@oxy.so/contracts 1.0.0

package/dist/esm/accountGraph.js

@@ -0,0 +1,480 @@
+/**
+ * Account graph wire contracts — the account-kind vocabulary, the account
+ * category taxonomy, and the create-account input.
+ *
+ * `accountCategories` classifies a NON-PERSONAL account — what it is about, what
+ * it does — without polluting `User.kind`. See the block above
+ * {@link ACCOUNT_CATEGORY_IDS} for the four rules that govern it.
+ */
+import { z } from 'zod';
+import { usernameSchema, usernameSchemaForAccountKind } from './username.js';
+/**
+ * The union is spelled out above and the array proves coverage BOTH ways
+ * (`satisfies` here, the `Gap` alias below) — the same shape this package's
+ * `ACCOUNT_CATEGORY_IDS` / `TRUST_TIERS` pairs use, and the one
+ * `db/schema/users.ts` mirrors to keep the `users_kind_check` CHECK honest.
+ *
+ * Deriving the union from the array instead would cost nothing here and be paid
+ * by consumers: `kind` travels into `@oxy.so/services` on every device-directory
+ * context (`deviceContextSchema.kind` → `DeviceContext` → the switcher rows),
+ * and an indexed-access type is materially more expensive to check there than a
+ * literal union.
+ */
+export const ACCOUNT_KINDS = [
+    'personal',
+    'organization',
+    'project',
+    'bot',
+    'channel',
+];
+export const accountKindSchema = z.enum(ACCOUNT_KINDS);
+export const CHILD_ACCOUNT_KINDS = [
+    'organization',
+    'project',
+    'bot',
+    'channel',
+];
+export const childAccountKindSchema = z.enum(CHILD_ACCOUNT_KINDS);
+/**
+ * Whether an account of this kind may be the SUBJECT OF A DELEGATION — an
+ * application acting as it on some person's authority. `POST /internal/accounts/
+ * :id/service-switch` and the OAuth delegated subject both gate on this.
+ *
+ * It is NOT the question an account switcher asks. See
+ * {@link isOperatorSwitchTargetKind}, and the note below on why the difference
+ * is `bot`.
+ *
+ * Two kinds are refused, for opposite reasons:
+ *
+ *  - `personal` is a human login, so assuming it would be impersonation.
+ *  - `channel` is a CONTENT identity, not an operating one. A channel exists so
+ *    that posts can be authored BY it; it is never a seat anybody occupies. Its
+ *    operators act on it through their own membership, and an application
+ *    publishes to it with its own credential. Refusing act-as is what makes
+ *    "no login, ever" structural rather than incidental: no session can be
+ *    minted whose subject is a channel, so no bearer exists that could add an
+ *    auth method to one (every auth-method write resolves its target from the
+ *    authenticated subject, never from a parameter).
+ *
+ * Consumers must gate on this predicate rather than testing `kind === 'personal'`,
+ * which silently admits every kind added after it was written.
+ */
+export function isDelegatedActAsEligibleKind(kind) {
+    return kind === 'organization' || kind === 'project' || kind === 'bot';
+}
+/**
+ * Whether a PERSON may switch into an account of this kind — become it, in an
+ * account switcher, for the rest of their session.
+ *
+ * ## Why this is not the same question as {@link isDelegatedActAsEligibleKind}
+ *
+ * The two differ on exactly one kind, `bot`, and that difference is the whole
+ * reason both exist.
+ *
+ * **A bot is not something you become. It is something that operates on your
+ * behalf.** Its whole purpose is to act while nobody is present: an application
+ * holds a credential, names the human whose authority it borrows, and speaks as
+ * the bot. That is delegation, and it is what
+ * {@link isDelegatedActAsEligibleKind} admits it for.
+ *
+ * Handing a person the bot's seat instead inverts that. It puts a human inside
+ * the identity that exists to act without one, and it does so on the human's own
+ * device, next to their personal login — which is precisely what happened: a
+ * `bot` account held a live session on a person's device, offered to them by a
+ * switcher that had asked the delegation question by mistake.
+ *
+ * `channel` is refused here as well, for the reason set out above, and
+ * `personal` because assuming somebody else's login is impersonation.
+ *
+ * ## This is the narrower predicate, deliberately
+ *
+ * Everything a person may become, a service may also act as; the reverse does
+ * not hold. A caller that is unsure which question it is asking wants THIS one:
+ * being wrong here withholds an affordance, while being wrong the other way
+ * hands out a seat.
+ */
+export function isOperatorSwitchTargetKind(kind) {
+    return kind === 'organization' || kind === 'project';
+}
+/**
+ * Narrow an unknown value to an {@link AccountKind}.
+ *
+ * The user-DTO serializers read from structurally-permissive `unknown` sources
+ * (a Drizzle row, a Mongo document, an already-formatted object), so each one
+ * would otherwise hand-roll this check and they would drift on what counts.
+ */
+export function isAccountKind(value) {
+    return typeof value === 'string' && ACCOUNT_KINDS.includes(value);
+}
+// ===========================================================================
+// Account categories
+//
+// Four rules govern this taxonomy. Each one is here because the obvious
+// alternative fails silently rather than loudly.
+//
+// 1. AN ID IS AN OPAQUE, IMMUTABLE SLUG — the LABEL is not stored anywhere.
+//    Every value below is an identifier that no rename may ever change. The
+//    human-readable label lives in each client's own translation catalogue,
+//    keyed by the id. Re-labelling `agency` from "Real estate agency" to
+//    "Agency" is therefore a one-line locale edit that touches no row; moving
+//    the label into the column or the DTO would make the same edit a data
+//    migration, and would pin every reader to the language of whoever chose it.
+//
+// 2. THE PRIMARY CATEGORY IS THE FIRST ELEMENT of an ordered list. Not a
+//    separate flag, and not a second field: a flag admits two primaries and
+//    admits none, and ordering makes both unrepresentable. The cost is that
+//    ORDER IS DATA — anything that re-serializes, sorts or de-duplicates this
+//    list can change which category is primary without erroring, so no layer
+//    between the column and the client may reorder it.
+//
+// 3. THE VOCABULARY IS APPEND-ONLY. `ACCOUNT_CATEGORY_IDS` may gain ids and may
+//    never lose one, because rows already carry the ids it holds. Withdrawing a
+//    category means listing it in {@link RETIRED_ACCOUNT_CATEGORY_IDS}, which
+//    removes it from the picker while leaving it readable and re-writable. The
+//    database enforces the same asymmetry: its CHECK is re-evaluated on EVERY
+//    update to a row, so narrowing the allowed set makes an unrelated write —
+//    saving a bio — fail on any account that had picked the withdrawn value.
+//    Measured on a real Postgres, `NOT VALID` included; it does not help.
+//
+// 4. THE SET IS CLOSED. An id nobody can validate is an id no client can render:
+//    the label comes from a translation key derived from the id, so an invented
+//    value paints as a blank or a raw slug in every app at once. Closing it also
+//    costs nothing that rule 3 does not already charge — adding a category needs
+//    a migration to widen the CHECK whether or not the enum exists, so an open
+//    list would buy back no work, only the validation.
+// ===========================================================================
+/**
+ * Every account category, by stable id.
+ *
+ * Grouped by comment for readability only; the storage, the wire and the picker
+ * all treat this as one flat list. `other` is the escape hatch for an account
+ * that fits nothing here.
+ *
+ * TO ADD ONE: append an id (lowercase ASCII, `snake_case`) here, publish
+ * `@oxy.so/contracts`, then ship a migration that widens
+ * `users_account_categories_check` — never edit an existing migration — and add
+ * an `accounts.accountCategory.<id>` label to each client's locales.
+ *
+ * TO WITHDRAW ONE: leave the id here and add it to
+ * {@link RETIRED_ACCOUNT_CATEGORY_IDS}. See rule 3 above.
+ */
+export const ACCOUNT_CATEGORY_IDS = [
+    // ---- media & public information -----------------------------------------
+    'news',
+    'politics',
+    // ---- business & economy --------------------------------------------------
+    'business',
+    'startup',
+    'finance',
+    'crypto',
+    'marketplace',
+    'retail',
+    // The four ids the single-valued `organizationCategory` field used to hold,
+    // carried forward VERBATIM. Their labels may be rewritten freely; their ids
+    // may not, because live rows hold them.
+    'real_estate',
+    'agency',
+    'landlord',
+    'cooperative',
+    'architecture',
+    // ---- technology ----------------------------------------------------------
+    'technology',
+    'software',
+    'ai',
+    'security',
+    'automation',
+    // ---- knowledge -----------------------------------------------------------
+    'science',
+    'education',
+    'books',
+    // ---- health --------------------------------------------------------------
+    'health',
+    'fitness',
+    // ---- sport & play --------------------------------------------------------
+    'sports',
+    'gaming',
+    // ---- culture & entertainment ---------------------------------------------
+    //
+    // There is deliberately NO generic `entertainment` here, and re-adding one is
+    // a regression rather than a gap. It is the only id this list ever carried
+    // that was dominated by its own specifics — `film`, `music`, `gaming` and
+    // `comedy` all exist — and a generic drawer sitting beside its four
+    // concretions collects the lazy pick, which degrades the data for all four at
+    // once: the accounts that would have said `film` say `entertainment` instead,
+    // and `film` stops meaning what it meant.
+    //
+    // The other overlaps in this vocabulary are NOT the same case and must not be
+    // merged on this reasoning: `sports`/`fitness`, `art`/`photography`,
+    // `business`/`startup`, `finance`/`crypto`, `home_garden`/`diy` and
+    // `technology`/`software`/`security` are genuinely different audiences, and
+    // with a cap of four the granularity is cheap.
+    'music',
+    'film',
+    'podcast',
+    'art',
+    'photography',
+    'comedy',
+    // ---- everyday life -------------------------------------------------------
+    'food',
+    'travel',
+    'fashion',
+    'home_garden',
+    'diy',
+    'automotive',
+    'animals',
+    'family',
+    // ---- society -------------------------------------------------------------
+    'nonprofit',
+    'government',
+    'community',
+    'activism',
+    'environment',
+    'religion',
+    // ---- fallback ------------------------------------------------------------
+    'other',
+];
+/**
+ * Accepts EVERY id, withdrawn ones included — see rule 3.
+ *
+ * A schema that rejected a withdrawn id would 400 the whole request whenever a
+ * client round-trips the categories it was served, so an account that had
+ * picked one could no longer save its bio either. That is the same failure the
+ * nullable `bio` / `avatar` fix addressed, wearing a different hat.
+ */
+export const accountCategoryIdSchema = z.enum(ACCOUNT_CATEGORY_IDS);
+/**
+ * Ids withdrawn from the picker. Empty today.
+ *
+ * A withdrawn id keeps working everywhere it is already stored: it validates,
+ * it survives a round-trip save, it still renders from its label key, and it
+ * stays PRIMARY if it was primary. Nothing rewrites a stored list — a read-time
+ * or migration-time demotion would silently replace a choice its owner made,
+ * which is precisely what stable ids exist to prevent. The owner drops it on
+ * their next edit; until then it is honoured.
+ *
+ * What withdrawal changes is only this: the id leaves
+ * {@link SELECTABLE_ACCOUNT_CATEGORY_IDS}, so no picker offers it, and
+ * {@link newlyAddedRetiredCategories} refuses to let a write ADD it to an
+ * account that did not already have it.
+ */
+export const RETIRED_ACCOUNT_CATEGORY_IDS = [];
+/** Whether a category may still be OFFERED. A stored one is readable either way. */
+export function isSelectableAccountCategoryId(id) {
+    return !RETIRED_ACCOUNT_CATEGORY_IDS.includes(id);
+}
+/** The ids a picker may offer, in declaration order. */
+export const SELECTABLE_ACCOUNT_CATEGORY_IDS = ACCOUNT_CATEGORY_IDS.filter(isSelectableAccountCategoryId);
+/**
+ * Which of `next` are withdrawn ids the account did not already carry — i.e.
+ * the ones a write must be refused for.
+ *
+ * `retired` is a parameter rather than a module read so the rule can be
+ * exercised against a non-empty set while the production one is empty; a test
+ * over `RETIRED_ACCOUNT_CATEGORY_IDS` alone would pass vacuously today and stay
+ * passing if the rule were deleted.
+ */
+export function newlyAddedRetiredCategories(next, previous, retired) {
+    return next.filter((id) => retired.includes(id) && !previous.includes(id));
+}
+/**
+ * How many categories one account may carry.
+ *
+ * Four, not "as many as you like". Three reasons, in the order they bind:
+ *
+ *  - The primary has to MEAN something. At ten categories the first element
+ *    reads as a sort artifact rather than a choice, and rule 2 above is the
+ *    entire mechanism by which a primary exists.
+ *  - The profile RENDERS them as a row of chips; four labels of this length is
+ *    what fits a phone-width profile header before the row wraps or truncates.
+ *  - Four is enough to place a genuinely compound account without a tag cloud:
+ *    a housing cooperative that is also a non-profit serving a local community
+ *    spends `cooperative`, `nonprofit`, `community`, `real_estate` — and is the
+ *    most compound real example in the ecosystem.
+ *
+ * One constant, read by the wire schema, the database CHECK and the picker, so
+ * changing it is one edit plus a migration.
+ */
+export const MAX_ACCOUNT_CATEGORIES = 4;
+/**
+ * An account's categories on the wire. ORDER IS MEANINGFUL — index 0 is the
+ * primary (rule 2).
+ *
+ * A duplicate is REJECTED rather than silently collapsed. De-duplicating would
+ * rewrite the caller's list, and any rewrite of this list can move which id sits
+ * at index 0 — so the one repair available here is the one that would break the
+ * property the list exists to carry. A duplicate only ever comes from a client
+ * bug, and a 400 naming the index is how that bug gets found.
+ */
+export const accountCategoriesSchema = z
+    .array(accountCategoryIdSchema)
+    .max(MAX_ACCOUNT_CATEGORIES)
+    .superRefine((ids, ctx) => {
+    const seen = new Set();
+    ids.forEach((id, index) => {
+        if (seen.has(id)) {
+            ctx.addIssue({
+                code: z.ZodIssueCode.custom,
+                message: `Duplicate account category "${id}"`,
+                path: [index],
+            });
+        }
+        seen.add(id);
+    });
+});
+/**
+ * Kinds that may carry categories: every kind EXCEPT `personal`.
+ *
+ * A person has interests, not a sector — and their interests are not a
+ * classification anybody else gets to read off their profile. Spelled out
+ * positively, like {@link isDelegatedActAsEligibleKind} and for the same reason: a `kind
+ * !== 'personal'` test silently admits every kind invented after it was
+ * written, whereas this list forces whoever adds one to decide.
+ */
+export const ACCOUNT_CATEGORY_KINDS = [
+    'organization',
+    'project',
+    'bot',
+    'channel',
+];
+/**
+ * Whether an account of this kind may carry categories.
+ *
+ * The API refuses the write and the `users_account_categories_kind_check`
+ * constraint makes it unrepresentable; both derive from
+ * {@link ACCOUNT_CATEGORY_KINDS}, so they cannot disagree.
+ */
+export function kindAcceptsAccountCategories(kind) {
+    return ACCOUNT_CATEGORY_KINDS.includes(kind ?? '');
+}
+/**
+ * An account's name on the create/update wire.
+ *
+ * `displayName` is EXPLICIT and stored, not derived. `first`/`last` model a
+ * human name, and composing a display string from them is right for a person —
+ * but a non-personal account has a TITLE, not a given and family name. Without
+ * this field the only way to name a channel "Notas de Nate" was to put the whole
+ * title in `first`, which renders correctly by accident while recording it as
+ * somebody's given name.
+ *
+ * When present it wins over the composed `first`/`last` (see the API's
+ * `composeDisplayName`, which already preferred an explicit value — only the
+ * storage for one was missing).
+ */
+const accountNameSchema = z
+    .object({
+    first: z.string().trim().max(100).optional(),
+    last: z.string().trim().max(100).optional(),
+    displayName: z.string().trim().max(100).optional(),
+})
+    .optional();
+/**
+ * POST /accounts — create a non-personal account under the caller's tree.
+ *
+ * No cross-field refinement guards `accountCategories`, and that is not an
+ * omission: `kind` here is a CHILD kind, and every child kind is in
+ * {@link ACCOUNT_CATEGORY_KINDS}, so `personal` is already unrepresentable on
+ * this route. The refinement the single-valued predecessor needed disappeared
+ * along with the restriction that made it necessary. A child kind that does NOT
+ * accept categories would break that reasoning silently, so
+ * `__tests__/accountGraph.test.ts` asserts the two lists agree.
+ */
+export const createAccountRequestSchema = z.object({
+    parentAccountId: z.string().trim().min(1).optional(),
+    kind: childAccountKindSchema,
+    /**
+     * The SAME policy a person's handle is held to. `users.username` is one unique
+     * index, so a managed account may not reserve a name a person could not ask
+     * for — and this route's predecessor (`.min(1).max(100)` here, `^[\w.-]+$`
+     * with no ceiling in the service) is how a one-character or dotted or
+     * 100-character handle became reachable for bots alone.
+     *
+     * A `bot` is held to that AND to the label its handle must end in. That half
+     * cannot live on this field — it depends on `kind`, a sibling — so it is in the
+     * `superRefine` below, which reports its issue against this path.
+     */
+    username: usernameSchema,
+    name: accountNameSchema,
+    bio: z.string().trim().max(500).optional(),
+    avatar: z.string().optional(),
+    description: z.string().trim().max(1000).optional(),
+    /**
+     * A named color preset KEY (`"blue"`, `"mint"`, …), never a hex value.
+     *
+     * Here at CREATION for the reason `isPrivateAccount` is, in miniature: for a
+     * managed account the color is a visual identity, and an account that is
+     * discoverable without one and acquires it on a second request is a face that
+     * changes by itself. One statement, one row, born looking like what its owner
+     * chose.
+     *
+     * The VALUE is checked in the API rather than here. The vocabulary is
+     * `USER_COLOR_PRESETS`, which is declared next to the `users_color_check` CHECK
+     * that is rendered from it — pinning the list a second time in this package
+     * would be a second source of truth for what the database accepts, and the two
+     * would drift apart silently. What this shape does is keep an over-long or
+     * non-string value from reaching the service at all.
+     */
+    color: z.string().trim().max(32).optional(),
+    /** Ordered, PRIMARY FIRST — see rule 2 above {@link ACCOUNT_CATEGORY_IDS}. */
+    accountCategories: accountCategoriesSchema.optional(),
+    /**
+     * Create the account already opted OUT of discovery.
+     *
+     * ## Why this belongs at CREATION and not only on the privacy route
+     *
+     * Every account is born discoverable: the column defaults to `false` and
+     * nothing on the create path wrote it, so a new account appears in people
+     * search the instant it exists. For a human signing themselves up that is the
+     * right default and it is NOT changed here. For an account a program creates
+     * on someone's behalf — an agent, an unlaunched project, an organization for
+     * something not yet announced — it publishes the thing before its owner ever
+     * decided to.
+     *
+     * The alternative is a second call right after create, which is a window in
+     * which the account IS public, and a window whose closing depends on a second
+     * request succeeding. A field here has neither: one statement, one row, born
+     * in the state the caller asked for.
+     *
+     * ## It reuses the existing flag deliberately
+     *
+     * This is `privacy_is_private_account`, the same one `PUT /users/:id/privacy`
+     * toggles — not a new "published" column. A second visibility flag would be a
+     * second source of truth for one question, and the two would disagree.
+     *
+     * Inherited semantics, stated because reusing a flag means inheriting ALL of
+     * it: the account is kept out of people search, out of the follow-graph lists
+     * (`followers` / `following` / `mutuals`), out of `/similar` and out of the
+     * recommendation candidate pools, and its non-public, non-unlisted media
+     * becomes follower-gated. It does NOT hide the profile from someone who knows
+     * the handle, and it carries NO follow-approval flow — following is immediate
+     * and unilateral whatever this says, so nothing here creates a request queue
+     * nobody attends.
+     *
+     * ## Not conditioned on `kind`, on purpose
+     *
+     * The same reasoning as `accountCategories` above: this object does not
+     * refine on kind, and an unlaunched organization has exactly the problem an
+     * unpublished agent does. The discovery predicate never reads `kind`, so the
+     * remedy must not either.
+     */
+    isPrivateAccount: z.boolean().optional(),
+}).superRefine((request, ctx) => {
+    // The ONE place `kind` and `username` arrive in the same object, so it is the
+    // only place a wire schema CAN apply the per-kind half of the policy: a bot's
+    // handle must end in `bot`. Not a second rule — it asks
+    // `usernameSchemaForAccountKind`, the same declaration the service asks.
+    //
+    // It is here rather than only in the API because this schema is exported for
+    // CLIENTS: an agent-creation flow that validates its request and is then 400ed
+    // by the server is the "propose, then refuse" defect the minimum length
+    // already caused once. The service check stays regardless — it also governs
+    // renames and the service-provisioned channel route, where the kind comes from
+    // the stored row and never from this object.
+    const parsed = usernameSchemaForAccountKind(request.kind).safeParse(request.username);
+    if (!parsed.success) {
+        ctx.addIssue({
+            code: z.ZodIssueCode.custom,
+            path: ['username'],
+            message: parsed.error.issues[0].message,
+        });
+    }
+});