diff --git a/server/db/sqlite.js b/server/db/sqlite.js
new file mode 100644
index 0000000..10d5aff
--- /dev/null
+++ b/server/db/sqlite.js
@@ -0,0 +1,50 @@
+// SQLite backend for the async DB adapter (dev + tests; also the current prod engine until pg cutover).
+//
+// Wraps the synchronous node:sqlite instance (schema applied at load in ../db.js) in the async interface
+// the repos call. Results are returned via resolved Promises, so the SAME repo code runs unchanged on this
+// synchronous engine and on asynchronous Postgres — the app never sees the difference.
+const raw = require('../db'); // DatabaseSync instance with the full schema already applied
+
+// node:sqlite re-prepares cheaply, but caching by SQL text avoids re-parsing on hot paths.
+const cache = new Map();
+function stmt(sql) {
+  let s = cache.get(sql);
+  if (!s) { s = raw.prepare(sql); cache.set(sql, s); }
+  return s;
+}
+const num = (v) => (typeof v === 'bigint' ? Number(v) : v);
+const runResult = (r) => ({ changes: num(r.changes), lastInsertRowid: num(r.lastInsertRowid) });
+
+function prepare(sql) {
+  return {
+    get: (...p) => Promise.resolve(stmt(sql).get(...p)),
+    all: (...p) => Promise.resolve(stmt(sql).all(...p)),
+    run: (...p) => Promise.resolve(runResult(stmt(sql).run(...p))),
+  };
+}
+
+function exec(sql) { raw.exec(sql); return Promise.resolve(); }
+
+// Transaction primitive. SQLite is single-connection, so BEGIN/COMMIT/ROLLBACK on `raw` is safe; the
+// callback gets a runner with the same async run/get shape. (The pg backend implements this on ONE pooled
+// client — the reason repos must use tx() rather than bare exec('BEGIN') for multi-statement atomicity.)
+async function tx(fn) {
+  raw.exec('BEGIN');
+  try {
+    const t = {
+      run: (sql, ...p) => Promise.resolve(runResult(stmt(sql).run(...p))),
+      get: (sql, ...p) => Promise.resolve(stmt(sql).get(...p)),
+      all: (sql, ...p) => Promise.resolve(stmt(sql).all(...p)),
+    };
+    const out = await fn(t);
+    raw.exec('COMMIT');
+    return out;
+  } catch (e) {
+    try { raw.exec('ROLLBACK'); } catch (_) {}
+    throw e;
+  }
+}
+
+function init() { return Promise.resolve(); } // schema already applied synchronously in ../db.js
+
+module.exports = { prepare, exec, tx, init, name: 'sqlite', _raw: raw };
diff --git a/server/dbx.js b/server/dbx.js
new file mode 100644
index 0000000..3e67c2f
--- /dev/null
+++ b/server/dbx.js
@@ -0,0 +1,6 @@
+// Async DB adapter facade. The backend is chosen by DB_BACKEND (default 'sqlite'); 'pg' is added at
+// cutover. Every backend implements the same async interface — prepare(sql).{get,all,run}, exec(sql),
+// tx(fn), init() — so repos and app code are engine-agnostic. Swapping engines is one backend file, no
+// repo changes. (This is the same "never hardwire the engine" principle we'll apply to the pub/sub layer.)
+const name = process.env.DB_BACKEND || 'sqlite';
+module.exports = require('./db/' + name);