@@ -0,0 +1,16 @@
+---
+"@objectstack/cli": patch
+---
+
+`os generate migration` gives the file family — `file` / `image` / `avatar` / `video` / `audio` — the **same column width in both formats**. The typescript format emitted a bare `table.string(name)`, knex's `varchar(255)`, while `--format sql` emitted `VARCHAR(2048)` for the same field, so one command answered one field with two widths depending on the flag (#17883).
+
+2048 is not a new number: ADR-0104 ruled the generator's `VARCHAR(2048)` the end-state for this family, `driver-sql` moved to it (`MEDIA_ID_VARCHAR_CHARS`, #15989), and `os migrate files-to-references --apply` retypes the column to `varchar(2048)`. The typescript format was the one producer left at 255 — so a deployment scaffolded from it declared a width the migration it will later run retypes away from.
+
+```diff
+- table.string('cover_image').nullable();
++ table.string('cover_image', 2048).nullable();
+```
+
+- **No regeneration is required of anyone.** `syncSchema` / `initObjects` are additive and never alter an existing column's type, and a `sys_file` id is far shorter than 255, so nothing stored today is at risk either way. What moves is the **declared** width of tables generated from now on.
+- **The width is now read from the sql format's own entry** instead of being retyped beside it, so the two formats cannot drift apart again; `generate-file-reference-width.pin.test.ts` measures both against `driver-sql`'s constant, which is what stops the two halves from "meeting in the middle" at some third value.
+- ⛔ **Nothing outside the family moved.** The `text` family, the reference types the file family used to share an arm with (`lookup` / `master_detail` / `user` / `tree`), `autonumber`, and every `--format sql` answer are byte-identical.
@@ -93,12 +93,14 @@
  *     `FIELD_TYPE_MAP` and a scalar column in `FIELD_TYPE_SQL_MAP` is a
  *     contradiction whoever is right, and `autonumber` was exactly that.
  *
- * ⚠️ Still NOT asserted, deliberately: the FILE_REFERENCE_TYPES family. Those
- * five ARE in the driver's `JSON_COLUMN_TYPES` today while this generator gives
- * them a varchar — but that is #14657's ADR-0104 D3 answer against a driver
- * that is still pre-D3, i.e. a decision about which side moves, not a wrong
- * value to correct. It is recorded below as a measured divergence so it cannot
- * be mistaken for coverage, and filed rather than fixed here.
+ * ⚰️ The FILE_REFERENCE_TYPES family used to be excluded from all of this,
+ * deliberately: those five were in the driver's `JSON_COLUMN_TYPES` while this
+ * generator gave them a varchar — #14657's ADR-0104 D3 answer against a driver
+ * that was still pre-D3, i.e. a decision about which side moves rather than a
+ * wrong value to correct. The ruling on #15041 decided it and #15989 landed it,
+ * so the exclusion below became coverage. [#17883] The width the two migration
+ * formats emit for the family is a second question and has a pin of its own —
+ * `generate-file-reference-width.pin.test.ts`.
  *
  * The runtime fallbacks (`|| 'unknown'`, `|| 'TEXT'`, `default:`) stay and are
  * NOT dead: they answer a `type` string that is not a `FieldType` at all, which
@@ -375,9 +377,18 @@ function sqlColumn(type: string): string | null {
   return m ? m[1] : null;
 }
 
-/** The `table.x('f_type')` call one field type contributes, or `null` for none. */
+/**
+ * The `table.x('f_type'[, …])` call one field type contributes, or `null` for
+ * none.
+ *
+ * [#17883] The argument list is part of the call and is captured with it. The
+ * reader used to stop at the name, so a SIZED call — `table.string(name, 2048)`,
+ * which is what the file family takes since #17883 — read as no column at all,
+ * and every assertion below would have been answering a question about absence
+ * instead of one about the column.
+ */
 function tsColumn(type: string): string | null {
-  const m = TS_OUT.match(new RegExp(`^ {4}(table\\.\\w+\\('f_${type}'\\)).*$`, 'm'));
+  const m = TS_OUT.match(new RegExp(`^ {4}(table\\.\\w+\\('f_${type}'(?:, [^)]*)?\\)).*$`, 'm'));
   return m ? m[1] : null;
 }
 
@@ -669,7 +680,12 @@ describe('#14828 — the SQL answers are the platform’s, not this file’s inv
 
     for (const type of FILE_REFERENCE_TYPES) {
       expect(sqlColumn(type)).toBe('VARCHAR(2048)');
-      expect(tsColumn(type)).toBe(`table.string('f_${type}')`);
+      // [#17883] Was a bare `table.string('f_type')` — knex's varchar(255),
+      // which left THIS generator's two formats disagreeing about the family
+      // after the driver had already moved to 2048. The width agreement itself
+      // is `generate-file-reference-width.pin.test.ts`; what stays here is the
+      // spelling, so the two pins fail in different sentences.
+      expect(tsColumn(type)).toBe(`table.string('f_${type}', 2048)`);
     }
   });
 });
@@ -0,0 +1,195 @@
+// Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
+
+/**
+ * THE #17883 PIN: one `os generate migration`, one column width for the file
+ * family — whichever `--format` the author picked.
+ *
+ * ## The defect
+ *
+ * `generate.ts` carries two migration formats, and they answered the same
+ * FILE_REFERENCE_TYPES field with two different widths:
+ *
+ * ```
+ *   --format sql   f_file  VARCHAR(2048)          FIELD_TYPE_SQL_MAP
+ *   typescript     f_file  table.string('f_file') knex's varchar(255)
+ * ```
+ *
+ * `generate.ts` states that second equality itself, in as many words on the
+ * `autonumber` entry: "`table.string(name)` = knex's `varchar(255)`
+ * (`DEFAULT_STRING_VARCHAR_CHARS`)". The typescript half reached it by riding
+ * the REFERENCE_VALUE_TYPES arm, whose derivation is the TARGET row's `id`
+ * column — a derivation that was never the file family's, whose value is an
+ * opaque `sys_file` id.
+ *
+ * ## Why the typescript half is the side that moves
+ *
+ * 2048 is not a width invented to settle a disagreement. It is the width
+ * ADR-0104 ruled and the rest of the tree has already shipped:
+ *
+ *   - ADR-0104, recording the maintainer ruling on #15041: "The driver is the
+ *     side that moves; the generator's `VARCHAR(2048)` already states the ruled
+ *     end-state and stands."
+ *   - #15989 moved `driver-sql`: `MEDIA_ID_VARCHAR_CHARS = 2048` and
+ *     `table.string(name, MEDIA_ID_VARCHAR_CHARS)` on the moved arm.
+ *   - `os migrate files-to-references --apply` retypes the column to
+ *     `varchar(2048)` (`MEDIA_ID_MOVE_WIDTH`, transcribed from the same width).
+ *
+ * So a deployment scaffolded from the typescript format was the one place left
+ * standing at 255 — its declared width disagreeing with the width the migration
+ * it will later run retypes to. ⛔ The repair is the typescript half joining
+ * the shipped target, never the two halves meeting in the middle, which is why
+ * the sweep below asserts the AGREED width against the driver's own constant
+ * rather than merely asserting the two formats equal. Two formats that had both
+ * drifted to 255 would satisfy an equality-only pin perfectly.
+ *
+ * ## The control, and where it comes from
+ *
+ * ⛔ Not invented here: `generate.ts` documents one. Its `text` entry carries a
+ * measured row for `f_text` — "sql gen f_text varchar(255) / ts gen f_text
+ * varchar(255)" — a NON-file family whose two halves already agree. Measured
+ * before and after the repair, `f_text` and every other non-family column in
+ * the same emission are byte-identical; only the five family rows move. The
+ * sweep pins that shape: the family agrees at the driver's width, and the
+ * neighbours it used to share an arm with keep the unsized call.
+ *
+ * ## What this pin does NOT assert
+ *
+ * Nothing about live DDL. `generate-string-family-width.pin.test.ts` carries
+ * the measured-on-PostgreSQL evidence for the character families; this file
+ * measures agreement between two emitters and the driver's own stated width,
+ * which is the whole content of the card. And nothing about the JSON-vs-varchar
+ * question that #15989 settled — `generate-field-type-vocabulary.pin.test.ts`
+ * holds that one, including the control that reddens if the driver reverts.
+ */
+
+import fs from 'node:fs';
+import path from 'node:path';
+import { fileURLToPath } from 'node:url';
+
+import { FILE_REFERENCE_TYPES, REFERENCE_VALUE_TYPES } from '@objectstack/spec/data';
+import { describe, expect, it } from 'vitest';
+
+import { generateMigrationSql, generateMigrationTs } from './generate.js';
+
+/** `packages/drivers/driver-sql/src` — declared for `@objectstack/cli#test` in `turbo.json`. */
+const DRIVER_SQL_SRC = path.resolve(path.dirname(fileURLToPath(import.meta.url)), '../../../drivers/driver-sql/src');
+const SQL_DRIVER_SOURCE = fs.readFileSync(path.join(DRIVER_SQL_SRC, 'sql-driver.ts'), 'utf8');
+
+/** The control the generator documents, plus the arm the family used to ride. */
+const CONTROL_TYPES = ['text', ...REFERENCE_VALUE_TYPES, 'autonumber'] as const;
+
+function probeConfig(): Record<string, unknown> {
+  const fields: Record<string, Record<string, unknown>> = {};
+  for (const type of [...FILE_REFERENCE_TYPES, ...CONTROL_TYPES]) fields[`f_${type}`] = { type };
+  return { objects: { probe: { name: 'probe', label: 'Probe', fields } } };
+}
+
+const SQL_OUT = generateMigrationSql(probeConfig());
+const TS_OUT = generateMigrationTs(probeConfig());
+
+/** The column declaration one field type contributes to the sql format. */
+function sqlColumn(type: string): string | null {
+  const m = SQL_OUT.match(new RegExp(`^ {2}"f_${type}" (.+?),?$`, 'm'));
+  return m ? m[1] : null;
+}
+
+/** The `table.x('f_type'[, …])` call one field type contributes to the typescript format. */
+function tsColumn(type: string): string | null {
+  const m = TS_OUT.match(new RegExp(`^ {4}(table\\.\\w+\\('f_${type}'(?:, [^)]*)?\\)).*$`, 'm'));
+  return m ? m[1] : null;
+}
+
+/** The character width one format states, or `null` when it states none. */
+function sqlWidth(type: string): number | null {
+  const m = /^VARCHAR\((\d+)\)$/.exec(sqlColumn(type) ?? '');
+  return m ? Number(m[1]) : null;
+}
+
+/**
+ * The width the typescript format's call carries. `null` for a bare
+ * `table.string(name)` — which is a WIDTH, knex's 255, and is exactly the
+ * answer this card is about; the sweep names it rather than reading it as
+ * absence.
+ */
+function tsWidth(type: string): number | null {
+  const m = /^table\.string\('f_[a-z_]+', (\d+)\)$/.exec(tsColumn(type) ?? '');
+  return m ? Number(m[1]) : null;
+}
+
+/** The driver's own width for a moved media column, read where it is declared. */
+function driverMediaChars(): number {
+  const m = SQL_DRIVER_SOURCE.match(/const MEDIA_ID_VARCHAR_CHARS = (\d+);/);
+  if (!m) {
+    throw new Error(
+      'MEDIA_ID_VARCHAR_CHARS is no longer declared in driver-sql/src/sql-driver.ts. That constant ' +
+      'is the width #15989 moved the driver to, and this pin has nothing to measure the generators ' +
+      'against without it — re-read ADR-0104 and #15989 before re-anchoring it.',
+    );
+  }
+  return Number(m[1]);
+}
+
+describe('#17883 — both migration formats give the file family one width', () => {
+  it('control — the probe really emitted both formats, and the readers discriminate', () => {
+    // Non-vacuity: readers that matched nothing would make every `toBe(null)`
+    // below pass while measuring literally nothing.
+    expect(FILE_REFERENCE_TYPES.size).toBe(5);
+    expect(SQL_OUT).toContain('CREATE TABLE IF NOT EXISTS "probe"');
+    expect(TS_OUT).toContain("db.schema.createTable('probe'");
+    expect(sqlColumn('file')).not.toBeNull();
+    expect(tsColumn('file')).not.toBeNull();
+    // …and they tell two different answers apart, in both formats.
+    expect(sqlColumn('text')).not.toBe(sqlColumn('file'));
+    expect(tsColumn('text')).not.toBe(tsColumn('file'));
+    // A width reader that always answered `null` would make the sweep vacuous.
+    expect(sqlWidth('lookup')).not.toBeNull();
+    expect(sqlWidth('text')).toBeNull();
+  });
+
+  it('the driver states a media width this pin can anchor to', () => {
+    // The anchor is READ, never transcribed: a literal here would be a third
+    // copy of the number the card exists to stop copying.
+    expect(driverMediaChars()).toBeGreaterThan(255);
+    expect(SQL_DRIVER_SOURCE).toContain('table.string(name, MEDIA_ID_VARCHAR_CHARS)');
+  });
+
+  for (const type of FILE_REFERENCE_TYPES) {
+    it(`${type} takes the same width in both formats, at the shipped target`, () => {
+      const width = driverMediaChars();
+      expect(
+        sqlWidth(type),
+        `os generate migration --format sql no longer states the ruled end-state width for ${type}. ` +
+        'ADR-0104 ruled that the GENERATOR states it and the driver moves to it — so a change here ' +
+        'is the two halves meeting in the middle, which #17883 forbids.',
+      ).toBe(width);
+      expect(
+        tsWidth(type),
+        `os generate migration (typescript) gave ${type} a different width from the sql format. A ` +
+        'bare `table.string(name)` reads as null here and is knex\'s varchar(255) — the #17883 ' +
+        'defect exactly: one command, one field, two widths depending on --format.',
+      ).toBe(width);
+      // Stated once more as the emitted text, so a reader of a failure sees the
+      // two lines rather than two numbers.
+      expect(sqlColumn(type)).toBe(`VARCHAR(${width})`);
+      expect(tsColumn(type)).toBe(`table.string('f_${type}', ${width})`);
+    });
+  }
+
+  it('the control does not move: the non-file families keep their own answers', () => {
+    // `f_text` is the control `generate.ts` documents on its own `text` entry —
+    // a non-file family whose two halves already agree. ⛔ Not invented here.
+    expect(sqlColumn('text')).toBe('TEXT');
+    expect(tsColumn('text')).toBe("table.text('f_text')");
+
+    // And the arm the family used to ride: a reference column holds the target
+    // row's id, `table.string('id').primary()` = knex's default width, so it
+    // takes the UNSIZED call. If the repair had widened this class too, it
+    // would have carried the file family's width into columns that never
+    // wanted it.
+    for (const type of [...REFERENCE_VALUE_TYPES, 'autonumber']) {
+      expect(tsWidth(type), `${type} was widened along with the file family`).toBeNull();
+      expect(tsColumn(type)).toBe(`table.string('f_${type}')`);
+      expect(sqlWidth(type)).not.toBe(driverMediaChars());
+    }
+  });
+});