prisma-composer-core-concepts

Prisma Composer core concepts

A Prisma App is a tree of typed declarations composed in TypeScript and handed to the prisma-composer CLI. This file covers structures, hierarchies, relationships, and workflows: the concepts you cannot observe from the code or the CLI's help output. It is not a CLI reference; discover any individual command and its flags with --help. Commands named here belong to the prisma-composer CLI itself; a host CLI that embeds Composer may not carry every verb, so confirm a command exists via --help rather than inferring it. The Prisma platform moves fast, so treat this file as the stable conceptual core and find current, fuller documentation at https://www.prisma.io/docs. For working code, read examples/ in the prisma/composer repo.

Two principles govern everything and are binding (docs/design/01-principles/):

1. Your code never reads its environment. Dependencies, configuration, credentials, and the port all arrive through the service node, typed. process.env is never the answer.
2. Composer never bundles or transforms your code. You build with your own bundler; the framework assembles the built output by deterministic steps and hands it to the configured deploy target.

Declarations are data

Everything you author is a declaration: plain data describing a piece of the app, executing nothing when imported. Three node kinds exist:

Nodes connect through ports: deps declares what a node requires, expose declares what it offers. Wiring happens in a Module's builder via provision(), and the root Module, handed to the CLI, is the App:

// module.ts
import { module } from '@prisma/composer';

export default module('store', ({ provision }) => {
  const catalog = provision(catalogModule);
  provision(storefrontService, { deps: { catalog: catalog.rpc } });
});

Because ports are typed, the compiler verifies every wire. A dependency wired to the wrong producer, a missing RPC handler, a literal input value of the wrong shape: all of it fails tsc, not the deploy. Env-bound input is the exception: those values exist only at deploy, so secret-binding mismatches and missing platform variables surface as early deploy-time refusals instead (see Two channels below). Typecheck, then build, then deploy; don't use the cloud to find out whether the wiring is correct.

Composer itself is target-agnostic: @prisma/composer carries authoring, testing, and the CLI, coupled to no platform. A deploy target is an extension registered in the deploy config; @prisma/composer-prisma-cloud is the Prisma Cloud target and the one this skill's deploy sections assume. Its root exports compute, rawPostgres, bucket, envSecret, and envParam; the ORM vocabulary (postgres, dataContract) lives under the /orm subpath, alongside the shared /cron, /storage, /streams, /auth, and /email modules. These are the only two Composer packages a basic Prisma Cloud app needs, and nothing installs them for you: a fresh project starts with neither, so add both as dependencies first. An extension adds its own prisma-composer-* package alongside them. Compose an existing Module before implementing a capability yourself; wiring one in is a couple of lines.

Within the entry graph (everything reachable from module.ts), write relative imports with explicit .ts extensions (./service.ts, with allowImportingTsExtensions in tsconfig): that form resolves everywhere. The prisma-composer CLI also maps ./service.js and extensionless ./service to the .ts source, but other hosts may not.

The service node is the only doorway

Your runtime code receives everything from the service declaration it imports:

1. service.load(): dependencies (typed RPC clients, database bindings).
2. service.input(): the whole input as one schema-validated object; credentials in it are redacting SecretString boxes.
3. service.port(): the reserved port to bind (default 3000).

A service declaration is pure data; the server entry is what your build produces and the platform boots:

// service.ts
export default compute({
  name: 'auth',
  deps: { db: rawPostgres() },
  build: node({ module: import.meta.url, entry: '../dist/server.mjs' }),
  expose: { rpc: authContract },
});

// server.ts
const { db } = service.load(); // { url }: you construct your own client
const handler = serve(service, {
  rpc: { verify: async ({ token }) => ({ ok: token.length > 0 }) },
});
Bun.serve({ port: service.port(), hostname: '0.0.0.0', fetch: handler });

The consumer declares deps: { auth: rpc(authContract) } and gets a typed client back from load().

Two channels: dependencies and input

The service declares its whole incoming configuration, plain values and credentials together, as one Standard Schema (arktype is the house choice). A credential is a field typed as secretString() from @prisma/composer/arktype; conditional legality ("no stripe key unless billing is on") is an ordinary schema union. The binding at provision() mirrors the schema's shape; envSecret('NAME') names the platform variable and never carries the value.

Rules that bite:

1. Secretness is enforced by validation. A literal bound where the schema expects SecretString fails the deploy; envSecret bound to a plain string field fails the same way.
2. envParam values arrive as raw strings; bind them to string fields. The stage's platform variable is the store; the deploying shell only seeds a missing name (and the deploy fails early, naming the variable, when both lack it). Changing the platform value needs a redeploy.
3. Absence is the schema's call. An env-bound field whose variable is unset or empty resolves to key omitted, which is legal only if the schema allows it (optional field, union arm). The deploy report prints the serialized input document (secrets ride as {"$secret":"VAR"} pointers) and every key that resolved absent.
4. The reserved port is outside the schema. Read it through service.port(), never process.env. The framework also exports PORT for Next.js standalone, which binds it itself.
5. A Module forwards a secret need without learning the platform name. Declare secrets: { signingKey: secret() } on the Module boundary and pass the forwarded ref as a binding leaf; the parent binds the real source.
6. input.apiKey.expose() is the only way to a secret's value; the box redacts everywhere else (logs, JSON, errors).

Contracts and RPC

A contract is the typed interface through which services communicate. It lives with the service that owns it, typed by any Standard Schema validator, and both provider (serve(), exhaustive over the contract's methods at compile time) and consumer (rpc(contract)) reference the same value. Calls travel as RPC over HTTP. Two behaviours are provisioned for you and must not be reimplemented:

1. Service keys. At deploy, Composer mints a distinct unguessable key per consumer→provider binding; serve() returns 401 to anything else before the handler runs. Nothing in your code declares it. Consequences: don't build your own service-to-service auth, and don't curl a deployed /rpc/<method> to check it works. An unwired caller always gets 401, which looks like a broken deploy and isn't. Debug through a consumer, or locally, where nothing is enforced. Keys are per binding (one leaking can't impersonate another consumer), service-scoped (any valid key reaches every method; split services to gate separately), rotated only by removing the binding or destroying the stack and redeploying, and stored in deploy-owned COMPOSER_* variables you never hand-edit.
2. Idempotency and retries. Every generated-client call carries an Idempotency-Key; dropped calls retry with backoff, and serve() runs one call per key, replaying the completed answer to late retries. Every method is therefore safely retryable and no contract declares anything about it (there is no "is this idempotent" flag; don't invent one). A handler may take an optional third argument (input, deps, ctx) and read ctx.idempotencyKey (string | undefined) if it needs exactly-once beyond one instance's memory; most don't. Locally and in tests nothing is provisioned, so serve() passes every call through: never supply a key in test inputs.

Builds are yours

You build, the framework assembles. For a plain server process, entry must point at a single self-contained ESM file: everything inlined except runtime built-ins (bun, bun:*, node:*). Deploy copies that one file and never ships node_modules, so anything left un-inlined fails at boot, not at deploy. Rules that bite:

1. Two services in one package means two separate builds, one per entry. A single multi-entry build splits shared code into a chunk neither output contains.
2. A directory build uses dir + entry (dir relative to the service module, entry a file inside dir; ../ is an error). The tree is copied verbatim, so the server must resolve siblings against import.meta.url, not the working directory. The tree must contain no symlinks: the packager rejects them, names the link, and assembly fails.
3. Next.js: next build with output: 'standalone' is the whole build; nextjs({ module, appDir }) names the app root. Any page or action that calls load() needs export const dynamic = 'force-dynamic', because the runtime environment doesn't exist at build time and Next ignores runtime env for prerendered routes.
4. Always build before deploy or dev. Neither builds for you.

Deploy configuration lives in prisma-composer.config.ts (or .mts, .mjs, .js; nearest ancestor of the entry wins, .ts first within a directory). It registers extensions (prismaCloud(), nodeBuild(), nextjsBuild() when the app has a Next.js service) and the deploy-state backend (prismaState()). It is read by the CLI's operations (deploy, destroy, and dev; a dev run without one refuses, naming the missing file) and never imported by app code.

Databases and migrations

Two kinds of Postgres dependency:

1. rawPostgres(): the binding is { url } and the app owns its client.
2. postgres(...): a Prisma-ORM-typed database. The binding is { url, client } (ADR-0040): the raw connection URL plus the typed client Composer constructs from your data contract, lazily on first access, so queries go through binding.client and are compile-time checked. Both postgres and dataContract import from @prisma/composer-prisma-cloud/orm, not the package root. One dataContract-wrapped value (emitted from contract.prisma by prisma contract emit) is referenced by both the dependency end (deps: { db: postgres(catalogData) }) and the resource end, which also names the prisma.config.ts path so the deploy's migration step can reload the emitted contract.json and find migrations/.

Deploys are replay-only: they apply the migrations committed under migrations/ and never create schema themselves. Every schema change, including the first schema of a new database, follows one loop:

1. Edit contract.prisma.
2. prisma contract emit regenerates contract.json + contract.d.ts.
3. prisma migration plan --name <slug> authors the migration (on an empty graph this authors the baseline).
4. Commit migrations/ with the change, then deploy. A fresh database replays the whole path from empty.

If no authored path reaches the target contract, deploy (and dev against a stale local database) refuses with MIGRATION_PATH_NOT_FOUND; its message lists the two ways out: author the missing migration, or, when iterating against a local database only, prisma db update. The tracked migration resource persists only compact contract identity in deploy state; if the emitted contract artifact named by prisma.config.ts is missing, unreadable, or no longer matches the declared dataContract(...) value, deploy fails before touching the database. Never skip step 3 before a deploy. See examples/store/modules/catalog for the complete pattern.

Deploy model: converge, don't script

Deploy compares the declared topology against recorded deploy state and applies only the difference. Re-deploying with nothing changed is a no-op; removing a node removes its deployed resource. The Prisma Cloud target requires exactly two environment variables: PRISMA_SERVICE_TOKEN and PRISMA_WORKSPACE_ID. There is no interactive login.

Stages. A stage is an environment name chosen on the command line at deploy time, never written in the topology. The identical graph deploys everywhere. On the Prisma Cloud target, a Prisma App is one Project and a stage is a Branch of it, with its own running services, its own empty database, its own configuration. A stage name must be a valid git ref name; an invalid name is a hard error.

Destroy always requires an explicit target: a bare destroy is an error, and naming a stage and production together is too. Destroying a stage deletes its Branch after removing its resources. Destroying production removes only the resources inside the production Branch, never the Branch itself directly; once the Project is empty it is deleted too, and that deletion takes the production Branch with it. A Project still holding another stage's resources is kept. Destroy never creates anything: destroying a never-deployed stage fails rather than standing one up.

The engine underneath is alchemy. Convergence is executed by alchemy, a third-party infrastructure-as-code engine that arrives as an ordinary, exactly-pinned npm dependency of @prisma/composer (2.0.0-beta.74 at this library version). Your code never imports or configures it; consult alchemy's own docs for the engine itself. What matters operationally:

Alchemy is resolved from the nearest node_modules/.bin, including hoisted ancestor directories. Windows resolves alchemy.exe, then alchemy.cmd, then the extensionless shim; POSIX resolves alchemy. No global Alchemy installation is needed.

1. Deploy and destroy write the pipeline's results to a generated, gitignored stack file at .prisma-composer/alchemy.run.ts, then run the alchemy CLI against it as a child process; dev does the same at .prisma-composer/dev/alchemy.run.ts with local providers. The file carries the computed values as literals but reads credentials via fromEnv(), so nothing sensitive lands on disk, and it is regenerated every run: output, not configuration, never edited.
2. Failures are bisectable through that file. A failing deploy names its path; running alchemy deploy .prisma-composer/alchemy.run.ts directly separates "the framework computed the wrong thing" from "the engine or platform rejected the right thing". An engine failure surfaces as DEPLOY.ENGINE_FAILED carrying the exit code and that reproduce command; the child's live output streams to the terminal either way.
3. Destroy evaluates the same stack program as deploy, and evaluating it packages the assembled bundles, so an app must be built before it can be torn down.
4. alchemy is why the effect pin exists: it resolves the effect constellation, and a hoisted newer effect halts every command (failure mode 1 below).

The deploy report ends with the app's own topology: authored names, the platform resource each became, and public URLs. Read ids out of it rather than hunting in the Console. A URL appears only where the address is genuinely public: a service prints one, a database never does, and a node whose product is secret material reports no resource line at all.

Connection contract refusals. A connection declares the values it needs by name; a producer that omits one fails the deploy, naming the edge, the param, and what the producer did supply:

Connection input "auth.db" declares param "url", but its producer "db" did not
supply it — the producer's outputs carry [host].

This is a deploy-time refusal, not a broken deploy, and it can appear on an app whose code didn't change (the gap used to pass silently as undefined and crash the consumer at boot). Fix whichever end is wrong; don't mark the param optional unless absent really is legal. Only reachable if you authored the connection or an extension on one side.

Driving deploys from code. @prisma/composer/control exposes typed deploy, destroy, dev, and log returning structured results. Failures come back as { ok: false, failure } with a dotted failure.code from a closed registry (e.g. ASSEMBLE.BUILD_FAILED, DEPLOY.ENGINE_FAILED, DEPS.EFFECT_VERSION_CONFLICT); branch on the code, not the message. A non-structured rejection out of an operation is a bug in composer, not an expected failure.

Local development

The dev command runs the whole app on this machine, wired as it deploys, against local emulators. No cloud credentials are needed or read. Concepts that surprise:

1. It runs the same pipeline as deploy, so build first, exactly like deploy. It watches built output and restarts a service when its build changes.
2. Ctrl-C stops the app's processes but leaves local databases, buckets, and their data up: the next dev is a warm start. Starting clean, wiping this app's local instances and data first, is an explicit opt-in flag.
3. dev does not print service logs; log is a separate, read-only command that follows the already-running app's merged logs. It never builds, provisions, starts, or stops anything.
4. An unset secret doesn't block a local run: it becomes a placeholder plus a warning, and only the code path that spends it fails, at the external service it calls.
5. Windows isn't supported yet.

Testing is an environment seam

A test is just another environment: one where you decide what load() and input() return, never by editing the code under test.

mockService returns a copy of the service whose load() yields your doubles (type-checked against the declared deps) and whose input() yields the object passed under the reserved input key (required exactly when the service declares an input schema; handed over as-is, not validated). Wiring the module substitution is your runner's job (vi.mock in Vitest, mock.module in bun test).

bootstrapService boots the service's real built entry in-process against a config you choose; drive it over real HTTP. Gotchas:

1. service.port must be concrete: the entry self-listens, and no OS-assigned port is reported back.
2. There is no close(); run each integration-test file in its own process (bun test does).
3. Next.js services take a third argument, a boot thunk, resolved with standaloneServerPath from @prisma/composer/nextjs/control.
4. A service with an input schema takes input in the config, a binding exactly like provision()'s, run through the real serialize/read path.

A dependency's type is its contract, so any value of that shape is a valid double: a bare object, the real client over an in-memory handler, or a real local server. Ship a dependency's fake from its own package as a /fake entry point, outside src/, so the fake and the real service share one contract.

Building blocks and extensions

First-party Modules ship inside @prisma/composer-prisma-cloud and provision exactly like your own:

bucket() (imported alongside rawPostgres) is a raw S3-compatible bucket: the dependency end receives { url, bucket, accessKeyId, secretAccessKey }, shape-compatible with /storage's s3() dependency, so a service wired to s3() can be rewired to a bucket resource unchanged.

An extension (a package bringing its own Modules, resources, or deploy target) is published on npm as prisma-composer-*. The ecosystem is new: today the blocks above plus your own Modules are the whole set, so verify a prisma-composer-* package exists on npm before reaching for it.

Failure modes quick reference

1. Every prisma-composer command halts at start-up on an effect version conflict (Dependency conflict: alchemy resolves effect@...). The app, or one of its dependencies, pins a different effect and the package manager hoisted it over Composer's pin. Match the app's own effect to @prisma/composer's exact pin, or force it with "overrides": { "effect": "<pin>" } in the app's package.json (yarn: resolutions; pnpm: pnpm.overrides), then reinstall. A plain Composer app never hits this: the public packages pin every effect-family package alchemy would float.
2. A deployed /rpc/<method> returns 401 to anything but a wired peer. Not a broken deploy; see Contracts above.
3. Scale-to-zero closes idle database connections. A persistent client crashes into a 502 restart loop unless the pool is small and reconnect-friendly (new SQL({ url, max: 1, idleTimeout: 10 }) for Bun) and the process logs uncaughtException/unhandledRejection instead of dying. Under dev watch-restarts against the local emulator, add prepare: false as well: restarted processes collide on prepared-statement names in the emulator's shared session.
4. Cold starts reset service-to-service connections. A call into a scaled-to-zero service can get ECONNRESET; retry it.
5. Bind 0.0.0.0, not loopback. The platform routes external HTTP to the VM; a loopback-only listener is unreachable.
6. The ingress buffers streaming responses. An open SSE tail delivers nothing and times out at 60s; don't build on streamed HTTP responses.
7. Naming rules fail at load, not typecheck. Provision ids and declared node names must be ASCII letters and digits only ([A-Za-z0-9]): they derive config keys and address segments, so a hyphenated name like my-db passes tsc and then fails the load. The root module's name is exempt. A provision id shorter than 3 characters is rejected by the platform (name the database 'database', not 'db'), and a service whose name equals its enclosing Module's reads as auth.auth unless given an explicit id.
8. MIGRATION_PATH_NOT_FOUND: see Databases above; author the missing migration, don't skip the plan step.
9. Date/time columns hand back Temporal.* values on read. Bun and stock Node ship no global Temporal, so a service with DateTime contract columns compiles and deploys, then fails on the first timestamp read. Provide the global at the server entry (import 'temporal-polyfill/global') or use string column types.

What Composer doesn't do yet

Name the gap instead of inventing an API:

1. No interactive auth in the prisma-composer CLI. Its deploys authenticate only via a static PRISMA_SERVICE_TOKEN; there is no login flow.
2. No in-memory contract bindings. A dependency can't yet be wired to a co-located handler without HTTP; use bootstrapService with a loopback fake.
3. RPC over HTTP is the only contract kind. No gRPC, WebSocket, or streaming contracts.

For anything else missing, check examples/, docs/design/10-domains/, and docs/design/90-decisions/ in the prisma/composer repo, then file an issue there rather than guessing.