Guard / How Guard works

How a request is decided in two steps

When a request arrives, your server resolves the signed-in user once and hands that principal to Policy. Policy resolves every permission in two steps:

  1. Decide principal guards once. The server runs each guard.principal() body for that user and replaces it with Always (allowed) or Never (denied).
  2. Simplify what remains into a Condition. A Condition is a tagged boolean tree (_tag: 'Always' | 'Never' | 'Row' | 'And' | 'Or'). Inside an array of guards (logical OR), any Always collapses the whole rule to Always, and any Never drops out.

Attach rules to an Entity with Guards.attach and ask Policy.readable what remains for an admin versus a member:

ts
const Note = NoteFields.pipe(Guards.attach(G, { row: { Read: [G.isAuthor, G.isAdmin] } }))

const forAdmin = Effect.runSync(
  Policy.readable(Note, Option.some({ id: 'lin', role: 'admin' })),
) // { _tag: 'Always' }

const forMember = Effect.runSync(
  Policy.readable(Note, Option.some({ id: 'ada', role: 'member' })),
) // { _tag: 'Row', guard: G.isAuthor, ... }

For Lin the admin, G.isAdmin becomes Always, so Read collapses to { _tag: 'Always' } and SQL adds no extra WHERE clause. For Ada the member, G.isAdmin becomes Never and disappears, leaving only G.isAuthor, which compiles to WHERE author_id = ?.

For a signed-out visitor (Option.none()), every row guard that reads a placeholder field of user collapses to Never. A row guard that reads only columns of the row (such as published = true) still runs for visitors.