Archestra Backend Development

Use this skill before changing files under platform/backend/ (except unit tests — see archestra-dev-backend-tests). Run all commands from platform/.

Adding or changing an API endpoint

1. Add a RouteId entry in platform/shared/routes.ts and set it as the route schema's operationId.
2. Add the route handler (see Route layout and Route conventions below).
3. Add the endpoint to requiredEndpointPermissionsMap in platform/shared/access-control.ts (see The 403 footgun).
4. Check the MCP-tool mirror (see below).
5. Run codegen, then validation (see Codegen and Validation).

Route layout

• New routes live in per-entity folders: backend/src/routes/<entity>/<entity>.routes.ts holds ALL of that entity's endpoints; tests are one file per endpoint in the same folder, named <action>.<entity>.route.test.ts (see routes/app/ for a full example).
• Canonical reference: copy the shape of backend/src/routes/virtual-api-key/virtual-api-key.routes.ts and create.virtual-api-key.route.test.ts.
• Legacy flat modules (routes/agent.ts, routes/user.ts, ...) still exist — extend them only for their own entity; new entities get a folder.
• Registration is automatic: registerApiRoutes in backend/src/server.ts iterates Object.values(routes) from routes/index.ts (and routes/index.ee.ts for enterprise routes), so the default re-export in the index file is mandatory or the route silently never registers.

The 403 footgun (deny by default)

• Every new endpoint MUST be added to requiredEndpointPermissionsMap in platform/shared/access-control.ts, keyed by its RouteId. The auth middleware (backend/src/auth/fastify-plugin/middleware.ts, isAuthorized) looks the route up by operationId and denies with 403 when the entry is missing.
• The map is Partial<Record<RouteId, Permissions>> — NOT compiler-enforced; forgetting it compiles fine and fails at runtime.
• An empty entry {} means "any authenticated user". Match permissions with similar existing routes.
• Evaluate RBAC from the database, never from the session-cookie cache — the cookie can carry a stale activeOrganizationId snapshot. Follow backend/src/auth/utils.ts (member role + custom roles resolved via models).

Route conventions

• Plugins are typed as FastifyPluginAsyncZod (fastify-type-provider-zod); schemas are Zod.
• Wrap response schemas with constructResponseSchema from @/types for consistent 400/401/403/404/500 responses.
• Errors: throw new ApiError(status, message) (from @/types) only — never reply.status().send(...); the central error handler formats { error: { message, type } }.
• Routes under /api/ are behind the auth middleware: request.user and request.organizationId are guaranteed — no redundant null checks.
• Pagination: PaginationQuerySchema + createPaginatedResponseSchema from @archestra/shared.
• Sorting: SortingQuerySchema or createSortingQuerySchema from @/types.

Data access

• All DB queries go through backend/src/models/ — never inline Drizzle in routes or services. Create a model file for new entities; business logic stays in services.
• Batch-load related data to avoid N+1 (e.g. AgentTeamModel.getTeamsForAgents in backend/src/models/agent-team.ts), never per-item queries in a loop.
• Entity types come from drizzle-zod (createSelectSchema / createInsertSchema / createUpdateSchema + z.infer), never hand-written interfaces. See the Database Types section in platform/CLAUDE.md.
• Schema changes: use the archestra-dev-migrations skill.

MCP-tool mirror

• When an endpoint's request/response schema changes, check for a mirrored archestra__* tool in backend/src/archestra-mcp-server/ and update its inputSchema and handler in sync.
• New tools need a TOOL_PERMISSIONS entry in backend/src/archestra-mcp-server/rbac.ts — that one IS compile-enforced (Record<ArchestraToolShortName, ...>).

Codegen

After any route/schema change, regenerate and commit the outputs — CI runs pnpm codegen and fails on uncommitted diffs (.github/workflows/on-pull-requests.yml):

pnpm codegen   # from platform/: everything (backend openapi + access-control docs + MCP-server docs, shared api-client + theme css, Grafana dashboard variants via python3)

Or piecewise, in this order: cd backend && pnpm codegen (writes the repo-root docs/openapi.json + docs), then cd shared && CODEGEN=true pnpm codegen:api-client. The CODEGEN=true is required: with it, shared/hey-api/openapi-ts.ts reads the committed docs/openapi.json; without it, it hits a live http://localhost:9000/openapi.json and silently ignores the spec you just regenerated.

Validation

pnpm type-check
pnpm lint
pnpm test
cd backend && pnpm knip   # runs knip:dev AND knip:production — CI runs both; --production ignores tests, so a test-only export fails it

Adding config / env vars

• Name: ARCHESTRA_<PRODUCT_AREA>_<THING>. Then: parse/validate in backend/src/config.ts (+ tests in config.test.ts for custom parsers) → list in platform/.env.example with a comment → document in ../docs/pages/platform-deployment.md → expose via backend/src/routes/config.ts + useFeature() if the frontend needs it.

Related skills

• archestra-dev-backend-tests — unit tests, mocking rules, DB fixtures.
• archestra-dev-migrations — Drizzle schema and migration changes.
• archestra-dev-frontend — consuming the regenerated API client.