Move a table to a new schema

Moving an uncertified public table into its feature schema is the useful response to a database finding on it — not patching the finding in place: canonical-first triage.

Relocate public.<table> → <new>.<table> with references intact. Postgres moves most things for you; the misses are predictable. Read ../db-change/TOOLKIT.md + ../db-change/SKILL.md first. Project: brsgrqvjdzwihsvnfqkf. Retiring a dead table → db-deprecate-table; bringing a table onto the platform standard → db-canonicalize-table.

What ALTER TABLE … SET SCHEMA carries automatically

The table's columns, PK, indexes, CHECK/UNIQUE/FK constraints (its own and inbound FK constraints — cross-schema FKs keep working), RLS policies, triggers, and owned sequences all follow. You do not re-create these.

What does NOT follow: schema-level USAGE and the schema's default privileges — these belong to the schema, not the table. The moved table's grants come along but are dead without schema USAGE: every authenticated/anon access throws permission denied for schema <new>, which wrapper RPCs swallow into a silent null (the cx_canvas_upsert returned null class). Step 2 grants it; Step 3 verifies it.

What you MUST update by hand

1. Registry rows — platform.entity_types.schema_name and platform.shareable_resource_registry.schema_name for this token.
2. Hardcoded public.<table> references in functions/RPCs and views (unqualified refs follow search_path; schema-qualified ones break). Find them, CREATE OR REPLACE repointed.
3. FE access = TWO separate doors. (a) PostgREST exposure + FE types — the target schema must be exposed to PostgREST and added to the pnpm db-types --schema list (TOOLKIT.md §0 trap), else the FE loses types and 404s. (b) Schema USAGE grant (Step 2) — SET SCHEMA does not grant it. A schema can be exposed yet USAGE-denied, which surfaces as a silent null, not a 404. Both are required. supabase-js calls change from .from('<table>') to .schema('<new>').from('<table>').
4. aidream ORM — the target schema must be in db/matrx_orm.yaml additional_schemas with a generate block; the model regenerates into that schema's models_<schema>.py. If a sub-package consumes the table, update aidream/package_integration.py.

Step 1 — Pre-flight discovery

-- functions & views referencing the qualified name
select n.nspname, p.proname from pg_proc p join pg_namespace n on n.oid=p.pronamespace
where pg_get_functiondef(p.oid) ilike '%public.<table>%' or pg_get_functiondef(p.oid) ilike '%<table>%';
select schemaname, viewname from pg_views where definition ilike '%<table>%';
-- inbound FKs (will keep working, but note them)
select conrelid::regclass as referencing, conname from pg_constraint where confrelid='public.<table>'::regclass and contype='f';
-- current policies/triggers (confirm they follow after the move)
select polname from pg_policy where polrelid='public.<table>'::regclass;
select tgname from pg_trigger where tgrelid='public.<table>'::regclass and not tgisinternal;

Grep both repos for <table> usages (FE .from, Python models/managers, package wiring).

Step 2 — Ensure the target schema is ready (the EXPOSURE BLOCKER)

create schema if not exists <new>; then — always, even if the schema already existed (this is the step that gets skipped when the schema is pre-existing, and it's what broke canvas/code/legal/scraper) — GRANT USAGE ON SCHEMA <new> TO authenticated, anon, service_role; plus ALTER DEFAULT PRIVILEGES IN SCHEMA <new> GRANT … TO … for future tables · add it to the db-types --schema list · add to aidream matrx_orm.yaml.

⛔ You CANNOT expose a new schema via the MCP. Supabase's PostgREST exposed-schema list is platform config, not a role GUC (verified: nothing on authenticator.rolconfig), so no SQL you can run reaches it — not pnpm db:apply, not a read through the database tool. It must be added via the dashboard (Settings → API → Exposed schemas) or the management API (PATCH /v1/projects/{ref}/postgrest, preserving the existing list). A FE-read table moved into an unexposed schema 404s for every user the instant it moves. So: get the schema exposed FIRST (ask the user / use the mgmt API), confirm, then move. Don't blind-ALTER ROLE authenticator SET pgrst.db_schemas — you don't know the full current list and will silently un-expose other schemas.

Step 3 — Move + verify it followed

alter table public.<table> set schema <new>;
select count(*) from <new>.<table>;                                    -- unchanged
select polname from pg_policy where polrelid='<new>.<table>'::regclass; -- policies followed
select tgname  from pg_trigger where tgrelid='<new>.<table>'::regclass and not tgisinternal; -- triggers followed
select has_schema_privilege('authenticated','<new>','USAGE'),
       has_schema_privilege('anon','<new>','USAGE');                   -- MUST be true (Step 2); SET SCHEMA does NOT grant it

Repo-wide audit for this whole class — any schema with table grants but no USAGE (run after every move):

select n.nspname,
  count(*) filter (where has_table_privilege('authenticated', format('%I.%I',n.nspname,c.relname),'SELECT')) as granted_tables,
  has_schema_privilege('authenticated',n.nspname,'USAGE') as usage
from pg_namespace n join pg_class c on c.relnamespace=n.oid and c.relkind in ('r','p','v','m')
where n.nspname not like 'pg_%' and n.nspname<>'information_schema'
group by n.nspname
having not has_schema_privilege('authenticated',n.nspname,'USAGE')
   and count(*) filter (where has_table_privilege('authenticated', format('%I.%I',n.nspname,c.relname),'SELECT'))>0;
-- expected leftovers: cron, deprecated (internal/retired — correctly NO usage). Anything else FE-facing = bug.

Step 4 — Repoint the misses

Update the registry schema_name rows; CREATE OR REPLACE any function/view that named public.<table>; re-verify those RPCs run.

Step 5 — Cross-repo finalize

db-change SOP: pnpm db-types (schema added) → swap .from('<table>') → .schema('<new>').from('<table>') everywhere → pnpm sync-types (fix TS). aidream: python db/generate.py → update imports to the new model module + package_integration.py → python db/detect_applied.py → python run.py clean boot. Ledger the migration. Commit + push main on both repos.

Clean cut — no silent shim (SKILL db-change → THE CUT)

The move makes public.<t> vanish, so every stale ref errors — that's correct and desired; do NOT soften it. Register the move FIRST in scripts/dead-relations.json + platform.deprecated_relations, then repoint using pnpm check:dead-relations as your checklist until it's green (it catches the raw-SQL/comment/Python refs tsc can't). If a table genuinely can't move in the window, tripwire it (platform.deprecate_relation), never a passthrough view.

NEVER

• Leave a compat VIEW at the old public.<t> name (the silent shim — the #1 disaster: reads/writes split silently across two tables). The old name must error or tripwire-RAISE, never pass through.
• Move a FE-read table into a schema that isn't exposed + in the db-types list (silent 404s), or that lacks schema USAGE for authenticated/anon (silent null — SET SCHEMA does not grant it).
• Forget the registry schema_name rows — verify_canonical/has_access resolve schema from entity_types, so a stale schema_name breaks RLS resolution.
• Recreate policies/triggers/constraints by hand — they moved with the table; recreating them risks drift.