prisma-guard

Schema-driven security layer for Prisma Generate input validation, query shape enforcement, and tenant isolation directly from your Prisma schema.

prisma-guard helps prevent three common classes of backend mistakes:

invalid input reaching Prisma
unsafe or overly broad query shapes
missing tenant filters in multi-tenant systems

client request
      ↓
.guard(shape)
      ↓
validated input + allowed query shape
      ↓
tenant scoped query
      ↓
database

Why this exists
What prisma-guard does
Architecture
Install
Quick start
Before / After prisma-guard
Schema annotations
The guard API
Logical combinators in where shapes
Relation filters in where shapes
Mutation return projection
Enforced projection mode
Upsert
Named shapes and caller routing
Context-dependent shapes
Automatic tenant isolation
Multi-root scope behavior
findUnique behavior
Output shaping
Strict Decimal mode
Security model
Limitations
Advanced: SQL-backed runtimes
Error handling
How it works internally
Why this approach
Performance characteristics
Security philosophy
Recommended production configuration
When to use prisma-guard
Version compatibility
Design principles
Comparison
Roadmap
License

Why this exists

Prisma is powerful, but most real backends eventually hit the same problems.

Missing input validation

await prisma.user.create({
  data: req.body,
})

The client controls the entire payload.

Dangerous query shapes

A client-controlled include chain can traverse relations and select sensitive fields from unrelated models:

await prisma.project.findMany({
  include: {
    tasks: {
      include: {
        comments: {
          include: {
            author: {
              select: {
                email: true,
                passwordHash: true,
                resetToken: true,
              },
            },
          },
        },
      },
    },
  },
})

Tenant isolation bugs

await prisma.project.findFirst({
  where: { id: { equals: projectId } },
})

If projectId belongs to another tenant, data can leak.

What prisma-guard does

prisma-guard turns your Prisma schema into a runtime data boundary layer.

Feature	Description
Input validation	Zod schemas generated from Prisma types
Query shape enforcement	Only allowed query shapes pass validation
Tenant isolation	Prisma extension injects tenant filters
Schema-driven	Rules come directly from your Prisma schema

The goal is simple:

Clients should not be able to accidentally or maliciously escape the data boundary defined by your schema.

prisma-guard is focused on data boundaries, not RBAC. Role-based access control is intentionally out of scope.

Architecture

prisma-guard sits between your application and Prisma Client. The .guard(shape) call defines the boundary; the chained Prisma method validates and executes in one step.

┌───────────────┐
│   Client      │
│  (API / RPC)  │
└───────┬───────┘
        │ request
        ▼
┌──────────────────────┐
│   .guard(shape)       │
│  validates input +    │
│  enforces query shape │
└─────────┬────────────┘
          │ validated args
          ▼
┌──────────────────────┐
│ Tenant Scope Layer    │
│ (Prisma extension)    │
│ injects tenant filter │
└─────────┬────────────┘
          │ scoped query
          ▼
┌──────────────────────┐
│    Prisma Client      │
└─────────┬────────────┘
          │ SQL
          ▼
┌──────────────────────┐
│       Database        │
└──────────────────────┘

Install

npm install prisma-guard zod @prisma/client

Peer dependencies:

zod ^4
@prisma/client ^6 || ^7

Both zod and @prisma/client must be installed before running prisma generate. The generator validates @zod directives against real Zod schemas at generation time.

Generated output files (index.ts, client.ts) are TypeScript. A TypeScript-capable build pipeline is required.

The generated output uses .js extension imports in TypeScript source (e.g. import { ... } from './index.js'). This requires ESM-aware module resolution in your TypeScript config — either "moduleResolution": "NodeNext" or "moduleResolution": "Bundler" in tsconfig.json. The classic "moduleResolution": "node" setting is not compatible.

Quick start

1. Add the generator

generator guard {
  provider = "prisma-guard"
  output   = "generated/guard"
}

2. Generate

npx prisma generate

This emits a ready-to-use client.ts with all type mappings and a pre-wired guard instance.

3. Set up the Prisma client

import { AsyncLocalStorage } from 'node:async_hooks'
import { PrismaClient } from '@prisma/client'
import { guard } from './generated/guard/client'

const store = new AsyncLocalStorage<{ tenantId: string }>()

const prisma = new PrismaClient().$extends(
  guard.extension(() => ({
    Tenant: store.getStore()?.tenantId,
  }))
)

4. Use it

await prisma.project
  .guard({
    data: { title: true },
  })
  .create({ data: req.body })

await prisma.project
  .guard({
    where: { title: { contains: true } },
    orderBy: { title: true },
    take: { max: 50, default: 20 },
  })
  .findMany(req.body)

That's it. Input is validated, query shape is enforced, tenant scope is injected. All in one chain.

Before / After prisma-guard

Without prisma-guard

await prisma.project.create({
  data: req.body,
})

await prisma.project.findMany(req.query)

await prisma.project.findFirst({
  where: { id: { equals: projectId } },
})

Problems:

unvalidated input
unrestricted query shape
missing tenant filter

With prisma-guard

await prisma.project
  .guard({ data: { title: true } })
  .create({ data: req.body })

await prisma.project
  .guard({
    where: { title: { contains: true } },
    take: { max: 50, default: 20 },
  })
  .findMany(req.body)

await prisma.project
  .guard({
    where: { id: { equals: true } },
  })
  .findFirst({ where: { id: { equals: projectId } } })

Risk	Mitigation
arbitrary input	`data` shape restricts writable fields + Zod
expensive queries	shape whitelist on where, include, take, orderBy
cross-tenant access	automatic scope injection via extension

Schema annotations

Mark tenant root models

Use @scope-root on models that represent tenant roots.

/// @scope-root
model Tenant {
  id   String @id @default(cuid())
  name String
}

Models with a single unambiguous foreign key to a scope root are auto-scoped. A model can be scoped by multiple roots if it has foreign keys to different scope root models — see Multi-root scope behavior.

If a model has multiple foreign keys to the same scope root, the ambiguous root is excluded from that model's scope entries when onAmbiguousScope is "warn" or "ignore", and causes a generation error when onAmbiguousScope is "error" (the default). Other non-ambiguous roots on the same model are still auto-scoped.

Add field-level validation with `@zod`

model User {
  id    String @id @default(cuid())
  /// @zod .email().max(255)
  email String
  /// @zod .min(1).max(100)
  name  String?
}

@zod chains apply automatically when the field appears in a data shape with true.

@zod directives are validated during prisma generate. The generator validates directive syntax, checks that each chained method is in the allowed list, and verifies method compatibility by advancing through the chain. Type-changing methods (such as .nullable(), .optional(), .default()) advance the schema type, so a chain like .nullable().email() is correctly rejected if .email() does not exist on the nullable wrapper. Note: some argument-level type mismatches may only be caught if Zod throws at schema construction time.

For list fields, @zod chains apply to the z.array(...) schema, not to individual elements. For example, .min(1) on a String[] field enforces a minimum array length of 1, not a minimum string length.

Supported `@zod` methods

The @zod DSL supports a restricted subset of Zod methods. These are the allowed methods:

String validations: min, max, length, email, url, uuid, cuid, cuid2, ulid, trim, toLowerCase, toUpperCase, startsWith, endsWith, includes, regex, datetime, ip, cidr, date, time, duration, base64, nanoid, emoji

Number validations: int, positive, nonnegative, negative, nonpositive, finite, safe, multipleOf, step, gt, gte, lt, lte

Array validations: min, max, length, nonempty

Field modifiers: optional, nullable, nullish, default, catch, readonly

Note on field modifiers: prisma-guard already handles optional and nullable based on Prisma field metadata (isRequired, hasDefault). Adding @zod .optional() or @zod .nullable() explicitly will apply the Zod method on top of what prisma-guard already does, which may cause double-wrapping. Use these only when you need to override prisma-guard's default behavior.

Note on default and catch: a @zod .default(...) chain adds a Zod-level default, which means Zod will fill in the value if it's undefined. A @zod .catch(...) chain provides a fallback value when parsing fails. The generator detects both .default() and .catch() in chains and emits a ZOD_DEFAULTS map. The create completeness check honors Prisma's @default attribute, @zod .default(...), and @zod .catch(...) — a required field with any of these sources of default is not flagged as missing from create data shapes. Prisma's @default remains the primary source of truth; @zod .default(...) and @zod .catch(...) are additional signals.

When a field has @zod .default(...) or @zod .catch(...) and appears in a data shape as true, prisma-guard preserves the Zod default/catch behavior in create mode by not wrapping the schema with .optional(). This ensures that omitting the field from client input triggers the Zod default or catch value rather than passing undefined through. If such a field is omitted from the data shape entirely (not listed as a key), the runtime auto-injects its default value as a forced field — the client cannot provide it, and the Zod default is always applied.

Note on chain ordering: type-changing methods like .nullable(), .optional(), .default(), and .catch() alter the wrapper type. Methods that follow a type-changing method must exist on the resulting wrapper, not on the original base type. For example, .email().nullable() is valid (.email() returns a string schema, .nullable() wraps it), but .nullable().email() is invalid (.nullable() returns a nullable wrapper that does not have .email()).

Supported argument types in `@zod` directives

The directive parser accepts these argument types: strings ('hello', "hello"), numbers (42, -3.14, 1e2), booleans (true, false), arrays ([1, 2, 3]), regex literals (/^[a-z]+$/i), and object literals ({offset: true}). Identifiers, template literals, null, NaN, Infinity, and function calls are not allowed.

`refine` replaces `@zod` chains

Both guard.input({ refine }) and inline refine functions in data shapes bypass @zod chains. The function receives the base Zod type (without @zod chains applied). This is by design — refine is a full override, not a modifier on top of @zod.

guard.input('User', {
  refine: {
    email: (base) => base.email().max(320),
  },
})

In this example, any @zod directive on the email field in the Prisma schema is ignored. The refine callback is the sole source of validation for that field.

Refine callbacks and inline refine functions must return a valid Zod schema. If the callback throws or returns a non-Zod value, a ShapeError is raised.

When a refine function returns a schema that handles undefined input (e.g. by including .default(...) or .catch(...)), prisma-guard detects this at runtime and preserves the default/catch behavior by not wrapping the schema with .optional() in create mode.

The guard API

.guard(shape) is available on every model delegate. It returns an object with all Prisma methods. The shape defines the boundary; the method validates and executes.

Data shape syntax

Each field in a data shape accepts one of four value types:

true — the client may provide this value; @zod chains apply automatically
literal value — the server forces this value; the client cannot override it
force(value) — the server forces this value; required when the value is literally true (see The force() helper)
function (base) => schema — the client may provide this value; the function receives the base Zod type (without @zod chains) and returns a refined schema

import { force } from 'prisma-guard'

await prisma.project
  .guard({
    data: {
      title: (base) => base.min(1, 'Title required').max(200),
      status: true,
      priority: (base) => base.refine(v => v >= 1 && v <= 5, 'Priority 1-5'),
      createdBy: currentUserId,
      isActive: force(true),
    },
  })
  .create({ data: req.body })

In this example, title and priority use inline refines for custom validation and error messages, status uses @zod chains from the Prisma schema, createdBy is forced to currentUserId, and isActive is forced to true using the force() helper.

Relation fields are not permitted in data shapes. Attempting to use a relation field in a data shape throws ShapeError. See Limitations.

The `force()` helper

The value true in a shape always means "client-controlled." This creates ambiguity when you need to force a boolean field to the literal value true. The force() helper resolves this:

import { force } from 'prisma-guard'

data: { isActive: true }         // client-controlled — client sends any boolean
data: { isActive: false }        // forced to false
data: { isActive: force(true) }  // forced to true
data: { isActive: force(false) } // also valid — equivalent to just false

force() works in both data shapes and where shapes:

where: {
  published: { equals: true },          // client-controlled — client sends any boolean
  isDeleted: { equals: false },          // forced to false
  isActive: { equals: force(true) },     // forced to true
}

force() wraps the value in a marker object. It can wrap any value type, not just booleans. Using force() on non-true values is allowed but unnecessary — only the literal true collides with the client-controlled sentinel.

Query shape syntax

For read operations, true means the client may provide this value and literal values are forced:

Reads

await prisma.project
  .guard({
    where: { title: { contains: true } },
    orderBy: { title: true },
    take: { max: 100, default: 25 },
  })
  .findMany(req.body)

The client can only filter by title, sort by title, and take up to 100 rows. Everything else is rejected.

Creates

await prisma.project
  .guard({
    data: { title: true, status: true },
  })
  .create({ data: req.body })

Only title and status are accepted from the client. @zod chains apply automatically.

For create operations, guard validates that all required fields without defaults are accounted for in the data shape — either as client-allowed (true or function), forced (literal value or force()), as scope foreign keys that the scope extension will inject automatically, or as fields with a @zod .default(...) or @zod .catch(...) directive. If a required field is missing from the shape and has no Prisma @default, no @zod .default(...) or @zod .catch(...), and is not a scope FK, guard throws ShapeError at shape evaluation time.

Fields with @zod .default(...) or @zod .catch(...) that are omitted from the data shape are automatically injected as forced values at runtime. The Zod schema is evaluated with undefined input and the resulting default or catch value is included in the create data. This ensures the field always has a value without requiring the client to provide one or the developer to list it in the shape.

Updates

await prisma.project
  .guard({
    data: { title: true },
    where: { id: { equals: true } },
  })
  .update({
    data: { title: 'New title' },
    where: { id: { equals: 'abc123' } },
  })

In update mode, all data fields are optional. The where shape enforces which filters the client can use.

Forced values

Literal values in the shape are forced by the server and cannot be overridden by the client:

import { force } from 'prisma-guard'

await prisma.project
  .guard({
    data: { title: true, status: 'draft', isActive: force(true) },
  })
  .create({ data: req.body })

status is always 'draft' and isActive is always true regardless of what the client sends.

The same applies to where:

await prisma.project
  .guard({
    where: {
      status: { equals: 'published' },
      isActive: { equals: force(true) },
      title: { contains: true },
    },
  })
  .findMany(req.body)

status = 'published' and isActive = true are always enforced. The client can only control the title filter.

Forced where conditions are conflict-checked during shape construction. If the same field and operator appear with different forced values in different parts of a shape (e.g. at the top level and inside a combinator), the shape is rejected with ShapeError. This prevents ambiguous security configurations where one forced value would silently overwrite another.

Deletes

await prisma.project
  .guard({
    where: { id: { equals: true } },
  })
  .delete({ where: { id: { equals: 'abc123' } } })

data is not valid for delete shapes.

Batch creates

await prisma.project
  .guard({
    data: { title: true, status: true },
  })
  .createMany({
    data: [
      { title: 'Project A', status: 'active' },
      { title: 'Project B', status: 'draft' },
    ],
  })

Each item in the array is validated against the same data shape.

In guarded mode, createMany and createManyAndReturn require data to be an array. Single-object data is not silently wrapped.

createMany and createManyAndReturn also accept skipDuplicates: boolean in the request body. This is passed through to Prisma without shape-level configuration.

Bulk mutations

updateMany, updateManyAndReturn, and deleteMany require a where shape in the guard definition. This prevents accidental unconstrained bulk writes.

await prisma.project
  .guard({
    data: { status: true },
    where: { status: { equals: true } },
  })
  .updateMany({
    data: { status: 'archived' },
    where: { status: { equals: 'draft' } },
  })

A guard shape without where on a bulk mutation method throws ShapeError. Additionally, if the client provides no where conditions at runtime (or the resolved where is empty), the request is rejected. Each where operator object must contain at least one operator with a value — empty operator objects like { status: {} } are rejected.

When combinators (AND, OR, NOT) are exposed in the where shape, the guard also prevents vacuous filters. Combinator arrays must contain at least one element, and each element must specify at least one condition with a defined value. Structures like { AND: [] }, { AND: [{}] }, and { NOT: [] } are rejected. See Logical combinators in where shapes for details.

Mutation body validation

Mutation bodies are strictly validated. The accepted keys depend on whether the shape defines a return projection (select or include):

Without projection in shape:

create: data
createMany, createManyAndReturn: data, skipDuplicates
update, updateMany, updateManyAndReturn: data, where
upsert: where, create, update, select, include
delete, deleteMany: where

With projection in shape (methods that support it):

create: data, select, include
createManyAndReturn: data, select, include, skipDuplicates
update, updateManyAndReturn: data, where, select, include
upsert: where, create, update, select, include
delete: where, select, include

For createMany and createManyAndReturn, skipDuplicates is also accepted as a body key. It must be a boolean if provided.

Unknown keys are rejected with ShapeError. If the body contains select or include but the shape does not define them, the request is rejected.

Guard shape keys are also validated per method:

create methods accept data, and optionally select/include (if the method supports projection)
update methods accept data, where, and optionally select/include
upsert accepts where, create, update, and optionally select/include
delete methods accept where, and optionally select/include

Shape keys not valid for the method throw ShapeError.

Supported shape keys

For reads: where, include, select, orderBy, cursor, take, skip, distinct, _count, _avg, _sum, _min, _max, by, having

For writes: data, where, select, include (select/include only on methods that return records)

For upsert: where, create, update, select, include

Where shapes accept scalar field filters, relation filters (some, every, none, is, isNot), and logical combinators (AND, OR, NOT).

Shape config value validation

Shape config values are strictly validated at construction time. Fields in orderBy, cursor, having, _count (object form), _avg, _sum, _min, _max must have the value true. The skip config must be exactly true. Passing any other value (including false, numbers, or strings) throws ShapeError. This prevents accidental misconfiguration where a developer writes { orderBy: { title: false } } expecting it to disable ordering — instead of silently enabling it, the shape is rejected.

Where DSL is a constrained Prisma subset

The where shape syntax supports a subset of Prisma's where filter API. Notable differences from raw Prisma where clauses:

The not operator accepts a scalar value only, not a nested filter object. Prisma's { not: { gt: 5 } } form is not supported.
AND and OR in client input must be arrays with at least one element. Prisma accepts a single object for AND; prisma-guard requires an array. Empty arrays are rejected.
NOT in client input accepts a single object or an array with at least one element. Empty arrays are rejected.
Each combinator member must specify at least one condition with a defined value. Empty objects inside combinators (e.g. { AND: [{}] }) are rejected when no forced values exist.
Relation filter operators (some, every, none, is, isNot) require at least one nested condition when all conditions are client-controlled. Empty relation filters like { posts: { some: {} } } are rejected.
Relation operator containers require at least one operator. { posts: {} } is rejected.
Empty combinator and relation filter definitions in shapes are rejected at shape construction time. A shape like where: { AND: {} } throws ShapeError.
Forced where conditions are conflict-checked at shape construction time. The same field and operator with different forced values across shape branches (e.g. top-level vs inside a combinator) is rejected with ShapeError.

These restrictions are intentional. They prevent clients from sending structurally valid but semantically vacuous filters that could broaden query scope, particularly in bulk mutation where clauses.

Body normalization

Read methods and mutation methods accept undefined or null as body input across all API surfaces. Missing bodies are consistently normalized to {} (empty object). This applies to single shapes, named shapes, and guard.query().parse(). An explicit body, when provided, must be a plain object.

Supported methods

Reads: findMany, findFirst, findFirstOrThrow, findUnique, findUniqueOrThrow, count, aggregate, groupBy

Writes: create, createMany, createManyAndReturn, update, updateMany, updateManyAndReturn, upsert, delete, deleteMany

Logical combinators in where shapes

Where shapes support AND, OR, and NOT to compose filter conditions. The combinator value is a where config defining allowed fields inside the combinator:

await prisma.project
  .guard({
    where: {
      OR: {
        title: { contains: true },
        description: { contains: true },
      },
    },
    take: { max: 50 },
  })
  .findMany({
    where: {
      OR: [
        { title: { contains: 'demo' } },
        { description: { contains: 'demo' } },
      ],
    },
  })

The shape defines which fields are allowed inside each combinator. The client sends arrays for AND/OR and an object or array for NOT.

Forced values inside combinators are lifted to the top-level query as AND conditions, regardless of the combinator type. This means a forced value inside an OR shape does not become an OR branch — it becomes an additional AND constraint on the entire query. This is consistent with the fail-closed design: forced values always restrict, never broaden.

import { force } from 'prisma-guard'

await prisma.project
  .guard({
    where: {
      title: { contains: true },
      NOT: {
        status: { equals: 'archived' },
      },
    },
  })
  .findMany({
    where: { title: { contains: 'demo' } },
  })

status = 'archived' is always excluded regardless of client input.

Combinators can be nested and mixed with scalar fields freely. The same field can appear both at the top level and inside a combinator.

If the same field and operator appear as forced values in different parts of the shape (e.g. top-level and inside an AND combinator), the forced values are conflict-checked. Identical values are deduplicated. Different values throw ShapeError — this prevents ambiguous security configurations from silently degrading.

Combinator validation rules

Combinator definitions in shapes must define at least one field. An empty combinator like where: { AND: {} } throws ShapeError at shape construction time. This prevents silent no-op branches that look restrictive but contribute nothing.

At runtime, combinator arrays from client input must contain at least one element. { AND: [] }, { OR: [] }, and { NOT: [] } are all rejected.

When no forced values exist inside a combinator, each member object must specify at least one condition with a defined value. { AND: [{}] } is rejected because the empty object carries no filtering constraint. This prevents clients from satisfying structural validation while bypassing semantic filtering, which is particularly important for bulk mutations where a vacuous where clause could affect all rows.

When a combinator branch contains forced values, client members may be empty because the forced values still provide meaningful filtering constraints.

Relation filters in where shapes

Where shapes support relation-level filters using Prisma's relation operators. To-many relations support some, every, and none. To-one relations support is and isNot.

await prisma.user
  .guard({
    where: {
      posts: {
        some: {
          title: { contains: true },
          published: { equals: true },
        },
      },
    },
  })
  .findMany({
    where: {
      posts: {
        some: {
          title: { contains: 'guide' },
          published: { equals: true },
        },
      },
    },
  })

Each relation operator value is a nested where config for the related model. All where features — scalar operators, forced values, logical combinators, and nested relation filters — work recursively inside relation filters.

Forced values in relation filters

import { force } from 'prisma-guard'

await prisma.user
  .guard({
    where: {
      posts: {
        some: {
          title: { contains: true },
          status: { equals: 'published' },
        },
      },
    },
  })
  .findMany({
    where: {
      posts: {
        some: { title: { contains: 'guide' } },
      },
    },
  })

status = 'published' is always enforced inside the some operator.

To-one relations

await prisma.post
  .guard({
    where: {
      author: {
        is: {
          role: { equals: true },
        },
      },
    },
  })
  .findMany({
    where: {
      author: {
        is: { role: { equals: 'ADMIN' } },
      },
    },
  })

Combined with logical combinators

await prisma.user
  .guard({
    where: {
      OR: {
        posts: {
          some: { published: { equals: true } },
        },
        profile: {
          is: { bio: { contains: true } },
        },
      },
    },
  })
  .findMany({
    where: {
      OR: [
        { posts: { some: { published: { equals: true } } } },
        { profile: { is: { bio: { contains: 'engineer' } } } },
      ],
    },
  })

Using an unsupported operator for the relation type throws ShapeError. For example, some on a to-one relation or is on a to-many relation is rejected.

Relation filter validation rules

Relation filter definitions in shapes must define at least one operator. An empty relation filter like where: { posts: {} } throws ShapeError at shape construction time.

Each operator's nested where config must define at least one field. An empty nested where like where: { posts: { some: {} } } throws ShapeError at shape construction time.

When all conditions inside a relation operator are client-controlled (no forced values), the client must provide at least one condition. Empty nested where objects are rejected:

where: {
  posts: {
    some: {
      title: { contains: true },
    },
  },
}

// Rejected — at least one condition required
{ where: { posts: { some: {} } } }

// Accepted
{ where: { posts: { some: { title: { contains: 'demo' } } } } }

When a relation operator contains forced values, the client may omit all client-controlled conditions. The forced values are still injected:

where: {
  posts: {
    some: {
      status: { equals: 'published' },
      title: { contains: true },
    },
  },
}

// Accepted — forced status is still applied
{ where: { posts: { some: {} } } }
// Becomes: { posts: { some: { status: { equals: 'published' } } } }

Mutation return projection

Mutations that return records can use select and include in the guard shape to control which fields and relations are returned. This uses the same shape syntax as reads — the shape whitelists what the client may request, and forced where conditions on nested includes work identically.

Which methods support projection

Method	Returns	select/include
`create`	record	yes
`createMany`	BatchPayload	no
`createManyAndReturn`	record[]	yes
`update`	record	yes
`updateMany`	BatchPayload	no
`updateManyAndReturn`	record[]	yes
`upsert`	record	yes
`delete`	record	yes
`deleteMany`	BatchPayload	no

Create with projection

await prisma.project
  .guard({
    data: { title: true },
    include: {
      members: true,
    },
  })
  .create({
    data: { title: 'New project' },
    include: { members: true },
  })

Update with select

await prisma.project
  .guard({
    data: { title: true },
    where: { id: { equals: true } },
    select: {
      id: true,
      title: true,
      members: {
        select: { id: true, email: true },
      },
    },
  })
  .update({
    data: { title: 'Updated' },
    where: { id: { equals: 'abc123' } },
    select: {
      id: true,
      title: true,
      members: {
        select: { id: true, email: true },
      },
    },
  })

Delete with include

await prisma.project
  .guard({
    where: { id: { equals: true } },
    include: { members: true },
  })
  .delete({
    where: { id: { equals: 'abc123' } },
    include: { members: true },
  })

Forced where on nested includes in mutations

Forced where conditions work the same as in reads. This is useful for ensuring tenant-scoped nested data in mutation responses:

import { force } from 'prisma-guard'

await prisma.project
  .guard({
    data: { title: true },
    include: {
      members: {
        where: { isActive: { equals: force(true) } },
      },
    },
  })
  .create({
    data: { title: 'New project' },
    include: { members: true },
  })

The returned members will always be filtered to isActive = true, regardless of what the client sends.

Projection is optional by default

If the shape does not define select or include, but the caller also omits them from the body, the mutation returns the full record (default Prisma behavior). Projection shapes only validate and constrain client-requested projections — they do not enforce a fixed output boundary unless enforced projection mode is enabled.

select and include are mutually exclusive

Same as reads: a shape (and a body) cannot define both select and include at the same level. Doing so throws ShapeError.

Batch methods do not support projection

createMany, updateMany, and deleteMany return BatchPayload (a count), not records. Passing select or include in the shape or body for these methods throws ShapeError.

Enforced projection mode

By default, projection shapes only constrain client-requested projections. If the client omits select/include from the body, Prisma returns its default full payload.

When enforceProjection is enabled, the shape's projection is always applied — even when the client does not request one. If the shape defines select or include and the client omits them, prisma-guard synthesizes a projection from the shape and passes it to Prisma.

Configuration

generator guard {
  provider           = "prisma-guard"
  output             = "generated/guard"
  enforceProjection  = "true"
}

Behavior

With enforced projection enabled:

await prisma.project
  .guard({
    data: { title: true },
    select: { id: true, title: true },
  })
  .create({ data: { title: 'New' } })

Even though the client omits select from the body, Prisma receives { select: { id: true, title: true } } and returns only those fields.

Without enforced projection (default): Prisma returns all fields.

Synthesized projection

When the client omits select/include, prisma-guard synthesizes a default projection body from the shape:

Scalar fields marked true in the shape produce true in the synthesized body
Nested relation shapes produce their structural equivalent (nested select/include)
_count configurations are preserved
Client-controllable args like where, orderBy, take, skip on nested includes are omitted from the synthesized body — only forced where conditions are applied through the existing forced-tree pipeline

When the client does provide select/include, behavior is identical regardless of this setting: the client's projection is validated against the shape.

This mode applies to all methods that support projection: create, update, upsert, delete, createManyAndReturn, and updateManyAndReturn.

Upsert

Upsert is supported with dedicated create and update shape keys that mirror Prisma's upsert API. The data key is not valid for upsert — use create and update instead.

await prisma.project
  .guard({
    where: { id: { equals: true } },
    create: { title: true, status: true },
    update: { title: true },
  })
  .upsert({
    where: { id: { equals: 'abc123' } },
    create: { title: 'New Project', status: 'active' },
    update: { title: 'Updated Title' },
  })

Shape requirements

Upsert shapes must define all three: where, create, and update. Missing any of them throws ShapeError. Using data instead of create/update throws ShapeError.

The create branch follows the same rules as regular create shapes: all required fields without defaults must be accounted for (as client-allowed, forced, scope FK, or @zod .default(...)/@zod .catch(...)). The update branch follows update rules: all fields are optional.

The where must satisfy a unique constraint with equality operators, same as update and delete.

All data shape value types work

import { force } from 'prisma-guard'

await prisma.project
  .guard({
    where: { id: { equals: true } },
    create: {
      title: (base) => base.min(1).max(200),
      status: 'draft',
      isActive: force(true),
    },
    update: {
      title: (base) => base.min(1).max(200),
    },
  })
  .upsert({
    where: { id: { equals: 'abc123' } },
    create: { title: 'New Project' },
    update: { title: 'Updated' },
  })

Projection support

Upsert returns a record and supports select and include:

await prisma.project
  .guard({
    where: { id: { equals: true } },
    create: { title: true, status: true },
    update: { title: true },
    select: { id: true, title: true, status: true },
  })
  .upsert({
    where: { id: { equals: 'abc123' } },
    create: { title: 'New', status: 'active' },
    update: { title: 'Updated' },
    select: { id: true, title: true },
  })

Scope behavior

On scoped models, upsert is fully supported:

Scope condition is merged into where using unique-preserving merge (same as update and delete)
Scope FK is injected into create data (same as regular creates)
Scope FK is stripped from update data (same as regular updates)
All scope roots must be present in context — missing roots throw PolicyError

Named shapes and context-dependent shapes

Upsert works with named shapes and context-dependent shapes:

await prisma.project
  .guard({
    '/admin/projects/:id': {
      where: { id: { equals: true } },
      create: { title: true, status: true, priority: true },
      update: { title: true, status: true, priority: true },
    },
    '/editor/projects/:id': {
      where: { id: { equals: true } },
      create: { title: true, status: 'draft' },
      update: { title: true },
    },
  }, req.headers['x-caller'])
  .upsert({
    where: { id: { equals: req.params.id } },
    create: req.body.create,
    update: req.body.update,
  })

Body keys

Upsert accepts: where, create, update, select, include. Unknown keys are rejected with ShapeError.

Named shapes and caller routing

Different pages or API consumers often need different shapes for the same model. Named shapes route requests to the right shape based on a caller value.

Caller is provided as the second argument to .guard() or via the context function — never in the request body. This keeps the method body clean and Prisma-compatible.

Caller keys must not collide with reserved shape config keys (where, data, create, update, include, select, orderBy, etc.). Using a reserved key as a caller path throws ShapeError.

Define named shapes

await prisma.project
  .guard({
    '/admin/projects': {
      where: { title: { contains: true }, status: { equals: true } },
      take: { max: 100 },
    },
    '/public/projects': {
      where: { title: { contains: true } },
      take: { max: 20, default: 10 },
    },
  }, req.headers['x-caller'])
  .findMany(req.body)

The frontend sends its current route as a header:

fetch('/api/projects', {
  headers: { 'x-caller': window.location.pathname },
  body: JSON.stringify({
    where: { title: { contains: 'demo' } },
  }),
})

The backend passes caller as the second argument to .guard(). The request body contains only Prisma-compatible fields.

Named mutation shapes

await prisma.project
  .guard({
    '/admin/projects/:id': {
      data: { title: true, status: true, priority: true },
      where: { id: { equals: true } },
    },
    '/editor/projects/:id': {
      data: { title: true },
      where: { id: { equals: true } },
    },
  }, req.headers['x-caller'])
  .update({
    data: req.body.data,
    where: { id: { equals: req.params.id } },
  })

Named shapes with inline refines

await prisma.project
  .guard({
    '/admin/projects': {
      data: {
        title: (base) => base.min(1).max(500),
        status: true,
      },
    },
    '/public/projects': {
      data: {
        title: (base) => base.min(1).max(100),
      },
    },
  }, req.headers['x-caller'])
  .create({ data: req.body })

All data shape value types (true, literal, force(), function) work in named shapes, context-dependent shapes, and single shapes.

Caller resolution order

Caller is resolved in priority order:

Explicit argument — .guard(shapes, '/admin/projects') always wins
Context function — if the context object has a caller string property, it is used as the default
None — if neither source provides a caller and the shape is a named map, CallerError is thrown

This enables three usage patterns:

// 1. Per-request via context (set once, used everywhere)
guard.extension(() => ({
  Tenant: store.getStore()?.tenantId,
  caller: store.getStore()?.caller,
}))
await prisma.project.guard(shapes).findMany(req.body)

// 2. Explicit override per call
await prisma.project.guard(shapes, '/admin/projects').findMany(req.body)

// 3. Single shape (no caller needed)
await prisma.project.guard({ where: { ... } }).findMany(req.body)

The caller key in the context object is not used for scope injection — it is only used for shape routing. Scope roots are identified by matching context keys against @scope-root model names.

Parameterized caller patterns

/org/:orgId/users
/org/:orgId/users/:userId

Matching is case-sensitive. Exact matches are checked first. If no exact match is found, parameterized patterns are evaluated. Parameters are routing-only and are not extracted into context.

Fail-closed behavior

If caller is missing or doesn't match any pattern, the request is rejected with a CallerError. If a caller matches multiple parameterized patterns, it is also rejected with a CallerError.

If a request body contains a caller field when using named shapes, it is rejected with a CallerError that directs the developer to use the second argument to .guard() or the context function instead.

Context-dependent shapes

Shapes can be functions that receive the context provided to guard.extension(). This is the same context used for tenant scoping and caller routing — no separate mechanism.

const prisma = new PrismaClient().$extends(
  guard.extension(() => ({
    Tenant: store.getStore()?.tenantId,
    role: store.getStore()?.role,
    caller: store.getStore()?.caller,
  }))
)

The context function returns an object with arbitrary keys. Keys whose values are string, number, or bigint and that match a scope root model name are used as scope context for tenant isolation. The caller key (if a string) is used as the default caller for named shape routing. Other keys (like role in the example above) are passed through to shape functions but are not used for scoping or routing.

The context function must return a plain object. If it returns null, undefined, an array, a primitive, or any non-plain-object value, a PolicyError is thrown. This is enforced consistently across all code paths that consume context — scope injection, caller resolution, and dynamic shape evaluation.

If a context key matches a known scope root model name but has a non-primitive value (e.g. an object or array instead of a string, number, or bigint), a PolicyError is thrown immediately. This prevents bugs in the context function from silently weakening scope enforcement.

Dynamic shape functions must return a plain guard shape object. If the function throws or returns a non-object value, a ShapeError is raised.

Single context-dependent shape

await prisma.project
  .guard((ctx) => ({
    where: {
      tenantId: { equals: ctx.Tenant },
      title: { contains: true },
    },
    take: ctx.role === 'admin' ? { max: 100 } : { max: 20 },
  }))
  .findMany(req.body)

Context-dependent data shapes with inline refines

import { force } from 'prisma-guard'

await prisma.project
  .guard((ctx) => ({
    data: {
      title: (base) => base.min(1).max(ctx.role === 'admin' ? 500 : 200),
      status: ctx.role === 'admin' ? true : 'draft',
      isActive: force(true),
    },
  }))
  .create({ data: req.body })

Context-dependent shapes can use the context both for structural decisions (which fields to expose, forced vs client-provided) and within inline refine functions (dynamic validation limits).

Named context-dependent shapes

await prisma.project
  .guard({
    '/admin/projects': (ctx) => ({
      where: {
        tenantId: { equals: ctx.Tenant },
        title: { contains: true },
      },
      take: { max: 100 },
    }),
    '/public/projects': {
      where: { title: { contains: true } },
      take: { max: 20 },
    },
  })
  .findMany(req.body)

Static shapes and function shapes can be mixed freely in the same shape map. In this example, the caller is resolved from contextFn().caller since no explicit caller is passed.

Automatic tenant isolation

Tenant predicates are injected into top-level scoped queries only. Nested reads via include or select and nested writes are not automatically scoped by the extension. See Limitations for details.

Input query:

await prisma.project
  .guard({ where: { id: { equals: true } } })
  .findFirst({ where: { id: { equals: projectId } } })

Actual enforced condition:

WHERE id = ?
AND tenantId = ?

This applies to all top-level operations on scoped models, including reads, writes, upserts, and deletes.

What is scoped

All top-level reads (findMany, findFirst, findFirstOrThrow, count, aggregate, groupBy)
All top-level creates (create, createMany, createManyAndReturn) — scope FK is injected into data
All top-level unique mutations (update, delete) — scope condition is merged into where
All top-level bulk mutations (updateMany, updateManyAndReturn, deleteMany) — scope condition is merged into where, scope FK is stripped from data
upsert — scope condition is merged into where, scope FK is injected into create data, scope FK is stripped from update data

What is NOT scoped

Nested reads loaded via include or select — use forced where conditions in the shape to restrict these (to-many relations only; see Limitations)
Nested writes — Prisma extension hooks operate on top-level operations only. Guarded data shapes reject relation fields entirely, so nested writes are only possible through raw (unguarded) Prisma calls.
$queryRaw and $executeRaw — raw SQL bypasses all guard protections

Scope relation writes

When a mutation includes data for a scoped model, the scope extension manages the foreign key field automatically. The onScopeRelationWrite generator config controls what happens if the mutation data also includes the Prisma relation field (e.g. writing tenant: { connect: { id: '...' } } alongside the managed tenantId FK):

Value	Behavior
`"error"`	Reject with `ShapeError` (default)
`"warn"`	Remove the relation field and log a warning
`"strip"`	Remove the relation field silently

This setting is configured in the generator block:

generator guard {
  provider              = "prisma-guard"
  output                = "generated/guard"
  onScopeRelationWrite  = "error"
}

Multi-root scope behavior

A model can be scoped by multiple scope roots simultaneously. If Project has a foreign key to both Tenant and Organization (both marked @scope-root), the scope extension enforces both.

On reads, both scope conditions are combined with AND. If onMissingScopeContext is "warn" or "ignore", only present roots are enforced — missing roots are skipped. If "error" (the default), all roots must be present.

On writes (including upsert), all scope roots must be present in the context. A missing root always throws PolicyError, regardless of onMissingScopeContext.

Scope foreign keys for all present roots are injected into create data and stripped from update/delete data.

If this behavior is not what you want, restructure your schema so the model references only one scope root, or handle scoping explicitly via shape rules.

findUnique behavior

Prisma findUnique only accepts declared unique selectors.

This is valid:

await prisma.project.findUnique({
  where: { id: { equals: projectId } },
})

This is not generally valid unless declared as a composite unique:

where: {
  id: { equals: projectId },
  tenantId: { equals: tenantId },
}

Because of this, prisma-guard supports two modes.

`findUniqueMode = "reject"` recommended

Scoped findUnique and findUniqueOrThrow are rejected.

Use findFirst instead:

await prisma.project
  .guard({ where: { id: { equals: true } } })
  .findFirst({ where: { id: { equals: projectId } } })

This allows tenant scope to be enforced at query time.

`findUniqueMode = "verify"`

The query runs first, then the result is verified against tenant scope.

This is weaker because:

it is post-read verification
it can require an extra query
it has a TOCTOU race window

For tenant isolation, "reject" is the safer production default.

Guard shapes for findUnique and findUniqueOrThrow must define where. A shape without where for these methods throws ShapeError.

Output shaping

guard.model() creates output schemas for validating and shaping returned data.

These schemas use base Prisma field types and do not apply @zod input constraints. This is intentional — @zod directives define input validation rules (e.g. .email(), .min(1)) that are not meaningful for validating data already stored in the database.

const userOutput = guard.model('User', {
  pick: ['id', 'email', 'name'],
  include: {
    profile: { pick: ['bio'] },
  },
  strict: true,
})

guard.model() produces a non-strict schema by default, meaning unknown fields in the data are silently stripped from the output. For output validation where unknown fields should be rejected instead of stripped, pass strict: true as shown above.

Notes:

pick and omit apply to scalar fields and are mutually exclusive — passing both throws ShapeError
relations must be added through include
include depth defaults to 5 and can be overridden with maxDepth
models may appear more than once in the include tree (e.g. User → posts → author) as long as the total depth does not exceed maxDepth

Strict Decimal mode

By default, Decimal fields accept JavaScript number, decimal string, and Decimal-like objects. Accepting number is convenient but carries a precision risk: floating-point precision may already be lost by the time the validator sees the value (e.g. 0.1 + 0.2 arrives as 0.30000000000000004).

Strict Decimal mode removes number from the accepted types, requiring decimal string or Prisma Decimal objects only.

Configuration

generator guard {
  provider       = "prisma-guard"
  output         = "generated/guard"
  strictDecimal  = "true"
}

Behavior

Mode	Accepted types
default	`number`, decimal string, Decimal-like object
strict	decimal string, Decimal-like object

With strict mode enabled:

// Accepted
{ price: "29.99" }
{ price: new Prisma.Decimal("29.99") }

// Rejected
{ price: 29.99 }

This applies globally to all Decimal fields across all models, in both data shapes and where filters.

For money or high-precision values, strict mode is recommended. Pass decimal strings (e.g. "0.30") or Prisma Decimal objects instead of JavaScript numbers.

Security model

prisma-guard enforces three layers.

Layer	Purpose
Input boundary	prevents invalid input
Query boundary	restricts allowed query shapes
Data boundary	injects tenant scope

These layers are complementary. Together they provide a fail-closed data boundary around Prisma usage at the top-level operation only. Nested reads and writes are not intercepted by the scope layer — see Limitations.

Limitations

These limitations are real and should be treated as part of the security model.

Raw SQL bypasses guard protections

$queryRaw and $executeRaw are not intercepted.

Nested writes are not intercepted

Prisma extension hooks operate on top-level operations. Nested writes do not trigger separate scope interception.

Guarded data shapes reject relation fields entirely — using a relation field in a data, create, or update shape throws ShapeError. This means nested writes (e.g. { author: { connect: { id: '...' } } }) are only possible through raw (unguarded) Prisma calls. The guard layer prevents them in guarded mutations, not by intercepting nested writes, but by refusing to include relation fields in the data shape.

Nested reads via include are not scope-filtered

The scope extension operates on the top-level operation only. If a query uses include or select to load a relation that is itself a scoped model, the nested results are not tenant-filtered by the extension. Use forced where conditions in the include/select shape to restrict nested reads. This applies to both read operations and mutation return projections.

Forced where on nested reads is limited to to-many relations

Prisma does not support where on to-one relation includes. Because of this, forced where conditions in nested include/select shapes only work on to-many relations.

For to-one relations (e.g. author on a Post), the available mitigations are: omit the relation from the include/select shape entirely, restrict which scalar fields are returned using nested select, or rely on database-level constraints (e.g. RLS, foreign key guarantees).

This is a Prisma API constraint, not a prisma-guard limitation.

`findUnique` cannot be safely scoped in Prisma extension mode

This is a Prisma API limitation, not a conceptual limitation of scoped unique lookups.

That is why findUniqueMode = "reject" is recommended.

Guard shapes for findUnique and findUniqueOrThrow must define where. A shape without where throws ShapeError.

Composite foreign keys to scope roots

If a model references a scope root through composite foreign keys, that specific root is excluded from the model's scope entries when onAmbiguousScope is "warn" or "ignore", and causes a generation error when onAmbiguousScope is "error" (the default). Other non-ambiguous roots on the same model are still auto-scoped.

Handle these models explicitly via shape rules.

Cursor fields must cover a unique constraint

Prisma requires cursor-based pagination to use uniquely-identifiable fields. Guard enforces this at shape construction time: cursor fields must cover at least one unique constraint from the model. Non-unique cursor shapes are rejected with ShapeError.

`@zod` on list fields applies to the array

@zod directives on list fields (e.g. String[]) apply to the z.array(...) schema, not to individual elements. For example, .min(1) on a String[] field enforces a minimum array length of 1, not a minimum string length per element.

Batch methods do not support return projection

createMany, updateMany, and deleteMany return BatchPayload (a count). Passing select or include in the shape or body for these methods throws ShapeError.

Generated output is TypeScript with ESM imports

The generator writes index.ts and client.ts using .js extension imports. A TypeScript-capable build pipeline with ESM-aware module resolution is required ("moduleResolution": "NodeNext" or "Bundler" in tsconfig.json). The classic "moduleResolution": "node" setting is not compatible.

`having` is limited to scalar fields

Guard having shapes support scalar field filters with their type-appropriate operators. Aggregate-level having expressions (e.g. _count, _avg inside having) are not supported.

Json fields accept any JSON-serializable value

Json fields are recursively validated as JSON-serializable values (string, number, boolean, null, plain objects, arrays). Values that are not JSON-serializable — including undefined, functions, symbols, class instances (such as Date), NaN, Infinity, and circular references — are rejected. This does not enforce any particular JSON structure. If you need structured JSON validation, use a context-dependent shape or validate before calling guard.

`refine` and inline refine functions replace `@zod` chains

When a refine callback is provided for a field in guard.input(), or when a function is used instead of true in a data/create/update shape, the callback receives the base Zod type without @zod chains. The @zod directive for that field is bypassed entirely. See Schema annotations.

If the refine function returns a schema that handles undefined input (produces a non-undefined value for undefined), prisma-guard detects this and preserves the behavior by not wrapping with .optional() in create mode.

`pick` and `omit` are mutually exclusive

Both guard.input() and guard.model() reject configurations that specify both pick and omit. This is enforced at both the type level and at runtime.

`@zod` field modifiers interact with prisma-guard nullability

Using @zod .optional(), .nullable(), or .nullish() applies the Zod method on top of prisma-guard's own nullability/optionality handling. This can cause double-wrapping. These modifiers are available but should only be used when intentionally overriding default behavior. Exception: when a chain contains .default() or .catch(), prisma-guard skips adding .optional() in create mode to preserve the default/catch behavior.

Inline refine functions are not cached

Data schemas containing inline refine functions are rebuilt on every request, since the function reference could be context-dependent (e.g. when used inside a dynamic shape that closes over context values). Static data shapes using only true and literal values are cached normally.

`take` does not support negative values

Prisma supports negative take for reverse cursor pagination. prisma-guard restricts take to positive integers (minimum 1). If you need reverse pagination, construct the query server-side using a context-dependent shape.

`skip` in shape config is a permission flag

skip: true in a shape config means the client is allowed to provide a skip value. The actual skip value must be a non-negative integer. The value must be exactly true — other truthy values are rejected with ShapeError. This is consistent with other shape flags but differs from take, which uses { max, default? } syntax.

`guard.input()` defaults to allowing null for nullable fields

guard.input() defaults allowNull to true, matching the behavior of .guard({ data: ... }) and Prisma's own nullable field handling. Pass allowNull: false to reject null values for nullable fields.

`Decimal` fields accept JavaScript numbers by default

The Decimal base type accepts JavaScript number, decimal string, and Decimal-like objects by default. Accepting number is convenient but carries a precision risk: by the time the validator sees the value, floating-point precision may already be lost. For example, 0.1 + 0.2 arrives as 0.30000000000000004. For money or high-precision values, enable strict Decimal mode or pass decimal strings (e.g. "0.30") or Prisma Decimal objects instead of JavaScript numbers.

`skipDuplicates` is supported for batch create methods

createMany and createManyAndReturn accept skipDuplicates: boolean in the request body. This is passed through to Prisma without shape-level configuration. It is not available on create.

Guarded data shapes do not permit relation fields

Relation fields in data, create, and update shapes are rejected with ShapeError. Nested writes (e.g. { author: { connect: { id: '...' } } }) are only possible through raw (unguarded) Prisma calls. The guard layer does not intercept or validate nested writes — it prevents them entirely in guarded mutations.

Conflicting forced where values are rejected

If the same field and operator appear as forced values in different parts of a where shape (e.g. at the top level and inside an AND combinator) with different values, the shape is rejected with ShapeError at construction time. Identical duplicate forced values are deduplicated silently. This prevents ambiguous security configurations from silently degrading.

Mutation projection shapes do not enforce a fixed output boundary by default

If a mutation shape defines select or include but the client omits them from the body, Prisma returns its default full payload. Projection shapes only validate and constrain client-requested projections. Enable enforced projection mode to always apply the shape's projection.

`create` and `update` are reserved shape keys

The bare words create and update cannot be used as caller keys in named shape routing, as they are reserved for upsert shape configuration. Full paths like '/admin/create' or '/api/users/update' are unaffected — only the bare words collide.

Empty `select` and `include` shapes are rejected

An empty select: {} or include: {} in a guard shape throws ShapeError at shape construction time. This is consistent with the fail-closed design applied to empty combinators, empty relation filters, and empty operator objects.

Shape config values must be exactly `true`

Fields in orderBy, cursor, having, _count (object form), _avg, _sum, _min, _max config objects must have the value true. The skip config must be exactly true. Passing false, numbers, strings, or any other value throws ShapeError. This prevents misconfiguration where false is silently treated as enabled.

`@zod .catch()` fields are tracked alongside `.default()` fields

Both @zod .default(...) and @zod .catch(...) are tracked in the generated ZOD_DEFAULTS map. Fields with either directive are exempted from create completeness checks and auto-injected as forced values when omitted from data shapes. The .catch() behavior (fallback on parse error) is preserved in create mode by not wrapping the schema with .optional().

Advanced: SQL-backed runtimes

The findUnique limitation exists in Prisma Client extension mode because Prisma requires a unique selector input type.

At the SQL level, scoped unique lookups are straightforward:

SELECT *
FROM "Project"
WHERE "id" = $1
  AND "tenantId" = $2
LIMIT 1

If your runtime controls SQL generation directly, it can enforce unique lookup plus tenant predicate in a single query.

Libraries like prisma-sql make this possible for advanced architectures.

Error handling

.guard(shape).method(body) may throw:

ZodError — Zod validation failures on data or query args (unless wrapZodErrors is enabled)
ShapeError — invalid shape config, unknown shape config keys, wrong method for shape, body format issues, unexpected body keys, incomplete create data shapes, invalid inline refine functions, dynamic shape functions returning invalid values, conflicting forced where values, empty combinator or relation filter definitions, empty projection shapes, vacuous combinator input, non-true config values in shape builders, or using data instead of create/update for upsert
CallerError — missing, unknown, or ambiguous caller in named shapes, or caller found in request body
PolicyError — denied scope, missing tenant context, invalid context function return value, invalid scope root value type, or rejected operations on scoped models (e.g. findUnique in reject mode)

All guard errors include status and code properties for HTTP response mapping:

Error	status	code
`ShapeError`	400	`SHAPE_INVALID`
`CallerError`	400	`CALLER_UNKNOWN`
`PolicyError`	403	`POLICY_DENIED`

ZodError wrapping

By default, Zod validation failures throw a raw ZodError. This means error handling code must check for both ZodError and guard error types.

To unify error handling, pass wrapZodErrors: true in the guard config:

const guard = createGuard({
  ...generatedConfig,
  wrapZodErrors: true,
})

When enabled, all ZodError thrown during validation is caught and rethrown as ShapeError with status: 400 and code: 'SHAPE_INVALID'. The original ZodError is preserved as the cause property. The error message includes a formatted summary of all Zod issues.

This applies to guard.input().parse(), guard.query().parse(), and all .guard(shape).* methods. guard.model() returns a raw z.ZodObject and is not affected.

How it works internally

prisma-guard has two main parts.

1. Generator

Runs during prisma generate. Requires zod and @prisma/client to be installed.

It reads the Prisma DMMF and emits:

TYPE_MAP — field metadata per model
ENUM_MAP — enum values
SCOPE_MAP — foreign key → scope root mappings
ZOD_CHAINS — @zod directive chains (validated for syntax, method allowlist, argument arity, and basic type compatibility (method existence and type advancement for wrapper-changing methods like .nullable(), .optional(), .default()). Argument type mismatches for non-type-changing methods (e.g. .min('x') on a number field) are not caught at generation time and will fail at runtime when the schema is built.)
ZOD_DEFAULTS — per-model list of fields that have @zod .default(...) or @zod .catch(...), used by the create completeness check and by runtime default injection for omitted fields
GUARD_CONFIG — generator config values (including strictDecimal and enforceProjection)
UNIQUE_MAP — unique constraint metadata per model
client.ts — pre-wired guard instance with typed model extensions

2. Runtime

At runtime, guard.extension() creates a Prisma extension that provides:

.guard(shape, caller?) on every model delegate — validates input, enforces query shapes, returns typed Prisma methods
$allOperations query hook — injects tenant scope into every top-level database operation

The .guard() call validates against the shape, merges forced values, and delegates to the underlying Prisma method. The scope layer runs transparently underneath.

For create operations, fields tracked in ZOD_DEFAULTS that are omitted from the data shape are auto-injected as forced values. The runtime evaluates the field's Zod schema with undefined input and uses the resulting value. This ensures @zod .default(...) and @zod .catch(...) produce correct data even when the field is not listed in the shape.

For fields that ARE listed in the data shape with true and have @zod .default(...) or @zod .catch(...), the runtime skips wrapping the schema with .optional() in create mode. This preserves the Zod default/catch behavior: omitting the field from client input triggers the default rather than passing undefined through.

Caller routing is resolved before method execution: the explicit caller argument takes priority, then contextFn().caller, then absent (which is fine for single shapes but throws CallerError for named shape maps).

The context function is validated on every code path that consumes it — scope injection, caller resolution, and dynamic shape evaluation all enforce the plain-object contract and throw PolicyError for invalid returns. Additionally, if a context key matches a known scope root but has a non-primitive value, PolicyError is thrown immediately rather than silently dropping the scope.

The `force()` helper

The force() function is exported from prisma-guard and creates a wrapper object with an internal symbol marker. At runtime, shape processing checks for this marker to distinguish forced values from the true sentinel. The wrapper is unwrapped before Zod validation, so the forced value is validated against the field's schema like any other literal. The symbol is not enumerable and does not interfere with serialization or inspection of shape objects.

Why this approach

Common alternatives have tradeoffs.

Approach	Tradeoff
ad hoc route validation	repetitive and inconsistent
middleware-only filtering	too easy to miss edge cases
runtime reflection-heavy systems	slower and harder to reason about
full ORM replacement	larger migration cost

prisma-guard focuses narrowly on data boundaries, not ORM replacement.

Performance characteristics

The runtime does lightweight argument rewriting and Zod validation.

In most real applications, overhead should be negligible relative to database round-trip time.

guard.query() caching: Static shapes passed to guard.query() are cached for the lifetime of the returned QuerySchema object. In a named shape map, each static entry is cached independently — a map with 9 static entries and 1 context-dependent entry will cache the 9 static entries. Context-dependent shapes (functions) resolve the function on each call because they depend on runtime context and are never cached.

.guard() caching: The .guard(shape).method(body) chain creates per-invocation caches. In typical usage, the cache is created, used for one method call, and discarded. If you store the result of .guard(shape) and call multiple methods on it, the cache is shared across those calls. For hot paths where the same static shape is used repeatedly, prefer guard.query() for persistent caching.

Data schemas containing inline refine functions are also not cached, since the function could close over runtime context values. Data shapes using only true and literal values are cached normally within their invocation scope. Projection schemas (select/include on mutations) are cached independently for static shapes within their invocation scope.

For upsert, create and update data schemas are cached independently under namespaced keys (upsert:create, upsert:update) to avoid cache collisions with regular create/update operations.

Security philosophy

prisma-guard is designed to fail closed.

Condition	Behavior
ambiguous scope mapping	error by default
missing scope context	error by default
invalid scope root value type	error always
`onMissingScopeContext = "ignore"`	scope bypassed for missing roots; present roots still enforced
unsafe scoped `findUnique`	reject recommended
invalid `@zod` directive	error by default
missing `caller` in named shapes	error always
`caller` in request body	error always
`data` in read shape	error always
`data` in upsert shape	error always (use `create`/`update`)
missing `create` or `update` in upsert	error always
missing `data` in write shape	error always
bulk mutation without `where` shape	error always
bulk mutation with empty `where`	error always
vacuous combinator input	error always
empty combinator/relation shape branch	error always
empty `select`/`include` shape	error always
unexpected keys in mutation body	error always
unknown keys in shape config	error always
non-`true` values in shape config	error always
cursor not covering unique constraint	error always
caller key collides with shape config	error always
empty operator objects in where	error always
empty relation filter (no forced)	error always
empty relation operator container	error always
empty relation filter shape definition	error always
`pick` and `omit` both specified	error always
scope relation in mutation data	controlled by `onScopeRelationWrite` (default: error)
incomplete create data shape	error always
invalid inline refine function	error always (ShapeError)
projection on batch method	error always
body projection without shape	error always
dynamic shape returns non-object	error always (ShapeError)
refine callback returns non-Zod schema	error always (ShapeError)
`findUnique` shape without `where`	error always (ShapeError)
invalid relation operator for type	error always (ShapeError)
context function returns non-object	error always (PolicyError)
conflicting forced where values	error always (ShapeError)
invalid context function return	error always (PolicyError)
`@zod .default()`/`.catch()` field omitted from shape	auto-injected as forced value

Recommended production configuration

generator guard {
  provider                = "prisma-guard"
  output                  = "generated/guard"
  onInvalidZod            = "error"
  onAmbiguousScope        = "error"
  onMissingScopeContext   = "error"
  findUniqueMode          = "reject"
  onScopeRelationWrite    = "error"
  strictDecimal           = "true"
  enforceProjection       = "true"
}

When to use prisma-guard

Best fit:

multi-tenant SaaS backends
Prisma-based microservices
RPC / internal API backends
systems that want schema-driven validation and scoping

Less suitable:

raw SQL-heavy systems
architectures that bypass Prisma Client
cases where another layer already owns validation and authorization comprehensively

Version compatibility

Supported Prisma versions:

Prisma 6
Prisma 7

Supported Node versions:

Node 20
Node 22

Design principles

Fail closed on ambiguous security conditions
Prefer query-time enforcement over verification
Generate minimal runtime metadata
Avoid automatic relation traversal
Keep scope rules explicit and schema-driven
One chain — shape defines the boundary, method executes
Method bodies stay Prisma-compatible — routing and context live in .guard()
No overloaded sentinel values — true always means client-controlled, force() for forced booleans
Upsert uses create/update keys, not data — matches Prisma's own API shape
Shape config values are validated strictly — true means enabled, anything else is rejected

Comparison

Feature	prisma-guard	raw Prisma
Input validation	yes	no
Query shape enforcement	yes	no
Automatic tenant scoping (top-level)	yes	no
Safe scoped `findUnique` in extension mode	reject	not handled
Schema-driven rules	yes	no
Caller-based shape routing	yes	no
Typed method chaining	yes	n/a
Bulk mutation safety	required where	not handled
Vacuous combinator rejection	yes	not handled
Mutation body validation	strict keys	no
Shape config validation	strict values	n/a
Create completeness validation	yes	no
Mutation return projection	yes	manual
Enforced projection mode	opt-in	no
Upsert support	yes	manual
Inline field refine in data shapes	yes	n/a
ZodError wrapping	opt-in	n/a
Logical combinators in where	yes	manual
Relation filters in where	yes	manual
Empty relation filter rejection	yes	n/a
Empty projection shape rejection	yes	n/a
Forced where conflict detection	yes	n/a
Forced boolean values via `force()`	yes	n/a
Strict Decimal mode	opt-in	n/a
`@zod .default()`/`.catch()` auto-injection	yes	n/a

Roadmap

Possible future improvements:

optional nested-write enforcement helpers
richer relation-level policies
adapter integrations for SQL-backed runtimes
model-specific generated types for stronger compile-time shape validation
structured JSON field validation via schema annotations

License

MIT

Name		Name	Last commit message	Last commit date
Latest commit History 23 Commits
.github/workflows		.github/workflows
.vscode		.vscode
dist		dist
src		src
test		test
.gitignore		.gitignore
.releaserc.json		.releaserc.json
CHANGELOG.md		CHANGELOG.md
README.md		README.md
og.png		og.png
package-lock.json		package-lock.json
package.json		package.json
prisma.config.ts		prisma.config.ts
tsconfig.json		tsconfig.json
tsup.config.ts		tsup.config.ts
vitest.config.e2e.ts		vitest.config.e2e.ts

Folders and files

Latest commit

History

Repository files navigation

prisma-guard

Table of contents

Why this exists

Missing input validation

Dangerous query shapes

Tenant isolation bugs

What prisma-guard does

Architecture

Install

Quick start

1. Add the generator

2. Generate

3. Set up the Prisma client

4. Use it

Before / After prisma-guard

Without prisma-guard

With prisma-guard

Schema annotations

Mark tenant root models

Add field-level validation with @zod

Supported @zod methods

Supported argument types in @zod directives

refine replaces @zod chains

The guard API

Data shape syntax

The force() helper

Query shape syntax

Reads

Creates

Updates

Forced values

Deletes

Batch creates

Bulk mutations

Mutation body validation

Supported shape keys

Shape config value validation

Where DSL is a constrained Prisma subset

Body normalization

Supported methods

Logical combinators in where shapes

Combinator validation rules

Relation filters in where shapes

Forced values in relation filters

To-one relations

Combined with logical combinators

Relation filter validation rules

Mutation return projection

Which methods support projection

Create with projection

Update with select

Delete with include

Forced where on nested includes in mutations

Projection is optional by default

select and include are mutually exclusive

Batch methods do not support projection

Enforced projection mode

Configuration

Behavior

Synthesized projection

Upsert

Shape requirements

All data shape value types work

Projection support

Scope behavior

Named shapes and context-dependent shapes

Body keys

Named shapes and caller routing

Define named shapes

Named mutation shapes

Named shapes with inline refines

Caller resolution order

Parameterized caller patterns

Fail-closed behavior

Context-dependent shapes

Single context-dependent shape

Add field-level validation with `@zod`

Supported `@zod` methods

Supported argument types in `@zod` directives

`refine` replaces `@zod` chains

The `force()` helper

`findUniqueMode = "reject"` recommended

`findUniqueMode = "verify"`

`findUnique` cannot be safely scoped in Prisma extension mode

`@zod` on list fields applies to the array

`having` is limited to scalar fields

`refine` and inline refine functions replace `@zod` chains

`pick` and `omit` are mutually exclusive

`@zod` field modifiers interact with prisma-guard nullability

`take` does not support negative values

`skip` in shape config is a permission flag

`guard.input()` defaults to allowing null for nullable fields

`Decimal` fields accept JavaScript numbers by default

`skipDuplicates` is supported for batch create methods

`create` and `update` are reserved shape keys

Empty `select` and `include` shapes are rejected

Shape config values must be exactly `true`

`@zod .catch()` fields are tracked alongside `.default()` fields

The `force()` helper