trails/openspec/changes/unify-auth-completion/design.md
Ullrich Schäfer d64c47614d
Implement completeAuth chokepoint + caller migration
Implements all of unify-auth-completion (12/14 tasks done; manual
smoke + archive-time spec sync remain).

Design refinement during implementation: completeAuth supports two
response shapes via a `mode` parameter:
- mode: 'redirect' (loaders / direct browser navigation; auth.verify.tsx)
- mode: 'json' (action handlers called by imperative fetch from
  client forms; api.auth.login, api.auth.register)

Both modes share createSession + safeReturnTo + Set-Cookie. JSON mode
carries `{ ok: true, step: "done", redirectTo }` (the `step` field
preserves the existing client-form check).

Why two modes: passkey ceremonies are inherently imperative
(start → browser WebAuthn API → finish), so action handlers can't
move to <Form>/useFetcher. Picking option (B) from the design grill —
the chokepoint owns destination selection while clients navigate —
required this dual shape. The 3 hardcoded client-side targets
(returnTo ?? "/", "/", "/?add-passkey=1") collapse into 1 server-side
sanitization pass (safeReturnTo) inside completeAuth.

New module:
- apps/journal/app/lib/auth/session.ts: cookie session storage
  (sessionStorage, createSession, getSessionUser, destroySession)
  moved from auth.server.ts. Legacy import path kept via re-exports
  with @deprecated JSDoc.
- apps/journal/app/lib/auth/completion.ts: completeAuth + safeReturnTo.
- apps/journal/app/lib/auth/completion.test.ts: 10 contract tests
  covering both modes, returnTo sanitization (path-relative, protocol-
  relative, absolute-URL, malformed), Set-Cookie attachment, redirect
  status, JSON shape.

Caller migration:
- api.auth.register.ts passkey-finish → completeAuth(json)
- api.auth.login.ts finish-passkey → completeAuth(json)
- api.auth.login.ts verify-code → completeAuth(json)
- auth.verify.tsx magic-link consumer → completeAuth(redirect)

Client form updates:
- auth.login.tsx: pass returnTo in fetch body, read result.redirectTo
  on done.
- auth.register.tsx: pass returnTo: "/?add-passkey=1" for the magic-
  link verify-code path (preserves the post-register passkey prompt
  via the chokepoint's safeReturnTo, instead of hardcoding it
  client-side).

Verified:
- pnpm typecheck && pnpm lint: green across all 15 workspaces.
- pnpm --filter @trails-cool/journal test: 126 passed.
- pnpm test:e2e auth: 4/4 passed without modification — confirms the
  refactor is behaviour-preserving for the user-facing flows that
  matter most (passkey register + login).

Spec delta in openspec/changes/unify-auth-completion/specs/ applies at
/opsx:archive time.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-08 02:38:15 +02:00

7.7 KiB
Raw Blame History

Context

apps/journal/app/lib/auth.server.ts is 489 lines covering passkey ceremonies, magic-link token lifecycle, session cookie storage, terms recording, and email-change tokens. Four route handlers (api.auth.register.ts, api.auth.login.ts × 2 branches, auth.verify.tsx) each repeat the same trailing sequence after their per-method identity verification:

if (isNewRegistration) await recordTermsAcceptance(userId, termsVersion);
const headers = await createSession(userId, request);
return redirect(safeReturnTo(returnTo) ?? "/", { headers });

The repetition is the entire asymmetry between methods at the post-verify layer. WebAuthn vs. magic-link diverge before this point (challenge persistence, token consumption, etc.) but converge here — every flow does the same thing.

ADRs already recorded:

  • ADR-0004: centralize this orchestration in a single completeAuth function.
  • ADR-0005: do not extract an AuthMethod interface; passkey + magic-link is the entire surface and OAuth2 is transport, not a peer.

Goals / Non-Goals

Goals:

  • One function (completeAuth) is the only place where a successful web auth records terms + creates session + redirects. Bug fixes apply once.
  • File reorganization: apps/journal/app/lib/auth/ directory containing completion.ts and session.ts. The existing auth.server.ts shrinks but stays for the per-method functions (passkey ceremony, magic-token lifecycle).
  • returnTo sanitization centralized inside completeAuth so no caller can ship an open-redirect.
  • Refactor is behaviour-preserving: existing e2e auth tests pass without modification.

Non-Goals:

  • No AuthMethod interface (per ADR-0005).
  • The OAuth-code issuance flow at /oauth/authorize is NOT migrated. It operates on already-authenticated users; only the trailing redirect overlaps, not enough leverage to fold in.
  • The add-passkey post-login ceremony does NOT need completeAuth (no session creation — user is already logged in).
  • The magic-link request-token step does NOT need completeAuth (only emits a token; verification happens later).
  • Session transport (cookie vs. bearer) — orthogonal; not touched.
  • Connected-devices UX — separate spec.
  • Any change to the Terms gate enforcement points (root loader for web, requireApiUser for API).

Decisions

Decision 1: Single function, no mode flag

completeAuth covers exactly the four web post-verify call sites listed in the proposal. There is no redirectMode: 'cookie' | 'oauth-code' flag because OAuth-code issuance is genuinely a different flow (already-authenticated user, no session creation, only a code + redirect). Folding both into one function would force a discriminated union and obscure that they don't share much. The OAuth-code path stays in oauth.authorize.tsx unchanged.

Decision 2: Function signature

async function completeAuth(input: {
  userId: string;
  request: Request;
  returnTo?: string | null;
}): Promise<Response>

Returns a Response (a redirect from React Router with the session Set-Cookie attached). Callers do return await completeAuth(...) from their action / loader. No headers-vs-redirect split: the chokepoint owns both halves.

Terms recording is NOT in completeAuth. Both registration paths already write terms_version + terms_accepted_at at user creation: finishRegistration writes them when the user row is inserted (passkey path); registerWithMagicLink writes them when the user row is inserted (magic-link path). Magic-link registration must write terms at user creation, before verification, because the user row exists between register and verify — if terms weren't already written, the Terms gate would bounce the user on first authenticated visit.

So completeAuth is purely session + redirect. No isNewRegistration flag, no termsVersion parameter. The earlier framing of an isNewRegistration discriminator was a mis-read — there is no code path that needs to record terms at completion time.

Decision 3: File layout

New directory apps/journal/app/lib/auth/ containing:

  • completion.tscompleteAuth + safeReturnTo helper (open-redirect rejection).
  • session.tssessionStorage, createSession, getSessionUser, destroySession. Moved from auth.server.ts.

auth.server.ts keeps the per-method functions (passkey ceremony, magic-token lifecycle, terms recording, email-change). It re-exports the moved session helpers from their new home so existing imports keep working without a touch:

// auth.server.ts
export { sessionStorage, createSession, getSessionUser, destroySession } from "./auth/session.ts";

That keeps the migration small. A future change can rip the re-exports and update callers across the app — out of scope here.

Decision 4: returnTo sanitization centralizes inside completeAuth

Today, several callers do their own returnTo checks of varying rigor (some check startsWith("/") && !startsWith("//"), others don't). Move the check into safeReturnTo inside completion.ts; every caller passes returnTo raw and gets the safe version applied uniformly. Reject anything that isn't a same-origin path → fall back to /.

Decision 5: Tests

Unit tests for completeAuth in completion.test.ts:

  • new-registration writes users.terms_version + users.terms_accepted_at (mock the DB or use an in-memory test DB — match existing test patterns).
  • non-registration skips the terms write.
  • response includes a Set-Cookie header from sessionStorage.commitSession.
  • returnTo of /foo redirects to /foo; returnTo of //evil.com/x redirects to /; returnTo of https://evil.com redirects to /; missing returnTo redirects to /.

Existing e2e/auth.test.ts (passkey register + login + logout) MUST pass without modification — that's the behaviour-preservation proof.

Risks / Trade-offs

  • [Risk] Re-exporting from auth.server.ts to maintain backwards compatibility means there are temporarily two paths to the same symbols. → Mitigation: comment marks the re-exports as transitional. A follow-up PR (small) can update import paths app-wide and remove the re-exports.

  • [Risk] The runtime assertion (isNewRegistration && !termsVersion → throw) is a programmer error, not a user-visible failure. If a future caller forgets to pass termsVersion, tests should catch it; if they don't, registration breaks loudly. → Accepted: the type-level alternative (discriminated union) costs more than it pays back at three internal call sites.

  • [Trade-off] Centralizing returnTo sanitization changes behaviour at any call sites that don't currently sanitize. The change is strictly safer (rejecting more inputs as /) but worth confirming there are no integration tests that pass a deliberately-unsafe returnTo and expect it to work.

Migration Plan

Single PR ("spec apply") containing:

  1. Create apps/journal/app/lib/auth/session.ts with the moved helpers.
  2. Add re-exports in auth.server.ts.
  3. Create apps/journal/app/lib/auth/completion.ts with completeAuth + safeReturnTo.
  4. Write completion.test.ts (unit tests).
  5. Migrate the 4 callers to completeAuth.
  6. Verify pnpm typecheck && pnpm lint && pnpm test && pnpm test:e2e green.

Rollback: revert the PR. No DB or schema changes; the file moves are pure code-organization, the call-site changes are mechanical.

Open Questions

  • Q1: Does sessionStorage get its secret from process.env.SESSION_SECRET directly, or from a config helper? Move-as-is and revisit only if it complicates the session.ts file. → Decide during implementation.
  • Q2: Should the auth.server.ts re-exports be marked @deprecated JSDoc-style so editors warn on use? Light touch — yes, with a comment pointing at ./auth/session.ts. Decide during implementation.