The image's init (migrate.sh) sent the supabase_admin password as SQL text while its postgresql.conf logged DDL, so the password reached the container log. The server now starts with -c log_statement=none, which the entrypoint also hands to the init server; a harness check pins it, and DEPLOY.md makes the post-init log scan a production prerequisite.
79 KiB
Supabase: local stack, migrations, tests
Status: Step 3 deliverable, extended at Step 4 (2026-10-08). Covers the
foundation (profiles, platform roles, audit log) and the anatomical sheet
tables (ANATOMICAL_SHEET_MODEL.md). Conventions: BACKEND_ARCHITECTURE.md §3–§9.
1. Layout
supabase/
├─ config.toml CLI config (public schema exposed only, PG 17, strong passwords)
├─ migrations/ forward-only SQL, <timestamp>_<name>.sql
│ ├─ 20261008100000_foundation.sql schemas private/audit, updated_at, audit log
│ ├─ 20261008100100_profiles.sql profiles + auth.users trigger + RLS
│ ├─ 20261008100200_platform_roles.sql platform_role, profile_roles, helpers, RPCs
│ ├─ 20261008110000_anatomical_sheets.sql species, sheets, immutable versions, assets, lifecycle RPCs
│ ├─ 20261008110100_private_execute_hardening.sql EXECUTE revoked from PUBLIC on the Step 3 helpers
│ ├─ 20261009100000_sheet_asset_storage.sql bucket anatomical-sheet-assets + read policy through the sheet
│ ├─ 20261010100000_sheet_document_schema_v2.sql document schema 2: references, review_status, count_range checks
│ ├─ 20261011100000_profiles_professional.sql practice_name, professional_title, siret
│ ├─ 20261013100000_files_foundation.sql assets table, seven private buckets, avatar policies through assets, purge queue, browser + worker RPCs (STORAGE.md)
│ ├─ 20261025100000_workspaces.sql V1-1.1 (M1): workspaces, workspace_members, personal workspace per account, access helpers, audit correlation cap (§3.3)
│ ├─ 20261025110000_clients.sql V1-1.2 (M2): clients, unaccent, client RPCs and search, write budget per workspace, N9 audit filter (§3.4)
│ └─ 20261026100000_animals_relations.sql V1-1.4 (M3): species rows, animals, relations with history and voiding, their RPCs and reads (§3.5)
├─ seed.sql LOCAL dev accounts + first owner (never run in prod)
├─ seed/anatomical-sheets/equine/ official equine sheet as a DRAFT, generated by `pnpm import:equine`
│ ├─ legacy-sources/ the historical Horse.svg, Horse_head.svg, pathologies.json, vertebrae.ts (importer input)
│ ├─ legacy-import/ baseline import (document, id map, SVG copies, IMPORT-REPORT.md), no SQL
│ ├─ curation/ authored changes.json + references.json
│ └─ draft/ curated working revision (document, horse.svg, sheet-ref.json, report-migration-map.json, CHANGELOG, REVIEW-REPORT, review/*.svg; bundled into the app with SHEET_BUNDLE=draft) + equine_draft.sql: idempotent, converging seed (sheet, 1 asset row, working draft), applied by db reset and by PGlite
└─ tests/ pgTAP, one file per concern, allowed AND denied paths
packages/supabase/
├─ src/client.ts createBrowserClient() (anon key, PKCE)
├─ src/database.types.ts `supabase gen types` output
├─ src/anatomical-sheets.ts publishSheetVersion(): the core's verdict (`sheet.check`) as gate before the RPC
├─ src/assets.ts avatar flow: begin / quarantine upload / confirm / remove / own state / download, closed AssetFailure kinds
├─ tests/db/ PGlite harness: runs the pgTAP files without Docker
└─ tests/e2e/ real-stack end-to-end (avatar flow with the worker container), `pnpm e2e:avatar` on the sandbox
crates/core/src/domain/sheet/ canonical sheet document and its integrity rules (Rust core; `packages/domain` deleted at R5.7)
2. Prerequisites
-
Node 24 + pnpm via corepack (
WORKSPACE.md). -
Docker daemon reachable by your user for the real stack. If
docker infosayspermission denied ... docker.sock, add yourself to thedockergroup and log in again:sudo usermod -aG docker "$USER" -
No Docker: the PGlite suite still runs (see §5), the CLI commands do not.
The Supabase CLI is a root devDependency (pnpm supabase <cmd>), pinned
like every other tool. Never install it globally.
3. Start, stop, reset
Every CLI call goes through scripts/supabase.sh (pnpm db:*, pnpm supabase <args>): it exports the DEV origin that config.toml reads with env()
(api.external_url, auth.site_url, studio.api_url). A direct
node_modules/.bin/supabase start would leave the literal env(...) text in
GoTrue's mail links. The CLI publishes API 54321, PostgreSQL 54322, Studio
54323 and Mailpit 54324 on 0.0.0.0 (no bind option); the host firewall
closes 54322-54324 on the LAN interface and the DEV gateway serves Mailpit
behind basic auth on 54325 (ENVIRONMENTS.md §3).
pnpm db:start # supabase start: pulls images, applies migrations + seed, prints URLs and keys
pnpm db:stop # supabase stop (keeps the volume)
pnpm db:reset # drop + replay every migration + seed from zero
Local endpoints (config.toml):
| Service | URL |
|---|---|
| API (PostgREST, Auth, Storage) | http://127.0.0.1:54321 |
| PostgreSQL | postgresql://postgres:postgres@127.0.0.1:54322/postgres |
| Studio | http://127.0.0.1:54323 |
| Mail catcher (confirmation / OTP mails) | http://127.0.0.1:54324 |
Addresses on the host itself. From another device of the network: app :5173,
API :54321, Mailpit through the gateway :54325 (basic auth); Studio by SSH
tunnel; PostgreSQL not at all (ENVIRONMENTS.md §3).
pnpm db:start prints the anon and service_role keys. Only the anon key
ever reaches the browser.
Containers, volumes and network are named supabase_<service>_faunosteo
(project_id of config.toml): the CLI imposes this pattern, so the
faunosteo_ prefix convention of DEPLOY.md ("Docker resource naming")
applies as a suffix here. Several projects can run side by side as long as
their project_id and ports differ.
Two kinds of seed (config.toml [db.seed] sql_paths):
| File | Scope | Content |
|---|---|---|
seed.sql |
local only | dev accounts + first owner |
seed/…/*.sql |
every environment | versioned reference data, idempotent (today: the official equine sheet as a draft); each file listed explicitly in sql_paths |
The PGlite harness applies every .sql under seed/ after the migrations,
seed.sql never (it needs the real auth schema); a test fails when a seed
file is missing from sql_paths. On another environment, apply a versioned
seed with psql -f from a privileged session, then upload its assets (§3.1).
Never run supabase db push --include-seed against a shared database: the
list also contains seed.sql and its known local password.
3.1 Sheet assets (Storage)
SVG files live in the private bucket anatomical-sheet-assets, keyed by
content (<sha256>.svg). Reading goes through the sheet (policy
anatomical_sheet_assets_read): a published version of a usable sheet, or
any version for its editors. No API role uploads; the operator does, with
the service key, from the generated seed folder. The tool is the Rust binary
upload-sheet-assets (crates/core, feature upload, R5.5). It targets the
local stack only: there is no default destination and a host that is not
this machine (or the configured FAUNOSTEO_DEV_HOST, a private address) is
refused before any request. Run it by hand; the assistant does not.
pnpm db:start # the seed creates the asset rows (pending)
export SUPABASE_URL=http://127.0.0.1:54321 # http, explicit port; the DEV address needs FAUNOSTEO_DEV_HOST
read -rs SUPABASE_SERVICE_ROLE_KEY && export SUPABASE_SERVICE_ROLE_KEY # typed, not in shell history
pnpm assets:upload --check supabase/seed/anatomical-sheets/equine/draft # read-only plan
pnpm assets:upload supabase/seed/anatomical-sheets/equine/draft # upload + confirm
What a run does, in order, and stops at the first refusal before the first write:
- The folder must lie under
supabase/seed/anatomical-sheets.assets.jsonis read under a size cap (strict keys, plain.svgfile names, lowercase sha256 and uuid, no repeated entry, 64 entries at most). Each drawing is read under a cap and must match its entry (size, sha256) and be a drawing the product accepts (the core's SVG reader, nothing removed). - The destination is checked, then a handshake: the bucket exists, is private, and the key is accepted.
- Each entry is planned from what the destination holds: an object under the key is read back and hashed; different bytes, a missing asset row or a row recording another hash refuse the run (nothing is overwritten, no row is invented; apply the seed first).
- Missing objects are created without replacing anything (
x-upsert: false), read back and verified, and only then is the row confirmed (confirm_sheet_asset_upload, service role only): the read policy ignores a row whose upload was never confirmed, so a user who registers a known hash on their own sheet opens nothing.
A rerun after any failure resumes: an object already verified is kept, a
confirmed row is left alone. A connection failure or a 5xx answer is retried
three times (four attempts); a refusal (401, 403, a state mismatch, an
oversized answer) is final. Exit codes: 0 done, 1 refused (usage, manifest,
destination, content, handshake, state, access), 2 an argument that is not
UTF-8, 3 network. The DEV address counts only if it is an address of this
machine. The key is
never printed and goes through no proxy. pnpm e2e:sheet-assets (= pnpm db:sandbox sheet-assets-e2e) runs the whole behaviour against the sandbox stack, never
DEV (upload, rerun, anon and wrong key, refused destinations, unreachable
port); the script refuses any URL but http://127.0.0.1:57800 and any key
whose issuer is not db-sandbox. The TypeScript tool
(packages/supabase/scripts/upload-sheet-assets.ts) was deleted at R5.7 (RUST.md §11.23).
3.2 Files worker (Storage, Step 8)
User files land in private buckets and are processed by the Rust worker
(crates/worker, docs/development/STORAGE.md). Against the local stack:
pnpm worker:image # faunosteo_worker:dev
pnpm e2e:avatar # = pnpm db:sandbox avatar-e2e: the suite and the worker on the throwaway stack (§5.1) only
docker compose --profile worker up worker # the same container from docker-compose.yml, reading .env.worker
.env.worker (not editable by the assistant): SUPABASE_URL=http://supabase_kong_faunosteo:8000,
SUPABASE_SERVICE_ROLE_KEY=<the service key of pnpm supabase status>, WORKER_ID=dev.
The service key stays in that file and in the shell; it never reaches the
app .env, env.js or a commit.
Seed accounts (local only, password LocalDev-only-password-1):
| Role | |
|---|---|
owner@faunosteo.local |
platform owner (bootstrapped by seed.sql) |
alice@faunosteo.local |
plain user |
bob@faunosteo.local |
plain user |
3.3 Workspaces (V1-1.1, migration M1)
Contract: docs/development/V1-1_OWNERS_ANIMALS.md §4, §8.1, §9. Every
account has exactly one personal workspace, created by the trigger
profiles_create_personal_workspace in the sign-up transaction (GoTrue →
handle_new_auth_user → profile → workspace + holder membership), through
the single write path private.ensure_personal_workspace(uuid) (one
statement, idempotent, race-safe). A deferred constraint trigger refuses at
commit a personal workspace without its holder or with any other member.
Reads: RLS through private.caller_workspace_ids() (member, active profile,
confirmed email, auth.uid() only). No API role writes either table; anon
and service_role hold no privilege on them. From M2 on, record tables use
the same helpers; the app passes its workspace id explicitly.
M1, M2 and M3 are written and tested on PGlite and on the isolated stack
(§5.1), none is applied to the DEV stack. Applying them there is an owner
action, never run by Claude, never with db reset, db push, migration repair, a volume removal or any other destructive command. The procedure
below takes DEV from its current state (every migration up to
20261013100000 applied, fictitious accounts present) to M3 in one run; the
upgrade path was rehearsed on the isolated stack (pnpm db:sandbox up upgrade 20261025100000). Steps 0 to 3 only read DEV (the SQL helpers force a
read-only session); step 3b writes on the sandbox only; step 4 is the one
write on DEV. Nothing below prints a key: the database password stays in the
containers' environment, and supabase status is read with -o env
filtered to the API URL, because its default output prints the service key
and the JWT secret.
Read-only preflight of 2026-10-26 (Claude; CLI 2.120.0, PostgreSQL 17.11,
branch at 9f585e9): history 9 applied, M1 to M3 pending, nothing
remote-only, no linked project; none of the tables, types, functions,
trigger, constraint or extension M1 to M3 create exists; types drift =
exactly the 24 names of step 1, nothing removed; 6 accounts = 6 profiles (4
confirmed), 1 platform role, 9 asset rows (1 clean), 2 storage objects (11
files in the storage volume, leftovers of earlier resets), 13 due rows in the
purge queue, 35 audit rows, species equine only, vault empty; unaccent
and pgtap available, not installed. Run steps 0 to 2 again on the day: DEV
changes as it is used.
# Same shell for every step: helpers, backup folder, private file mode.
devsql() { docker exec -i supabase_db_faunosteo sh -c 'PGOPTIONS="-c default_transaction_read_only=on" \
PGPASSWORD="$POSTGRES_PASSWORD" psql -X -q -A -t -v ON_ERROR_STOP=1 -h 127.0.0.1 -U postgres -d "$1"' sh "$1"; }
boxsql() { docker exec -i faunosteo_dbsandbox_db sh -c 'PGOPTIONS="-c default_transaction_read_only=on" \
PGPASSWORD="$POSTGRES_PASSWORD" psql -X -q -A -t -v ON_ERROR_STOP=1 -h 127.0.0.1 -U postgres -d "$1"' sh "$1"; }
B=~/faunosteo-backups
umask 077
# 0. Target: the local CLI stack of this project, nothing remote.
git status -sb # the V1-1 branch, clean, equal to origin
test ! -f supabase/.temp/project-ref && echo "no linked project"
pnpm -s supabase status -o env | grep '^API_URL=' # http://<DEV host>:54321
docker port supabase_db_faunosteo 5432 # 0.0.0.0:54322: the DEV database
docker inspect -f '{{.Name}} {{.Config.Image}}' supabase_db_faunosteo faunosteo_dbsandbox_db
# the same postgres:17.11.0.004 image twice; no sandbox
# container → `pnpm db:sandbox up fresh` first (§5.1)
mkdir -p "$B" && chmod 700 "$B" # outside Git, never synced, never attached to a report
# 1. History and drift: only the V1-1 migrations may be pending.
pnpm supabase migration list --local # local = remote on every row up to 20261013100000, remote empty on
# 20261025100000 (M1), 20261025110000 (M2), 20261026100000 (M3), nothing else
pnpm -s db:types:check > "$B/types-before.diff"
grep -c '^<' "$B/types-before.diff" # 0: nothing removed
grep -oE '^> [a-z_]+:' "$B/types-before.diff" | sort -u | wc -l
# 24 names, only these: workspaces, workspace_members, workspace_kind, workspace_member_role (M1);
# clients, client_kind, create_client, update_client, set_client_archived, search_clients (M2);
# animals, animal_client_relations, animal_sex, animal_relation_role, create_animal, update_animal,
# set_animal_archived, create_/end_/update_/void_animal_client_relation, search_animals,
# list_client_animals, client_archive_preview (M3); any other name = drift, STOP
# 2. State before: counts, row fingerprints (md5 only, no row printed) and structure.
cat > "$B/counts.sql" <<'SQL'
set timezone = 'UTC';
select 'users ' || count(*) from auth.users;
select 'identities ' || count(*) from auth.identities;
select 'profiles ' || count(*) from public.profiles;
select 'profile_roles ' || count(*) from public.profile_roles;
select 'species ' || count(*) from public.species;
select 'sheets ' || count(*) from public.anatomical_sheets;
select 'sheet_versions ' || count(*) from public.anatomical_sheet_versions;
select 'assets ' || count(*) from public.assets;
select 'purge_queue ' || count(*) from private.object_purge_queue;
select 'buckets ' || count(*) from storage.buckets;
select 'objects ' || count(*) from storage.objects;
select 'audit_logs ' || count(*) from audit.audit_logs;
select 'migrations ' || string_agg(version, ',' order by version) from supabase_migrations.schema_migrations;
select 'md5 users ' || md5(coalesce(string_agg(t::text, '|' order by t::text), '')) from auth.users t;
select 'md5 profiles ' || md5(coalesce(string_agg(t::text, '|' order by t::text), '')) from public.profiles t;
select 'md5 profile_roles ' || md5(coalesce(string_agg(t::text, '|' order by t::text), '')) from public.profile_roles t;
select 'md5 assets ' || md5(coalesce(string_agg(t::text, '|' order by t::text), '')) from public.assets t;
select 'md5 sheet_versions ' || md5(coalesce(string_agg(t::text, '|' order by t::text), '')) from public.anatomical_sheet_versions t;
select 'md5 species equine ' || md5(t::text) from public.species t where t.id = 'equine';
select 'tables ' || n.nspname || ' ' || count(*) from pg_class c join pg_namespace n on n.oid = c.relnamespace
where c.relkind in ('r', 'p') and n.nspname in ('public', 'private', 'audit', 'auth', 'storage') group by n.nspname order by n.nspname;
select 'functions ' || n.nspname || ' ' || count(*) from pg_proc p join pg_namespace n on n.oid = p.pronamespace
where n.nspname in ('public', 'private') group by n.nspname order by n.nspname;
select 'policies ' || count(*) from pg_policies where schemaname in ('public', 'storage', 'audit');
select 'rls tables public ' || count(*) from pg_class c join pg_namespace n on n.oid = c.relnamespace
where n.nspname = 'public' and c.relkind = 'r' and c.relrowsecurity;
SQL
devsql postgres < "$B/counts.sql" > "$B/counts-before.txt" && cat "$B/counts-before.txt"
# users = profiles = N (6 on 2026-10-26); the other lines are the reference of steps 3b and 5
# Privileges (read only), the reference of step 3b: effective ACLs of the application schemas, grants
# of auth and storage to the API roles, policies, triggers, constraints, extensions, postgres defaults.
cat > "$B/privs.sql" <<'SQL'
select 'acl relations ' || n.nspname || ' ' || md5(string_agg(c.relname || ':' || coalesce(c.relacl, acldefault(case c.relkind when 'S' then 's'::"char" else 'r'::"char" end, c.relowner))::text || ':' || c.relrowsecurity || c.relforcerowsecurity, '|' order by c.relname))
from pg_class c join pg_namespace n on n.oid = c.relnamespace
where n.nspname in ('public', 'private', 'audit', 'supabase_migrations') and c.relkind in ('r', 'v', 'S', 'p') group by n.nspname order by n.nspname;
select 'acl functions ' || n.nspname || ' ' || md5(string_agg(p.oid::regprocedure::text || ':' || coalesce(p.proacl::text, '-') || ':' || p.prosecdef || ':' || coalesce(array_to_string(p.proconfig, ','), '-'), '|' order by p.oid::regprocedure::text))
from pg_proc p join pg_namespace n on n.oid = p.pronamespace where n.nspname in ('public', 'private', 'audit') group by n.nspname order by n.nspname;
select 'acl schemas ' || md5(string_agg(nspname || ':' || coalesce(nspacl::text, '-'), '|' order by nspname))
from pg_namespace where nspname in ('public', 'private', 'audit');
select 'grants ' || n.nspname || ' ' || md5(string_agg(c.relname || ':' || pg_get_userbyid(a.grantee) || ':' || a.privilege_type || ':' || a.is_grantable, '|'
order by c.relname, pg_get_userbyid(a.grantee), a.privilege_type))
from pg_class c join pg_namespace n on n.oid = c.relnamespace cross join lateral aclexplode(c.relacl) a
where n.nspname in ('auth', 'storage') and a.grantee not in (c.relowner, 'postgres'::regrole) group by n.nspname order by n.nspname;
select 'policies ' || md5(string_agg(schemaname || '.' || tablename || '.' || policyname || ':' || cmd || ':' || roles::text || ':' || coalesce(qual, '-') || ':' || coalesce(with_check, '-'), '|'
order by schemaname, tablename, policyname)) from pg_policies where schemaname in ('public', 'storage', 'audit');
select 'triggers ' || count(*) || ' ' || md5(string_agg(tgrelid::regclass::text || '.' || tgname || ':' || tgenabled::text || ':' || tgfoid::regprocedure::text, '|' order by tgrelid::regclass::text, tgname))
from pg_trigger t join pg_class c on c.oid = t.tgrelid join pg_namespace n on n.oid = c.relnamespace
where not tgisinternal and n.nspname in ('public', 'private', 'audit', 'auth', 'storage');
select 'constraints ' || count(*) || ' ' || md5(string_agg(conrelid::regclass::text || '.' || conname || ':' || pg_get_constraintdef(oid), '|' order by conrelid::regclass::text, conname))
from pg_constraint where connamespace::regnamespace::text in ('public', 'private', 'audit');
select 'extensions ' || string_agg(extname || '@' || extversion, ',' order by extname) from pg_extension;
select 'default acl postgres ' || count(*) || ' ' || md5(string_agg(defaclnamespace::regnamespace::text || ':' || defaclobjtype::text || ':' || defaclacl::text, '|' order by defaclnamespace::regnamespace::text, defaclobjtype::text))
from pg_default_acl where defaclrole = 'postgres'::regrole;
SQL
devsql postgres < "$B/privs.sql" > "$B/privs-before.txt" && wc -l < "$B/privs-before.txt" # 15
cat > "$B/managed.sql" <<'SQL'
select 'vault secrets ' || count(*) from vault.secrets;
select 'managed default acl ' || count(*) from pg_default_acl
where defaclrole in ('supabase_admin'::regrole, 'supabase_auth_admin'::regrole);
SQL
devsql postgres < "$B/managed.sql"
# vault secrets 0, else STOP: no backup, no M1 to M3; vault rows need a procedure of their own (keys)
# managed default acl D (21 on 2026-10-26): step 3b leaves out exactly D + 8 entries
# 3. Backup outside Git: the whole `postgres` database (auth, storage metadata, public, private, audit,
# migration history, grants) and the bytes of the Storage files (the storage container's volume).
docker exec supabase_db_faunosteo sh -c 'PGPASSWORD="$POSTGRES_PASSWORD" pg_dump -h 127.0.0.1 -U postgres -d postgres -Fc' \
> "$B/dev-before-v1-1.dump" && chmod 600 "$B/dev-before-v1-1.dump"
docker cp supabase_storage_faunosteo:/mnt - > "$B/dev-before-v1-1-storage.tar" && chmod 600 "$B/dev-before-v1-1-storage.tar"
ls -l "$B" # both files non-empty, -rw-------
docker exec -i supabase_db_faunosteo pg_restore --list < "$B/dev-before-v1-1.dump" \
| grep -cE "TABLE DATA (auth users|public profiles|storage objects|supabase_migrations schema_migrations)" # 4
tar -tvf "$B/dev-before-v1-1-storage.tar" | grep -c '^-' # the volume's file count (11 on 2026-10-26)
# 3b. The dump restores, on the isolated stack (never on the DEV server), with exit status 0: a scratch
# database of the sandbox, restored in one transaction that stops at the first error, compared with
# step 2, then dropped. `restore_check_v1_1` is the only database created or dropped here.
# Prerequisite, sandbox only: DEV owns the cluster role `supabase_functions_admin` (schema
# `supabase_functions`), a fresh sandbox does not, and the dump's ACL entries name it. Create
# it once there, NOLOGIN with no global privilege (what a no-owner replay needs); never on DEV:
docker exec -i faunosteo_dbsandbox_db sh -c 'PGPASSWORD="$POSTGRES_PASSWORD" psql -X -q -v ON_ERROR_STOP=1 -h 127.0.0.1 -U postgres -d postgres' <<'SQL'
do $$ begin
if not exists (select 1 from pg_roles where rolname = 'supabase_functions_admin') then
create role supabase_functions_admin nologin nosuperuser nocreaterole nocreatedb noreplication nobypassrls noinherit;
end if;
end $$;
SQL
# The restore leaves out the D + 8 TOC entries `postgres` may not replay; the dump file itself stays
# complete and unchanged (the list is a copy of its table of contents):
# - D « DEFAULT ACL … supabase_admin | supabase_auth_admin »: default privileges for objects those
# roles create (platform defaults; `postgres` may not change another role's defaults; no data,
# no application privilege; the sandbox image holds its own);
# - 5 « ACL vault »: secrets, decrypted_secrets, create_secret, update_secret,
# _crypto_aead_det_decrypt; with --no-owner their extension makes them postgres's own, and
# postgres cannot grant itself a grant option; the service_role grants of these entries come
# out identical from the extension (rehearsal of 2026-10-26);
# - 1 « ACL pg_catalog » pg_reload_conf, 1 « ACL extensions » pg_stat_statements_reset: owned by
# supabase_admin, postgres may not grant on them;
# - 1 « TABLE DATA vault secrets »: postgres may not write vault.secrets; nothing is lost only
# because step 2 read 0 rows. With rows there, this procedure does not apply.
docker exec -i faunosteo_dbsandbox_db pg_restore --list < "$B/dev-before-v1-1.dump" > "$B/full.list"; echo "list exit $?" # 0
grep -E '^[0-9]+; [0-9]+ [0-9]+ (DEFAULT ACL [a-z_]+ DEFAULT PRIVILEGES FOR (SEQUENCES|FUNCTIONS|TABLES) (supabase_admin|supabase_auth_admin)|ACL pg_catalog FUNCTION pg_reload_conf\(\) supabase_admin|ACL extensions FUNCTION pg_stat_statements_reset\(.*\) supabase_admin|ACL vault (FUNCTION|TABLE) .* supabase_admin|TABLE DATA vault secrets supabase_admin)$' \
"$B/full.list" > "$B/excluded.list"
cut -d' ' -f4-5 "$B/excluded.list" | sort | uniq -c
# exactly: D « DEFAULT ACL », 1 « ACL extensions », 1 « ACL pg_catalog », 5 « ACL vault », 1 « TABLE DATA »;
# any other count = STOP (never widen the pattern to make it fit)
grep -vxFf "$B/excluded.list" "$B/full.list" > "$B/restore.list"
wc -l < "$B/full.list"; wc -l < "$B/restore.list" # the second = the first minus D + 8
docker cp "$B/restore.list" faunosteo_dbsandbox_db:/tmp/restore.list # names only, no data; gone with the sandbox
docker exec faunosteo_dbsandbox_db sh -c 'PGPASSWORD="$POSTGRES_PASSWORD" createdb -h 127.0.0.1 -U postgres restore_check_v1_1'
docker exec -i faunosteo_dbsandbox_db sh -c 'PGPASSWORD="$POSTGRES_PASSWORD" pg_restore -h 127.0.0.1 -U postgres \
-d restore_check_v1_1 --no-owner --use-list=/tmp/restore.list --single-transaction --exit-on-error' \
< "$B/dev-before-v1-1.dump" 2> "$B/restore-check.log"; echo "pg_restore exit $?"
wc -c < "$B/restore-check.log" # 0: the log, rewritten by this run, holds no message
# exit 0 and an empty log, else STOP: the single transaction rolled back, the scratch database is
# empty; keep the log (a COPY error can quote a row: report only its ERROR and Command was lines),
# drop the database (last line), report
boxsql restore_check_v1_1 < "$B/counts.sql" > "$B/counts-restored.txt"
diff "$B/counts-before.txt" "$B/counts-restored.txt" && echo "rows identical" # the 28 lines
boxsql restore_check_v1_1 < "$B/privs.sql" > "$B/privs-restored.txt"
diff "$B/privs-before.txt" "$B/privs-restored.txt" && echo "privileges identical" # the 15 lines
docker exec faunosteo_dbsandbox_db sh -c 'PGPASSWORD="$POSTGRES_PASSWORD" dropdb -h 127.0.0.1 -U postgres restore_check_v1_1'
# 4. Apply without reset (M1's backfill runs inside the migration); no app session open on DEV meanwhile.
# `migration up` applies every pending migration of the checkout, in order.
pnpm supabase migration up --local; echo "migration up exit $?"
# "Applying migration 20261025100000_workspaces.sql...", then ..._clients.sql, then
# 20261026100000_animals_relations.sql; exit 0. Keep the whole output for the report.
Step 3b was rehearsed on the sandbox with fictitious data, never with DEV
data (2026-10-26, a 156 MB dump of the sandbox's postgres database at M3,
2 144 accounts, vault empty; D = 18 there, the sandbox has no
supabase_functions schema). Four variants, exit status read from the shell:
--no-owneralone (the earlier procedure): exit 1, 26 errors, the 26 entries above (18 default privileges, 5 vault ACL, 2 ACL on supabase_admin functions, the vault.secrets COPY); rows identical;--no-acl: exit 1 (the vault.secrets COPY), and every grant of the application schemas lost: rejected, it removes what the check must prove;--exit-on-erroron the whole list: exit 1 at the vault.secrets COPY, before the policies: rejected;- the filtered list with
--single-transaction --exit-on-error: exit 0, empty log, the 28 lines and the 15 privilege lines identical. Retained.
First run on the real DEV dump (2026-10-27): the 29 expected entries were
left out (D = 21), and the restore stopped with exit 1 on
GRANT USAGE ON SCHEMA supabase_functions TO postgres: role
supabase_functions_admin missing from the sandbox (DEV: LOGIN, CREATEROLE,
search_path=supabase_functions, owns the supabase_functions objects).
With the NOLOGIN role above and nothing else changed: exit 0, empty log, the
28 and 15 lines identical, schema supabase_functions restored with its 2
tables. The DEV role's login, CREATEROLE and password are not copied: no
privilege is exercised by a --no-owner replay.
A table whose stored ACL is the explicit owner-only one comes back with no
ACL, the same privileges (pg_dump leaves out an ACL equal to the default):
privs.sql therefore compares effective ACLs (acldefault). DEV differs
from the sandbox (read-only check, 2026-10-26): D = 21, the three more rows
being the defaults of its supabase_functions schema, so 29 entries left
out; its other own objects (supabase_functions, _realtime) grant
postgres no grant option. An entry that would still fail on the DEV dump
stops the single transaction: exit status not 0, nothing kept, STOP. This
backup is no full cluster backup. What it covers and what it does not:
- the dump holds the whole
postgresdatabase: auth (accounts, identities, sessions, password hashes: the files are as sensitive as DEV itself), profiles, roles, assets, storage metadata, audit, the migration history and the grants. It does not hold the cluster roles (created by the image), the_supabasedatabase (services' internal metadata), nor the Storage file bytes. It holds the vault rows, but a restore aspostgrescannot reload them (left out of step 3b; it only proves anything because DEV has none, step 2, and says nothing of a restore with vault rows); - the tar holds the Storage file bytes; restoring it is not rehearsed, and M1
to M3 touch neither Storage nor
assets; - a restore of this dump into the DEV stack is not rehearsed: step 3b proves the dump is complete and readable, not that DEV can be rolled back with it (see below).
After applying (read only, before anyone signs in):
pnpm supabase migration list --local # M1, M2, M3 now have their remote column; nothing pending
cat > "$B/post.sql" <<'SQL'
select 'workspaces ' || count(*) from public.workspaces;
select 'members ' || count(*) from public.workspace_members;
select 'profiles without exactly one personal workspace ' || count(*) from public.profiles p
where (select count(*) from public.workspaces w where w.personal_user_id = p.id) <> 1;
select 'personal workspaces without their sole holder ' || count(*) from public.workspaces w
where w.kind = 'personal'
and ((select count(*) from public.workspace_members m where m.workspace_id = w.id) <> 1
or not exists (select 1 from public.workspace_members m
where m.workspace_id = w.id and m.user_id = w.personal_user_id and m.member_role = 'holder'));
select 'duplicate memberships ' || count(*) from (
select workspace_id, user_id from public.workspace_members group by 1, 2 having count(*) > 1) d;
select 'active species ' || count(*) from public.species where active;
select 'species ' || string_agg(id, ',' order by sort_order) from public.species;
select 'clients ' || count(*) from public.clients;
select 'animals ' || count(*) from public.animals;
select 'relations ' || count(*) from public.animal_client_relations;
select 'rls off on V1-1 tables ' || count(*) from pg_class c join pg_namespace n on n.oid = c.relnamespace
where n.nspname = 'public' and not c.relrowsecurity
and c.relname in ('workspaces', 'workspace_members', 'clients', 'animals', 'animal_client_relations');
select 'anon, service_role or PUBLIC grants on V1-1 tables ' || count(*) from information_schema.role_table_grants
where table_schema = 'public' and grantee in ('anon', 'service_role', 'PUBLIC')
and table_name in ('workspaces', 'workspace_members', 'clients', 'animals', 'animal_client_relations');
select 'anon EXECUTE on V1-1 RPCs ' || count(*) from information_schema.role_routine_grants
where routine_schema = 'public' and grantee in ('anon', 'PUBLIC')
and routine_name in ('create_client', 'update_client', 'set_client_archived', 'search_clients',
'create_animal', 'update_animal', 'set_animal_archived', 'search_animals', 'list_client_animals',
'client_archive_preview', 'create_animal_client_relation', 'end_animal_client_relation',
'update_animal_client_relation', 'void_animal_client_relation');
select 'unaccent ' || count(*) from pg_extension where extname = 'unaccent';
select 'pgtap ' || count(*) from pg_extension where extname = 'pgtap';
SQL
devsql postgres < "$B/post.sql"
# workspaces N, members N, then 0, 0, 0; active species 10;
# species equine,asinine,canine,feline,bovine,caprine,ovine,porcine,leporine,other;
# clients 0, animals 0, relations 0; then 0, 0, 0; unaccent 1; pgtap 0
devsql postgres < "$B/counts.sql" > "$B/counts-after.txt"; diff "$B/counts-before.txt" "$B/counts-after.txt"
# only these lines differ (2026-10-26 values): species 1 → 10; migrations + the three V1-1 versions;
# tables private 1 → 2, public 7 → 12; functions private 41 → 54, public 26 → 40; policies 15 → 20;
# rls tables public 7 → 12. Every count and md5 of the existing rows is unchanged.
pnpm db:types:check # no diff
pnpm db:test; echo "db:test exit $?" # pgTAP 001-014, 921 assertions, exit 0 (see its effects and the 2026-10-27 run below)
echo "select 'pgtap ' || count(*) from pg_extension where extname = 'pgtap';" | devsql postgres
# 0, as before the run; 1 = left installed (interrupted run): report it, do not drop it by hand
devsql postgres < "$B/counts.sql" | diff "$B/counts-after.txt" - # no diff: the tests left no row
Effects of the test commands on DEV, read from the installed CLI 2.120.0 and the scripts:
pnpm db:test(supabase test db): runs pg_prove aspostgres; every file ofsupabase/testsis onebegin … rollback, so no row stays. It creates thepgtapextension before the run and drops it after when it was absent (DEV: absent), a transient schema change; an interrupted run can leavepgtapinstalled (post.sqlshows it). Run it before the browser acceptance, with no app session open and the worker stopped (010 leases the queue rows it finds until its rollback). The files share the stack with the data it already holds: 007 and 010 count only their own keys, and 010 first leases the base's own queued work (due purge entries, avatar jobs, uploads begun more than 24 h ago) inside its transaction. Assumptions that stay: the avatars bucket far below its 10 GiB ceiling; fewer than about 90 quarantine leftovers of the base (the sweep returns at most 100, oldest first); the official equine sheet still at its seeded single draft (007 checks the seed as delivered).- First run on DEV after M1 to M3 (2026-10-27, suites at
6c36e2f): exit 1, 919 assertions planned, 915 run, 914 passed. 007 stopped at its storage insert (the drawing was already uploaded and confirmed on DEV: duplicate key, 4 assertions never run); 010 failed « nothing is due before its grace elapsed » (13 due purge entries of DEV counted). 012 to 014 (V1-1) passed; data, privileges andpgtap(absent) identical before and after. Both files were fixed on the branch (007: 20 assertions, 010: 141) and validated on the sandbox, clean and with those conditions reproduced (scenarios and mutations in the journal of 2026-10-27). DEV is not validated bydb:testuntil a run there passes. pnpm e2e:avatar: not part of this procedure. Since 2026-11-01 it runs on the sandbox only (owner decision; the DEV scriptscripts/avatar-e2e.shis removed). Its DEV run, as recorded on 2026-10-26, wrote on DEV beyond its own fixture: it creates a temporary confirmed accounte2e-avatar-<time>@faunosteo.local(and, after M1, its workspace), deletes it at the end (a failed run leaves it), signs in asbob@faunosteo.local(sessions andlast_sign_in_atof an existing account), and its worker drains every due purge-queue row and quarantine leftover of the stack (13 rows on 2026-10-26). M1 to M3 touch neither Storage norassets; the avatar stays covered by the browser acceptance below.
Stop and report instead of continuing when: migration list shows another
pending migration or a remote-only one; db:types:check before step 4 shows a
removed line or another name than the 24 of step 1; the dump check misses a
line; vault.secrets is not empty; the excluded entries differ from D + 8;
the restore check exits other than 0, logs anything or shows a differing
line; migration up applies anything else than the three V1-1 migrations
or fails; any check after it differs from the expectation.
When something fails:
- backup or restore check fails: nothing was applied; fix the cause (space, container, sandbox) and start again from step 2. Never apply without a restore check that passed;
- history diverges (another pending version, a remote-only version, an object
present without its history): stop; no
migration repair, no hand-made insert into the history. The divergence is analysed first, then resolved by its own owner-validated procedure; migration upfails: run nothing else. Read its message, thenpnpm supabase migration list --local(which of the three versions are recorded) and the object query below (which objects exist; before M1: 0 everywhere and species 1; after M3: tables 6, types 5, functions 27, trigger 1, constraint 1, unaccent 1, species 10). The CLI applies the files one after the other and stops at the first failure: M1 can be recorded with M2 not. Whether a failed file leaves part of its statements behind was not verified on this CLI version: check the objects, not only the history. Do not assume a full rollback, and do not runmigration upagain before the state is understood;- profiles not caught up (a profile without exactly one workspace): stop;
private.backfill_personal_workspaces()is idempotent, but running it by hand is a write on DEV, decided after the cause is known; - an RLS policy refuses a legitimate access in the browser: note the account
(fictitious), the screen and the time; read
pnpm db:testandpost.sql; the fix is a new migration on the branch, tested on the sandbox first, never a grant typed on DEV; - the generated types differ: compare with the committed file; a difference means the applied schema is not the branch's: stop;
- the browser acceptance finds a regression: record it per line, device and language; the fix goes through the branch, its gates and a fresh verifier.
Object query (read only), for a failed or doubtful step 4:
cat > "$B/objects.sql" <<'SQL'
select 'tables ' || count(*) from pg_class c join pg_namespace n on n.oid = c.relnamespace
where (n.nspname, c.relname) in (('public', 'workspaces'), ('public', 'workspace_members'), ('public', 'clients'),
('public', 'animals'), ('public', 'animal_client_relations'), ('private', 'workspace_write_budgets'));
select 'types ' || count(*) from pg_type t join pg_namespace n on n.oid = t.typnamespace where n.nspname = 'public'
and t.typname in ('workspace_kind', 'workspace_member_role', 'client_kind', 'animal_sex', 'animal_relation_role');
select 'functions ' || count(*) from pg_proc p join pg_namespace n on n.oid = p.pronamespace
where (n.nspname = 'private' and p.proname in ('backfill_personal_workspaces', 'caller_workspace_ids',
'check_personal_workspace', 'create_personal_workspace', 'ensure_personal_workspace', 'writable_workspace_ids',
'consume_write_budget', 'fold_text', 'is_clean_text', 'lock_relation_for_write', 'parse_birth_year',
'parse_record_date', 'utc_tomorrow'))
or (n.nspname = 'public' and p.proname in ('create_client', 'update_client', 'set_client_archived', 'search_clients',
'client_archive_preview', 'create_animal', 'update_animal', 'set_animal_archived', 'search_animals',
'list_client_animals', 'create_animal_client_relation', 'end_animal_client_relation',
'update_animal_client_relation', 'void_animal_client_relation'));
select 'trigger ' || count(*) from pg_trigger where tgname = 'profiles_create_personal_workspace';
select 'constraint ' || count(*) from pg_constraint where conname = 'audit_logs_correlation_id_format';
select 'unaccent ' || count(*) from pg_extension where extname = 'unaccent';
select 'species ' || count(*) from public.species;
SQL
devsql postgres < "$B/objects.sql"
Going back is a decision of its own, never a reflex. No down migration
exists. The preferred way is forward: a new migration that corrects. A
restore of the step 3 dump into DEV is not rehearsed: it would replace the
postgres database the services use, and it loses every write made after
the backup (accounts created or confirmed, password and profile changes,
sessions, which would end on every device, audit rows, clients, animals and
relations created during the acceptance, assets and purge-queue changes);
a Storage file uploaded after the backup would remain without its row, and a
file deleted after it would come back as a row without bytes. Rehearse it on
the sandbox before relying on it.
3.4 Clients (V1-1.2, migration M2)
Contract: V1-1_OWNERS_ANIMALS.md §5.1, §8.2, §9, §10. public.clients
holds the clients of a workspace; the browser reads them through RLS (members
of the workspace, active profile, confirmed email) and writes them only
through create_client, update_client and set_client_archived (security
definer); search_clients (invoker, POST) searches one workspace. Refusals
are closed codes: 42501 not_found (also for a foreign id), P0001
invalid_input, archived, conflict, request_conflict, limit_reached,
rate_limited. Writes are counted per workspace in
private.workspace_write_budgets (UTC hour and day). M2 creates the
extension unaccent in extensions (available on the CLI image and in
PGlite 0.3.16, loaded by the harness). An account whose workspace holds a
client can no longer be deleted (D9a); in DEV, delete its clients first from
a privileged SQL session, or reset.
3.5 Animals and relations (V1-1.4, migration M3)
Contract: V1-1_OWNERS_ANIMALS.md §5.2 to §5.4, §8.3, §9, §10. M3 adds the
species rows (asinine to leporine and other; equine unchanged; a row
never means a sheet exists), public.animals and
public.animal_client_relations. The browser reads them through RLS and
writes them only through create_animal (with an optional first relation),
update_animal, set_animal_archived, create_animal_client_relation,
end_animal_client_relation, update_animal_client_relation and
void_animal_client_relation (security definer); search_animals,
list_client_animals and client_archive_preview read (invoker, POST), and
search_clients now returns current_animal_count. A relation joins an
animal and a client of one workspace (composite keys); it is current, ended
(ended_at, ended_on when known) or voided (entered by mistake, kept,
frozen, shown nowhere); it is never deleted. New refusal codes:
duplicate_relation, already_ended. Writes share the workspace budget of
M2. An account whose workspace holds an animal can no longer be deleted
either (D9a): in DEV, delete its relations, animals and clients first from a
privileged SQL session, in that order.
The browser reaches these tables through packages/supabase/src/{workspaces,clients,species,records}.ts
since V1-1.5 (contract §11.7) and {animals,relations}.ts since V1-1.6
(§11.8). The client and animal screens run against DEV only once the owner has
applied M1 to M3 with §3.3; until then they are developed and tested on the
isolated stack, and no DEV acceptance is recorded.
Browser acceptance on DEV (owner, after §3.3)
A checkpoint of its own: the Playwright suites never run against DEV, so this
is the owner's check, on pnpm dev (DEV origin), from a computer and a phone
of the private network, with fictitious accounts only: alice@faunosteo.local
(account A) and bob@faunosteo.local (account B), password of §3.2. No real
profile is changed for a test and nothing is created outside these two
workspaces; records carry names that say they are tests (« Recette Client
1 »). Each journey is done in French and in English, on a computer and at
375 px (a phone, or the browser's responsive mode); the main actions once more
with the keyboard only (Tab, Enter, Escape in the dialogs). Results are
recorded per line, device and language; a line not done is reported as not
done.
Sign-in and workspace:
- A signs in; the dashboard opens; Clients and Animaux open from the sidebar (computer) and from the drawer (phone);
- A signs out and signs in again: the records of the lines below are still there, after a reload too;
- signed out,
/app/clientsand a copied record URL lead to the sign-in page, then back to that page after sign-in.
Clients:
- create a client; find it by name, by city, then by phone digits; open its page;
- edit its details (address, phone), save, reload: the values are kept;
- archive it (the confirmation lists its sole-owned animals), find it under « Afficher les archivés », restore it.
Animals and relations:
- create an animal with no client: « Propriétaire non renseigné »;
- from a client's page, « Ajouter un animal »: owner role chosen, back on the client page with the animal;
- link an existing animal to a second client (co-ownership), then a structure as keeper and a third client as contact: each role shows on both pages; linking a client already linked in the same role is refused (no second current relation);
- end a relation with a date: « Ancien propriétaire » on that client's page, the relation under the former ones on the animal's page;
- correct the dates of a relation;
- remove a relation entered by mistake (« Erreur de saisie »): gone from both pages;
- « Autre » with « Furet »; a birth year alone, then a full date;
- archive the animal, find it with the archived toggle, restore it;
- from the phone: find an animal, open it, the owner's
tel:link opens the dialer; - Mon compte: A's photo still shows (no V1-1 change; checks the Storage path after the migrations).
Isolation:
- A signs out and B signs in on the same browser: no client and no animal of A in the lists or the searches, not even for an instant, nor with the back button;
- B opens a copied URL of A's client and of A's animal, then the same URL with one character changed: the same « introuvable » screen each time;
- optional:
owner@faunosteo.local(platform owner) sees none of A's records either (no staff access in V1-1).
Not covered: Safari iOS and Firefox unless such a device is used (the browser suites are Chromium only until V1-5); the refusal of an unconfirmed account (only with a fictitious unconfirmed account whose password is known, which means a sign-up on DEV, the owner's choice); the back and forward guard (accepted debt). This acceptance is no production security certification: RR-1 to RR-10 stay open.
Run of 2026-10-27 (the owner's results; the detailed records stay outside Git): the 19 lines validated in French one by one; English validated as a whole, four lines at 375 px recorded one by one; isolation A, B and owner checked. Two anomalies:
- RESP-01, sideways scroll at 375 px: the global scroll first seen came from
the browser zoom (the owner's measures without zoom:
/appand/app/clientsat 375 and 320 px, no overflow; no other page is declared conform by them). A real defect found apart, on the isolated stack: an account e-mail longer than the card (from about 31 characters at 320 px) widened the dashboard cards past a phone screen (implicit grid track); fixed, covered bye2e/shell/responsive.spec.ts. Owner counter-check on the fixed code: 375 px FR and EN, card readability, client list at 375 px, dashboard at 320 px, console measure at 320 px (0 px overflow): six PASS. Closed, fixed. - AVATAR-01, a new photo ends on « Le traitement prend plus de temps que prévu »: expected on DEV without a running worker (STORAGE.md §4 step 4); the pipeline completes once a worker runs (reproduced on the isolated stack, never on DEV). Owner decision: class A, an operating constraint that does not block V1-1; the DEV worker stays stopped on purpose, Alice's waiting upload is kept as is and may be processed whenever a worker is started later. No V1-1 regression shown. The reload gap found on the way (STORAGE.md §4 step 4) is ticket AVATAR-02, out of V1-1.
4. Writing a migration
pnpm supabase migration new <name>createssupabase/migrations/<timestamp>_<name>.sql.- A migration that adds a table adds, in the same file: RLS enable, explicit
revoke all ... from public, anon, authenticated(andservice_rolefor V1-1 record tables, contract N7) then the exact grants, separate SELECT / INSERT / UPDATE / DELETE policies, indexes, comments. - Add or extend a pgTAP file under
supabase/tests/: the allowed path and every denied path (other user, anon, direct write bypassing an RPC). Scope counts to the fixture ids: on the real stack the seed accounts exist, in PGlite they do not. pnpm --filter @faunosteo/supabase test:run(PGlite, seconds), thenpnpm db:reset && pnpm db:teston the real stack.pnpm db:typesand commitpackages/supabase/src/database.types.ts.- Never edit an applied migration. Corrections are new migrations.
Rules enforced by the schema-types test for every future public table:
RLS enabled, no using (true) / with check (true) policy for anon or
authenticated.
5. Tests
Two runners, one SQL source of truth (supabase/tests/*.sql, pgTAP):
| Command | Engine | Needs Docker | What runs |
|---|---|---|---|
pnpm --filter @faunosteo/supabase test:run |
PGlite (PostgreSQL 17 in WASM) | no | shim of the Supabase base schema (auth + storage tables) + all migrations + seed/** + every pgTAP file; schema ↔ types check; client factory; publish gate |
pnpm db:test |
supabase test db (pg_prove in the local stack) |
yes | the same pgTAP files against the real Supabase image |
pnpm test:run at the root runs the PGlite suite together with the app
suite. The shim (packages/supabase/tests/db/supabase-shim.sql) recreates
the API roles, auth.users, auth.uid() and the permissive default
privileges of Supabase so that a forgotten revoke fails a test.
Two parity traps met so far: counts must be scoped to fixture ids (the real
stack carries seed.sql rows), and order by on text must pin
collate "C" (the real image sorts with a locale collation, PGlite with C).
5.1 Isolated stack (pnpm db:sandbox)
A throwaway Supabase stack for migration steps, so the DEV stack is never
reset or migrated by a test (owner rule since V1-1.1). scripts/db-sandbox.sh
starts the images the CLI stack uses (PostgreSQL 17.11.0.004, GoTrue v2.197.0,
Storage v1.79.36, PostgREST v16.4, Mailpit v1.31.3, pinned in the script) on
the network faunosteo_dbsandbox_net, containers faunosteo_dbsandbox_*,
ports on 127.0.0.1 only and distinct from DEV (db 57432, auth 57999, rest
57300, storage 57500, mail 57025). Nothing is mounted and the images declare
no volume, so down removes the containers and the network by their literal
names and can never reach a supabase_*_faunosteo volume. The password, JWT
secret and keys are generated at each up into a 600 file under
~/.cache/faunosteo-db-sandbox (FAUNOSTEO_SANDBOX_DIR, refused inside the
repository), with 600 env files that hand them to the containers; deleted by
down, never printed, never on a command line (the CLI and psql read
PGPASSWORD from the environment). An incomplete secrets file stops the
script before any CLI call, and a wrong up mode or version (fourteen
digits) is refused before any container or secret exists. Works with rootless
Docker.
| Command | What it does |
|---|---|
pnpm db:sandbox up fresh |
every migration through supabase migration up --db-url, then the seeds of config.toml (as db reset) |
pnpm db:sandbox up upgrade <version> [<then>] |
the migrations before <version>, the seeds, two GoTrue accounts (confirmed, unconfirmed), then migration up: the path a DEV database takes; with <then>, the migrations before <then> first (the workspaces count is printed), then the rest |
pnpm db:sandbox test-db |
supabase test db --db-url: the pgTAP files on the real image (rows unchanged afterwards) |
pnpm db:sandbox types [--check] |
supabase gen types --db-url ... --schema public through Prettier, written or diffed with the committed file |
pnpm db:sandbox test |
packages/supabase/tests/sandbox/*.sandbox.ts: real GoTrue, PostgREST and Storage sessions, concurrent SQL sessions (two-session lock order, V1-2.1), timing at the caps, the PGlite shim against the image; re-runnable on one stack |
pnpm db:sandbox e2e-shell (= pnpm e2e:shell) |
the authenticated /app shell in Chromium (apps/app/e2e/shell, docs/development/BROWSER_TESTS.md): a loopback dev server whose Supabase calls are replayed on the sandbox, fictitious accounts per test |
pnpm db:sandbox verify-e2e |
V1-2.3: a gateway (nginx:1.31-alpine, sandbox network only, no host port) and the real verification worker faunosteo_worker:dev run --jobs sheet-verify against the sandbox, then packages/supabase/tests/sandbox/sheet-verification.sandbox.ts; both containers removed at the end, the service key passed through a 600 file deleted at start |
pnpm db:sandbox avatar-e2e |
V1-2.3: the same gateway, published on 127.0.0.1:57800 only and routing /auth/v1/ too, the real worker faunosteo_worker:dev run (default mode: avatars and sweep, 5 s sweep, no quarantine grace, 512 MiB as the compose worker) against the sandbox, then packages/supabase/tests/e2e/avatar-flow.e2e.ts through that port; the suite's psql calls go to the sandbox database container, which it requires by name; its two accounts are made per run with random passwords and deleted after; both containers removed at the end |
pnpm db:sandbox trial / trial-account <email> / trial-grant <email> <kind> |
V1-2.6 owner trial (docs/development/V1-2-6_OWNER_TRIAL.md): the app on http://$FAUNOSTEO_TRIAL_HOST:5193 (default the LAN address of this machine, 127.0.0.1 = host-only; loopback or private IPv4 of an interface only, Host header allowlisted) behind a trial gateway on 127.0.0.1:57801 (refuses to start when faunosteo_dbsandbox_trial_gateway exists), a confirmed account (hidden password prompt), a catalogue right (email and kind checked, passed as psql variables; editor and publisher also get the admin role) |
pnpm db:sandbox sql / down |
psql as postgres on stdin / remove the stack |
Every CLI call passes --db-url with sslmode=disable (the sandbox has no
TLS); none can fall back to the local DEV database. Differences from the CLI
stack: no Kong (GoTrue and PostgREST are called directly), no Studio, no
realtime or edge runtime, and the database init keeps only the role
passwords the services use.
How a test impersonates a user (same technique in both runners):
set local role authenticated;
select set_config('request.jwt.claims', '{"sub":"<user uuid>","role":"authenticated"}', true);
5.2 Dedicated trial stack (pnpm db:trial)
A persistent stack for the owner's manual trials, apart from DEV and from the sandbox, which the
automated suites keep using (docs/development/TRIAL_STACK.md, arbitrations A11 to A15). The script
scripts/db-trial.sh and its checks exist; the stack is created only on the owner's go, one operation
at a time, and never by a test.
| Container | Image | Port (127.0.0.1 only) | Data | Caps |
|---|---|---|---|---|
faunosteo_trial_db |
supabase/postgres:17.11.0.004 |
58432 | volume faunosteo_trial_db (PGDATA) |
1 GiB, shared_buffers 128 MB, 512 pids |
faunosteo_trial_auth |
supabase/gotrue:v2.197.0 |
58999 | 128 MiB, 256 pids | |
faunosteo_trial_rest |
supabase/postgrest:v16.4 |
58300 | 384 MiB, 256 pids | |
faunosteo_trial_storage |
supabase/storage-api:v1.79.36 |
58500 | volume faunosteo_trial_storage (/mnt) |
256 MiB, 256 pids |
faunosteo_trial_mail |
supabase/mailpit:v1.31.3 |
58025 | 64 MiB, 64 pids | |
faunosteo_trial_gateway |
nginx:1.31-alpine |
58801 | 64 MiB, 64 pids |
Every container: network faunosteo_trial_net, --cpus 1, restart policy no. The app (pnpm db:trial app)
runs Vite on http://<trial host>:5194 with a 768 MiB Node heap; the trial host is the LAN address by default,
FAUNOSTEO_TRIAL_HOST=127.0.0.1 for host-only, decided again at each run (nothing to edit after a change of the LAN
address). GoTrue's mail links carry the host of the moment GoTrue was started (create, restore,
rotate-secrets). GoTrue's sign-up is closed and the gateway answers 403 on /auth/v1/admin/.
State: ~/.cache/faunosteo-trial (or FAUNOSTEO_TRIAL_DIR), 700, refused inside any git work tree:
secrets.env and the services' env files (600, written whole then renamed), stack.conf (images),
migrations.sha256 (applied files and their hashes), backups/. Keys are signed for the issuer faunosteo-trial
and valid one year. Never run a bare docker inspect on a trial container: it prints the env-file values (database
password, JWT secret, keys); use docker inspect -f with a named non-secret field. Every command but plan needs a
reachable Docker daemon. account, grant, restore and destroy need a terminal: on a piped stdin they refuse
before any Docker call. Their typed confirmations: destroy asks faunosteo_trial, restore <backup> asks
restore <backup>, grant <email> <right> asks grant <right> for every right, catalogue rights included.
| Command | What it does |
|---|---|
plan |
read-only: the exact resources, each port free or taken, volume and backup places, memory, the refusals create would meet |
create |
after the Rust pipeline check of the equine seed (pnpm r5:import-equine --check): secrets, network, volumes, services, the branch's migrations, then equine_draft.sql only (never seed.sql); checks that the historical sheet is the only official one and that its draft equals document.json |
start / stop |
the same containers, data kept; start recreates a missing service container from stack.conf and secrets.env |
status |
containers, keys' expiry, migrations, equine sheet, accounts with roles, rights and verified factors, backups |
migrate |
a cold backup, then the branch's new migrations; refuses while the app runs (it stops the stack), an applied file edited since, or a new file older than the last applied; a migrate that stops part-way names its backup |
backup |
refuses while the app runs; stops the stack a few seconds (and refuses when a container still runs after the stop), refuses when free space is below 110 % of the volumes, archives each volume read-only through a --rm helper with no network, then lists it back and compares its entry count with the volume's (an archive cut short is refused), copies the state files, writes checksums last, restarts; a failed backup keeps no folder (a folder without SHA256SUMS is listed as incomplete) |
restore <backup> |
checksums, archives and the recorded database image checked first, typed confirmation restore <backup>, a backup of the current volumes, then the backup's volumes and secrets and the containers recreated; also rebuilds a destroyed stack |
rotate-keys |
a backup, then new anon and service keys signed with the same secret (one more year), Storage recreated and the gateway restarted; works on a stopped stack too (Storage is then recreated by the next start); the previous keys stay valid until their own expiry. Expired keys never block start or rotate-keys; app, assets and account refuse them and name rotate-keys; status reports them |
rotate-secrets |
a backup, then a new JWT secret, keys and role passwords set in one transaction as supabase_admin (values through psql variables, statement and error-statement logging off, a digest marker proving the commit); the new values wait in secrets.next.env (600) until the commit is known; every container recreated; access tokens signed with the old secret are refused, while refresh tokens and sessions stay valid (a signed-in browser keeps its pre-rotation access token and gets 401 from PostgREST, even after an app restart, until auth-js refreshes it near its expiry, up to 3600 s, or the user signs out and in again; the script revokes no session, which is an owner decision not taken; to end sessions after a rotation, sign in again, then sign out: see the paragraph below the table), account passwords and TOTP factors stay |
account <email> |
a confirmed account through GoTrue's admin API on 58999, password typed twice at a hidden prompt (12 characters min.); the address is lowercased |
grant <email> <right> |
owner / admin: the operator path of the V1-2 contract §5.7.1 (verified TOTP factors shown, two for owner, one for admin; typed confirmation grant <right>; private.assign_platform_role(null, …) and an audit row); a platform role already active is refused. Catalogue rights mirror public.grant_catalog_right in operator SQL (the RPC needs an owner session at aal2): a confirmed active account; a verified TOTP factor for every right but catalog_reader; admin held first for sheet_editor and sheet_publisher (the refusal names pnpm db:trial grant <email> admin); one active row per right; typed confirmation; the insert and an audit row catalog_grant.operator_bootstrap in one transaction; no platform role is granted implicitly. granted_by is the holder's own id (the column is NOT NULL and the operator has no profile), so the rights screen shows a self-grant; the audit row names the operator. A catalogue right already active exits 0 with nothing written |
assets |
the equine drawings with pnpm assets:upload through the gateway (no overwrite, rerun = nothing to do), then every row confirmed with its object |
app |
the app, foreground; Ctrl+C stops Vite and waits for it, so 5194 is free on return; the stack keeps running |
destroy |
inventory, typed confirmation faunosteo_trial, a backup of every trial volume checked as a restorable backup, then the containers, the network, the volumes and the live state files; backups kept. When that backup cannot be made or fails the check, it removes nothing and names the backup kept (or none), the exact docker rm, docker network rm and docker volume rm lines for what exists, then pnpm db:trial destroy to clear the state files; the script never runs those lines. With no trial resource left, it removes the live state files only |
First use, after the owner's go: plan, create, assets, then the two accounts.
- Owner/admin account, for the administration trials:
account <email>,app, sign in, enrol two TOTP factors in Mon compte > Sécurité, thengrant <email> owner,grant <email> admin, thengrant <email> sheet_editor,sheet_reviewerandsheet_publisher; sign in again and pass the step-up. - Practitioner account:
account <email>,app, enrol one TOTP factor,grant <email> admin, thengrant <email> sheet_editor(refused beforeadmin). This is a trial account, not a production practitioner profile: the technicaladminrole is there only because rule O1 activatessheet_editorfor a platform admin (hence the grant askingadminfirst, asgrant_catalog_rightdoes); O1 is unchanged. A production practitioner holds neitheradminnor a catalogue editing right by default.
Rotation and recovery: rotate-keys before the keys expire (status shows the date); rotate-secrets refuses the access tokens signed with the old secret (refresh tokens and sessions survive it).
Ending a session after a rotation: signing out in the app right after rotate-secrets does not end it. The app calls
auth.signOut() with auth-js's default scope global, which revokes every session of the account, but GoTrue refuses
the /logout call of a token signed with the old secret, and auth-js 2.117.3 ignores that 401 or 403 and only drops the
local session: the server session and its refresh token stay valid (a token refreshed with the new secret lasts up to
jwt_exp, 3600 s). To end the sessions, sign in again (the token is then signed with the new secret), then sign out:
that revokes all sessions of the account, other browsers included. The script revokes nothing itself (an owner
decision not taken).
If rotate-secrets loses psql, the marker the transaction wrote tells whether it committed and only the matching
file stays secrets.env (password probes prove nothing here: loopback connections are trust in this image); if a
rotation stops after the commit, pnpm db:trial start recreates the containers; if it stops before the commit is
known, the script names the backup taken just before, and restore <that backup> returns to the previous secrets
and data. Every backup holds the secrets.env of its time: restoring a backup taken before a rotate-secrets
brings the old secrets back (the script says so before the confirmation). A restore needs the same Postgres image
as the backup (recorded in its stack.conf).
No automated suite reaches this stack: db-trial.sh refuses a shell holding SANDBOX_* variables; the sandbox
suites (packages/supabase/tests/sandbox/target.ts, imported by stack.ts and the shell Playwright config)
accept only the sandbox's container, loopback ports and key issuer db-sandbox; the LAN Playwright config accepts
port 5193 only; e2e:avatar and e2e:sheet-assets require the sandbox URL and issuer; db-sandbox.sh refuses a
state folder holding the trial marker; every secret differs from the sandbox's. bash scripts/db-trial-check.sh,
run by hand, proves these refusals and the script's Docker calls with stubs, no daemon reached; it is not a
pnpm check:all stage because it reads the host's real ports and memory, and it refuses to run (exit 2) while the
trial stack or app holds a port.
Identification of the target (owner 2026-11-10): the trial dev server (apps/app/vite/trial.config.ts, shared with
pnpm db:sandbox trial) serves exactly one of the two stacks listed in apps/app/vite/trial-targets.ts. The exact
gateway string picks it; the app port (with --strictPort), VITE_SUPABASE_URL and the anon key's role and
iss must then match that same stack, and the db:trial gateway must answer GET /__faunosteo_stack with
faunosteo_trial; anything else stops Vite before it listens, with no fallback to the sandbox, DEV or another
local service. db-trial.sh checks, before acting: for app and assets, the gateway container published exactly
on 127.0.0.1:58801 and its identity answer; for account, GoTrue published exactly on 127.0.0.1:58999; for
migrate, Postgres exactly on 127.0.0.1:58432; and every key it hands over (Vite, the upload tool, GoTrue's admin
API, Storage's env) must be signed with this stack's JWT secret for the issuer faunosteo-trial and the expected
role. The identity answer is a static name, not authentication: the exact port, the container checks and the key
signatures carry the rest.
The absent sentences db-trial.sh recognises (absent_answer) were measured on Docker 29.9.3: a Docker upgrade
that rewords them makes every command refuse with "not the daemon's own 'absent' sentence", until absent_answer is
updated.
Limits: one host; no verification worker on this stack, so a draft submitted for verification stays waiting, and no avatar worker either, so a profile photo stays pending; the
historical sheet allows one working copy at a time; Mailpit is reachable on the host only; backups sit on the same
disk (copy them elsewhere by hand). app takes no lock: stop, backup, migrate, destroy, restore, rotate-keys or
rotate-secrets run in another terminal in the second before Vite listens on 5194 goes ahead. After rotate-keys
the app keeps the previous anon key until its expiry; after rotate-secrets a signed-in browser keeps its pre-rotation access
token and PostgREST answers 401 to it, even after the app restarts, until auth-js refreshes it near its expiry (up to
3600 s) or the user signs out and in again; after stop or destroy it proxies to whatever later listens on 127.0.0.1:58801. Launch no
other db:trial command while app is starting.
Proven only at the first create (the stub harness cannot prove them): the database init under a volume, the
supabase_admin password used by rotate-secrets, the pids caps, tar and du in the database image (checked by
create before anything is made), the image's default logging settings, the backup helper's tar restoring
ownership in the database image, set local pgaudit.log and set local pg_stat_statements.track_utility accepted by
the image, the init's session-level set lines (log_statement, log_min_duration_statement,
log_min_error_statement, pgaudit.log, pg_stat_statements.track_utility) accepted by it, whether the image's own
init runs an ALTER USER supabase_admin ... PASSWORD before the schema file (where the init's settings cannot help; the first create, 2026-11-11, proved it does: migrate.sh runs the statement on the init temp server and the image's postgresql.conf sets log_statement = 'ddl', so the statement with the password reached the db container's log; create now starts the server with -c log_statement=none, which the entrypoint hands to the init temp server and to the final server; not closed, named only: a failing init statement is still logged at the image's log_min_error_statement = error, migrate.sh puts the password on psql's command line inside the container during init, and db-sandbox.sh starts its database the same way without the flag; the fix is proven only by a create on a fresh volume with the container log scanned for the password, counts only, with the scan commands below), loopback trust in its pg_hba.conf
(the rotation's commit marker relies on it), psql's echo of an error statement (a failing rotation statement must not print a
secret), and the grant and factor SQL, which the harness checks as text and which first
runs at the first real grant. Checks to make right after that first create, printing no value, only counts.
<secrets.env> is $FAUNOSTEO_TRIAL_DIR/secrets.env (default ~/.cache/faunosteo-trial/secrets.env), <db> is
faunosteo_trial_db:
# roles holding a password: exactly the five the script sets (postgres is not a superuser in this image)
docker exec <db> psql -X -h 127.0.0.1 -U supabase_admin -d postgres -At \
-c 'select rolname from pg_authid where rolpassword is not null order by 1'
# the server log: where it goes is read at that moment, not assumed
docker exec <db> psql -X -h 127.0.0.1 -U supabase_admin -d postgres -At \
-c 'show logging_collector; show log_directory; show log_destination; show data_directory;'
# secrets in the container log, in the in-container log files and in pg_stat_statements (counts only, the
# secret never on argv: grep reads the pattern from a process substitution)
for key in TR_PW TR_JWT_SECRET; do
printf '%s in docker logs: ' "$key"
docker logs <db> 2>&1 | grep -cFf <(sed -n "s/^$key=//p" <secrets.env>)
printf '%s in the log files: ' "$key"
docker exec <db> sh -c 'cat "$1"/*' _ <log_path> | grep -cFf <(sed -n "s/^$key=//p" <secrets.env>)
printf '%s in pg_stat_statements: ' "$key"
docker exec <db> psql -X -h 127.0.0.1 -U supabase_admin -d postgres -At \
-c 'select query from extensions.pg_stat_statements' | grep -cFf <(sed -n "s/^$key=//p" <secrets.env>)
done
<log_path> is the log_directory printed above when it is absolute, else <data_directory>/<log_directory>
(a relative log_directory is relative to data_directory, both printed by the show line). With
logging_collector off there are no log files: the container log is the only place and the second count does not
apply (a cat that prints "No such file" means a wrong path: that count is meaningless). pg_stat_statements is
the store track_utility = off protects: the image's ACL on pg_stat_statements_reset sits in schema extensions
(see the dump check above), so the view is extensions.pg_stat_statements; if the query answers "relation does not
exist", look up the schema with select extnamespace::regnamespace from pg_extension where extname = 'pg_stat_statements' and use it.
Every count must be 0 and the role list must be the five roles (supabase_admin, postgres, authenticator,
supabase_auth_admin, supabase_storage_admin). A count above 0: stop, tell the owner, do not use the stack for
anything else. After the owner's go: when pg_stat_statements holds a value, run select pg_stat_statements_reset();
as supabase_admin (same psql line), then pnpm db:trial rotate-secrets, which renews both scanned secrets
(rotate-keys renews only the two API keys); scan again afterwards. What the old values leave behind depends on
where the log is. A log inside PGDATA (the log files under data_directory) lives in the volume: it stays until
the volume is recreated and is in every later backup's db.tar, the pre-rotate one included. A log in the
container's own filesystem is gone once the container is recreated, which rotate-secrets does. Existing backups
keep any text they already hold: scan the backups too, and delete by hand the ones that carry a secret you no longer
want (the script prunes nothing).
6. Generated types
pnpm db:types # regenerate packages/supabase/src/database.types.ts (local stack running)
pnpm db:types:check # CI: fail on any diff
The CLI emits unformatted TypeScript; both scripts pipe it through Prettier (repo config) so the committed file is stable byte for byte. The schema-types test checks tables, columns, enum labels and RPC argument names against the migrated schema, so a stale file fails before CI.
7. First platform owner in a real environment
The RPCs refuse self-grants and need an existing owner or admin. The first
owner is created once, from a privileged SQL session (service connection,
psql on the host), never from the browser:
select private.assign_platform_role(null, '<auth.users.id of the owner>', 'owner');
The call is audited with a null actor. Afterwards every role change goes
through grant_platform_role / revoke_platform_role.
8. Manual validation (Definition of Done: "an authenticated request hits a protected table safely")
URL=http://127.0.0.1:54321
ANON=<anon key printed by pnpm db:start>
token() { # $1 = email
curl -s -X POST "$URL/auth/v1/token?grant_type=password" \
-H "apikey: $ANON" -H "Content-Type: application/json" \
-d "{\"email\":\"$1\",\"password\":\"LocalDev-only-password-1\"}" | jq -r .access_token
}
ALICE=$(token alice@faunosteo.local); OWNER=$(token owner@faunosteo.local)
ALICE_ID=10000000-0000-0000-0000-00000000000a
# 1. alice reads exactly one profile (hers)
curl -s "$URL/rest/v1/profiles?select=id,display_name" -H "apikey: $ANON" -H "Authorization: Bearer $ALICE"
# 2. anon reads nothing (permission denied)
curl -s "$URL/rest/v1/profiles" -H "apikey: $ANON"
# 3. alice cannot deactivate herself (column privilege) nor self-grant admin
curl -s -X PATCH "$URL/rest/v1/profiles?id=eq.$ALICE_ID" -H "apikey: $ANON" \
-H "Authorization: Bearer $ALICE" -H "Content-Type: application/json" -d '{"active":false}'
curl -s -X POST "$URL/rest/v1/rpc/grant_platform_role" -H "apikey: $ANON" \
-H "Authorization: Bearer $ALICE" -H "Content-Type: application/json" \
-d "{\"target_user_id\":\"$ALICE_ID\",\"role\":\"admin\"}"
# 4. the owner grants moderator to alice; the grant is audited
curl -s -X POST "$URL/rest/v1/rpc/grant_platform_role" -H "apikey: $ANON" \
-H "Authorization: Bearer $OWNER" -H "Content-Type: application/json" \
-d "{\"target_user_id\":\"$ALICE_ID\",\"role\":\"moderator\"}"
psql postgresql://postgres:postgres@127.0.0.1:54322/postgres \
-c "select actor_user_id, action, target_user_id, metadata from audit.audit_logs order by occurred_at"
Expected: 1 → one row; 2 → 42501; 3 → both 42501; 4 → a uuid, then two
audit rows (bootstrap with null actor, grant by the owner).
9. Wiring the app
Development: pnpm dev (scripts/dev.sh) sets VITE_SUPABASE_URL to the DEV
origin (http://<FAUNOSTEO_DEV_HOST>:54321, the LAN address of the machine:
ENVIRONMENTS.md §2) and the anon key from supabase status; the app is
served on http://<host>:5173 for every device of the private network.
pnpm dev:offline runs the local tool without backend. Docker deployments use
.env:
APP_SUPABASE_URL=<Supabase origin reachable by the browser>
APP_SUPABASE_ANON_KEY=<anon key> # public by design
Both empty keeps the app in its local-only mode:
getSupabaseClient() returns null, the auth screens show an "unavailable"
notice and /app redirects to /login. APP_AUTH_PROVIDERS /
VITE_AUTH_PROVIDERS (google,apple) lists the OAuth buttons to show.
Sign-in, sign-up, email confirmation, password recovery, the /app shell
and Mon compte are documented in IDENTITY.md (routes, session handling,
provider setup, MFA gap, manual validation).
10. Self-hosted parity
The local stack runs the same images as a self-hosted deployment
(supabase/docker compose), so migrations, policies and tests transfer
unchanged. The self-hosted runbook (VPS sizing, secrets, backups, TLS) is a
separate deliverable once the hosting question (OPEN_QUESTIONS Q2) is
settled; it reuses pnpm supabase db push --db-url <...> for migrations.