`@pgsql/utils` constraint builder + `pgsql-deparser@18` produce non-enforcing constraints

[jefflub-ashby]

Summary

t.nodes.constraint() does not set is_enforced. pgsql-deparser@18.x interprets an absent is_enforced as false and emits NOT ENFORCED.

The result is that building a CHECK or FOREIGN KEY constraint with the project's own AST builder, and rendering it with the project's own deparser, produces a constraint that PostgreSQL will never enforce. No error is raised at any stage — the SQL is valid and executes successfully.

This is a disagreement between two first-party packages rather than a bug in either one alone. The deparser's interpretation is correct for parser output; constructed ASTs don't follow the same convention, and the deparser has no way to tell the two apart.

Affected versions

Package	Version
@pgsql/utils	18.2.8	builder omits is_enforced
pgsql-deparser	18.0.0 → 18.3.6	reads absent as NOT ENFORCED
pgsql-deparser	17.18.5	✅ unaffected

Present from the first pgsql-deparser@18 release through current latest.

Reproduction

Using only first-party packages, following the pattern in the pgsql-deparser README:

import * as t from "@pgsql/utils";
import { deparseSync as deparse } from "pgsql-deparser";

const stmt = t.nodes.alterTableStmt({
 relation: t.nodes.rangeVar({ relname: "t", inh: true, relpersistence: "p" }).RangeVar,
 objtype: "OBJECT_TABLE",
 cmds: [t.nodes.alterTableCmd({
 subtype: "AT_AddConstraint",
 behavior: "DROP_RESTRICT",
 def: t.nodes.constraint({
 contype: "CONSTR_CHECK",
 conname: "x_positive",
 raw_expr: t.nodes.aExpr({
 kind: "AEXPR_OP",
 name: [t.nodes.string({ sval: ">" })],
 lexpr: t.nodes.columnRef({ fields: [t.nodes.string({ sval: "x" })] }),
 rexpr: t.nodes.aConst({ ival: t.ast.integer({ ival: 0 }) }),
 }),
 }),
 })],
});

console.log(deparse([{ RawStmt: { stmt } }]));

Expected:

ALTER TABLE t ADD CONSTRAINT x_positive CHECK (x > 0)

Actual:

ALTER TABLE t ADD CONSTRAINT x_positive CHECK (x > 0) NOT ENFORCED

The builder output is:

t.nodes.constraint({ contype: "CONSTR_CHECK", conname: "c" })
// { Constraint: { contype: 'CONSTR_CHECK', conname: 'c' } } <- no is_enforced

Passing is_enforced: true explicitly is preserved by the builder and produces correct SQL.

Scope

Affects the two constraint types PostgreSQL permits NOT ENFORCED on:

CHECK, builder default ALTER TABLE t ADD CONSTRAINT c CHECK (x > 0) NOT ENFORCED ❌
CHECK, is_enforced: true ALTER TABLE t ADD CONSTRAINT c CHECK (x > 0) ✅
FK, builder default ALTER TABLE t ADD CONSTRAINT f FOREIGN KEY(x) REFERENCES o (id) NOT ENFORCED ❌
FK, is_enforced: true ALTER TABLE t ADD CONSTRAINT f FOREIGN KEY(x) REFERENCES o (id) ✅

CONSTR_UNIQUE and CONSTR_PRIMARY are unaffected. Inline constraints in CREATE TABLE are affected the same way as ALTER TABLE ... ADD CONSTRAINT.

skip_validation is not part of the trigger — is_enforced alone determines it. Where skip_validation: true is set, NOT ENFORCED replaces the expected NOT VALID rather than accompanying it:

skip_validation	is_enforced	Output
absent	absent	NOT ENFORCED ❌
absent	true	(none) ✅
true	absent	NOT ENFORCED ❌ — expected NOT VALID
true	true	NOT VALID ✅
true	false	NOT ENFORCED ✅

Why the deparser's logic is not itself wrong

libpg_query omits false booleans from its JSON output, so for parsed ASTs an absent is_enforced genuinely does mean not-enforced. Confirmed with pgsql-parser@18.2.6:

Parsed SQL	is_enforced	skip_validation
CHECK (x > 0)	true	—
CHECK (x > 0) NOT ENFORCED	undefined	true
CHECK (x > 0) ENFORCED	true	—
CHECK (x > 0) NOT VALID	true	true

So parse() → deparse() round-trips correctly today, including NOT ENFORCED, and the deparser's is_enforced !== true check is right for that path.

Changing the deparser to treat absent as enforced would break this. Verified by applying is_enforced === false to 18.3.6 and re-running round-trips:

✗ in : ALTER TABLE t ADD CONSTRAINT c CHECK (x > 0) NOT ENFORCED
 out: ALTER TABLE t ADD CONSTRAINT c CHECK (x > 0) NOT VALID

That silently converts a non-enforced constraint into an enforced one — the same class of failure in the opposite direction. The deparser cannot distinguish "false, omitted by convention" from "not specified by the caller"; both arrive as undefined.

Why this matters

NOT ENFORCED (PG18) means the constraint applies to neither existing nor new rows. Both plausible caller intents are silently violated:

• A bare CHECK or FK constraint should be fully enforced. It is inert instead.
• NOT VALID should enforce for new rows while deferring validation of existing ones — the standard pattern for adding a constraint to a large table. It is inert instead, and the intended clause is dropped.

Nothing surfaces the problem: the SQL is valid, applies cleanly, and enforces nothing. For a migration tool this means shipped constraints that never take effect, discoverable only by inspecting the database afterwards.

This is reachable through the documented path — the pgsql-deparser README's primary example constructs an AST with @pgsql/utils and notes it "could have been obtained from any JSON or AST."

Suggested fix

Preferred: default is_enforced: true in the @pgsql/utils constraint builder for CONSTR_CHECK and CONSTR_FOREIGN. This aligns the builder with the serialization convention the deparser already assumes, fixes existing callers with no code change on their side, and leaves the deparser's parser-output handling untouched. It's clear that this is a large, invasive change given how pgsql/utils is built, so it's understandable if you don't want to pursue it.

Alternatives, if that isn't desirable:

• Add a Typescript type constraint that requires an explicit decision in the builder signature for CONSTR_CHECK and CONSTR_FOREIGN. This would be a breaking change for consumers (though probably fixing a bug in their construction). I can provide a PR for this if you're interested.
• Add a deparser option (e.g. deparse(ast, { assumeEnforced: true })) for callers constructing ASTs rather than round-tripping.
• Document in packages/deparser/README.md under ## Options, and as TSDoc on Constraint.is_enforced, that omitting the field on a constructed AST yields NOT ENFORCED.

Documentation alone would leave every consumer needing to know an implicit serialization convention in order to avoid emitting non-enforcing constraints, so a default in the builder seems the more robust resolution.