1. Architecture chosen
Manager pushes environment variables into every business's .env, rather than businesses fetching config at runtime. Manager has no database of its own today (confirmed via code — no Prisma, no ORM dependency anywhere in manager/); every piece of its state is either an env var or a flat file under STACKS_DIR. A "runtime shared-config source" design would make every single email send (including OTP login) depend on Manager being reachable — a new failure mode this app has never had, for no real benefit. Instead:
Storage: one JSON file, stacks/manager/global-settings.json (gitignored), shaped for future sections (SMS, storage, backups, branding...) to slot in as sibling keys.
Secret at rest: the SMTP password is AES-256-GCM encrypted with a new MANAGER_SETTINGS_ENCRYPTION_KEY, never returned to the browser.
Rollout: saving complete settings automatically pushes the 7 EMAIL_* keys into every business's .env via the exact same surgical upsertEnvValues helper updateBusiness() already uses for name/domain edits, then force-recreates each container (composeUp(dir, {forceRecreate:true})) — the identical mechanism that already applies domain changes today.
New businesses get the current global settings baked into their .env at provisioning time — no manual step, ever.
2. Files changed
New (Manager): global-settings.ts (storage+encryption), email-rollout.ts (push-to-all-businesses), test-email.ts (standalone nodemailer test-send), dashboard/settings/ (page.tsx, actions.ts, email-settings-form.tsx).
Modified (Manager): compose-template.ts (optional email block in generated .env), provisioning.ts (reads global settings at creation), dashboard/layout.tsx (Settings nav link), package.json/package-lock.json (added nodemailer).
Modified (main app): brevo-email-provider.ts (added encryption mode + replyTo, backward-compatible), its test file (8 new tests).
Modified (docs): root .gitignore, .env.example, .env.business.example, manager/.env.example.
3. New volumes/databases/config files
None — no new Docker volume, no database. One new gitignored file, stacks/manager/global-settings.json, living inside the STACKS_DIR bind-mount Manager already has.
4. New environment variables
MANAGER_SETTINGS_ENCRYPTION_KEY (Manager only, required to use Settings) — 32-byte hex, generate with openssl rand -hex 32.
EMAIL_SMTP_ENCRYPTION and EMAIL_REPLY_TO (main app, both optional) — unset EMAIL_SMTP_ENCRYPTION preserves today's exact port-based inference, so existing deployments are unaffected.
Reply-to is now business-specific with global fallback,
5. How settings are persisted
stacks/manager/global-settings.json, written by saveEmailSettings(). The password is encrypted before writing; every other field is plain text (host, port, username, from-address, from-name, reply-to, encryption mode, last-test result) since none of those are secrets.
6. How business containers receive global settings
Via their own .env file (Docker Compose's env_file:, unchanged mechanism) — Manager writes the 7 EMAIL_* keys into it and recreates the container. Nothing in the business app itself needed a new runtime dependency; getEmailProvider() still just reads process.env once at process start, exactly as before.
7. What happens to existing businesses when settings change
Saving complete settings (host+username+password+from-address all present) automatically loops every business, updates its .env, and force-recreates its container in one action — the UI reports "applied to N of M businesses" with per-business failure reasons. A separate "Reapply to all businesses" button retries this without re-entering anything, for recovering from a business that was down during the last save. Saving incomplete settings persists them as a draft and touches no business (verified live — this is what stops a business from being flipped into a broken "brevo with blank credentials" state).
8. Manual deployment/update steps on the VPS
Generate a key: openssl rand -hex 32, add it as MANAGER_SETTINGS_ENCRYPTION_KEY to stacks/manager/.env.
Rebuild/redeploy the Manager image and the main app image (both changed) as usual.
Open Manager → Settings, fill in real SMTP details, send a test email, then Save — this rolls out to every existing business automatically. Nothing else needs to run by hand.
9. Security considerations
Password never logged, never returned to the browser after save (only a hasPassword boolean) — confirmed live: the form re-shows a masked placeholder, not the real value.
Test-email errors are categorized (auth/connection/timeout) before display — the raw SMTP error is never forwarded.
Encrypted at rest with AES-256-GCM; losing/rotating the key makes the stored password undecryptable (documented in manager/.env.example).
Settings routes sit under manager/src/proxy.ts's existing auth gate — no new auth surface.
Validated before rollout: malformed addresses/ports are rejected before Save, and rollout only fires when configuration is actually complete.
On JSON vs. database for Manager: JSON is the right call here. Reasoning:
Manager has zero DB dependencies today, deliberately — its own package.json description says an earlier attempt at giving it business-data access "repeatedly broke." Introducing SQLite/Prisma just for one settings blob breaks that design principle for not much gain.
The access pattern is a poor fit for a DB anyway: read on page-load/provisioning/rollout, written by a human clicking Save occasionally. No queries, no joins, no need for transactions beyond "read-modify-write one small file."
Reconsider if: multiple admins need to edit concurrently with conflict detection, you want an audit history of who-changed-what-when beyond "last test result," or the settings graph grows relational (e.g., per-business overrides tied by foreign key to businesses) rather than a flat global blob. None of that applies yet.
1. Summary of changes
The installer-created account is no longer special. "Owner" is now literally renamed Administrator and behaves exactly like the existing role/permission system already allowed — the codebase turned out to already define "administrator" structurally (a StaffProfile whose role holds MANAGE_STAFF), not as a name or a business.ownerId-style field, and last-administrator protection already existed. The real work was: closing the gap between that and a full admin-management UI, fixing two OTP-migration bugs the new spec's requirements exposed, and making the existing protection transactional.
Businesses can now have any number of Administrators, added via Invite Administrator (name + email → OTP-verified before access), promoted/demoted via the existing staff role dropdown (now intercepted with a confirmation dialog when it crosses the Administrator boundary), or removed via the existing deactivate flow.
Two real bugs found and fixed, both required for the spec's "unverified email must not get access":
Invitation acceptance (/invite/accept) never actually verified an OTP — it silently relied on a sign-in path that would have failed for every real invitee. It now sends and verifies a real code, reusing the shared OtpCodeStep UI.
A newly-invited (not-yet-accepted) account could sign in via plain email OTP before ever touching the invite link, because sign-in only checked "does a User row exist," not whether it was verified. Fixed in the shared OTP service, benefiting every invitation, not just administrators'.
Last-administrator protection (already existed) is now transactional — a SELECT ... FOR UPDATE lock on the business row closes the race where two concurrent removals of two different administrators could both succeed.
Administrators (any staff member, in fact) can now change their own email — new address must be OTP-verified before it replaces the login email.
No hidden installer privilege: confirmed via code audit and a live install → the account is a plain Administrator-role staff member, provable identically via the Team UI as any later-added one.
NextChair Manager and platform Super Administrator are untouched.
2. Prisma/database changes
No schema changes. One data-only migration: prisma/migrations/20260905214944_administrator_role_rename — UPDATE Role SET name = 'Administrator' WHERE name = 'Owner' AND isSystem = true. Runs through the normal migration-deploy process; not something to run by hand.
3. How Administrator roles are represented
A per-business Role row named "Administrator" (was "Owner"), isSystem: true, holding every Permission — created identically at install and in dev seed data. This is what "Invite"/"Make Administrator" assign.
Whether a StaffProfile currently counts as an Administrator is permission-based (holds MANAGE_STAFF), not a name check — the same definition the codebase already used for last-owner protection (isStaffAdministrator in staff.service.ts). This is deliberately reused rather than introduced, and stays correct if a business ever gets a custom role that also grants MANAGE_STAFF.
No Business.adminId-style field anywhere — confirmed none existed before this work either.
4. How administrator invitations work
Reuses the existing StaffInvitation model and createStaffProfile/createInvitation machinery verbatim — a new inviteAdministratorAction just pre-selects the Administrator role and uses a minimal Name+Email form (invite-administrator-form.tsx). Acceptance (accept-invitation.ts) now requires a real OTP code sent to the invited email before the account activates — expiry (7 days), resend, and revoke are all the pre-existing invitation mechanics (revoke = the existing deactivate action, relabeled "Revoke Invitation" while pending).
5. How last-administrator protection is enforced
Server-side only, inside staff.service.ts's updateStaffAsOwner (role change) and deactivateStaff (remove), each wrapped in prisma.$transaction that first locks the business row (SELECT ... FOR UPDATE), then counts other active Administrators before allowing the change. Verified with a genuine concurrency test (two simultaneous removals against a 2-admin business) and manually in-browser (see below).
6. Installation flow changes
Only the role name and one line of wizard copy changed ("This becomes the first Administrator..."). No structural UX change — still collects name/email, verifies via OTP, creates/activates the account, assigns the role, exactly as before.
7. Files changed
New: the migration; email-change.service.ts + change-email.ts action + change-email-section.tsx; invite-administrator-form.tsx + add-administrator/page.tsx; 3 new integration test files.
Modified — core: staff.service.ts (transactional protection, isStaffAdministrator), staff-invitation.service.ts, otp.service.ts (onlyIfUserExists fix, email-change purpose), default-permissions.ts, permissions.ts, auth.ts (JWT email-refresh, simplified finalize path).
Modified — actions/UI: owner/actions/staff.ts, staff-detail-panel.tsx, team/page.tsx, add-staff-form.tsx, accept-invitation.ts/form/page, otp-code-step.tsx, profile-form.tsx, installer files, seed.ts, plus a terminology sweep ("owner" → "administrator" copy in breaks.ts/time-off.ts/staff-emails.ts/profile page).
8. Manual testing steps
Sign in as the seeded owner → /owner/team shows an Administrator badge and an "Invite Administrator" icon next to "+". Invite someone, accept via /invite/accept?token=... printed to the terminal (code also prints there) — they land signed in with full Manage access. Promote/demote any staff member via the role dropdown on their detail page — crossing the Administrator boundary shows a confirmation dialog. With only one Administrator left, attempting to demote or deactivate them shows "You can't remove the only administrator. Add another administrator first." Change your own email from Profile settings — the new address must be OTP-verified before it replaces the old one.
9. Deployment steps on the VPS
None beyond the normal deploy process — the migration is a standard Prisma migration that prisma migrate deploy (whatever already runs it there) will pick up automatically. No new environment variables. No changes to NextChair Manager or its authentication.
Verification performed: tsc/lint/unit — all clean; full test:integration — 219 passed (up from 207, all new); full test:e2e — same 36/37 as before this work (the 1 failure is a pre-existing, auth-unrelated flake unrelated to any of this); a full live manual pass covering every scenario above, including the last-administrator block and the email-change JWT-refresh behavior.
1. Summary of changes
Password auth → email OTP, everywhere it existed for the tenant/business app: sign-in, customer self-registration (/register), staff-invitation acceptance, and the installer's administrator step. Flow is always: email → 6-digit code emailed → code entered → session created, via one shared service (otp.service.ts) reused by every entry point.
NextChair Manager's own auth was not touched — this only affects tenant-side (business admin/staff/customer) authentication, per the spec's explicit instruction.
Installer improvements: Step 2 (Administrator) now asks only Name + Email + optional Phone, verified inline via the same OTP widget (no redirect). Step 3 (Business Details) pre-fills Business Name from the Manager-provisioned BUSINESS_NAME, shows the booking URL (NEXT_PUBLIC_APP_URL) as a read-only info box instead of asking for it, and has a distinctly-labeled "Business Website (optional)" field with copy clarifying it's not the booking URL.
Installation cannot complete with an unverified admin email — enforced server-side (finalizeOtp runs inside the same DB transaction as user/business creation), not just trusted from client state. A mid-wizard refresh still can't create duplicate admins or partial records — that property was preserved by design, not incidentally.
2. Prisma/database changes
New EmailOtp model (email as PK — one row per address, so a new request naturally invalidates the previous code): codeHash, purpose, expiresAt, attempts, verifiedAt, consumedAt, timestamps. Only a bcrypt hash of the code is ever stored.
User.passwordHash — unchanged schema (already nullable), just annotated as deprecated. No destructive migration; historical rows are untouched.
Migration: prisma/migrations/20260905190022_email_otp_authentication — adds EmailOtp only. Generated and applied against throwaway containers only; the real next-chair-mysql production database was never touched (confirmed not even running on this host during this session).
3. New environment variables
None required for production. One new variable, E2E_TEST_MODE, exists purely for the Playwright suite (set only in playwright.config.ts's webServer.env) — it unlocks a /api/test/otp-code debug endpoint and bypasses OTP rate limiting so the same seeded users can sign in dozens of times across one test run. It must never be set on a real deployment; nothing in the app sets it anywhere else.
4. Email delivery configuration
Nothing new to configure. OTP emails reuse the existing EmailProvider abstraction and EMAIL_PROVIDER env var exactly as configured today — console in dev (prints to server logs), brevo in production via the same SMTP credentials already documented. The OTP email is a new template (otp-email.ts) sent directly (bypassing the async notification/cron pipeline, since a 10-minute-expiry code can't wait for a delivery retry loop).
5. Migration considerations for existing users/installations
Existing users lose password sign-in — the only way to authenticate now is email OTP. There is no in-app fallback; this is an intentional, sharp cutover. If existing customers/staff/admins need advance notice, that's a business/communications decision outside this change.
No account data is deleted or migrated — passwordHash rows are simply never read for auth anymore.
prisma/seed.ts / scripts/ensure-super-admin.ts still set a passwordHash for seeded dev users — harmless and vestigial now, not cleaned up (flagging it rather than touching seed data silently).
Installations already completed are unaffected; only new installer runs go through the OTP admin-verification step.
6. Files changed
New:
src/lib/auth/otp.service.ts — the core requestOtp/verifyOtp/finalizeOtp service
src/lib/auth/__tests__/otp.service.integration.test.ts
src/services/notifications/templates/otp-email.ts
src/features/auth/components/otp-code-step.tsx — shared code-entry UI
src/features/auth/actions/request-sign-in-otp.ts
src/app/api/test/otp-code/route.ts — e2e-only, gated on E2E_TEST_MODE
prisma/migrations/20260905190022_email_otp_authentication/
Modified — auth core: prisma/schema.prisma, src/auth.ts, src/features/auth/schemas.ts
Modified — sign-in/register/invitations: sign-in-form.tsx, register.ts, register-form.tsx, accept-invitation.ts, accept-invitation-form.tsx, staff-invitation.service.ts
Modified — installer: installer.service.ts, schema.ts, install.ts, install-wizard.tsx, install/page.tsx, installer.service.integration.test.ts
Modified — email / e2e infra: console-email-provider.ts, playwright.config.ts, e2e/helpers.ts, and the four e2e/*.spec.ts files (dropped the now-unused password argument from every signIn() call).
7. Verification performed
npx tsc --noEmit, npm run lint, npm test (276 unit) — all clean.
npm run test:integration (207 tests, throwaway DB) — includes 11 new OTP-service tests and a rewritten installer suite; a genuine race condition in requestOtp (concurrent requests for a brand-new email double-hitting the upsert's create branch) was caught and fixed here.
Full npm run test:e2e against a fresh throwaway database: 36/37 passed. The one failure (owner-journeys "opening hours conflict" test) is a pre-existing, auth-unrelated ordering flake — appointment-count ambiguity from earlier tests' bookings — not something this change touches or introduced.
Live manual verification in-browser: existing seeded customer sign-in via OTP; new customer registration via OTP (session created, landed on Home); installer wizard end-to-end through Step 3, confirming email-verification gating, Business Name pre-population from BUSINESS_NAME, and the read-only Booking URL / distinct Business Website fields.
All throwaway Docker containers/DBs used for this work have been removed; the real production database was never touched.
A Delete button now sits next to Edit on each business's detail page, styled in red to distinguish it from the safe actions.
It opens a dedicated confirmation page (matching the Add/Edit page pattern) that clearly lists exactly what gets destroyed — container + media volume, the database by name, the Caddy site/domain, and the stack directory — and requires typing the business's exact slug before the Delete button even becomes clickable. This is a stronger guard than a plain "are you sure?" popup, given what we just learned about how easy it is to lose data here.
deleteBusiness() in provisioning.ts reuses the same primitives as create/edit (composeDown, reloadCaddy, the shared MySQL root connection), and is deliberately best-effort past the first step — if a piece is already gone (matching exactly the kind of half-broken state we just fixed for blake-heritage), it still finishes removing the rest and reports what it couldn't clean up automatically, rather than getting stuck.
Verified live: provisioned a real test business, confirmed its container/database/user/volume/Caddy site all existed, deleted it, and confirmed every single one was actually gone — including checking Caddy's running config (not just the file on disk) to prove the reload took effect. Also confirmed the confirm-button correctly stays disabled for a wrong slug and only enables on an exact match. Your real production database was never touched throughout.
New files:
manager/src/lib/env-file.ts — a shared, surgical .env reader/updater. It only ever touches the specific keys you tell it to (never regenerates the file), so anything hand-edited into a business's .env since provisioning is preserved exactly.
manager/src/app/dashboard/[slug]/edit/ — the edit page, its form component, and its server action.
Modified:
provisioning.ts gained updateBusiness() alongside the existing provisionBusiness(), reusing the same Caddy-site-writing and container-recreation logic (refactored into a shared writeCaddySiteAndReload helper).
stacks.ts/compose-template.ts — extended to track a BUSINESS_NAME field (previously collected on the Add Business form but silently discarded — that gap is now fixed too, so newly created businesses get their name persisted from the start).
The detail page and dashboard list now show the business name, and the detail page has an Edit button.
Per the requirements:
Slug is read-only in the edit form (disabled input, never accepted by the server action — it's bound in from the route, not form data).
Domain uniqueness is validated against every other business (and correctly self-excludes the business being edited).
A domain change rewrites the Caddy site file, reloads Caddy, and force-recreates the container; a name-only change does neither.
No secrets (DATABASE_PASSWORD, AUTH_SECRET, etc.) are ever read into the edit UI — only name/domain/databaseName/status are exposed.
Clear success banners ("Changes saved" vs. "...the container was restarted...") and inline error display, matching the existing Add Business pattern.
The redirect loop was caused by the Manager cookie configuration:
1. secure was based on NODE_ENV=production, so HTTP deployments silently dropped the cookie.
2. Path=/manager excluded /api/manager/*, preventing authenticated Manager API requests.
Implemented
- Root cause fixed in [src/proxy.ts]: Auth.js is now initialized lazily and never invoked when MANAGER_MODE=true.
- Manager requests call handleProxyRequest directly, preserving redirects and CSP.
- Business requests continue through Auth.js with existing protected-route behavior.
- Added coverage in [proxy-manager-mode.test.ts]
1. Root cause of the database error
The Manager proxy was bypassing installation checks, but / still resolved to src/app/page.tsx, which statically imported:
- @/auth
- Prisma
- business branding
- business permissions
- primary business data
The root route now redirects to /manager inside the proxy before the business homepage is evaluated.
2. Root cause of the Auth.js error
The global business SessionProvider was mounted around Manager pages. It automatically requested:
The business homepage is not evaluated first.
5. Provider and authentication separation
Manager mode:
- Uses ManagerProviders
- Uses the signed Manager cookie session
- Does not mount SessionProvider
- Does not call /api/auth/session
- Does not import business Auth.js or Prisma
Business mode:
- Dynamically loads the existing business Providers
- Continues using Auth.js and SessionProvider
- Preserves installation, /owner, and /staff behavior
6. Database independence
Manager mode now avoids database initialization for:
- Proxy requests
- Root metadata and viewport
- Root routing
- Manager pages
- Manager APIs
- Manifest requests
- Health requests
- Business Auth.js requests
- Business cron requests
No fake DATABASE_* variables were added.
7. Tests
Added coverage for:
- Manager / routing
- Manager /manager access
- No installation-state check in Manager mode
- Prisma import protection
- Business installation checks
- /owner protection
- Business Auth.js 404 boundary
- Manager health route without Prisma
- Manager manifest without business branding
8. Validation
Root cause
src/proxy.ts statically imported installation-state.ts, which statically imported Prisma. Additionally, the root layout statically imported business branding, which also loads Prisma. The pathname check was also incorrect for Manager deployments served at /.
Key additions:
* Separate platform Manager stack: [stacks/manager/compose.yaml]
* Dedicated manager authentication with signed sessions.
* Business overview dashboard.
* Add Business workflow with slug suggestions and deployment progress.
* Business detail page with start, stop, restart, update, logs, Dockge, and conservative removal actions.
* Explicit management API endpoints under [src/app/api/manager]
* Shared provisioning service layer in [provisioning.service.ts]
* CLI and UI now use the same [create-business.sh] provisioning logic.
* Registry metadata, database status, container health, and media usage reporting.
* Manager-only entrypoint mode that does not require a business database.
* Updated deployment documentation with Manager setup and Caddy routing.
Validation passed:
* Infra, Caddy, Manager, and generated business Compose configurations.
* Shell syntax checks.
* Registry generation with registry image tags.
* git diff --check.
Stefan Lazic