objectstack-ai
diff --git a/‎.changeset/sql-driver-init-objects-indexes-param.md‎
Lines changed: 34 additions & 0 deletions b/‎.changeset/sql-driver-init-objects-indexes-param.md‎
Lines changed: 34 additions & 0 deletions
diff --git a/‎packages/drivers/driver-sql/src/sql-driver-16570-init-objects-indexes-param.test.ts‎
Lines changed: 149 additions & 0 deletions b/‎packages/drivers/driver-sql/src/sql-driver-16570-init-objects-indexes-param.test.ts‎
Lines changed: 149 additions & 0 deletions
diff --git a/‎packages/drivers/driver-sql/src/sql-driver.ts‎
Lines changed: 22 additions & 7 deletions b/‎packages/drivers/driver-sql/src/sql-driver.ts‎
Lines changed: 22 additions & 7 deletions
@@ -0,0 +1,34 @@
+---
+"@objectstack/driver-sql": minor
+"@objectstack/driver-sqlite-wasm": minor
+---
+
+`SqlDriver.initObjects()` and `SqlDriver.registerObjectMetadata()` now declare the `indexes` key they have always read.
+
+Both entry points took `Array<{ name; fields?; tenancy? }>`, with no `indexes` in the type. The key was read out of those very objects one call deep anyway, through an `as any`, in `registerManagedObjectMetadata` — and the map it fills, `managedObjectIndexes`, is what `syncDeclaredIndexes` renders every declared UNIQUE from. So the driver's whole index-sync path was driven by a key its own signature said did not exist, while the sibling `detectManagedDrift` on the same class had always declared `indexes?: any[]`: the two halves of one class disagreed about the shape of the same input.
+
+That is the shape #4311 already addressed for `tenancy`, one key over, and the comment it left above `initObjects` described `indexes` word for word.
+
+**Why nothing tripped over it.** TypeScript's excess-property check fires on a fresh object literal and not on one bound to a variable first, so the same object was accepted or rejected by nothing but where it was spelled — `await driver.initObjects([{ ...bare, indexes: [] }])` was rejected with TS2353, `const o = { ...bare, indexes: [] }; await driver.initObjects([o])` was accepted, and the index was synced either way. Every caller happened to bind first, so the package typechecked green for a reason unrelated to correctness.
+
+**Why this matters beyond a compile error.** The loud symptom was a rejected correct call. The quiet one is the reachable branch: an author — or an AI — reading the signature concludes `indexes` is not accepted and drops the key, and a declared UNIQUE is then never synced, with no error at authoring time and no error at boot. The schema says those rows cannot collide; they can.
+
+What changed, all inside `SqlDriver`:
+
+- `registerObjectMetadata(objects)`, `initObjects(objects)` and the shared `registerManagedObjectMetadata(obj)` helper each gained `indexes?: any[]`, spelled exactly as `detectManagedDrift` already spells it.
+- Every `(obj as any)` cast reading `indexes` off those parameters is gone — the one at the `managedObjectIndexes.set` site and the two inside `initObjects`' own create/alter path. The cast was the evidence that the declaration and the read disagreed; leaving any of them would have fixed the signature while keeping the "the type does not admit me but I read it anyway" path alive. That path is now closed on this parameter.
+
+**What the accept set does, precisely — it moves in both directions.** For a **fresh object literal**, which is what an author writes and what the excess-property check judges, this is purely a widening: `{ ...bare, indexes: [...] }` was rejected and is now accepted. For a **variable-bound** argument, which bypasses that check and is judged by ordinary assignability, it is a narrowing: `indexes` spelled as a record, as a `readonly` tuple (`as const`), or as `null` compiled under the previous signatures and is now rejected with TS2322. Measured in both directions, all three shapes, on this package's own `tsc`.
+
+That narrowing is deliberate, and the three shapes did **not** all behave the same way before it — the difference is worth stating exactly, because only one of them ever worked:
+
+- A **record** and **`null`** never survived the `Array.isArray(obj.indexes)` guard the driver has always applied. That author got no index and no diagnostic — silently, at run time. Rejecting those two at compile time is precisely the failure this change exists to make impossible.
+- A **`readonly` tuple (`as const`)** is a different case, and the only one with anything to lose. `as const` is type-only: at run time the value is a plain array, `Array.isArray` returns `true`, and the index **was** synced. That caller compiled and worked, and is now rejected at compile time. Nothing about its run-time behaviour changed — the rejection is entirely on the type surface.
+
+No migration is owed even so. No caller in this repository is affected, and the shape could never have reached `detectManagedDrift` on the same class either, which publishes the very same `any[]` spelling for the very same key — so a `readonly` caller was already unable to use half of this driver's declared-index surface. A caller in that position spells the array without `as const`, or widens it at the call site.
+
+The disposition on that corrected ground, recorded here because the ground itself moved: **no `BREAKING` banner and no ADR-0087 disposition**, resting on grounds (i) and (iii) alone — zero affected callers, and the `any[]` spelling already published on `detectManagedDrift` for the same key on the same class. The ground that every newly rejected shape had already been discarded at run time is **not** among them: it is false for the `readonly` tuple, and nothing here leans on it.
+
+`@objectstack/driver-sqlite-wasm` is named because `SqliteWasmDriver extends SqlDriver` and overrides neither method, so both widened signatures land in its own published `.d.ts` and its consumers see the identical change. The two packages are in the same fixed version group, so this is a CHANGELOG effect rather than a version one.
+
+The `IDataDriver` contract itself did not move: `registerObjectMetadata?(schemas: unknown[])` in `@objectstack/spec` already accepted `unknown[]`, and `SqlDriver` narrowed it on its own. What grew is `SqlDriver`'s own published accept set.
@@ -0,0 +1,149 @@
+// Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
+
+/**
+ * [#16570] `initObjects` and `registerObjectMetadata` accept `indexes` — the
+ * key they have always READ — spelled as a **fresh object literal**.
+ *
+ * ## The defect this pins
+ *
+ * Both entry points declared `Array<{ name; fields?; tenancy? }>`, with no
+ * `indexes`. The key was read one call deep anyway, through an `as any`, in
+ * `registerManagedObjectMetadata`:
+ *
+ * ```ts
+ * this.managedObjectIndexes.set(tableName, (obj as any).indexes);
+ * ```
+ *
+ * and `managedObjectIndexes` is what `syncDeclaredIndexes` renders every
+ * declared UNIQUE from — so the driver's whole index-sync path was driven by a
+ * key its own signature said did not exist. `detectManagedDrift`, on the same
+ * class, had always declared `indexes?: any[]`: the two halves of one class
+ * disagreed about the shape of the same input. That is the shape #4311 already
+ * fixed for `tenancy`, and the comment it left above `initObjects` described
+ * `indexes` word for word.
+ *
+ * ## Why the FORM of this pin is the whole point
+ *
+ * TypeScript's excess-property check fires on a **fresh object literal** and
+ * not on one bound to a variable first, so the same object was accepted or
+ * rejected by where it was spelled:
+ *
+ * ```ts
+ * await driver.initObjects([{ ...bare, indexes: [] }]);   // TS2353
+ * const withoutIndex = { ...bare, indexes: [] };
+ * await driver.initObjects([withoutIndex]);               // accepted
+ * ```
+ *
+ * Every existing caller in this package happened to bind first — one of them
+ * (`sql-driver-11794-richtext-text-family.test.ts`) even wrote the workaround
+ * down: *"Hoisted (not an inline literal) … `indexes` rides through
+ * `initObjects` beyond its narrow parameter type"*. So the package typechecked
+ * green for a reason unrelated to correctness, and a **variable-bound pin
+ * cannot go red on this defect** — it measures nothing. Every call below is
+ * therefore an inline literal in argument position, which is what makes
+ * `tsc --noEmit` (this package's `typecheck` script) the instrument that
+ * measures it: revert either signature and these lines stop compiling with
+ *
+ *   TS2353: Object literal may only specify known properties, and 'indexes'
+ *   does not exist in type '{ name: string; fields?: Record<string, any>
+ *   | undefined; tenancy?: any; }'.
+ *
+ * The runtime assertions are the other half: they prove the key is not merely
+ * *admitted* by the type but still *read* — recorded in `managedObjectIndexes`
+ * (§1), rendered into a physical UNIQUE (§2), and cleared when withdrawn (§3).
+ * A signature relaxation that quietly stopped reading the key would pass the
+ * compile leg alone.
+ *
+ * Runs on the always-available in-memory SQLite cell: the defect is in a
+ * parameter type and in the registry it feeds, neither of which is
+ * dialect-specific.
+ */
+
+import { describe, it, expect, afterEach } from 'vitest';
+import { SqlDriver } from './sql-driver.js';
+import { dialectCell } from './live-dialect-matrix.testkit.js';
+
+const SQLITE = dialectCell('sqlite');
+
+/**
+ * The un-indexed base object, deliberately WITHOUT `indexes` — every call site
+ * below spreads it and writes `indexes` inline, so the literal being checked is
+ * fresh in argument position. `tenancy: { enabled: false }` keeps the declared
+ * UNIQUE on the plain (non-tenant-scoped) path, and the bounded `maxLength`
+ * keeps `v` a keyable varchar rather than an unbounded TEXT.
+ */
+const bareObject = (name: string) => ({
+  name,
+  tenancy: { enabled: false },
+  fields: { v: { type: 'text', maxLength: 64 } },
+});
+
+/** What the driver recorded for `table`, read off the protected registry. */
+const recordedIndexes = (driver: SqlDriver, table: string): unknown =>
+  (driver as unknown as { managedObjectIndexes: Map<string, unknown> }).managedObjectIndexes.get(table);
+
+describe('initObjects / registerObjectMetadata accept `indexes` as a fresh object literal (#16570)', () => {
+  let driver: SqlDriver | undefined;
+  afterEach(async () => {
+    await driver?.disconnect().catch(() => {});
+    driver = undefined;
+  });
+
+  it('§1 registerObjectMetadata: the inline literal compiles AND the key is recorded', async () => {
+    const T = 'os16570_register';
+    driver = new SqlDriver(SQLITE.config());
+    const declared = [{ fields: ['v'], unique: true as const, name: `uniq_${T}_v` }];
+
+    // Fresh literal in argument position — not hoisted to a variable first.
+    driver.registerObjectMetadata([{ ...bareObject(T), indexes: declared }]);
+
+    expect(recordedIndexes(driver, T)).toEqual(declared);
+  });
+
+  it('§2 initObjects: the inline literal compiles AND the declared UNIQUE is physically synced', async () => {
+    const T = 'os16570_init';
+    driver = new SqlDriver(SQLITE.config());
+
+    // Fresh literal in argument position.
+    await driver.initObjects([
+      { ...bareObject(T), indexes: [{ fields: ['v'], unique: true as const, name: `uniq_${T}_v` }] },
+    ]);
+
+    const knex = (driver as unknown as { knex: (t: string) => any }).knex;
+    await knex(T).insert({ id: 'a', v: 'same' });
+    // If `indexes` had been dropped at authoring time — the silent failure mode
+    // this card is about — this second row would be accepted.
+    await expect(knex(T).insert({ id: 'b', v: 'same' })).rejects.toThrow();
+  });
+
+  it('§3 initObjects: the exact `{ ...bare, indexes: [] }` spelling from the card compiles, and withdraws the entry', async () => {
+    const T = 'os16570_withdraw';
+    driver = new SqlDriver(SQLITE.config());
+    const bare = bareObject(T);
+
+    await driver.initObjects([{ ...bare, indexes: [{ fields: ['v'], unique: true as const, name: `uniq_${T}_v` }] }]);
+    expect(recordedIndexes(driver, T)).toHaveLength(1);
+
+    // The spelling named in the card, verbatim: an empty array must CLEAR the
+    // entry, not leave the previous one standing.
+    await driver.initObjects([{ ...bare, indexes: [] }]);
+    expect(recordedIndexes(driver, T)).toEqual([]);
+  });
+});
+
+/**
+ * The other half of the accept set. A variable-bound argument bypasses the
+ * excess-property check and is judged by ordinary assignability, so `indexes`
+ * must be an array — the two `@ts-expect-error`s ARE the assertion here: if
+ * either line stops erroring, `tsc` fails it as TS2578. Compile-time only,
+ * deliberately never called.
+ */
+export async function pinsTheNarrowingAxis(driver: SqlDriver): Promise<void> {
+  const asRecord = { ...bareObject('os16570_narrow'), indexes: { uniq_v: { fields: ['v'] } } };
+  // @ts-expect-error TS2322 — a record is not `any[]`, and never synced an index at run time.
+  await driver.initObjects([asRecord]);
+
+  const asNull = { ...bareObject('os16570_narrow'), indexes: null };
+  // @ts-expect-error TS2322 — `null` is not `any[]`, and never synced an index at run time.
+  await driver.initObjects([asNull]);
+}
@@ -9735,7 +9735,7 @@ export class SqlDriver implements IDataDriver {
    *          which `initObjects` goes on to use for its DDL.
    */
   protected registerManagedObjectMetadata(
-    obj: { name: string; fields?: Record<string, any>; tenancy?: any },
+    obj: { name: string; fields?: Record<string, any>; tenancy?: any; indexes?: any[] },
   ): { tableName: string; tenantField: string | null } {
     const tableName = StorageNameMapping.resolveTableName(obj);
     // #2186: remember the authoritative metadata field set for this table so
@@ -9744,8 +9744,8 @@ export class SqlDriver implements IDataDriver {
     // Always overwrite — a metadata change that REMOVES `indexes` must clear
     // the previous entry, or drift detection keeps expecting an index nobody
     // declares any more (and never reports it as orphaned).
-    if (Array.isArray((obj as any).indexes)) {
-      this.managedObjectIndexes.set(tableName, (obj as any).indexes);
+    if (Array.isArray(obj.indexes)) {
+      this.managedObjectIndexes.set(tableName, obj.indexes);
     } else {
       this.managedObjectIndexes.delete(tableName);
     }
@@ -9865,7 +9865,7 @@ export class SqlDriver implements IDataDriver {
    * Idempotent: pure metadata assignment, safe to re-drive on every reload.
    */
   registerObjectMetadata(
-    objects: Array<{ name: string; fields?: Record<string, any>; tenancy?: any }>,
+    objects: Array<{ name: string; fields?: Record<string, any>; tenancy?: any; indexes?: any[] }>,
   ): void {
     for (const obj of objects) this.registerManagedObjectMetadata(obj);
   }
@@ -9876,7 +9876,22 @@ export class SqlDriver implements IDataDriver {
   // undeclared here until #4311 (`registerExternalObject` and
   // `computeAndRecordTenantField` both had it), so a caller spelling the key
   // correctly was rejected by the type while the driver read it regardless.
-  async initObjects(objects: Array<{ name: string; fields?: Record<string, any>; tenancy?: any }>): Promise<void> {
+  //
+  // `indexes` is the same story, one key over, and it went undeclared here for
+  // longer: `registerManagedObjectMetadata` fills `managedObjectIndexes` from
+  // it, and that map is what `syncDeclaredIndexes` renders every declared
+  // UNIQUE from — so the whole index-sync path was driven by a key this
+  // signature said did not exist, reached through an `as any`. The sibling
+  // `detectManagedDrift` on this class had always declared it, so the two
+  // halves disagreed about the shape of the same input. Nothing tripped over
+  // it because TypeScript's excess-property check fires on a FRESH object
+  // literal and not on one bound to a variable first, and every caller here
+  // happened to bind first — a green that held for a reason unrelated to
+  // correctness. `src/sql-driver-16570-init-objects-indexes-param.test.ts`
+  // pins the fresh-literal form so it cannot silently go back.
+  async initObjects(
+    objects: Array<{ name: string; fields?: Record<string, any>; tenancy?: any; indexes?: any[] }>,
+  ): Promise<void> {
     // In-memory registration FIRST, and deliberately ahead of the DDL gate
     // below: being refused permission to alter a schema is not a reason to stay
     // ignorant of the objects we were just told about. On a datasource we are a
@@ -9974,7 +9989,7 @@ export class SqlDriver implements IDataDriver {
         table: tableName,
         fields: obj.fields ?? {},
         tenantField,
-        declaredIndexes: (obj as any).indexes,
+        declaredIndexes: obj.indexes,
       });
 
       if (!exists) {
@@ -10049,7 +10064,7 @@ export class SqlDriver implements IDataDriver {
       // referenced column physically exists — which is also why field-level
       // `unique` can no longer be emitted inline by `createColumn`: a composite
       // needs the tenant column to already be there.
-      const declaredIndexes = (obj as any).indexes;
+      const declaredIndexes = obj.indexes;
       const uniqueFields = Object.values<any>(obj.fields ?? {}).some((f) =>
         isUniqueScopeDeclared(f?.unique),
       );