Termix/scripts/generate-dialect-schema.cjs

/**
 * Generates the Postgres and MySQL schema modules from the SQLite one.
 *
 * ## These files produce DDL. They are not used at runtime.
 *
 * drizzle-kit reads them to emit the migrations in drizzle/postgres and
 * drizzle/mysql. Nothing imports them to run a query.
 *
 * That is not an oversight. The query builder needs two things from a table
 * object — the identifiers to interpolate, and the encoders that turn JS values
 * into driver values — and the sqlite definitions supply both correctly for
 * every engine, which is why all 44 repositories import schema.ts directly:
 *
 *   - text and integer encode as themselves everywhere
 *   - integer({ mode: "boolean" }) writes 1/0, which Postgres and MySQL both
 *     accept for a boolean column, and reads back through `Number(v) === 1`,
 *     which is true for JS `true` as well as for 1
 *   - real is a plain number on all three
 *
 * What genuinely differs between the dialects is DDL — column types, key
 * lengths, autoincrement syntax — and DDL is exactly what these files exist to
 * generate. See scripts/verify-dialects.mjs, which asserts the round-trips
 * above against real servers rather than trusting this comment.
 *
 * The schema is declared once, in sqlite-core, and the other two dialects are
 * derived. Hand-maintaining three copies of 52 tables would mean a renamed
 * table has to land in three places consistently or a foreign key silently
 * points at the wrong one — and the schema is regular enough that the mapping
 * is mechanical.
 *
 * What varies between dialects is small and closed:
 *   - booleans are integers on sqlite, native elsewhere
 *   - autoincrement keys are `integer primary key autoincrement`, `serial`,
 *     and `int auto_increment`
 *   - MySQL cannot index unbounded TEXT, so any column that is a primary key,
 *     is unique, or participates in a foreign key must be varchar
 *   - MySQL rejects a bare DEFAULT CURRENT_TIMESTAMP on a text column, so it is
 *     written as a parenthesised expression default
 *
 * Usage: node scripts/generate-dialect-schema.cjs [--check]
 *   --check verifies the committed files match what would be generated,
 *   for CI to catch a schema edit that forgot to regenerate.
 */

const fs = require("fs");
const path = require("path");

const ROOT = path.join(__dirname, "..");
const SOURCE = path.join(ROOT, "src/backend/database/db/schema.ts");
const TARGETS = {
  postgres: path.join(ROOT, "src/backend/database/db/schema.pg.ts"),
  mysql: path.join(ROOT, "src/backend/database/db/schema.mysql.ts"),
};

const KEY_LENGTH = 255;

/**
 * Columns that must be varchar rather than text on MySQL, per table. A column
 * qualifies if it is a primary key, is unique, is either end of a foreign key
 * or is part of an index in its own table. Keyed per table: a column name that
 * happens to be indexed in one table says nothing about another.
 */
function splitTables(source) {
  const starts = [];
  const header = /export const (\w+) = sqliteTable\(\s*"([a-z0-9_]+)"/g;
  let match;
  while ((match = header.exec(source)) !== null) {
    starts.push({ variable: match[1], table: match[2], index: match.index });
  }
  return starts.map((entry, i) => ({
    ...entry,
    end: i + 1 < starts.length ? starts[i + 1].index : source.length,
  }));
}

function collectKeyColumns(source) {
  const tables = splitTables(source);
  const byVariable = new Map(tables.map((t) => [t.variable, t.table]));
  const keyed = new Map(tables.map((t) => [t.table, new Set()]));
  const add = (table, column) => keyed.get(table)?.add(column);

  for (const { table, index, end } of tables) {
    const block = source.slice(index, end);
    let match;

    // `name: text("col")....primaryKey()` / `.unique()` / `.references(...)`
    const declaration =
      /(\w+):\s*text\("([a-z0-9_]+)"\)((?:\s*\.\w+\([^)]*\))*)/g;
    while ((match = declaration.exec(block)) !== null) {
      if (/\.(primaryKey|unique|references)\(/.test(match[3])) {
        add(table, match[2]);
      }
    }

    // Multi-line form: the modifiers land on following lines.
    const multiline =
      /(\w+):\s*text\("([a-z0-9_]+)"\)\s*\n(\s*\.\w+\([\s\S]*?\),)/g;
    while ((match = multiline.exec(block)) !== null) {
      if (/\.(primaryKey|unique|references)\(/.test(match[3])) {
        add(table, match[2]);
      }
    }

    // A referenced column implies the target side too; both must match.
    const reference = /\.references\(\(\)\s*=>\s*(\w+)\.(\w+)/g;
    while ((match = reference.exec(block)) !== null) {
      const target = byVariable.get(match[1]);
      if (target) add(target, camelToSnake(match[2]));
    }

    // foreignKey({ columns: [table.a], foreignColumns: [other.b] })
    const foreign =
      /columns:\s*\[([^\]]*)\],\s*foreignColumns:\s*\[([^\]]*)\]/g;
    while ((match = foreign.exec(block)) !== null) {
      for (const column of match[1].split(",")) {
        const name = column.trim().replace(/^\w+\./, "");
        if (name) add(table, camelToSnake(name));
      }
      for (const column of match[2].split(",")) {
        const [variable, name] = column.trim().split(".");
        const target = byVariable.get(variable);
        if (target && name) add(target, camelToSnake(name));
      }
    }

    // Table-level indexes: `(table) => [uniqueIndex("x").on(table.a, table.b)]`,
    // and the same for the plain `index("x")` used by the performance indexes.
    // MySQL 8 rejects a TEXT column in a key without a length; MariaDB takes it.
    const tableIndex = /\b(?:unique)?[iI]ndex\("[a-z0-9_]+"\)\.on\(([^)]*)\)/g;
    while ((match = tableIndex.exec(block)) !== null) {
      for (const column of match[1].split(",")) {
        const name = column.trim().replace(/^\w+\./, "");
        if (name) add(table, camelToSnake(name));
      }
    }
  }

  return { tables, keyed };
}

function camelToSnake(value) {
  return value.replace(/[A-Z]/g, (c) => `_${c.toLowerCase()}`);
}

function keyTables(source, tables, keyed, dialect) {
  let out = "";
  let cursor = 0;
  for (const { table, index, end } of tables) {
    out += source.slice(cursor, index);
    const columns = keyed.get(table);
    out += source
      .slice(index, end)
      .replace(
        /\btext\("([a-z0-9_]+)"(?:,\s*\{\s*enum:\s*(\[[^\]]*\]),?\s*\})?\)/g,
        (whole, col, values) => {
          // Closed sets stay compact without shortening identifiers or keys.
          if (values && dialect === "mysql")
            return `mysqlEnum("${col}", ${values})`;
          const options = values ? `, enum: ${values}` : "";
          return columns.has(col)
            ? `varchar("${col}", { length: ${KEY_LENGTH}${options} })`
            : whole;
        },
      );
    cursor = end;
  }
  return out + source.slice(cursor);
}

function transform(source, dialect) {
  const { tables, keyed } = collectKeyColumns(source);
  const isPg = dialect === "postgres";
  // Key-bearing strings must be indexable. First, while the table offsets
  // still match the source.
  let out = keyTables(source, tables, keyed, dialect);

  // Autoincrement primary keys, before the plain integer rule below.
  out = out.replace(
    /integer\("([a-z0-9_]+)"\)\.primaryKey\(\{\s*autoIncrement:\s*true\s*\}\)/g,
    (_, col) =>
      isPg
        ? `serial("${col}").primaryKey()`
        : `int("${col}").autoincrement().primaryKey()`,
  );

  // Integer-backed booleans become native ones. Prettier wraps the longer
  // declarations across lines, so this has to span newlines too.
  out = out.replace(
    /integer\(\s*"([a-z0-9_]+)",\s*\{\s*mode:\s*"boolean",?\s*\},?\s*\)/g,
    (_, col) => `boolean("${col}")`,
  );

  // Remaining integers.
  if (!isPg) {
    out = out.replace(
      /\binteger\("([a-z0-9_]+)"\)/g,
      (_, col) => `int("${col}")`,
    );

    // Timestamps are stored as text (see sql-timestamp.ts). MySQL only accepts
    // DEFAULT CURRENT_TIMESTAMP on a DATETIME or TIMESTAMP column — on a TEXT
    // one it is ER_INVALID_DEFAULT, "Invalid default value". Since 8.0.13 an
    // expression default works on any type, and an expression is written
    // parenthesised. MariaDB accepts the bare form, which is why this only
    // surfaces against real MySQL.
    out = out.replace(/sql`CURRENT_TIMESTAMP`/g, "sql`(CURRENT_TIMESTAMP)`");
  }

  // Floating point.
  out = out.replace(/\breal\("([a-z0-9_]+)"\)/g, (_, col) =>
    isPg ? `doublePrecision("${col}")` : `double("${col}")`,
  );

  // text("x", { length: n }) is sqlite-only sugar; drop the length.
  out = out.replace(
    /\btext\("([a-z0-9_]+)",\s*\{\s*length:\s*\d+\s*\}\)/g,
    (_, col) => `text("${col}")`,
  );

  out = out.replace(/\bsqliteTable\(/g, isPg ? "pgTable(" : "mysqlTable(");

  // Self-referencing FK callbacks are typed against the source dialect's
  // "any column" helper so TS can resolve the circular table reference.
  out = out.replace(
    /\bAnySQLiteColumn\b/g,
    isPg ? "AnyPgColumn" : "AnyMySqlColumn",
  );

  // Only the helpers the output uses, so an unused one never fails lint.
  const helpers = isPg
    ? [
        "pgTable",
        "text",
        "varchar",
        "integer",
        "serial",
        "boolean",
        "doublePrecision",
        "index",
        "uniqueIndex",
        "foreignKey",
      ]
    : [
        "mysqlTable",
        "mysqlEnum",
        "text",
        "varchar",
        "int",
        "boolean",
        "double",
        "index",
        "uniqueIndex",
        "foreignKey",
      ];
  const body = out.replace(
    /import\s*\{[^}]*\}\s*from\s*"drizzle-orm\/sqlite-core";/,
    "",
  );
  const used = helpers.filter((name) =>
    new RegExp("\\b" + name + "\\(").test(body),
  );
  const anyColumn = isPg ? "AnyPgColumn" : "AnyMySqlColumn";
  if (body.includes(anyColumn)) used.push(`type ${anyColumn}`);
  const imports = `import {\n${used.map((name) => `  ${name},`).join("\n")}\n} from "drizzle-orm/${isPg ? "pg-core" : "mysql-core"}";`;

  out = out.replace(
    /import\s*\{[^}]*\}\s*from\s*"drizzle-orm\/sqlite-core";/,
    imports,
  );

  return `${header(dialect)}\n${out}`;
}

function header(dialect) {
  return `// GENERATED FILE — do not edit.
//
// Produced from schema.ts by scripts/generate-dialect-schema.cjs.
// Edit the sqlite schema and re-run \`node scripts/generate-dialect-schema.cjs\`.
// Target dialect: ${dialect}.
//
// DDL source for drizzle-kit. NOT imported to run queries — repositories use
// schema.ts on every dialect. See the generator header for why that is correct.
`;
}

function main() {
  const check = process.argv.includes("--check");
  const source = fs.readFileSync(SOURCE, "utf8");
  let drift = false;

  for (const [dialect, target] of Object.entries(TARGETS)) {
    const generated = transform(source, dialect);

    if (check) {
      const current = fs.existsSync(target)
        ? fs.readFileSync(target, "utf8")
        : "";
      if (current !== generated) {
        console.error(
          `[generate-dialect-schema] ${path.relative(ROOT, target)} is out of date`,
        );
        drift = true;
      }
      continue;
    }

    fs.writeFileSync(target, generated);
    console.log(
      `[generate-dialect-schema] wrote ${path.relative(ROOT, target)}`,
    );
  }

  if (drift) {
    console.error(
      "[generate-dialect-schema] run `node scripts/generate-dialect-schema.cjs` and commit the result",
    );
    process.exit(1);
  }
}

module.exports = { transform, collectKeyColumns };

if (require.main === module) {
  main();
}