Backend Patterns

Backend Patterns — Yosemite Crew

Description

Use this skill when working on apps/backend. Covers Express.js architecture, service/controller patterns, validation, error handling, and healthcare-specific integrations.

TRIGGER: any task in apps/backend — new endpoints, services, models, or integrations.

Architecture

apps/backend/src/
  routers/        ← Express route definitions (thin — just register handlers)
  controllers/    ← Request/response handling, input validation
  services/       ← Business logic (no req/res objects here)
  models/         ← legacy data models; no new files here (use Prisma via packages/database)
  queues/         ← BullMQ job definitions
  workers/        ← BullMQ worker processors
  integrations/   ← External services (IDEXX, Merck, Stripe, Firebase, AWS)

Pattern: Router → Controller → Service → Model

Controllers call services. Services call models. Never put business logic in controllers or routers.

Validation

Use Zod for request validation. Never trust raw req.body.

import { z } from 'zod';

const CreateAppointmentSchema = z.object({
  patientId: z.string().uuid(),
  date: z.string().datetime(),
});

// In controller:
const data = CreateAppointmentSchema.parse(req.body);

Database

• PostgreSQL via Prisma is the database for all new code; Prisma owns schema + migrations (see @yosemite-crew/database).
• Prisma only for new persistence. All new models and queries go through Prisma (packages/database); do not add new files under src/models/ — this matches apps/backend/AGENTS.md.
• Never access data directly from controllers — always go through services/models.

Authentication

SuperTokens is the auth provider behind the provider-neutral boundary in packages/auth (#1672), initialized by initSuperTokens in app.ts. Product code uses the session guards in src/middlewares/auth.ts and never imports a provider SDK (eslint-enforced). Pick the guard by product surface:

• requireWebAuth - staff / PIMS web routes.
• requireMobileAuth - pet-parent mobile routes.
• requireAnyAuth - routes genuinely shared by both.

Never roll custom auth.

Authorization: derive the tenant from the resource

Authentication only proves who is calling. withOrgPermissions() then proves the caller belongs to the organisation named by the request (route param, x-org-id, query, or body). On a route addressed by a resource id that is not enough on its own - a caller can name an organisation they legitimately belong to while addressing another tenant's record.

• On an id-addressed route, use the resource-derived middleware so the organisation comes from the record: withAppointmentOrgPermissions, withInvoiceOrgPermissions, withPaymentOrgPermissions, withPaymentIntentOrgPermissions, withTaskOrgPermissions, withInventoryItemOrgPermissions, withEncounterOrgPermissions, withCaseOrgPermissions, withRenderedDocumentOrgPermissions, withRoomUnitOrgPermissions, withRoomUnitGroupOrgPermissions. Add a new one via withResourceOrgPermissions rather than hand-rolling a lookup.
• In a controller, scope on (req as OrgRequest).organisationId - the value the middleware authorized. If the request also carries an organisation, it must agree with that value; reject a mismatch rather than preferring it.
• Take the organisation as a required service argument, never organisationId?. Prisma drops undefined where-fields, so an optional scope silently becomes an unfiltered, cross-tenant query. Required turns that into a build error.
• Document has no organisation column - scope document queries through documentWhereForOrg() (src/services/document-scope.ts), which expresses the patientOrganisation join and the PMS visibility flag once.
• requirePermission([a, b]) is any-of. That is the intended idiom for an :any/:own pair of the same resource. An array naming two different resources grants each to holders of the other - require both instead.
• Read the acting user from the verified session (req.userId). Do not use resolveUserIdFromRequest for an authorization decision: it falls back to a client-supplied x-user-id header. It is fine for attribution only.

Background Jobs

BullMQ is the queue system. Jobs go in queues/, processors in workers/.

// Never process jobs inline in a request handler
// Always enqueue and let a worker handle async operations
await emailQueue.add('send-reminder', { appointmentId });

Healthcare Integrations

• FHIR types from @yosemite-crew/fhir — use these, never invent custom health data shapes.
• IDEXX and Merck integrations live in src/integrations/ — extend there, never inline.

Logging

Use Winston for all logging. Never use console.log in production code.

import logger from 'src/utils/logger';
logger.info('Appointment created', { appointmentId });
logger.error('Payment failed', { error, userId });

Gotchas

• Do not refactor backend architecture unless explicitly asked — the user's CLAUDE.md is explicit about this.
• Zod .parse() throws on invalid input — use .safeParse() when you want to handle errors gracefully.
• BullMQ jobs are persisted in Redis — make job processors idempotent.
• All Stripe webhook handlers must verify the signature before processing.
• Firebase Admin SDK is initialized once — never re-initialize it in a handler.