Every `just deploy` now takes a pg_dumpall to
/opt/colectivo/backups/predeploy-<ts>-<sha>.sql.gz BEFORE rebuilding the
app or applying migrations, aborts the deploy if the dump is < 1 KiB,
and (on success) appends `<iso-ts> \t <sha> \t <backup-file>` to
/opt/colectivo/.deploys.log on ambrosio. Keeps the newest 10 backups
(prunes older predeploy-* files).
New `infra/scripts/rollback-erosi.sh` reads .deploys.log and pairs a
code rollback with a DB restore atomically:
just rollback-list # show recent deploys
just rollback # roll back to N-1
just rollback-to <sha> # roll back to a specific deploy
just rollback-code # roll back code only, keep current DB
Rollback safety:
- 5-second abort window.
- Verifies the target SHA exists locally + the backup is still on prod
(warns if it's been pruned past the 10-deploy window).
- Uses a temporary git worktree so the user's working tree isn't
disturbed.
- Stops app/auth/rest/realtime/storage before the gunzip|psql restore.
- Rebuilds the app image at the rolled-back SHA with the same GIT_SHA
build-arg path the deploy uses (so __APP_COMMIT__ in the bundle
matches the running code).
- Does NOT write a new .deploys.log entry — rollback is intentionally
not a deploy event; the next `just deploy` is from current HEAD.
- Storage volume (/var/lib/storage user uploads) is NOT rolled back.
Justfile `deploy` recipe repointed from the stale `deploy.sh` (which
referenced GHCR pull) to `deploy-erosi.sh` (the active prod path).
New project-scoped skill: `.claude/skills/deploy/SKILL.md` documents the
flow + safety boundaries for Claude-assisted invocations. `.gitignore`
keeps `.claude/settings.local.json` ignored but tracks `.claude/skills/`.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
`apps/web/vite.config.ts` now injects three build-time constants via
`define`: `__APP_VERSION__` (root package.json), `__APP_COMMIT__`
(GIT_SHA env > local `git rev-parse` > 'dev'), `__BUILD_TIME__` (ISO
timestamp at build). They're declared in `apps/web/src/global.d.ts`
and re-exported through `$lib/version` — components consume the named
exports rather than the magic globals so an accidental tree-shake
would break the test, not the page silently.
`/settings` gains an "About" section at the bottom: a small chip
`v{version} · {commit} · {date}` in muted text. Not interactive,
purely diagnostic — useful when a user reports a bug from a stale
PWA install.
Deploy pipeline:
- `apps/web/Dockerfile` accepts a `GIT_SHA` build arg and exports
it as an env var for the SvelteKit build step.
- `infra/docker-compose.erosi.yml` forwards `${GIT_SHA:-dev}` into
the build args.
- `infra/scripts/deploy-erosi.sh` captures `git rev-parse --short
HEAD` locally (the deploy host has .git; the remote doesn't —
rsync excludes it) and forwards it via `ssh ... env GIT_SHA=…`
so the remote `docker compose build` sees it.
V-01/V-02 vitest: stub the three globals via `vi.stubGlobal` (vitest
doesn't run through the main vite.config so `define` isn't applied),
assert the named exports thread through and never collapse to
undefined or the 'dev' sentinel under a real build.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Migration 024 introduces the orthogonal global role and the append-only
audit log. server_admins is a separate table (not a column on users) so
that promote/revoke don't require a JWT re-issue; is_server_admin() is a
STABLE SECURITY DEFINER helper matching the shape of is_admin/is_member.
admin_actions has only a SELECT policy (admin-only) — no INSERT/UPDATE/
DELETE policies, so only the SECURITY DEFINER RPCs in migration 025 can
write to it. actor_id FK uses RESTRICT so the audit trail outlives
admins.
Bootstrap via infra/db-init/10-server-admin-seed.sh on first volume
init (reads SERVER_ADMIN_EMAIL); supabase/seed.sql also pre-seeds Ana
for dev/test because docker-entrypoint-initdb.d does not re-fire on
long-lived dev volumes. Documented in .env.erosi.example.
20 new pgTAP assertions cover schema shape, RLS posture, FK behaviour,
and the SECURITY DEFINER + STABLE attributes on is_server_admin().
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The stack no longer talks to any specific external proxy. Removed traefik.*
labels and the shared `traefik` external network from kong + app, and dropped
TRAEFIK_NETWORK / TRAEFIK_ENTRYPOINT / TRAEFIK_CERTRESOLVER from the .env
template and the deploy-script bootstrap.
Internal edge is a `caddy:2-alpine` container on the `colectivo` network,
publishing host port 3000. `infra/Caddyfile.erosi` does the path split:
/rest|/auth|/realtime|/storage|/graphql|/pg -> kong:8000, everything else
-> app:3000. Global block sets `auto_https off` + `admin off` so :3000 is
the only listener.
Whatever proxy fronts the host (NetBird's Traefik, nginx, anything) just
terminates TLS for erosi.limonia.net and forwards plain HTTP to
ambrosio:3000. Off-stack and out of this repo.
Docs (CLAUDE.md + docs/deployment.md) updated to describe the two-layer
edge. Includes the gotcha that bind-mounted file edits (Caddyfile) need
`docker compose restart caddy` since `up -d` won't recreate the container
on file-only changes.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Ambrosio's TLS is now handled by the Traefik that NetBird's self-hosted
installer deploys. kong + app attach to that Traefik network and expose
themselves via Docker-provider labels (erosi-kong priority 100 for
Supabase API paths, erosi-app priority 1 for everything else on
erosi.limonia.net). No container publishes host ports; host Caddy is
out of the deploy path permanently.
Keycloak is no longer bundled — PUBLIC_KEYCLOAK_URL points at an
external IdP. The deploy script writes .env with FILL_IN_* placeholders
for PUBLIC_KEYCLOAK_URL, KEYCLOAK_CLIENT_SECRET, and TRAEFIK_NETWORK,
and fails fast until they are filled.
New domain: erosi.limonia.net. realm-export.erosi.json is retained
as reference for the external Keycloak operator; it is no longer
imported by this stack.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Symptom observed twice (migrations 012 and 013): deploy-erosi.sh ran the
migration loop's first iteration (the `CREATE TABLE IF NOT EXISTS
_applied_migrations` step) then silently skipped every file in
supabase/migrations/*.sql without printing a single skip/apply line. Had
to `psql -f <new-migration>` by hand on ambrosio after every deploy.
Root cause: the deploy uses `ssh host bash -s << 'REMOTE' ... REMOTE` so
the outer bash on the remote reads its own script from stdin. Any
`docker compose exec -T db psql ...` inside that script — particularly
inside a command substitution like `already=$(docker compose exec -T ...)`
— inherits that stdin and consumes it, eating the rest of the heredoc.
Fix: pass `</dev/null` (or the migration file for the actual `apply`
step) to every `docker compose exec -T` call so docker gets an empty /
file-scoped stdin instead of the parent's heredoc.
Verified on ambrosio: deploy now prints all 13 `skip ... (applied)` lines
as expected, and will apply new migrations going forward without manual
intervention.
- CLAUDE.md: gotcha #20 documenting the pattern; applies to any
heredoc-delivered remote script using `docker compose exec -T` or
`kubectl exec`.
Self-contained compose for a single-VPS deploy behind the host's Caddy.
App + Supabase API share erosi.oier.ovh; Keycloak on auth.oier.ovh.
- infra/docker-compose.erosi.yml — full stack bound to 127.0.0.1 only
(db, auth, rest, realtime, storage, imgproxy, kong, meta, keycloak, app)
- infra/caddy/erosi.Caddyfile — host-Caddy snippet with TLS + path routing
(/auth /rest /storage /realtime /graphql → kong, rest → app)
- infra/scripts/deploy-erosi.sh — rsync + first-run secret generation +
build + migrations + Caddy snippet install
- keycloak/realm-export.erosi.json — prod redirect URIs, registration on,
client secret as __KEYCLOAK_CLIENT_SECRET__ placeholder (deploy sub'd)
- .env.erosi.example — env template with placeholders
Supporting fixes to make the deploy work on a clean Supabase-postgres volume:
- infra/db-init/00-role-passwords.sh — rewrote as idempotent bootstrap:
CREATE the Supabase service roles if missing, set passwords on pre-existing
ones, enable pg_cron/pgcrypto/uuid-ossp, create auth/storage/graphql_public/
_realtime/realtime schemas, create empty supabase_realtime publication,
set per-role search_path (auth→auth, storage→storage). Old version only
ALTERed passwords and relied on the roles already existing — worked for a
grandfathered dev volume, failed on a fresh prod init.
- infra/db-init/00-role-passwords.sql — removed (folded into .sh).
- apps/web/vite.config.ts — PWA strategy generateSW (was injectManifest).
The plugin's injectManifest hardcodes the SW source filename to
service-worker.js and ignores the filename: 'sw.ts' override, making the
production build fail. generateSW's auto-generated precache-only SW is
functionally equivalent to our 5-line src/sw.ts.
Tested end-to-end on ambrosio: all 9 containers up, 11 migrations applied,
https://erosi.oier.ovh returns 200, Keycloak OIDC discovery serves the
correct issuer, /auth/v1/settings lists Keycloak as the sole external
provider.
Closes the deferrals from Fase 4 that gated a public deploy of the MVP.
PWA install
- Custom SW at apps/web/src/sw.ts (named sw.ts, not service-worker.ts, so
SvelteKit doesn't intercept it and block Workbox imports).
- vite.config.ts switched to injectManifest strategy with precache of the
built shell; no runtime caching of Supabase (offline writes stay on the
existing $lib/sync pending_ops queue).
- Root +layout.svelte onMount() calls registerSW({ immediate: true }) from
virtual:pwa-register — the plugin does not auto-register.
- Web manifest with 192/512/512-maskable icons + iOS meta tags
(apple-mobile-web-app-capable, apple-touch-icon, etc.) in app.html.
- Placeholder icons generated via ImageMagick; replace before public launch
(noted in apps/web/static/icons/README.md).
Kong rate-limit
- /rest/v1/collective_invitations POST: 20/hour per Authorization header.
Anti-spam on invitation creation is the only rate-limit that survived —
auth rate-limits were dropped during execution (see plan §6.3): Keycloak
already provides bruteForceProtected=true on the realm, and stacking a
Kong limit on /auth/v1/token throttled legitimate PKCE code exchanges
during E2E.
- Added 'rate-limiting' to KONG_PLUGINS in docker-compose.dev.yml.
- Discovered: strip_path=true on a full-table path sends "/" upstream and
PostgREST rejects POST to root with PGRST117. Route carries a
request-transformer plugin that rewrites the URI back.
JWT rotation + prod template
- infra/scripts/rotate-jwt.sh signs anon + service_role JWTs with a fresh
48-byte secret via openssl. Prints three values for the operator to
paste; does not touch files.
- Dev secret rotated as a dry-run. Updated both .env (local, gitignored)
and apps/web/.env.development. Existing sessions are invalidated — all
users must re-login once.
- .env.production.example committed with placeholders + rotation notes.
Lighthouse + RLS audit tooling
- `just lighthouse` boots the prod build + Supabase stack and runs
lighthouse CLI against PWA/A11y/Best-Practices.
- `just rls-audit` runs infra/scripts/rls-audit.sh which signs an Eva JWT
(non-member) and GETs 8 REST endpoints — all must return []. Complements
the Vitest rls-audit.test.ts suite.
Tests
- 236 green: 34 pgTAP + 140 Vitest integration + 15 unit + 46 Playwright.
- 1 rate-limit test gated behind RUN_RATE_LIMIT_TESTS=1 (Kong counters are
shared state); run via `just test-rate-limit` which force-recreates Kong
first to reset counters.
- 3 skipped in `just test-all` (2 Realtime presence upstream bug, 1 gated
rate-limit).
Deliberately out of scope: push notifications, CI ephemeral Docker, final
production icons, runtime caching of Supabase (SyncBanner + pending_ops
already covers offline).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The dev stack previously mixed two hostnames (`localhost` in some places,
`keycloak` in others). That only worked because the laptop's /etc/hosts
maps `keycloak` to 127.0.0.1. Phones, tablets, or a second laptop had no
such mapping, so the OAuth flow broke the moment you tried to log in from
any non-laptop device.
Switched the SPA, Kong, GoTrue, and Keycloak client to all agree on a
single external host — the laptop's LAN IP (192.168.1.167) — so the full
Keycloak → GoTrue → Supabase → SvelteKit round-trip works from any device
on the same Wi-Fi.
Changes
.env (root, ungitignored): PUBLIC_SUPABASE_URL / PUBLIC_KEYCLOAK_URL /
PUBLIC_APP_URL now use the LAN IP (updated locally; tracked via
apps/web/.env.development below).
apps/web/.env.development: mirror the LAN IP so `$env/static/public`
resolves the same on laptop and phone.
apps/web/vite.config.ts: `server.host: true` — bind Vite to 0.0.0.0.
apps/web/playwright.config.ts: `baseURL: PUBLIC_APP_URL ?? localhost`
so E2E uses the same origin the stack is configured with.
apps/web/tests/fixtures/login.ts:
- wait for the collective button in the sidebar to render before
returning (LAN-IP latency surfaced a pre-existing race in
`handleCreate` where Enter fired before `$currentCollective`
hydrated, silently no-op'ing).
- derive expected app origin from PUBLIC_APP_URL.
apps/web/tests/e2e/items.test.ts: scope D-04 delete assertion to the
`[role="listitem"]` locator — with undoQueue enabled the toast text
"Deleted <name>" also matches `getByText(itemName)` and shadowed the
original check.
infra/docker-compose.dev.yml:
- GOTRUE_EXTERNAL_KEYCLOAK_URL and REDIRECT_URI now read from
PUBLIC_KEYCLOAK_URL / PUBLIC_SUPABASE_URL so they follow the same
host as the rest of the stack.
- NEW: GOTRUE_URI_ALLOW_LIST with `/auth/callback` on both the LAN
IP and localhost. Without an explicit allow-list GoTrue rejected
`redirect_to=.../auth/callback` and fell back to SITE_URL root,
stranding `?code=` at `/` and re-triggering signIn → infinite loop.
keycloak/realm-export.json: `colectivo-web` client gains LAN-IP
redirectUris and webOrigins (alongside the existing localhost
entries). Persisted + live-applied via admin API.
.gitignore: add .claude/ (per-project scheduler runtime state).
Verification
just test-all → 224 passed, 4 skipped:
34 pgTAP
140 Vitest integration + 2 skipped (Realtime presence, upstream bug)
11 Vitest unit
39 Playwright + 2 skipped (mobile-swipe touch, WebKit-only)
Phone sanity check: open http://192.168.1.167:5173 on a LAN device,
log in as a seed user, full RLS-respecting session.
Caveats
LAN IP (192.168.1.167) is DHCP-assigned — if it rotates, rerun the
Keycloak admin API update (realm-export.json needs a new entry) and
update the three PUBLIC_* URLs. Consider a Tailscale magic-DNS name
as a stable replacement for the prod-deploy sprint.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Extend the 2b.0 red-phase coverage: presence, RLS isolation, offline-queue
contract, and Playwright scaffolds for the shopping-session UX. Only the
presence test is blocked on an upstream Realtime bug (documented below); all
other Realtime paths are proven end-to-end.
Tests
- realtime-postgres-changes.test.ts: now uses afterEach socket disconnect
(Vitest singleFork kept Phoenix sockets alive between files, leaking state)
- realtime-isolation.test.ts (new): R-I-01 David (guest, same collective)
receives events; R-I-02 Eva (non-member) receives NONE — RLS is enforced
server-side per subscribed JWT
- realtime-presence.test.ts (new, describe.skip): two-client roster + leave
assertions ready, gated on upstream bug
- sync-queue.test.ts (new, describe.skip): 7 placeholders pinning the
apps/web/src/lib/sync/queue.ts contract (Q-01..Q-04 enqueue / in-order
flush / retry; F-01..F-03 online-event flush + last-write-wins)
- apps/web/tests/e2e/{session,realtime,offline}.test.ts (new, describe.skip):
S-01..S-03, R-E-01, R-E-02, O-01, O-02 pinning the Modo Compra UI contract
- vitest.config.ts: fileParallelism=false, drop singleFork so each file gets
a fresh worker fork — prevents RealtimeClient singleton leakage
Infra
- db-init/00-role-passwords.sh: pre-create `realtime` schema with
AUTHORIZATION supabase_admin. Realtime's per-tenant migrator creates tables
INSIDE the schema but does not create the schema itself; without this the
first tenant connect fails with "schema realtime does not exist"
- docker-compose.dev.yml: adopt the upstream supabase/supabase Realtime env
shape — SEED_SELF_HOST=true + RUN_JANITOR=true + RLIMIT_NOFILE=10000 +
drop the custom `command: eval seeds` (that bypasses the normal supervisor
startup and was observed to cause GenServer crashes under load).
Version pinned to v2.76.5 to match the official docker-compose.
Known upstream bug
- Realtime v2.76.5 and v2.83.0 both crash on presence_diff:
`(UndefinedFunctionError) RealtimeChannel.handle_out/3 is undefined`.
postgres_changes + isolation unaffected (they don't traverse handle_out).
Presence tests stay `describe.skip` until a fixed upstream is released.
Totals: 59 Vitest passed (+ 9 skipped awaiting implementation/upstream),
16 pgTAP, 15 Playwright (+ 7 E2E scaffolded for 2b UI) = 90 green.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Enable postgres_changes broadcasts on shopping_items / shopping_lists and prove
the path end-to-end with Vitest Realtime tests and a pgTAP configuration check.
Infra changes that were silently blocking events
- Migration 007: shopping_items + shopping_lists added to supabase_realtime
publication with REPLICA IDENTITY FULL (UPDATE/DELETE payloads need the full
row so list_id-based filters work and RLS can evaluate DELETE against OLD)
- Realtime service: DB_USER=supabase_admin (superuser; supabase_replication_admin
lacks CREATE on the `realtime` schema that the service auto-creates per tenant)
- Realtime service: SELF_HOST_TENANT_NAME=realtime so the seed tenant name
matches the default supabase-js resolves for localhost URLs (was realtime-dev
→ TenantNotFound)
Tests
- pgTAP 004_realtime_publication.sql — P-01..P-04: publication membership and
REPLICA IDENTITY FULL for both shopping tables
- Vitest realtime-postgres-changes.test.ts — R-01 INSERT broadcast,
R-02 UPDATE carries full row, R-03 list_id filter isolates events
- test-utils realtime-helpers.ts — subscribePostgresChanges() returns a
waitFor(predicate, timeout) helper that captures the SUBSCRIBED handshake
before returning so mutations issued right after are not lost
- createClientAs now calls realtime.setAuth(token) so the server evaluates
RLS against the test user's JWT
- Justfile test-db now globs supabase/tests/*.sql so new pgTAP files are picked
up automatically
Totals: 54→57 Vitest, 12→16 pgTAP, 15 Playwright = 88 tests green.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Auth: `getSession()` in a SvelteKit load function races with Supabase's async
localStorage init and can return null for a valid session, triggering a spurious
login() redirect. Moved auth redirect to (app)/+layout.svelte via $effect, which
correctly waits for onAuthStateChange to fire. Also fixed collective data never
loading on page refresh (INITIAL_SESSION event, not SIGNED_IN).
PWA: SvelteKit blocks Workbox imports in src/service-worker.ts, preventing
@vite-pwa/sveltekit from finding the compiled SW output. Switched to generateSW
strategy (plugin auto-generates the SW). Custom SW deferred to Fase 2b using
src/sw.ts (a filename SvelteKit does not intercept).
Docs: added gotchas 6–8 to CLAUDE.md covering both root causes so they are not
reintroduced. Updated plan/fase-2b with the sw.ts workaround note.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
