Drizzle ORM Schema Style Guide

Adding a Model or Repository? Ship a sibling test in the same PR — every new file under packages/database/src/models/** or src/repositories/** needs a matching __tests__/<name>.test.ts. See the testing skill (.agents/skills/testing/references/db-model-test.md) for the getTestDB() integration pattern, user-isolation tests, the BM25 describe.skipIf(!isServerDB) guard, and schema gotchas. CI's coverage patch gate won't reliably catch a brand-new untested file, so this is on you.

Configuration

• Config: drizzle.config.ts
• Schemas: packages/database/src/schemas/
• Migrations: packages/database/migrations/
• Dialect: postgresql with strict: true

Helper Functions

Location: packages/database/src/schemas/_helpers.ts

• timestamptz(name): Timestamp with timezone
• createdAt(), updatedAt(), accessedAt(): Standard timestamp columns
• timestamps: Object with all three for easy spread

Naming Conventions

• Tables: Plural snake_case (users, session_groups)
• Columns: snake_case (user_id, created_at)
• New tables: Check nearby existing tables before naming a new one. Preserve the established noun family and suffix. For example, if the user-scoped table is user_xxx_logs, the workspace-scoped counterpart should be workspace_xxx_logs, not workspace_xxx_records or another new synonym.

// ✅ Good: follows the existing user/workspace table family (names are
// illustrative — swap `widget` for the real noun).
export const userWidgetLogs = pgTable('user_widget_logs', { ... });
export const workspaceWidgetLogs = pgTable('workspace_widget_logs', { ... });

// ❌ Bad: introduces a new suffix for the same concept.
export const workspaceWidgetRecords = pgTable('workspace_widget_records', { ... });

Column Definitions

Primary Keys

Do not use auto-incrementing primary keys (serial, bigserial, generated identity columns). They create sequence-state problems during cross-database migrations, restores, and data copy jobs. Prefer text IDs from application generators (idGenerator, createNanoId) or uuid for internal tables.

Keep $defaultFn(...) when a table normally owns ID generation. Callers can still pass an explicit id; the default only runs when the insert omits it. Do not remove the default just because one flow needs to supply a request-scoped ID.

// ✅ Good: app-generated text ID; explicit inserts can still override it.
id: text('id')
  .primaryKey()
  .$defaultFn(() => idGenerator('agents'))
  .notNull(),

// ❌ Bad: sequence state is fragile across DB migrations and restores.
id: serial('id').primaryKey(),

ID prefixes make entity types distinguishable. For internal tables, use uuid.

Do not use composite primary keys on new tables. Give every table a single-column surrogate PK and carry business uniqueness in a uniqueIndex instead. PK columns cannot be nullable, so when the uniqueness scope later grows by a nullable dimension the composite PK must be torn down and rebuilt — exactly what happened when ai_providers / ai_models were workspace-scoped (migrations 0110–0111 replaced their composite PKs with a surrogate _id plus partial unique indexes). A unique index still works as the arbiter for onConflictDoUpdate upserts.

// ✅ Good: surrogate PK; uniqueness scope can evolve without a PK rebuild.
export const workspaceUserSettings = pgTable(
  'workspace_user_settings',
  {
    id: uuid('id').defaultRandom().notNull().primaryKey(),
    workspaceId: text('workspace_id').references(() => workspaces.id, { onDelete: 'cascade' }).notNull(),
    userId: text('user_id').references(() => users.id, { onDelete: 'cascade' }).notNull(),
    ...timestamps,
  },
  (t) => [uniqueIndex('workspace_user_settings_workspace_id_user_id_unique').on(t.workspaceId, t.userId)],
);

// ❌ Bad: locked to exactly these columns; adding a nullable scope column
// (workspaceId, deviceId, …) later forces a full PK rebuild migration.
(t) => [primaryKey({ columns: [t.workspaceId, t.userId] })],

Existing composite PKs are legacy — leave them alone unless they block a scope change, then migrate them the 0110–0111 way.

Foreign Keys

userId: text('user_id')
  .references(() => users.id, { onDelete: 'cascade' })
  .notNull(),

Timestamps

...timestamps,  // Spread from _helpers.ts

Optional and Undefined Values

Do not introduce artificial sentinel strings for missing values, such as unknown, unless the domain already has that explicit state and existing code uses it consistently. Prefer nullable columns, optional TypeScript fields, or a separate concrete status enum when the value is genuinely absent.

// ✅ Good: absent until a custom-execution attempt actually runs
// (packages/database/src/schemas/agentIntervention.ts).
export const AGENT_INTERVENTION_CUSTOM_EXECUTION_STATES = ['pending', 'executing', 'completed'] as const;
export type AgentInterventionCustomExecutionState =
  (typeof AGENT_INTERVENTION_CUSTOM_EXECUTION_STATES)[number];

customExecutionState: text('custom_execution_state').$type<AgentInterventionCustomExecutionState>(),

// ❌ Bad: invents a new state that callers now need to handle everywhere.
export type AgentInterventionCustomExecutionState = 'pending' | 'executing' | 'completed' | 'unknown';

customExecutionState: text('custom_execution_state')
  .$type<AgentInterventionCustomExecutionState>()
  .notNull()
  .default('unknown');

Database Enums

Default to not using PostgreSQL/Drizzle pgEnum. Database enums are expensive to evolve safely: adding members needs migrations, removing or renaming members is awkward, and deployment order becomes more fragile.

For product/business states, use text() or varchar() with a TypeScript value type via $type<...>(). Keep those TS-only value types in the domain/shared type module, then import them into the schema. For cloud DB schemas, that usually means cloudDB/types.ts.

Do not copy existing DB enums as a pattern. Treat them as legacy or explicitly reviewed exceptions. If a new pgEnum seems necessary, stop and justify why the value set is effectively immutable and why the migration cost is acceptable.

Field Descriptions

For columns whose meaning is not obvious from the name alone, add JSDoc on the schema field. Include a concrete example when it clarifies the stored value or the lifecycle moment that writes it. This is especially important for external IDs, lifecycle statuses, denormalized snapshots, JSONB signals, and fields whose name could mean either a request ID or a persisted row ID.

// ✅ Good: explain the table's business object first, then only document
// non-obvious lifecycle fields
// (packages/database/src/schemas/agentIntervention.ts).
/**
 * Private resolution/outbox records. Unlike `agent_interventions`, this table
 * may retain user-edited arguments so a claimed decision can be delivered
 * reliably after the app disconnects. It is never a notification payload.
 */
export const agentInterventionResolutions = pgTable('agent_intervention_resolutions', {
  /** Owner of the durable operation, distinct from the resolving actor. */
  userId: text('user_id').references(() => users.id, { onDelete: 'cascade' }).notNull(),

  /** Private executor result; never selected into Review or notification DTOs. */
  customExecutionResult: jsonb('custom_execution_result').$type<AgentInterventionCustomExecutionResult>(),

  /** Monotonic lease attempt, starting at zero while pending. */
  customExecutionAttempt: integer('custom_execution_attempt'),
});

// ❌ Bad: comments restate obvious column names without adding domain meaning.
/** User email */
email: text('email'),

JSONB Types

Avoid Record<string, unknown> or similarly loose JSONB types for schema columns. Define a concrete interface that describes the expected JSON shape, even when most properties are optional. This keeps callers, migrations, and review queries aligned on the same data contract.

// packages/types/src/agent/agentIntervention.ts +
// packages/database/src/schemas/agentIntervention.ts
interface AgentInterventionCustomExecutionResult {
  content: string;
  pluginState: Record<string, unknown>;
}

customExecutionResult: jsonb('custom_execution_result').$type<AgentInterventionCustomExecutionResult>(),

// ❌ Bad: hides the contract and makes downstream access untyped.
customExecutionResult: jsonb('custom_execution_result').$type<Record<string, unknown>>(),

A loosely-typed JSONB column is often a symptom of a deeper problem: the column was reserved speculatively ("for future extension") and nothing actually writes it. Don't add metadata / extra JSONB columns for hypothetical future needs — a column earns its place only when a concrete writer ships alongside it. When review finds such a column, the fix is to delete the column, not to invent an interface for data that doesn't exist; add a properly-typed column once the real requirement arrives.

Indexes

// Return array (object style deprecated)
(t) => [uniqueIndex('client_id_user_id_unique').on(t.clientId, t.userId)],

Type Inference

export const insertAgentSchema = createInsertSchema(agents);
export type NewAgent = typeof agents.$inferInsert;
export type AgentItem = typeof agents.$inferSelect;

Example Pattern

export const agents = pgTable(
  'agents',
  {
    id: text('id')
      .primaryKey()
      .$defaultFn(() => idGenerator('agents'))
      .notNull(),
    slug: varchar('slug', { length: 100 })
      .$defaultFn(() => randomSlug(4))
      .unique(),
    userId: text('user_id')
      .references(() => users.id, { onDelete: 'cascade' })
      .notNull(),
    clientId: text('client_id'),
    chatConfig: jsonb('chat_config').$type<LobeAgentChatConfig>(),
    ...timestamps,
  },
  (t) => [uniqueIndex('client_id_user_id_unique').on(t.clientId, t.userId)],
);

Common Patterns

Junction Tables (Many-to-Many)

The surrogate-PK rule above applies to junction tables too — pair uniqueness goes in a uniqueIndex, not a composite PK (many existing junction tables still use composite PKs; that is legacy, not the template):

export const agentsKnowledgeBases = pgTable(
  'agents_knowledge_bases',
  {
    id: uuid('id').defaultRandom().notNull().primaryKey(),
    agentId: text('agent_id')
      .references(() => agents.id, { onDelete: 'cascade' })
      .notNull(),
    knowledgeBaseId: text('knowledge_base_id')
      .references(() => knowledgeBases.id, { onDelete: 'cascade' })
      .notNull(),
    userId: text('user_id')
      .references(() => users.id, { onDelete: 'cascade' })
      .notNull(),
    enabled: boolean('enabled').default(true),
    ...timestamps,
  },
  (t) => [
    uniqueIndex('agents_knowledge_bases_agent_id_knowledge_base_id_unique').on(
      t.agentId,
      t.knowledgeBaseId,
    ),
  ],
);

Query Style

Always use db.select() builder API. Never use db.query.* relational API (findMany, findFirst, with:).

The relational API generates complex lateral joins with json_build_array that are fragile and hard to debug.

Select Single Row

// ✅ Good
const [result] = await this.db.select().from(agents).where(eq(agents.id, id)).limit(1);
return result;

// ❌ Bad: relational API
return this.db.query.agents.findFirst({
  where: eq(agents.id, id),
});

Select with JOIN

// ✅ Good: explicit select + leftJoin
const rows = await this.db
  .select({
    runId: agentEvalRunTopics.runId,
    score: agentEvalRunTopics.score,
    testCase: agentEvalTestCases,
    topic: topics,
  })
  .from(agentEvalRunTopics)
  .leftJoin(agentEvalTestCases, eq(agentEvalRunTopics.testCaseId, agentEvalTestCases.id))
  .leftJoin(topics, eq(agentEvalRunTopics.topicId, topics.id))
  .where(eq(agentEvalRunTopics.runId, runId))
  .orderBy(asc(agentEvalRunTopics.createdAt));

// ❌ Bad: relational API with `with:`
return this.db.query.agentEvalRunTopics.findMany({
  where: eq(agentEvalRunTopics.runId, runId),
  with: { testCase: true, topic: true },
});

Select with Aggregation

// ✅ Good: select + leftJoin + groupBy
const rows = await this.db
  .select({
    id: agentEvalDatasets.id,
    name: agentEvalDatasets.name,
    testCaseCount: count(agentEvalTestCases.id).as('testCaseCount'),
  })
  .from(agentEvalDatasets)
  .leftJoin(agentEvalTestCases, eq(agentEvalDatasets.id, agentEvalTestCases.datasetId))
  .groupBy(agentEvalDatasets.id);

Raw SQL and Advanced Queries

Prefer Drizzle builders whenever the query reads clearly with select, insert().select(), update().from(), joins, CTEs, and groupBy — this keeps table/column references tied to schema, so changes surface as TypeScript errors. Within a builder, expression-level sql<T> is fine for features lacking a helper (JSON path, casts, aggregates, CASE, NOW()). Row locks are clauses, not expressions — use .for('update'), never raw FOR UPDATE.

Use COALESCE only when null-handling is part of required DB semantics (nullable JSONB append/merge, "keep first non-null"). Don't scatter COALESCE(excluded.col, current.col) across ordinary upsert scalars just to avoid an update object — build set from defined values only, and hide any remaining SQL behind named helpers (appendJsonbArray, mergeJsonbObject, keepFirstValue) so the method reads as business intent, not SQL plumbing.

// ✅ Scalars included only when present; SQL hidden behind a named helper.
// (`userWidgetLogs` is illustrative — swap for the real table.)
const updateValues = compactUndefined({
  email: record.email ?? undefined,
  ip: record.ip ?? undefined,
});
await db.insert(userWidgetLogs).values(values).onConflictDoUpdate({
  set: { ...updateValues, events: appendJsonbArray(userWidgetLogs.events, event), updatedAt: now },
  target: userWidgetLogs.id,
});

// ❌ Every scalar becomes SQL plumbing.
set: {
  email: sql`COALESCE(excluded.email, ${userWidgetLogs.email})`,
  ip: sql`COALESCE(excluded.ip, ${userWidgetLogs.ip})`,
}

When refactoring raw SQL:

• Preserve query shape on latency-sensitive paths. If raw SQL is one roundtrip, don't split it into multiple depth-based queries just to drop execute.
• Use $with(...) + insert().select() / update().from() for multi-step single-roundtrip writes Drizzle can express.
• Don't rely on execute<MyRow>(sql...) for safety — it types rows but doesn't keep selected columns in sync with schema changes.
• If only a PostgreSQL feature Drizzle can't express works, keep the raw SQL and tighten it: schema refs in interpolations, explicit user scope, a narrow row interface, and regression tests.

Recursive CTEs are the canonical "keep raw" case — there's no clean WITH RECURSIVE builder, and a rewrite would add depth-based roundtrips:

interface TaskTreeRow {
  id: string;
  parent_task_id: string | null;
}

// execute<T> acceptable: no clean WITH RECURSIVE builder. Keep schema refs in the
// interpolations and scope every leg to the user.
const { rows } = await db.execute<TaskTreeRow>(sql`
  WITH RECURSIVE task_tree AS (
    SELECT ${tasks.id}, ${tasks.parentTaskId}
    FROM ${tasks}
    WHERE ${tasks.id} = ${rootTaskId} AND ${tasks.createdByUserId} = ${userId}
    UNION ALL
    SELECT ${tasks.id}, ${tasks.parentTaskId}
    FROM ${tasks}
    JOIN task_tree ON ${tasks.parentTaskId} = task_tree.id
    WHERE ${tasks.createdByUserId} = ${userId}
  )
  SELECT * FROM task_tree
`);

One-to-Many (Separate Queries)

When you need a parent record with its children, use two queries instead of relational with::

// ✅ Good: two simple queries
const [dataset] = await this.db
  .select()
  .from(agentEvalDatasets)
  .where(eq(agentEvalDatasets.id, id))
  .limit(1);

if (!dataset) return undefined;

const testCases = await this.db
  .select()
  .from(agentEvalTestCases)
  .where(eq(agentEvalTestCases.datasetId, id))
  .orderBy(asc(agentEvalTestCases.sortOrder));

return { ...dataset, testCases };

Database Migrations

See the db-migrations skill for the detailed migration guide.