring Session JDBC**:
+
+- A `POST /api/auth/login` endpoint in a new `auth` package authenticates with `email +
+  password`, creates a server-side session record in PostgreSQL, and returns the `AppUser`
+  JSON in the response body.
+- The response sets an **opaque** `fa_session` cookie (`HttpOnly`, `SameSite=Strict`,
+  `Secure` in non-dev profiles, `Max-Age=28800` — 8 h idle timeout) that contains only the
+  session ID, never a credential.
+- A `POST /api/auth/logout` endpoint invalidates the session record immediately. Subsequent
+  requests carrying the same cookie return 401.
+- `AuthTokenCookieFilter` is deleted in the same PR. No transitional coexistence period.
+- Cookie name `fa_session` (not the default `SESSION`) minimises framework fingerprinting.
+
+Session storage uses the canonical `spring_session` / `spring_session_attributes` tables,
+re-introduced via `V67__recreate_spring_session_tables.sql` (dropped by V2 when the
+dependency was previously removed as unused).
+
+**Idle timeout:** 8 h (`MaxInactiveIntervalInSeconds = 28800`). No 24 h absolute cap is
+implemented in Phase 1 — the 8 h idle bound contains the risk to one workday. A weekend-long
+active session is acceptable given the family-archive threat model. The absolute cap and
+additional revocation paths (password-change, admin force-logout) land in Phase 2 (#524).
+
+---
+
+## Alternatives Considered
+
+### Stay on Basic cookie + add a server-side revocation table
+
+Keeps the credential-in-cookie problem. Implementing a revocation table would re-invent
+Spring Session badly — we'd write bespoke session storage that already exists and is
+well-tested upstream.
+
+### JWT (stateless)
+
+Opaque revocation is simpler than JWT revocation (token introspection or short-lived tokens
++ refresh). The cluster is single-node; session affinity is not a constraint. Stateless tokens
+buy complexity without benefit here. JWKS infrastructure and refresh-token rotation are
+unnecessary for a family archive with < 50 concurrent users.
+
+### Keep `auth_token` cookie but add `AuthTokenCookieFilter` improvements
+
+The root problem is that the cookie contains the credential. No amount of filter hardening
+fixes that. Nora's P1 flag stands until the credential leaves the cookie.
+
+---
+
+## Consequences
+
+- **One breaking deploy.** All existing sessions (the `auth_token` cookies) become inert
+  on the next request after the deploy. The SvelteKit `handleAuth` hook redirects to
+  `/login?reason=expired`; a banner renders. Users re-login. No data loss.
+- **~2 KB per active session** in PostgreSQL (`spring_session_attributes` stores the
+  serialised `SecurityContext`). With < 50 family members, this is immaterial.
+- **Session cleanup task** runs on the default Spring Session JDBC schedule (every 10 min).
+  No custom job needed.
+- **Caddy / infrastructure unchanged.** `forward-headers-strategy: native` already ensures
+  `Secure` cookies work correctly behind the reverse proxy.
+- **Dev profile:** `application-dev.yaml` sets `secure: false` on the session cookie so
+  local HTTP dev (port 5173 → 8080) works without TLS.