objectstack-ai
diff --git a/‎.changeset/driver-sql-pg-introspection-search-path.md‎
Lines changed: 14 additions & 0 deletions b/‎.changeset/driver-sql-pg-introspection-search-path.md‎
Lines changed: 14 additions & 0 deletions
diff --git a/‎packages/drivers/driver-sql/src/live-dialect-matrix.globalsetup.ts‎
Lines changed: 100 additions & 0 deletions b/‎packages/drivers/driver-sql/src/live-dialect-matrix.globalsetup.ts‎
Lines changed: 100 additions & 0 deletions
@@ -0,0 +1,14 @@
+---
+"@objectstack/driver-sql": patch
+---
+
+**Bug fix:** on Postgres, index and schema introspection now resolve tables the way the session does, instead of assuming the `public` schema (#9350).
+
+`introspectIndexes` pinned `n.nspname = 'public'` and `introspectSchema` pinned `table_schema = 'public'`. For a driver whose connection carries a `searchPath` pointing anywhere else, both returned **empty** — not an error, an empty result. Measured on a live Postgres 16: for a table carrying a primary key *and* a declared unique index, `introspectIndexes` returned `[]` and `introspectSchema` listed no tables at all.
+
+Empty does not read as "I could not see" downstream; it reads as "there are no indexes". `assertConflictTargetHonoured` turns that into a refusal, so an `upsert` against a perfectly well-indexed table would be rejected with *no PRIMARY KEY or UNIQUE index backs them* — and index-drift detection would propose creating indexes that already exist.
+
+- `introspectIndexes` now resolves the table with `to_regclass(?)` and reads `pg_index` by OID. That is the same resolution every other statement in the session performs — first match along `search_path` — and it removes an ambiguity a schema list would introduce, since two schemas on the path can hold the same table name and only one of them is the one a query reaches.
+- `introspectSchema` now lists `table_schema = ANY (current_schemas(false))`.
+
+**No change for a default deployment.** With the default `search_path`, `current_schemas(false)` is exactly `{public}` and `to_regclass` resolves into `public`, so both queries return what they returned before. The behaviour only differs where the old queries returned nothing.
@@ -0,0 +1,100 @@
+// Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
+
+/**
+ * #9350 — create the per-file schemas before any test opens a connection.
+ *
+ * ## Why this runs here and not in a hook
+ *
+ * The isolation names each file's database in the CONNECTION (see
+ * `mysqlUrlForSchema`), which is what keeps knex's `client.database()` and the
+ * session the same value. The cost of that choice is an ordering constraint:
+ * connecting to a MySQL database that does not exist fails at the handshake, so
+ * the databases have to exist before the first pool opens.
+ *
+ * A `beforeAll` cannot do it. `cell.config()` is called from inside `beforeEach`
+ * in most of the eleven consumers, which is too late to register a hook, and the
+ * testkit module is cached PER WORKER rather than per file — so a hook
+ * registered at its module scope would attach to whichever file that worker
+ * collected first and to no other. `globalSetup` runs once, in the main process,
+ * before any worker starts, and can await. That is exactly the shape of the
+ * constraint.
+ *
+ * ## Deliberately total, and deliberately cheap
+ *
+ * It creates a schema for EVERY test file in the package rather than for the
+ * live ones only. Deciding which files are live would mean parsing them, and a
+ * wrong answer is a handshake failure in a required check. A schema costs one
+ * dictionary row on both dialects (`create database` on MySQL is not a template
+ * copy the way Postgres' `createdb` is), and the teardown removes them.
+ *
+ * Without either URL this does nothing at all: a developer running without
+ * servers sees no connection attempt, exactly as before.
+ */
+
+import knex from 'knex';
+import { liveSchemaLedger } from './live-dialect-matrix.testkit.js';
+
+const PG_URL = process.env.OS_TEST_POSTGRES_URL;
+const MYSQL_URL = process.env.OS_TEST_MYSQL_URL;
+
+/**
+ * Statements are built from the ledger's names, never from anything a caller
+ * supplies, and `liveSchemaNameFor` refuses to emit a name outside
+ * `/^[a-z][a-z0-9_]*$/` — so the interpolation below cannot carry a quote.
+ */
+async function withServer<T>(
+  client: 'pg' | 'mysql2',
+  connection: string,
+  run: (db: ReturnType<typeof knex>) => Promise<T>,
+): Promise<T> {
+  const db = knex({ client, connection, pool: { min: 0, max: 1 } });
+  try {
+    return await run(db);
+  } finally {
+    await db.destroy();
+  }
+}
+
+export async function setup(): Promise<void> {
+  const ledger = liveSchemaLedger();
+  if (PG_URL) {
+    await withServer('pg', PG_URL, async (db) => {
+      for (const { schema } of ledger) {
+        await db.raw(`create schema if not exists "${schema}"`);
+      }
+    });
+  }
+  if (MYSQL_URL) {
+    await withServer('mysql2', MYSQL_URL, async (db) => {
+      for (const { schema } of ledger) {
+        await db.raw(`create database if not exists \`${schema}\``);
+      }
+    });
+  }
+}
+
+/**
+ * Drop what the setup created.
+ *
+ * Best-effort by design: a failed drop must not turn a green run red — the
+ * schemas are re-created idempotently next time, and CI's servers are thrown
+ * away with the job. It exists for the developer running against a long-lived
+ * local server, who would otherwise accumulate one schema per test file.
+ */
+export async function teardown(): Promise<void> {
+  const ledger = liveSchemaLedger();
+  if (PG_URL) {
+    await withServer('pg', PG_URL, async (db) => {
+      for (const { schema } of ledger) {
+        await db.raw(`drop schema if exists "${schema}" cascade`).catch(() => {});
+      }
+    }).catch(() => {});
+  }
+  if (MYSQL_URL) {
+    await withServer('mysql2', MYSQL_URL, async (db) => {
+      for (const { schema } of ledger) {
+        await db.raw(`drop database if exists \`${schema}\``).catch(() => {});
+      }
+    }).catch(() => {});
+  }
+}