orm

Declare and query a table

Handwritten declarations bind concrete models and keys to the shared SQL runtime. Use this package for model reads, writes, lifecycle callbacks and relations; orm/expr and orm/rows provide expressions, joins and typed result selections.

Keep the model and its table declaration together. The table's second type parameter is the primary key type, including named natural keys.

type Product struct {
	Code  string
	Price int64
}

type ProductTable struct {
	orm.Table[Product, string]
	Code  orm.TextPrimary[Product, string]
	Price orm.Number[Product, int64]
}

var Products = orm.Define("products", func(b orm.Builder[Product], p *Product) ProductTable {
	table := orm.TableOf(b, &p.Code)
	return ProductTable{
		Table: table,
		Code:  orm.TextKey(orm.PrimaryOf(b, table.Key())),
		Price: orm.NumberOf(b, &p.Price),
	}
})

Call orm.Check(Products) during startup to report declaration errors together. Each query or write also rejects an invalid table.

Column handles are methodless. Per-kind functions retain the concrete model, value and role types. Use them with the query and table package functions:

products, err := orm.All(ctx, db, Products.Query().
	Where(orm.NumberBetween(Products.Price, 100, 500)).
	OrderBy(orm.TextAsc(Products.Code)))

product, err := orm.Find(ctx, db, Products.Table, "SKU-1")
created, err := orm.Insert(ctx, tx, Products.Table, orm.TextSet(Products.Code, "SKU-1"), orm.NumberSet(Products.Price, 100))
updated, err := orm.Update(ctx, tx, Products.Query(), Products.Key(), "SKU-1", orm.NumberSet(Products.Price, 0))
page, err := orm.Paginate(ctx, db, Products.Query(), orm.Page(2, 20))

Omit an assignment to leave the field unwritten. orm.NumberSet(column, 0) writes zero. Nullable families provide explicit NULL operations, such as orm.NullableTextSetNull(column). Use NullableTextOf, NullableNumberOf, NullableTimeOf, NullableIDOf, or NullableColumnOf for value.Nullable fields.

Declare every persisted field. Tag transient fields with foundry:"-". Relation fields are slots and do not become columns in this prototype. foundry:"column=email_address" overrides the SQL name, and foundry:"default=database" permits omission on insert.

Use EnumOf for a generated enum descriptor and BytesOf for owned byte buffers. Use ColumnOf(b, &p.Field, orm.Codec(custom)) for a custom codec.

Built-in codecs carry a private framework capability and avoid per-value callback isolation. Custom codecs, validators, scanners and JSON methods remain contained: panic or runtime.Goexit fails the read, preserves safe classification and closes the result stream after the callback exits. Applications cannot grant that capability with WithParameterType or by embedding a framework scanner. A failed collection returns no partial model slice; successful rows retain independent values.

A persisted CreatedAt and UpdatedAt pair uses the existing timestamp convention. Both fields must be time.Time or temporal.DateTime. orm.NoTimestamps() disables that convention. A nullable instant named DeletedAt enables soft deletion. Explicit timestamp constructors support other field names.

Use orm.AllLocked(ctx, tx, orm.Lock(q)) or orm.AllLocked(ctx, tx, orm.ForShare(q)) for row locks. The locked terminal functions require *database.Tx. Ordinary reads accept database.Executor; writes accept database.Transactor and retain nested transaction behavior. A *orm.WriteError[M] preserves a hydrated candidate when transaction confirmation is uncertain. The candidate does not prove persistence.

Declare lifecycle hooks with orm.WithHooks and field mutators on the model. Direct relations, joins and projections are described below.

The key field is bound only by TableOf. Derive ordered numeric capabilities with orm.NumberKey(table.Key()) or text capabilities with orm.TextKey(table.Key()); a second field binding is an error. Primary keys are comparable, cannot use sensitive codecs, and a model.ID[M] key must belong to its model. Text and NullableText retain named string types.

Numeric adjustments require a strictly positive, finite amount:

product, err := orm.Update(ctx, tx, Products.Query(), Products.Key(), "SKU-1", orm.NumberDecrement(Products.Price, 10, orm.AtLeast(int64(0))))
n, err := orm.UpdateAll(ctx, tx, Products.Query().Where(orm.NumberLt(Products.Price, 500)), orm.NumberIncrement(Products.Price, 10, orm.AtMost(int64(500))))

The bound is checked by the same SQL UPDATE. Per-key adjustments return orm.ErrBoundExceeded wrapped in fault.Conflict for a bound-rejected visible row, and database.NotFound for a missing or hidden key, with or without hooks/observers. An empty guarded write or lifecycle locking read adds one key-only locking read in the same transaction, retaining all visibility predicates except the bound. Successful writes keep their statement count; see the batch-write guide. Set-based writes return the number of rows satisfying the bound. Adjustments preserve SQL NULL and accept one numeric adjustment plus ordinary assignments. validation.Positive[N]() applies the same strictly positive, finite rule to scalar numeric input; validation.DecimalPositive() checks exact decimal input. Both rules currently run only on the server.

Construct bounds with AtLeast or AtMost; a zero-value Bound is rejected. PostgreSQL evaluates column +/- amount before comparing the bound. An int64 or float64 overflow returns a database out-of-range error rather than silently passing the guard.

orm.InsertMany(ctx, tx, table, [][]orm.Assignment[M]{...}) inserts an atomic bounded batch. NoSoftDeletes() disables the deletion convention and conflicts with explicit SoftDeleteOf. orm.ForceDelete(ctx, db, table.Query(), soft, table.Key(), key) physically deletes active records by default; use package ForceDelete with a WithTrashed or OnlyTrashed query for deleted records. orm.ForceDeleteAll respects query visibility. Explicit timestamp roles win over names; explicit timestamp constructors conflict with NoTimestamps(). CreatedAt, UpdatedAt, and SoftDelete retain the instant type as their second type parameter. Upserts use orm.OnConflict with typed targets and assignments.

orm.Chunk(ctx, ex, query, size, func([]M) error) reads bounded batches with stable ordering and preserves Limit/Offset. Later batches advance from the last ordered values captured before the callback. Nullable ordering uses the same keyset predicate as the existing query API; explicit NULL placement falls back to OFFSET. Rows close before callbacks run. EachChunked uses the same batches and yields one model at a time. Callback errors and cancellation stop traversal.

Relations and projected rows (Stage 1b)

Bind slots during Define, then declare relationships after both tables exist:

var OrderRelations = struct { Buyer orm.BelongsTo[Order, User] }{
    Buyer: orm.BelongsToOf(Orders.Buyer, orm.ColumnExpr(Orders.BuyerID), orm.KeyExpr(Users.ID)),
}
orders, err := orm.All(ctx, db, Orders.Query().With(OrderRelations.Buyer))
loaded, err := orm.LoadFrom(ctx, db, Orders.Query().With(OrderRelations.Buyer), orders)

OneOf and ManyOf bind relation.One and relation.Many fields. Nullable foreign IDs retain their non-null key type. The loader uses the shared engine's 500-key batches and 10,000-row fetched/attachment budgets. orm.LoadFrom returns copies; errors publish no partial result. Target soft deletes and context scopes remain active. orm.Each closes each parent batch before loading relations.

For joins, rows.As[Tag] creates a distinct scope, expr.At binds an explicitly adapted model column to that scope, and rows.LeftScope/RightScope move fields into the joined scope. rows.Define returns a schema embedding rows.Row[R] from b.Row(); rows.IDOf, TextOf, NumberOf, EnumOf, TimeOf, BytesOf and their Nullable...Of counterparts take field pointers and choose their typed codecs. Custom values use rows.FieldOf(b, &r.Field, orm.Codec(custom)) or rows.NullableFieldOf with the underlying value codec. rows.Map unifies that field's value type with its scoped expression. rows.Select requires every field once. rows.SelectValue(Orders.Query(), Orders.TotalCents) selects a scalar. The complete example is the reporting fixture.

Enum descriptors resolve lazily on Check and first use, after package initialization. Model files can sort before generated enum files; no manual initialization references are needed. Call orm.Check and rows.Check during application startup to collect typed fault.Invalid declaration errors.

Scoped expr.Expression[S,V,P] keeps comparison value V separate from projected value P. Nullable fields compare against V and project value.Nullable[V]; use IsNull/IsNotNull for nullness. rows.On unifies comparable unwrapped keys, and rows.InnerJoin accepts only an aliased table on its right side.

Column handles have no methods. Per-kind functions, such as orm.TextEq and orm.NumberGt, accept the concrete handle first and preserve its model, value, input and declaration-role types. Query chains remain methods; read terminals, lock transitions and table operations are package functions.

Expression and source APIs accept explicit sealed adapters, for example expr.Value(orm.TextExpr(Users.Email)). Each adapter is a defined type with the column's existing layout and directly declared private capability methods. Conversion allocates no wrapper and adds no capabilities: input columns remain ineligible for source mappings, primary roles retain their restrictions, and EncryptedExpr permits assignments without expression or equality access.

The surface test pins methodless columns and the retained query/table method sets. Fixture budgets cover individual kinds and model-count growth. There are no compatibility methods or promoted .Column fields; callers use the per-kind operators and explicit adapters.

Stage 3.1a adds JSONOf/NullableJSONOf for value.JSON[T] and IntervalOf/NullableIntervalOf for temporal.Interval. JSON handles keep the payload type through equality, membership, containment and assignment. They support JSONAsc/JSONDesc and JSONIsJSONNull, but no range comparisons or text patterns. expr.JSONType and expr.JSONLength inspect JSON expressions. Interval handles add ordered comparisons and sorting, without numeric adjustments or text patterns. Nullable handles distinguish SQL NULL from values; JSON null is represented by value.JSON[value.Nullable[T]].

Use TableOf(b, &m.Period, orm.Codec(codec.Interval())) for an interval natural key. PostgreSQL compares intervals using 30-day months and 24-hour days: an orm.IntervalEq(Users.Period, temporal.Months(1)) predicate also matches stored 30 days. Interval natural keys therefore collide for 1 mon and 30 days; hydration preserves their distinct calendar and elapsed components. Use expr.SumInterval and expr.AvgInterval for interval aggregates. BytesOf/NullableBytesOf retain buffer ownership and distinguish empty bytes from NULL. orm.TextIContains and orm.NullableTextIContains treat %, _ and \ as literal substring characters and match case-insensitively. For example, orm.TextIContains(Users.Email, "@EXAMPLE.com") binds a value; it never interpolates SQL. Stage 3.1b adds PasswordOf/NullablePasswordOf, which preserve typed hash scalar comparisons, sorting and assignment while rejecting plaintext and text patterns. The password codec retains its sensitivity and redaction.

EncryptedTextOf/EncryptedJSONOf and their nullable constructors expose typed Set; nullable handles also expose SetNull, IsNull and IsNotNull. They do not support ordering, value comparisons or row projections. Single-model writes seal after UUID generation, inside the transaction using its keyring. Every set-based encrypted assignment is rejected, even with a primary-key equality. Hydration opens using each row's table, column and key through the same shared owner as the generated API. Failures publish no partial model. Table/column/key renames require re-encryption; see encryption. Unknown or committed callback failures retain reconciliation candidates through WriteError; a candidate never proves successful commit.

Encrypted Go values (including value.Nullable wrappers) must use the EncryptedTextOf/EncryptedJSONOf constructors or their nullable forms. ColumnOf with an explicit encrypted codec is rejected at startup. Row projections reject encrypted values because they cannot open them yet.

Hydration publishes only complete, successfully opened models. Buffered reads, page reads, eager loads, and each chunk discard the failing batch. Streaming orm.Each retains rows already delivered before an error; it never delivers the failing row or any later row. This matches the old API's streaming behavior.

Stage 3.1c adds cursor pagination and primary-key chunking as package functions:

page, err := orm.CursorPaginate(ctx, db,
    models.Users.Query().OrderBy(orm.NullableTextAsc(models.Users.Nickname)), request.Page)
// request is pagination.CursorRequest[Path, Filters, models.User].
// Map the returned orm.CursorPage with pagination.MapCursorPage.
err = orm.ChunkByKey(ctx, db, models.Orders.Query(), models.Orders.ID, 500,
    func(batch []models.Order) error { return process(batch) })
err = orm.EachByKey(ctx, db, models.Orders.Query(), models.Orders.ID, 500, processOrder)

Cursor requests/results reuse the existing HTTP pagination adapters and data/meta/links JSON shape. Tokens retain the old unsigned v1 format and concrete model, filter and ordering scope. They are positions, not credentials: a client may move its own position, but query predicates (tenant, soft-delete and scopes) always still apply. Nullable and mixed orders use the same keyset predicate, with the primary key appended as a unique tiebreaker. Existing Limit/Offset windows are rejected.

Key chunks accept only primary-key ordering and zero Offset; Limit bounds total delivery. Rows close before callbacks and the next key is captured before callback edits. Deleting rows behind the boundary does not shift later batches. Keep keys stable; iteration does not provide a snapshot across callbacks.

Locked keyed reads compose orm.RequireLocked(ctx, tx, orm.Lock(Orders.Query().Where(orm.KeyEq(Orders.ID, id)))). The executor remains *database.Tx; unlocked reads retain read-pool routing.

Typed raw hydration

orm.Raw and rows.Raw apply no automatic scopes: callers must write tenant, soft-delete and other scope predicates themselves. Both use the supplied executor exactly; raw SQL may write, so they do not choose a read replica. For explicit read routing, the existing level-1 db.QueryRead or database.ReadQuery API is available; an executor wrapper can deliberately forward Query to QueryRead when its SQL is known to be read-only. There is no implicit SQL classification.

orders, err := orm.Raw(ctx, tx, Orders.Table, "SELECT * FROM orders WHERE id=$1", id)
report, err := rows.Raw(ctx, db, OrderRows,
    "SELECT id, label FROM orders WHERE id=$1", id)

Results must contain exactly the declaration's column names, in any order. Use SQL aliases for declared names. Missing, duplicate or extra names fail even when there are zero rows. Arguments go unchanged to the executor. Existing codecs and model encryption opening are reused; row schemas still reject encrypted types. Any decoding/opening failure discards the whole result and closes the stream. For UPDATE … RETURNING, a hydration error does not prove rollback: use an owning transaction and preserve the executor's outcome classification.

Direct eager trees and existence

Declare BelongsToOf, HasOneOf or HasManyOf once, then derive immutable target constraints and finite trees:

orders, err := orm.All(ctx, db, Orders.Query().With(
    OrderRelations.Buyer.Where(orm.NumberGt(Users.Age, 18)).With(UserRelations.Orders),
))

q := Users.Query().Where(orm.Has(UserRelations.Orders, orm.NumberGt(Orders.TotalCents, 100)))
q = orm.WithRelationLimits(q, orm.RelationLimits{
    BatchSize: 100, MaxRows: 5000, MaxDepth: 4,
})
missing, err := orm.LoadMissingFrom(ctx, db, q.With(UserRelations.Orders), users)

Each query load executes one parent statement plus, for each tree branch, ceil(distinct eligible non-null keys / batch size) statements. Nested branches batch the complete fetched target collection. orm.LoadFrom and orm.LoadMissingFrom issue no parent statement; the latter preserves already-loaded slots, including empty ones. LoadFrom and LoadMissingFrom accept a query with With relations and WithRelationLimits overrides. They validate that query without fetching or filtering the supplied parents; parent predicates, order and window do not change the input collection. Both functions share the same copy-and-load engine. Supplied parent provenance/authorization belongs to the caller. Related soft-delete and declared scopes still apply. orm.Has/orm.Not(orm.Has(...)) are correlated EXISTS predicates, compose recursively, and add zero statements or eager loads.

Defaults come from the shared relation-limits owner: batches of 500, depth 8, and a shared maximum of 10,000 fetched rows and 10,000 attached values across the whole tree. Query overrides apply to every branch, pagination and each chunk window. Invalid graphs fail before parent SQL; budget/cardinality/decoding errors publish no partial collected parents. Singular relations reject multiple matches. Relation Limit/Offset is unavailable: a global batch limit would truncate collections across parents. Target ordering appends the primary key as a tie-breaker.

For a statement-count regression, install one detector on the pool and begin a scope around the read. Set a threshold appropriate to the expected batch count:

detector, err := database.NewStatementDetector(database.StatementDetectorConfig{
    Threshold: 4, MaxShapes: 32,
})
// Handle err; supply database.WithStatementDetector(detector) when opening db.
ctx, scope := detector.Begin(t.Context())
_, err = orm.All(ctx, db, Users.Query().With(UserRelations.Orders))
// Handle err, then check both the operation's exact statement count and shapes.
dbassert.AssertStatementShapes(t, scope)

Close rows before checking. Only compiler-produced shapes appear as SQL text; raw statements use keyed fingerprints and violations include bounded call sites. See statement instrumentation.

Joined relation families

ThroughOf(b, &m.Groups) binds a relation.Through[Group, Membership] slot. ManyToManyOf unifies the source/pivot and pivot/target key pairs without retaining public generic queries in the descriptor:

links := orm.ManyToManyOf(Users.Groups, orm.KeyExpr(Users.ID),
    orm.ColumnExpr(Memberships.UserID), orm.ColumnExpr(Memberships.GroupID), orm.KeyExpr(Groups.ID))
users, err := orm.All(ctx, db, Users.Query().With(
    links.Where(orm.ColumnEq(Groups.Active, true)).WherePivot(orm.InputTextEq(Memberships.Role, "admin")).
        With(GroupRelations.Owner).WithPivot(MembershipRelations.Inviter),
))
for _, user := range users {
    for _, link := range user.Groups.All() {
        use(link.Model, link.Pivot)
    }
}

Targets and pivots remain separate typed values in each relation.Link. Distinct pivot rows remain distinct links. A pivot matching multiple targets fails with database.TooManyRows; NULL keys, missing targets and filtered targets create no link. Target and pivot soft-delete scopes apply independently. Use WithTrashedPivot, OnlyTrashedPivot or WithoutTrashedPivot to change the latter. OrderBy and OrderByPivot retain their combined call order, with target and pivot primary keys appended as tie-breakers.

HasManyThroughOf and HasOneThroughOf accept a slot followed by four columns: source local key, intermediate first key, intermediate second key, target foreign key. Both key pairs and the intermediate owner must match. Intermediates retain their scopes. A singular descriptor's OfMany(orders...), LatestOfMany() and OldestOfMany() select one target per parent through DISTINCT ON, with a primary key tie-breaker. OfMany requires an order; latest and oldest use the primary key.

model.MorphName and model.Morphable reuse the old API's persisted discriminator contract. MorphOneOf and MorphManyOf filter targets by the source's declared name. MorphToOf declares one concrete slot per possible target and loads only matching source rows. MorphToManyOf and MorphedByManyOf constrain the pivot discriminator. Names must follow the identifier grammar. Discriminator values are bound parameters.

Each joined family uses one SELECT per key batch. Nested target and pivot branches batch their complete collections under the same depth and fetched/attached-value budgets. Each attached link consumes two values. Polymorphic inverses batch each concrete discriminator group; an empty eligible key set sends no statement. orm.Has and orm.Not(orm.Has(...)) also accept these descriptors without extra round trips. Relation writes use orm.Attach, orm.Detach, orm.Sync and their batch variants; pivot lifecycle callbacks retain the declared model ownership.

Field mutators and explicit accessors

orm.Check reports invalid declarations before application resources start. Mutate<Field>(Input) (Stored, error) and Access<Field>() (Result, error) use value receivers and exact concrete types. Their suffix must name a persisted Go field, including after a rename. Binding happens once during Define; writes invoke typed functions without reflective dispatch. Mutators use a zero receiver, so use the input argument rather than receiver state. Accessors are ordinary explicit calls and never replace stored hydration or DTO mappings.

For a same-type transformation, use an input column with a typed transform:

Email: orm.InputTextOf(b, &m.Email, normalizeEmail),

Use the same input-column family for distinct input and storage types:

Email: orm.InputTextOf(b, &m.Email, normalizeEmailInput),
Note:  orm.NullableInputTextOf[Member, LabelInput](b, &m.Note, nil),

A nil constructor transform selects the matching Mutate<Field> method. An explicit transform and a named mutator for the same field are a declaration error. InputColumn, InputText, InputNumber, InputTime, InputEnum, InputJSON and InputInterval retain their stored comparison, ordering and expression capabilities. Each also has a nullable form. InputBytesOf uses an owned binary codec with the InputColumn kind. Set takes only the concrete input type; predicates take only the stored type. Nullable SetNull keeps SQL NULL distinct from an omitted assignment or an assigned zero value.

Mutators run in declaration order inside the write transaction, after automatic timestamps and before encryption and binding. They run once per assigned value, including values in a batch or a set update. Omitted fields and explicit SQL NULL skip scalar callbacks. Reusing assignments starts from the original input; direct binary inputs own their buffers. Other custom inputs retain ordinary Go copy semantics. Return an error to abort the transaction or savepoint. Panic and Goexit use the existing transaction isolation boundary. Keep transforms pure and free of I/O. Encrypted fields reject mutators; transform plaintext before creating the encrypted value. SQL adjustments cannot bypass a field's Go mutator.

The independent mutator consumer and its PostgreSQL source exercise distinct inputs, explicit getters, omission, NULL and rollback.

Primary keys with distinct mutation inputs reuse the one primary declaration:

table := orm.TableOf(b, &user.ID)
id := orm.InputKeyOf[User, FreshID](b, table.Key(), nil)

orm.InputKeySet(id, FreshID{Value: storedID}) takes fresh input. orm.InputKeyEq(id, storedID) and orm.Find(ctx, db, table, storedID) retain the stored key type. Declare this adapter inside Define. Its nil transform selects the matching named method. A distinct input disables automatic UUID creation, so an omitted required key fails before SQL. Ordinary same-type UUID key mutators retain automatic UUID preparation.

Mutator callback errors expose only a fixed field diagnostic through ordinary formatting. Their original cause remains available through errors.Is and errors.As. Required and nullable password constructors bind named mutators; encrypted fields continue to reject them.

Write lifecycle

Declare a per-write factory with WithHooks(func() Hooks[Order] { ... }) as a Define option. The factory runs inside the owning transaction, after a locked original has been hydrated and its rows closed. It runs once per normal write. Saving precedes Creating or Updating. Mutators and conventions follow these callbacks. The effective draft is validated before mutators and encryption; omitted required fields can be supplied with WriteSet(write, orm.NumberSet(Orders.TotalCents, 0)).

Draft getters accept typed setter functions. Declare callbacks with the concrete model and input types:

setTotal := func(v int64) orm.Assignment[Order] { return orm.NumberSet(Orders.TotalCents, v) }
setNote := func(v LabelInput) orm.Assignment[Member] { return orm.NullableInputTextSet(Members.Note, v) }

orm.WriteGet(write, setTotal) returns Optional[int64]; orm.WriteNullableGet(write, setNote) returns Optional[Nullable[LabelInput]]. orm.WriteIsSet(write, setNote) also reports an explicit NULL as assigned. These functions preserve input values without running mutators. Use the nullable getter for nullable fields. Getter/assignment errors abort the draft even if a callback ignores them. Draft pointers are valid only during before hooks.

orm.WriteUnset(write, setNote) removes the pending assignment in Saving/Creating/Updating. Getters then report omission, including after an explicit zero or SQL NULL. It does not assign a zero or NULL, invoke mutators, or undo a previously persisted value. Unsetting an omitted field is harmless. Later hooks can set it again; automatic defaults/timestamps still follow their ordinary conventions. The setter keeps the model and input types; foreign declarations and expired drafts fail validation. Caller-owned assignment slices remain reusable.

After RETURNING has hydrated and closed its rows, Created or Updated precedes Saved. Deletion runs Deleting then Deleted; force deletion adds ForceDeleting before Deleting and ForceDeleted after Deleted. Restore runs Restoring then Restored. A callback error, panic, Goexit or cancellation aborts the owning transaction/savepoint. Hook and comparison errors expose fixed framework text while retaining the original cause for errors.Is/errors.As. The supplied context shares the generated API's 32-level lifecycle nesting bound; propagate it to nested operations.

Changes[Order] contains only persisted fields, with locked originals and actual RETURNING results (including defaults, triggers and managed timestamps). ChangeOf(changes, Orders.TotalCents) returns the existing lifecycle.FieldChange[int64]; NullableChangeOf(changes, Members.Note) retains Nullable[StoredLabel]. Assigned reflects effective assignments, even if the canonical stored value did not change. Before is absent on create; After is absent after physical deletion. Snapshot getters return owned binary buffers and exclude ignored fields and relation slots. Compare(Orders.Table, before, after, assignments...) provides the same comparison without a write or operation tag.

Observers use NewObserver[Order]("orders.audit") and RegisterObserver through the existing database Module/provider registration at application.New. Dependency constructors return func() Hooks[Order]. Each application's pool owns an immutable observer set. At every stage local hooks run first, followed by observers in declaration order, including models without a local factory. NewDeletionObserver limits dispatch to deletion operations.

AfterCommit callbacks merge through successful savepoints, are discarded on rollback, and run after the outer commit releases its connection. All callbacks run even if one fails; failures preserve the committed outcome and reconciliation candidate. Use the transaction supplied to after hooks for durable outbox/audit writes: in-memory after-commit work is not crash durable.

Set writes skip per-model hooks, observers and change capture. A declaration with hooks or observers requires WithoutModelHooks(Orders.Query()).UpdateAll(...) (and likewise delete/force-delete); unknown executor wrappers require this explicit acknowledgement too. Per-model writes reject the acknowledgement. InsertMany retains its existing bulk semantics: no per-model hooks/observers. Observer discovery shares the generated API's concrete framework-owner lookup; it never invokes a wrapper's observer method or gives it an internal capability. Per-model writes through wrappers dispatch observers from the supplied transaction.

orm.ForceDelete(ctx, db, Orders.Query(), Orders.Soft, Orders.Key(), id) preserves default active visibility. For deleted rows use ForceDelete(ctx, db, Orders.Query().WithTrashed(), Orders.Soft, Orders.Key(), id); OnlyTrashed and other predicates are retained through locking and deletion.

Typed relation aggregate slots

Declare a non-persisted relation.Value[V] field on the model and an orm.Value[Model,V] handle on its table. Bind the handle inside Define with orm.ValueOf(b, &model.Count). Like relation slots, aggregate slots cannot carry persistence tags and do not enter write drafts or ordinary row hydration.

Bind computations after tables and relations have initialized:

var UserAggregates = struct {
    OrdersCount orm.AggregateRelation[User, int64]
    OrdersTotal orm.AggregateRelation[User, value.Nullable[decimal.Decimal]]
}{
    OrdersCount: Users.OrdersCount.Using(
        expr.Related(UserRelations.Orders, expr.Count[Order]())),
    OrdersTotal: Users.OrdersTotal.Using(
        expr.Related(UserRelations.Orders, expr.Sum(orm.NumberExpr(Orders.TotalCents)))),
}

users, err := orm.All(ctx, db, Users.Query().With(UserAggregates.OrdersCount))
count := UserAggregates.OrdersCount.Using(expr.Related(
    UserRelations.Orders.Where(orm.NumberGt(Orders.TotalCents, minimum)),
    expr.Count[Order](),
))
users, err = orm.LoadMissingFrom(ctx, db, Users.Query().With(count), users)

Keep dependencies flowing from tables to relation namespaces to aggregate namespaces. Table initializers must not reference relation or aggregate namespaces; relation initializers must not reference aggregates. Compose nested eager loads at use sites, for example OrderRelations.Buyer.With(UserAggregates.OrdersCount). Wrapping a reference in a closure does not remove a Go initialization cycle.

expr.Related preserves the source model, measured model and exact aggregate result. Using retains the destination slot when replacing its computation. Computations without slot binding fail validation before execution. Different declarations of the same model remain distinct owners and cannot be mixed.

The expression aggregate rules apply unchanged: counts return int64, existence returns bool, exact sums/averages return nullable decimal, float summaries return nullable float64, interval summaries return nullable interval, and extrema retain the present field type with one nullable layer. Use the existing SumFloat/AvgFloat and SumInterval/AvgInterval constructors for those families. expr.Filter takes expression row predicates, for example expr.Filter(expr.Sum(orm.NumberExpr(Orders.TotalCents)), expr.Gt(expr.OrderRow(expr.Value(orm.NumberExpr(Orders.TotalCents))), 0)). It filters the measure without hiding cardinality errors in the input relation. Relation Where restricts that input before aggregation.

A joined relation measures targets. orm.Pivot(UserRelations.Groups) measures its concrete pivot, retaining target and pivot filters and soft-delete visibility. Counts include duplicate edges; use expr.CountDistinctField(orm.KeyExpr(Groups.ID)) for distinct targets. Missing targets and NULL join keys contribute no edges. Singular relations still reject multiple matching rows; many-to-many relations reject a pivot matching multiple targets. Direct, through and concrete morph families are supported. Aggregate sources must be framework descriptors: zero, nil and application wrappers (including embedded AggregateSource interfaces) fail validation before SQL. One-of-many and MorphTo inputs are rejected at plan validation. Inputs requesting eager target or pivot children are also rejected; load those as separate branches.

Loading uses existing With, orm.LoadFrom and orm.LoadMissingFrom, including nested target and pivot branches, transactions, locked reads and chunk windows. Each selected slot executes one grouped statement per batch of distinct eligible non-null source keys. Slots are not fused. A query adds its parent statement; explicit loads add none. Empty key sets execute no aggregate SQL. Relation ordering is omitted from grouped SQL. Shared row and attachment limits apply: each returned group consumes one row, and each computed parent slot consumes one attachment, including empty results. These limits bound returned data, not the rows examined by the database.

Get() returns the typed value and a loaded flag. A loaded zero/false/SQL NULL is preserved by orm.LoadMissingFrom; orm.LoadFrom refreshes it. No related models are hydrated to compute a summary. Errors or cancellation discard the collected result and leave caller-owned parents untouched. All statements use the supplied executor; choose transaction isolation when separate slots require one consistent snapshot.

Complete-model retrieval and global scopes

WithRetrievalHooks(func() RetrievalHooks[Order] { ... }) declares a separate read factory. Retrieved(ctx, executor, order) runs after complete stored hydration and successful row closure. Each nonempty fetch/batch constructs the local factory first, then retrieval observer factories in registration order; each model invokes local callbacks before observers. Register instance-owned read observers with NewRetrievalObserver[Order]("orders.read") and RegisterRetrievalObserver, using the existing database Module at application.New. Write and retrieval observer kinds remain distinct.

Reads use the actual rows' frozen observer owner. Unknown executor wrappers cannot prove absence and receive no internal capability. orm.Each keeps streaming when callbacks cannot run; otherwise it reads bounded, stable, closed batches before callbacks and delivery. All, pagination, chunks and loaded relation roles dispatch retrieval. A failed batch publishes no pending models. Earlier delivered batches remain delivered. Write hydration, DTO projections and scalars skip retrieval, even when their source has read callbacks. orm.Raw dispatches retrieval for its complete models while bypassing all scopes and read routing. A callback failure does not roll back SQL supplied to Raw.

Attach tenant defaults through a lazy source receiving the completed typed declaration. This avoids Go initialization cycles and snapshots declaration handles before consumers can replace their public fields:

var Orders = orm.Define("orders", declareOrders,
    orm.WithGlobalScopes(func(t OrderTable) []orm.GlobalScope[Order] {
        return []orm.GlobalScope[Order]{
            orm.Scope(t.Table, "tenant", func(ctx context.Context) (orm.Predicate[Order], error) {
                id, err := tenantID(ctx)
                return orm.NumberEq(t.TenantID, id), err
            }),
        }
    }),
)

The source runs once when first needed. It may construct relation predicates but must not compile or execute queries. Each scope returns an additive predicate and an error; a zero predicate, foreign declaration, missing context/tenant, invalid name, duplicate name or callback failure fails closed before SQL. Source failures, including panic and Goexit, are retained by the declaration; every later scoped operation returns the failure without executing SQL. Context values are explicit typed request metadata. There is no service registry in the context. Every terminal uses its executing context; WithScopeContext(q, ctx) only supplies context for Compile and inspection.

Scopes apply through the shared effective-predicate owner to reads, keyed and set writes, relation targets/pivots, aggregates, joins and subqueries. Inserts are unfiltered. Remove defaults for one query with WithoutGlobalScope(q, scope) or WithoutGlobalScopes(q); explicit predicates and soft-delete visibility remain. Apply these before taking a row lock. Relation package functions preserve the concrete descriptor and its other side's defaults: WithoutRelationGlobalScope, WithoutRelationGlobalScopes, WithoutPivotGlobalScope and WithoutPivotGlobalScopes. Bypasses do not mutate the source query, descriptor or declaration. Treat them as explicit authorization choices, especially for tenant scopes.

Audit capture and history

Declare model-owned policy with orm.AuditPolicyOf(Orders.Table, orm.AuditField(orm.NullableTextExpr(Orders.Note), record.Exclude)). Unspecified fields use record.Automatic; record.Redact retains change flags and hides values. A zero orm.AuditPolicy[Order] applies Automatic everywhere. Policies cannot exclude or redact the primary key while retaining it as the subject identity. Foreign columns, duplicate rules and invalid disclosures fail registration or capture. Encrypted fields use AuditEncryptedField or AuditNullableEncryptedField for explicit exclusion; Automatic redacts them.

Register orm.Auditing(Orders.Table, policy) through audit.Register(r, pool, recorderKey, config, declaration) in an application module. The recorder and write observer belong to that application instance. Created, Updated, Deleted and Restored capture inside the business transaction, after local callbacks. Force deletion uses Deleted once. An audit failure or a later write failure rolls back both business and audit rows. Set writes still require explicit WithoutModelHooks acknowledgement and produce no automatic audit entries. Install audit storage through the existing audit migrations.

orm.Audit(Orders.Table, changes, policy) also captures a lifecycle callback's stored changes without SQL. Standalone Compare results cannot be audited. Updates, soft deletes and restores retain assigned or changed fields plus the key; creates and physical deletes retain complete snapshots subject to policy. Values come from locked originals and RETURNING, after mutators and database changes; accessors never replace them. Sensitive codecs, encrypted fields and sensitive names are redacted before their capture binder runs. Nested JSON keys are redacted by the shared record owner before binding the audit row. Unchanged and excluded fields do not invoke their capture codec.

History remains record.Model[M,K]; use orm.AuditRead(history, orm.NumberExpr(Orders.Total)) or orm.AuditReadNullable(history, orm.NullableTextExpr(Orders.Note)) with stored column types. AuditReadEncrypted and AuditReadNullableEncrypted expose redaction state. An omitted field, an absent model snapshot and SQL NULL remain distinct. Redacted fields and partially redacted JSON cannot hydrate fabricated values; inspect Snapshot() for deliberate export. The same declared codecs restore persisted keys and historical field values. Audit adapters are package functions and add no methods or per-column closures to ordinary reads.

Pivot relation writes

Relation writes are package functions. Relation descriptors retain their own chain methods; model queries and tables use the package terminals. Inputs and results retain the concrete source, target and pivot types:

pivot, err := orm.Attach(ctx, db, UserRelations.Groups, user, group,
    orm.InputTextSet(Memberships.Role, "member"))
changes, err := orm.Sync(ctx, db, UserRelations.Groups, user, groups,
    []orm.Assignment[Membership]{orm.InputTextSet(Memberships.Role, "member")}, nil)
removed, err := orm.Detach(ctx, db, UserRelations.Groups, user, group)

AttachMany, SyncWithoutDetaching, Toggle, DetachMany, DetachAll and UpdatePivot cover bounded bulk operations. Sync accepts separate create and update assignment slices; a nil update leaves retained links unchanged. Endpoint keys and morph discriminators fill omitted assignments. Explicit assignments and hooks may change them, but inserted pivots must satisfy the relation's endpoint and filter postconditions. Attach permits duplicate existing links; constraints remain the migration's responsibility. Target inputs are deduplicated by refreshed relation keys in input order. Nullable keys use their stored value type even when assignment inputs differ. If a pivot key or morph discriminator has a distinct mutator input, supply that typed assignment explicitly: its stored key cannot synthesize an input. Missing distinct inputs fail before pivot hooks and mutation SQL, matching the old API. Explicit assignments and hook replacements run the mutator once; typed draft getters retain the original input type. Single Attach rechecks both original endpoint identities and their scopes after pivot hooks, inside the same transaction.

Writes refresh and share-lock endpoint identities, skip retrieval callbacks and retain global scopes, target filters, pivot filters and soft-delete visibility. Bulk operations that read current links take FOR NO KEY UPDATE on the source to serialize competing changes. Existing links are locked in primary-key order. Like the old API, bulk sync/detachment selects existing links by pivot filters; target filters constrain requested endpoints and inserted-link checks. Detach soft-deletes when declared; ForceDetach requires a soft-delete pivot and preserves selected pivot visibility, including WithTrashedPivot().

Pivot hooks or observers select the ordinary per-model B5b lifecycle. Otherwise, create/update/removal uses one set statement per change kind. No per-row savepoints are added: each operation owns one transaction, or one savepoint inside the caller's transaction. With primary endpoint keys, attachment uses four statements for any hook-free batch (source, targets, INSERT, postcondition). Sync adds a current-links read and one statement for each nonempty update/removal; inserts also add their postcondition. Hooked inserts cost one statement per pivot; hooked updates/removals also select candidates and lock each original. Internal reads do not call Retrieved.

WithRelationWriteLimit(links, n) lowers the existing atomic row bound. Oversized inputs and current/candidate sets fail; set RETURNING rows are bounded and an excess rolls back the operation with fault.Invalid. Endpoint cardinality violations remain database.TooManyRows. Ordering and eager-child options are rejected on write relations. Empty target batches still refresh the source, as the old API does.

Errors never return successful partial results. Unknown commit and committed callback failures preserve WriteError[P], WriteError[[]P] or WriteError[PivotChanges[P]] according to the function's result. Candidate() is reconciliation data, not proof of persistence or permission to retry. After-commit callbacks remain owned by the outer transaction.

Conflicts and upserts

Upserts use package functions and an explicit immutable conflict policy:

policy := orm.OnConflict(orm.ConflictKey(orm.TextExpr(Users.Email))).DoUpdate(
    orm.ConflictIncoming(orm.TextExpr(Users.Name)), // EXCLUDED, already transformed
    orm.ConflictNull(orm.NullableTextExpr(Users.Nickname)),
)
item, err := orm.Upsert(ctx, db, Users.Query(), policy,
    orm.TextSet(Users.Email, "ada@example.test"), orm.TextSet(Users.Name, "Ada"))

Conflict captures are framework-owned. ConflictSet accepts explicit declared-column adapters; application embedding wrappers cannot replace their setters. ConflictNull uses sealed nullable metadata and never calls an application SetNull override.

orm.ConflictSet(orm.InputTextExpr(column), input) accepts the column's ordinary write input, including its distinct mutator input. Literal updates transform once per policy execution; ConflictIncoming copies the destination's own proposed value without transforming it twice. Omitted proposed fields use database defaults or NULL. Encrypted fields cannot be conflict update targets. OnConflictConstraint[User]("users_email_key") selects a named constraint. PostgreSQL owns constraint/index existence.

UpsertMany accepts a bounded batch of assignment slices. Both APIs skip local hooks, observers, retrieval callbacks and automatic audit, matching the generated upsert API; WithoutModelHooks is unnecessary. Mutators, encryption of proposed insert fields, timestamps and complete RETURNING hydration still apply. Managed update timestamps copy the normalized proposed instant. A skipped conflict has no returned model; batch result order does not map results to input positions.

policy.Where(...) conditions the update using the stored row. Destination global scopes also constrain DO UPDATE. An unconditional update suppressed by those scopes returns a conflict error and rolls back the whole batch. Inserts remain unfiltered. Conflicting repeated update targets are PostgreSQL errors. Caller transactions use the existing savepoint behavior; reconciliation failures retain typed candidates in WriteError. Column/constraint targets ship in slice 6e1; calculated index targets and stored/proposed row calculations belong to the next slice.

Advanced conflict expressions use orm/rows and the existing orm/expr functions:

records := rows.ConflictRowsOf(Products.Query())
stored := expr.At(records.Stored(), orm.NumberExpr(Products.Price))
proposed := expr.At(records.Proposed(), orm.NumberExpr(Products.Price))
policy := rows.OnConflictKeys(expr.Group(expr.Value(orm.TextExpr(Products.Code)))).
    TargetWhere(orm.NumberGt(Products.Price, 0)).
    DoUpdate(rows.SetConflictValue(orm.NumberExpr(Products.Price), expr.Add(stored, proposed)))
policy = rows.ConflictWhereRows(policy, expr.Greater(proposed, stored))

OnConflictKeys accepts row-phase index expressions. TargetWhere (or rows.ConflictTargetWhere for computed predicates) describes a partial index; its fixed migration values are rendered as schema constants. Ordinary insert inputs, update calculations and conditions remain bound parameters. PostgreSQL owns physical index compatibility. Subqueries and selected-phase expressions cannot be index targets.

ConflictRowsOf exposes complete stored and proposed records in a distinct rows.ConflictRow[M] scope. Query filters do not become conflict conditions. The proposed record includes database defaults and trigger changes. Conflict rows cannot execute or serve as standalone SELECT sources. Use rows.Correlate with them for correlated scalar calculations and existence conditions; rows.ConflictScalarRowQuery and rows.ConflictScalarNullableRowQuery handle independent scalar selections. Source declarations and global scopes remain validated, and missing tenant context fails closed before SQL.

SetConflictValue requires a declared destination and its exact stored type, including nullability. Use expr.NullableRow to widen a required result. Encrypted destinations remain excluded. A field with a Go mutator accepts only its own proposed value as a calculation; other SQL calculations cannot bypass that mutator. Upsert lifecycle, transaction and bounded RETURNING behavior are unchanged from the literal conflict API.

Insert-select and source writes

Source writes use orm/rows package functions and immutable builders. Source expressions keep their scope and exact stored type, including nullability. Mapping, match-key and operation types cannot be converted into one another:

source := rows.From(Products.Query().Where(orm.NumberGt(Products.Price, 0)).Limit(100))
insert := rows.InsertFrom(source, Products.Query(),
    rows.MapInsert(orm.TextExpr(Products.Code), expr.Value(orm.TextExpr(Products.Code))),
    rows.MapInsert(orm.NumberExpr(Products.Price), expr.Value(orm.NumberExpr(Products.Price))))
inserted, err := insert.Returning(ctx, db, 100)

matched := rows.MatchSource(orm.KeyExpr(Products.Key()), expr.Value(orm.TextExpr(Products.Code)))
update := rows.UpdateFrom(Products.Query(), source).Match(matched).
    Select(rows.MapUpdate(orm.NumberExpr(Products.Price), expr.Value(orm.NumberExpr(Products.Price))))
count, err := update.Exec(ctx, db)
count, err = rows.DeleteUsing(Products.Query(), source).Match(matched).Exec(ctx, db)

Values(assignments...) replaces fixed inputs and normalizes them once inside the owning transaction. SQL mappings cannot bypass Go mutators, copy encrypted envelopes, or replace a managed update timestamp. Updates cannot assign primary keys. Literal encrypted inputs are rejected for these set-based writes.

Omitted insert fields keep database defaults or NULL. Required fields must be mapped or supplied; UUID keys require a source/literal value or database default. No single Go-generated UUID is shared across selected rows. Managed creation time is supplied only when omitted; update time always comes from the owning clock.

Source filters, windows and tenant scopes remain intact. Destination scopes apply to updates and deletes; inserts are unfiltered. Missing tenant context fails closed. MatchNullableSource allows nullable keys; NULL matches no destination. Mapped updates require at most one source match per affected model and roll back on ambiguity. Literal-only updates and deletes tolerate repeated source keys.

Exec returns an affected count. Returning accepts a limit from 1 through orm.MaxInsertRows (1000), hydrates complete models, and rolls back on overflow, decode failure or row-close failure. The limit bounds retained models, not SQL work or field sizes; it neither truncates the source nor promises result order.

These explicit set-based APIs skip local hooks, observers, automatic audit and retrieval callbacks, matching the old source-write API. No WithoutModelHooks acknowledgement is required. They retain the existing transaction/savepoint and write-outcome reconciliation behavior. DeleteUsing applies soft deletion when configured; ForceDeleteUsing requires a soft-delete declaration and preserves the supplied destination visibility (WithTrashed includes deleted rows).

Scoped keyed writes

Use package functions when a primary-key write must also retain an existing selection, such as a tenant boundary:

visible := Users.Query().Where(orm.NumberEq(Users.TenantID, tenantID))
user, err := orm.Update(ctx, tx, visible, Users.Key(), id, orm.TextSet(Users.Name, "Ada"))
removed, err := orm.Delete(ctx, tx, visible, Users.Key(), id)

Both predicates constrain the write. The key must come from that exact table declaration. Ordinary lifecycle hooks, observers, complete returning hydration, savepoints and single-row cardinality checks remain active. A scope mismatch returns database.NotFound. Delete uses soft deletion when declared; otherwise it physically removes the row. Unknown commit outcomes retain WriteError reconciliation candidates and must not be retried automatically.

UUID seeks and bounded lifecycle writes

Use UUIDLess, UUIDLessOrEqual, UUIDGreater and UUIDGreaterOrEqual for PostgreSQL bytewise UUID comparisons. Both the query model and model.ID[T] owner remain checked by Go. Nullable UUID columns compare present values; SQL NULL does not match. The corresponding expr functions accept scoped row expressions; use their Nullable variants for nullable row values. They add no column methods.

page := Files.Query().Where(orm.UUIDGreater(Files.Key(), after)).
    OrderBy(orm.KeyAsc(Files.Key())).Limit(100)

UpdateEach(ctx, writer, query, limit, draft) accepts a callback of type func(context.Context, *database.Tx, M) ([]orm.Assignment[M], error). DeleteEach(ctx, writer, query, limit) performs normal physical or soft deletion. RestoreEach(ctx, writer, query, soft, limit) restores only trashed candidates. ForceDeleteEach(ctx, writer, query, soft, limit) physically deletes candidates, preserving the query’s active/trashed visibility. Both require the declaration’s SoftDeletable[M] token, such as Users.Soft; foreign or zero tokens fail before I/O.

All four operations lock the complete candidate set in primary-key order, require a positive limit no greater than MaxPerModelWriteRows, and reject overflow and duplicate identities before any draft or lifecycle callback. Incoming lock policies are rejected; candidate selection always uses a waiting FOR UPDATE. Ordering, pagination, eager-loading options and WithoutModelHooks are rejected.

Each row uses the normal hooks, observers, mutators, encryption, timestamps and change tracking. Internal candidate hydration skips retrieval hooks. Rows are closed before callbacks; propagate the supplied callback context and transaction to nested work. One atomic transaction (or one enclosing savepoint on a supplied transaction) covers all candidates, with no savepoint per model. Cancellation, callback failures and later write failures roll back the operation and discard its after-commit work. Successful results follow primary-key order.

An uncertain commit or committed after-commit failure returns no successful rows. *orm.WriteError[[]M] retains reconciliation candidates; Candidate() is not proof of persistence. There is no automatic retry. These bounded lifecycle operations differ from set-based UpdateAll/DeleteAll, which explicitly skip per-model behavior.

Declaration roles

Declare managed timestamp pairs with orm.CreatedAt/orm.CreatedAtOf and orm.UpdatedAt/orm.UpdatedAtOf. These columns retain ordinary time predicates, ordering, projections and aggregates. A managed field declared through TimeOf fails startup validation with the field name and the required constructor. A lone, unmanaged creation timestamp keeps TimeOf and its existing database-default behavior; NoTimestamps continues to disable conventional timestamp management.

Mutated fields use Input* declarations even when input and stored types match: orm.InputTextOf[User, string](b, &m.Name, nil) discovers MutateName; passing a transform function declares it explicitly. Ordinary columns with mutators fail startup. Each Input capability family has a distinct private layout: converting a column or enum into an ordered input field is rejected by the compiler.

Table.Key() is for lookup and ordinary key assignments. NumberKey and TextKey preserve that role. Applications need PrimaryOf only when mapping a primary key in insert-select: declare ID orm.Primary[User, int64] and bind it with orm.PrimaryOf(b, table.Key()). Numeric/text primary adapters retain their role. PrimaryOf cannot coexist with InputKeyOf; neither lookup keys nor mutated input keys can be used as insert-select targets. All primary roles are excluded from source-update targets. KeyColumn remains a read/relation interface and cannot restore mapping eligibility after role erasure.

Expose soft-delete authority as an ordinary declaration field:

type UserFields struct {
    orm.Table[User, int64]
    DeletedAt orm.SoftDelete[User, time.Time]
    Soft orm.SoftDeletable[User]
}
// Inside Define:
deleted := orm.SoftDeleteOf(b, &m.DeletedAt)
fields := UserFields{table, deleted, orm.SoftDeletes(deleted)}
// After declaration:
_, err := orm.ForceDelete(ctx, db, Users.Query().WithTrashed(), Users.Soft, Users.Key(), id)

Every force-delete entry point requires this token, including ForceDeleteAll, rows.ForceDeleteUsing and pivot ForceDetach. Tokens are bound to the exact model declaration; zero and foreign-declaration tokens fail before SQL. Query visibility remains unchanged. Normal soft deletion does not require the token.

rows.MapInsert and rows.MapUpdate accept explicit sealed adapters for eligible column layouts, such as orm.TextExpr(column). Both reject Input fields and managed update timestamps; creation timestamps remain eligible. Embedding a column in an application wrapper does not grant mapping eligibility. Runtime declaration, ownership and SQL-mapping checks remain active.

Framework adapters

Extension fields remain ordinary translations.Text, attachments.One[M], attachments.Many[M] and metadata.Value[V] values on the model. Bind each with orm.SlotOf(b, &m.Title) inside Define; it adds no persisted column. Declare policies separately with translations.TextOf, attachments.OneOf / ManyOf and metadata.ValueOf, passing extensions.BindSlot(Articles.Key(), Articles.Title). The owner comes from extensions.DefineORMOwner("articles", orm.IdentityOf(Articles.Table, Articles.Key()), extensions.OwnerOptions{StorageModel: "articles"}). Keep that storage identity pinned across table renames.

Bind each descriptor to its concrete manager once, then pass bound.Title.Relation(Articles.Title) to a query with With, then use orm.LoadFrom or orm.LoadMissingFrom. This uses the existing extension fetcher and shared relation row budgets, including loaded-empty values and batch splitting. Register slots.DeclareORM(owner, parts) through the existing model-extension assembly to retain deletion cleanup. Runtime managers never live in the table.

HTTP binding accepts modelbinding.ByKey(db, orm.Keyed(Users.Query(), Users.Key()), pathKey). Use ByColumn for a unique alternate key, ThroughRelation for a direct parent relationship, and ThroughSelectedRelation to retain earlier nested models. Duplicate matches fail before eager loading; missing or out-of-scope matches keep the existing 404 behavior. Binding does not acquire authorization or locks.

factory.Define(Users.Table, build) accepts a builder returning []orm.Assignment[User]. Existing states, relationship composition and after-create hooks work with that concrete draft type. Batch creation uses orm.InsertEach, which invokes every row's lifecycle in one bounded transaction, with one enclosing savepoint for a caller-owned transaction. Unknown outcomes remain failures with reconciliation candidates; no automatic retry occurs. Factories and binding adapters use handwritten declarations.

Pagination adapters consume orm.PageResult, orm.SimplePage, orm.CursorRequest and orm.CursorPage. These expose the shared SQL pagination contracts in internal/querycore. Validation, field shapes, cursor bytes and redacted formatting are unchanged; the reflected defining package is now the shared owner. Applications continue using the public aliases. HTTP pagination retains the published cursor schema identity, so manifests and generated clients do not inherit the implementation package move.

The extension namespace uses TextHandle, OneHandle, ManyHandle and ValueHandle: direct typed methods over private, non-generic callback state. These handles reuse the existing managers and preserve publication outcomes; they do not expose legacy Field, Collection or Key descriptor families. OneHandle and ManyHandle remain distinct, non-convertible capabilities. Lookup diagnostics (Keyed, Unique, RelatedUnique) redact scoped values under ordinary, detailed and Go-syntax formatting.

Diagnostic redaction

ORM query handles redact retained predicate literals, bound data values, assignment inputs, numeric guard bounds, cursor tokens, plan contents and fixed pivot bindings from fmt formatting and interpolating errors such as fmt.Errorf("query: %+v", q). This includes %v, %+v and %#v through private and nested wrapper fields. Structural query metadata, including LIMIT/OFFSET, identifiers and SQL shape, is outside this guarantee and may remain visible, even when the compiler sends it as a parameter. Compiled statement formatting omits its retained SQL text too, because explicit raw SQL can contain literals.

Explicit accessors such as Statement.SQL(), Statement.Arguments(), Cursor.Token(), mutation input readers and Plan.Root()/Plan.JSON() still return their inspection data. Formatting the returned data is an explicit disclosure; the guarantee does not redact hydrated models or page items.

Lifecycle terminals and query plans

CreateEach(ctx, db, query, drafts) runs one normal creation lifecycle per input, atomically, in input order. A valid empty batch does not acquire a transaction. CreateOrFirst(ctx, db, query, assignments...) tries insertion first in a savepoint. Only a unique violation from that INSERT on the model's own table allows recovery: after rollback, it reads the first matching competitor. Recovery requires server-reported schema, table and constraint metadata matching the resolved destination; missing metadata preserves the violation. Hooks, observers, hydration errors and uncertain outcomes never trigger recovery or retry. The query must select the conflicting row; snapshot isolation can hide a competitor, in which case the violation is returned. Internal reads skip Retrieved. FirstOrCreate locks the first primary-ordered match, while UpdateOrCreate accepts separate creation assignments and a typed update callback. Only the chosen branch prepares assignments. Omitted creation fields inherit unambiguous top-level equality predicates, including global scopes; mutated fields require explicit input. A created row must satisfy the original query. Concurrent absent rows rely on physical unique constraints; these functions do not retry. Internal model reads skip retrieval callbacks. Read ordering, windows and eager loads are rejected.

PruneBatch(ctx, db, query, size, mass) physically removes a bounded batch of ordinary models. Soft-delete models use ForcePruneBatch(ctx, db, query, soft, size, mass) with their declaration token. Selection locks rows in primary-key order and preserves filters and visibility. With mass=false, every model runs its deletion lifecycle. mass=true explicitly acknowledges one DELETE without per-model hooks, observers or audit. Each batch is atomic. prune.ORMModel and prune.ORMSoftModel register these operations with the bounded prune registry. The registry uses orm.MaxPerModelWriteRows.

Explain returns a bounded typed PostgreSQL Plan without running the SELECT; ExplainAnalyze explicitly executes it. Neither hydrates models, dispatches retrieval callbacks nor runs eager loads. Locked models use orm.ExplainLocked or orm.ExplainAnalyzeLocked, requiring *database.Tx. The corresponding rows functions accept ordinary Result, locked results or transaction-owned records and scalars. Adapt projections, scalars and sets through the existing rows.Read* functions. rows.ExplainCursor/ExplainValueCursor and their Analyze variants plan canonical cursor ordering without adding a page boundary. Plan formatting is redacted; explicit Root and JSON accessors return owned snapshots and may expose server predicates and bound values.

Numeric adjustments use orm.NumberIncrement, orm.NumberDecrement and their nullable equivalents. Amounts are positive and finite; AtMost guards increments and AtLeast guards decrements. Numeric primary and lookup-key adapters expose comparisons, ordering and identity assignment, but no numeric adjustment functions. Managed timestamps and input columns also have no numeric adjustment capability. A guarded per-key update returning no row distinguishes visible bound rejection (ErrBoundExceeded) from a missing or hidden key (database.NotFound).

Projection callbacks use rows.EachRead(ctx, db, rows.Read(report), consume). They receive one fully decoded value at a time while rows remain open. Callback errors stop further decoding and close the result; values already consumed remain consumed if a later decode or close fails. Use report.All when the caller needs an all-or-error slice.

Advisory validation lookups

validation/database accepts an ORM model query and one typed stored column:

available := databasevalidation.Unique(db, Users.Query(), Users.Email)
active := databasevalidation.Exists(db, Users.Query().Where(orm.EnumEq(Users.Status, StatusActive)), Users.Key())

orm.Lookup(Users.Query(), orm.TextExpr(Users.Email)) is the underlying observation handle. It preserves filters, global scopes, soft-delete visibility and stored codecs, without hydrating models, retrieval hooks or implicit transactions. Nullable and mutated columns accept their non-null stored value type. Batch observations use SQL equality in groups of 64; empty lists perform no I/O. Ordered, paginated, eager-loaded and locked sources cannot be used as complete lookup scopes.

databasevalidation.Scoped derives a scope at check time; databasevalidation.UniqueIgnoring derives a typed key exclusion from a supplied validation.Slot. Keep tenant predicates and derive exclusions from trusted route or actor data. Errors never prove availability. These reads are advisory: database constraints still govern writes. See validation for composition and safe metadata exports.

Shared extension loading

Declaration-based attachment, translation and metadata handles use the shared external-slot loader in internal/extensionslot. Public handles retain concrete owner, key and value types through direct method shells over private callback state. Loading keeps per-slot parent batches, loaded-empty values, separate fetched and attached budgets, and the caller's executor. Failed fetches do not attach an incomplete batch tree.

FoundryExtension requires a validated query-family capability and accepts concrete framework slot state. Capabilities never reach slot accessors or application callbacks. Owners retain concrete ORM identity and scope declarations.

Bounded Explain execution and plan decoding also share querycore. The public orm.Plan/PlanNode retain their methods and data contract; their defining Go package is now the internal shared owner.

Consumer resource limits

Use the public constants in orm when bounding application requests or batches. They refer to the compiler's shared limits; expr and rows use the same limits and do not publish alternate spellings.

Constants	What applications can bound
MaxInsertRows, MaxParameters	Rows in an insert/upsert and parameters in the entire statement; wide rows can require smaller batches.
MaxPerModelWriteRows, MaxRelationWriteRows	One atomic lifecycle batch or relation write.
MaxIdentityBatch	Keys supplied to an identity observation or extension scope.
MaxChunkSize, DefaultChunkSize	Explicit model chunks and the default closed batches for eager/observed iteration.
MaxPageSize	Requested page size.
MaxCursorBytes, MaxCursorFields	Encoded cursor size and ordered cursor-key fields.
MaxMembershipValues	Values in a membership predicate.
MaxExpressionDepth, MaxExpressionNodes, MaxScalarSQLBytes	Dynamic expression construction and computed scalar SQL size.

These are upper bounds, not a promise that every combination is valid. Query compilation still validates ownership, phases, ordering and the combined statement. Independent CTE scopes share statement-wide compilation budgets. Ordinary bind values count toward MaxParameters; they are not interpolated into SQL.

Plan-decoder budgets and lifecycle callback recursion guards remain internal implementation bounds.