Policies — one declaration, three enforced layers

Caravela's policy block is the framework's core authorization primitive. A single declaration compiles into three simultaneous enforcement targets:

1. Ecto WHERE clauses — row-level scope on every read.
2. Field projection — invisible fields are redacted from every record before it leaves the context (JSON + GraphQL ride on this automatically).
3. Typed Svelte field_access prop — the generated Svelte components gate columns/fields/inputs with {#if field_access.x} so the UI never renders controls the actor shouldn't see.

No three-way sync. No drift. The client could ignore the prop and the data still wouldn't be there — the server already stripped it.

DSL

defmodule MyApp.Domains.Library do
  use Caravela.Domain, multi_tenant: true
  version "v1"

  import Ecto.Query, only: [where: 3]

  entity :books do
    field :title, :string, required: true
    field :isbn, :string
    field :published, :boolean, default: false
    field :price, :decimal, precision: 10, scale: 2
    field :internal_notes, :text
    field :cost_basis, :decimal
    field :author_email, :string
  end

  policy :books do
    # Row-level: which records can this actor see?
    scope fn query, actor ->
      case actor.role do
        :admin -> query
        _ -> where(query, [b], b.published == true)
      end
    end

    # Field-level: which fields can this actor see on accessible records?
    field :price,          visible: fn actor -> actor.role in [:admin, :editor] end
    field :internal_notes, visible: fn actor -> actor.role == :admin end
    field :cost_basis,     visible: fn actor -> actor.role == :admin end

    # Record-dependent rules — the field drops per row.
    field :author_email,   visible: fn actor, record ->
      actor.role == :admin or actor.id == record.author_id
    end

    # Action-level: who can create/update/delete?
    allow :create, fn actor -> actor.role in [:admin, :editor] end
    allow :update, fn actor, record ->
      actor.role == :admin or actor.id == record.author_id
    end
    allow :delete, fn actor -> actor.role == :admin end
  end
end

Rules

• scope fn query, actor -> query end — exactly one per entity. Applied to every read in the generated context.
• field :name, visible: fn actor -> bool end — arity 1. Constant per request. The Svelte field_access prop carries the resolved boolean.
• field :name, visible: fn actor, record -> bool end — arity 2. Record-dependent. The prop resolves to the sentinel 'per_record' and each row is evaluated server-side; denied rows simply have the field set to null.

• allow :create | :update | :delete, fn actor[, record] -> bool end — gates the context's create_* / update_* / delete_* functions.

The block is plain Elixir

The body of policy :entity do … end is a normal Elixir block, so for, if, helper function calls, and module-attribute splicing all work — the compiler expands them before our macros run:

@admin_fields [:price, :cost_basis, :internal_notes]

policy :books do
  scope fn q, actor ->
    if actor.role == :admin, do: q, else: where(q, [b], b.published)
  end

  for f <- @admin_fields do
    field f, visible: fn actor -> actor.role == :admin end
  end

  if Mix.env() == :dev do
    allow :delete, fn _ -> true end
  end
end

Multiple policy :entity blocks are additive — useful for splitting rules across files or wrapping one in an environment guard. Duplicate rules within the same entity (two scopes, two field :xs, two allow :creates) still raise at compile time, because silent clause-ordering resolution is bewildering.

One limitation: field :x, @some_opts where @some_opts is a module attribute reference raises a pointed error. The macro dispatches on AST shape at compile time and can't see through the attribute. Either inline the kw list or iterate with for over field names.

Actor lookup

The actor is context.current_user (or the string key "current_user"). The authenticatable trait from docs/auth.md sets it for you via on_mount hooks and the fetch_current_user plug.

What the compiler generates

For each policy :entity block, the compiler emits three dispatch functions on the domain module:

• __caravela_policy_scope__/3 — (entity, query, actor)
• __caravela_policy_field_visible__/3 — (entity, field, actor)
• __caravela_policy_field_visible__/4 — (entity, field, actor, record)
• __caravela_policy_allow__/3 — (entity, action, actor)
• __caravela_policy_allow__/4 — (entity, action, actor, record)

Each unpolicied entity falls through to a no-op default (query unchanged, visibility true, allow true). Adding or removing a policy never breaks existing call sites.

The generated context then exposes:

• field_access(entity, context) — the typed map passed to Svelte via LiveSvelte.
• list_* / get_* — read paths apply the scope and the field projection before returning.
• create_* / update_* / delete_* — write paths hit both the existing can_* check AND the new allow gate.

TypeScript + Svelte shape

Generated assets/svelte/[v<N>/]types/<context>.ts:

export interface BookFieldAccess {
  title: true;                // no rule → constant true
  price: boolean;              // arity-1 → boolean
  internal_notes: boolean;
  cost_basis: boolean;
  author_email: 'per_record';  // arity-2 → per-row, field may be null
}

Generated Svelte components accept it as a typed prop:

<script lang="ts">
  import type { Book, BookFieldAccess, LiveHandle } from '../types/library';

  let {
    books = [],
    field_access = { title: true, isbn: true, published: true, price: true, ... },
    live
  }: {
    books?: Book[];
    field_access?: BookFieldAccess;
    live: LiveHandle;
  } = $props();
</script>

<table>
  <thead>
    <tr>
      <th>Title</th>
      {#if field_access.price}<th>Price</th>{/if}
      {#if field_access.internal_notes}<th>Notes</th>{/if}
    </tr>
  </thead>
  <!-- ... -->
</table>

Fields without a rule are rendered unconditionally; fields with a rule are wrapped in {#if field_access.<name>}.

Security model

• The server is the authority. Field projection strips data before it hits the wire. Row scoping limits what the DB ever returns. A malicious client ignoring field_access still cannot see redacted fields — the bytes were never sent.
• Per-record rules evaluate per row. The TypeScript prop says 'per_record' so the client knows the field is sometimes present, sometimes null. Your Svelte code typically renders it with a fallback ({book.author_email ?? '—'}).
• Deny-by-default. Domains default to default_policy: :deny: any entity without a declared policy block has its scope filtered to zero rows, every field marked invisible, and every write denied. Forgetting to add a policy on a new entity can never silently leak data.

default_policy: :deny | :allow

# strict (default): entities without a policy block are fully denied
use Caravela.Domain

# explicit strict
use Caravela.Domain, default_policy: :deny

# opt out to permissive — entities without a policy block behave as
# if every field were visible and every action allowed
use Caravela.Domain, default_policy: :allow

Regardless of the domain default, any entity that declares a policy block is implicitly permissive for rule types it didn't declare. Writing policy :books do field :price, visible: … end grants full scope + allow access for :books — only :price is gated. This keeps policy blocks additive: you can start with a single field rule and grow into scope/allow over time without accidentally locking the entity out.

The decision tree for any (entity, rule) pair:

┌─ entity has a declared rule for this exact (action|field)? ──┐
│            ├── yes → use the declared rule                   │
│            └── no → entity has any `policy` block?           │
│                    ├── yes → permissive (true / pass-through)│
│                    └── no  → domain-level `default_policy`   │
└──────────────────────────────────────────────────────────────┘

Authorization model

policy is the only authorization primitive. Earlier versions shipped a separate can_read / can_create / can_update / can_delete set of hooks — those were removed in 0.7.0 because they ran alongside policy with no coordination (see the 0.7.0 changelog entry). If you're migrating from an earlier release, move every can_* rule into a policy :entity do … end block:

The rule function now receives the actor (context.current_user) directly instead of the raw context map, so references to context.current_user.role become just actor.role.