feat(sidebar): browse PostgreSQL user-defined types, edit enum labels and offer them in the type picker

[datlechin]

Fixes #2484

What this adds

A Types section under each PostgreSQL schema beside Functions and Triggers, listing enums, composites, domains and ranges. Opening one shows its CREATE statement rebuilt from pg_type, with the catalog's properties beside it. An enum's labels sit above the definition and can be added, at the end or before or after a neighbour, and renamed in place. The structure editor's type picker gains a User-Defined group of the table's own types, Open Quickly indexes types, and the MCP server gets list_types.

PostgreSQL and PGlite in this PR. CockroachDB and Redshift are sibling driver classes in the same plugin and keep the empty default. DuckDB, Cassandra, MSSQL and Oracle each have a comparable catalog, and the PluginKit surface makes each one a single +Types.swift, but each needs a probe against a live server, so they are follow-ups.

How it is built

PluginKit, additive: PluginUserDefinedTypeInfo, PluginUserDefinedTypeKind, PluginUserDefinedTypeField, PluginEnumLabelPlacement, and five defaulted requirements on PluginDatabaseDriver (fetchUserDefinedTypes, fetchUserDefinedType, createTypeTemplate, generateAddEnumLabelSQL, generateRenameEnumLabelSQL) plus DriverPlugin.supportsUserDefinedTypeBrowse. PluginKit 21 is still unreleased on main (#2607's bump has no tag), so these ride that bump and every registry plugin is rebuilt with it anyway.

PostgreSQL: one catalog query over pg_type gated on typtype IN ('e','c','d','r'). Composites are kept only where the backing relation is a stand-alone type (relkind = 'c'), so every table's row type stays out; extension members are excluded through pg_depend the way routines are; arrays and the multirange 14 creates beside every range are other typtypes. PostgreSQL 17 records a domain's NOT NULL as a pg_constraint row, so the constraint list keeps contype = 'c' alone. There is no pg_get_typedef, so PostgreSQLTypeDefinition rebuilds the statement from the row and is pure so every shape is pinned by a test; it carries a domain's own collation, a composite field's collation, a range's collation, its non-default operator class qualified, and a multirange name the creator chose (PostgreSQL's default replaces a trailing range rather than appending). The listing also carries the server's own quote_ident spelling of each type, which is what the picker offers. Version gates: pg_range at 9.2, json_build_object at 9.4, ADD VALUE BEFORE/AFTER at 9.1, RENAME VALUE at 10. A literal carrying a backslash is written as an E'' string so it reads the same under standard_conforming_strings=off; the oid predicate accepts only an integer.

App: a fourth SidebarObjectCategory, .type, threaded through the same seams routines and triggers use: SchemaService for the flat list, DatabaseTreeMetadataService per database and schema, DatabaseTreeNode.Kind.userType, the row view, menu spec, type-select, double-click, selection, filter buckets and the quick switcher. SidebarObjectListPresentation.resolve takes one hasSideObjects instead of two parallel bools. DatabaseObjectRef gains .userType and an optional typeKind, so a persisted tab from before this change still decodes.

One fix outside the new paths: the tree's empty-container check counted every side fetch as pending, and a fetch an engine never runs stays idle for good. Types made that universal, so sideLoadStates now counts only the kinds the engine declares. The same gap already held triggers on a tree-layout engine without trigger browsing.

Enum editing runs each edit as one statement the moment the field commits, authorised through ExecutionGateProvider as a schema mutation, recorded in history, then the listings reload and the open grids get the same refresh a trigger edit sends. It runs on the metadata route, a dedicated autocommit connection wherever the engine can pool one, not the session driver: that one holds whatever transaction the user opened in a query tab, and a label added inside it is unusable until the commit, gone on a rollback, and invisible to the reload. ADD VALUE is written IF NOT EXISTS from 9.3 on, so the driver's one reconnect-and-resend cannot fail on a label that already landed; rename has no such form and stays the same class as every other DDL the driver sends. ALTER TYPE … ADD VALUE cannot run inside a transaction block before 12 anyway, so there is nothing to stage. PostgreSQL cannot drop or reorder a label; the UI does not offer it and the docs say why.

Type picker: fetched for the table's own database on popover open through DatabaseManager.withMetadataDriver, across every non-system schema, because the tree service's cache is expansion-driven and empty in the default flat layout. Every entry is schema-qualified, the table's own schema included, because PostgreSQL searches pg_catalog ahead of the search path and a bare text names the built-in even where the schema holds a domain called text. The spelling is the server's quote_ident, so reserved and mixed-case names arrive quoted; the client-side fallback quotes anything that is not a plain lower-case identifier.

Verified

• Catalog query probed with psql against PostgreSQL 17.11 on a throwaway cluster holding an enum, a composite, two domains, a range, a table, hstore and tablefunc: only the user types came back, with labels, fields, constraints and the range's subtype diff parsed as expected.
• Debug build driven against that cluster: the Types section lists the schema's types, filters by name, Open Quickly opens the definition tab, and adding a label through the + control ran ALTER TYPE … ADD VALUE (confirmed in psql) and refreshed the labels and the statement.
• verify.sh generate, build, test (the suites owning every touched type), build PostgreSQLDriver, lint, docs: see the verdicts in the thread. verify.sh plugins is inconclusive on this machine for the known oracle-nio @TaskLocal macro issue; CI runs the aggregate.
• Codex review (working tree) returned nine findings; all four P1s and three P2s are fixed here. The remaining P2, keying the flat-list type fetch by scope rather than connection, is the same shape routines and triggers already use and is listed below as a follow-up.
• Codex adversarial-review returned six findings. Fixed here: enum edits joining the user's open transaction (now the metadata route), the non-idempotent ADD VALUE under the driver's reconnect resend (now IF NOT EXISTS), a same-schema type resolving to a built-in and reserved names left bare (server-side spelling, always qualified), the catalog state the rebuilt definition dropped (collations, opclass namespace, multirange name), and the identity lookup that widened a malformed oid into a schema listing and took the first row (fails closed, verifies the oid, no schema predicate). Dismissed: the PluginKit version bump. Adding defaulted requirements does break a plugin rebuilt against them under an older host, but PluginKit 21 has not shipped: v0.71.0 carries 20 and the registry holds 19 and 20 only, so every registry plugin is rebuilt for 21 with the release that feat(plugins)!: order SQL exports by foreign key dependency and report what the order cannot fix (#2517) #2607 already requires.
• The extended catalog query, both the schema listing and the by-oid reload, was compiled in a standalone harness and run verbatim against the cluster after these changes.
• Not driven live: the type picker popover itself, because the drawn grid cell does not take an accessibility click; its entries are pinned by unit tests.
• No TableProUITests case: the flow needs a live PostgreSQL, which CI does not have.

Before / After

Before: a column of type order_status shows the type name and nothing else; nothing lists the schema's types.

After: the docs screenshot pair committed at docs/images/user-defined-types-sidebar(-dark).png shows the Types section, the definition tab and the label list. They were captured on a 1x display, so they are 1512x861 rather than the 3024x1722 the other docs images carry.

Follow-ups noticed, not changed here

• PostgresColumnTypeResolver.swift:53: any USER-DEFINED column type without enum labels is labelled ENUM(name), so a composite or range column reads as an enum in the structure grid.
• SchemaService.swift:36: the routine, trigger and type fetches are deduplicated per connection rather than per scope, so two windows loading different databases at once can share one result.

https://claude.ai/code/session_014THhVdsUjbCboxNLnQB1dR

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
TablePro	🟢 Ready	View Preview	Sep 3, 2026, 2:51 AM

💡 Tip: Enable Automations to automatically generate PRs for you.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.