Found by running the delivery flow against the actual built container rather
than a local dev process.
DEPLOY-02 (Critical) — ADMIN LOGIN RETURNED 500 FOR EVERY USER IN DOCKER.
docker-compose.yml passes each optional variable as ${X:-}, so an unconfigured
one arrives as "" rather than undefined. The code read them with
`process.env.X ?? 'default'`, which only falls back on undefined, so the empty
string reached jsonwebtoken and it threw
"expiresIn" should be a number of seconds or string representing a timespan
on every sign. Invisible to every existing suite, all of which run against a
local process where the variable is genuinely absent.
Added src/common/env.ts (envOr / envOrNull), which treat blank as unset, and
applied it to every auth, port, upload and SMS read. Three were not cosmetic:
a blank UPLOAD_DIR would have served static files from the FILESYSTEM ROOT, a
blank PORT would have bound a random ephemeral port (Number('') === 0), and a
blank JWT_SECRET would have become the verifying key. Recorded as D25, with a
standing rule to use these readers instead of `??`.
Pinned by src/common/env.spec.ts (9 tests), which reproduces the original throw
against the real jsonwebtoken and asserts the reader can no longer produce it.
DEPLOY-03 (process, documented) — the API image is `COPY release/ ./`, so it
ships a bundle pre-built on the developer machine. `docker compose build` will
happily produce an image from a STALE release/app.js and report success; at the
start of this run that bundle was four days old and the rebuilt container was
missing every route added since. DELIVERY.md's deploy section now leads with
`pnpm build:release`.
New: scripts/test-compose-e2e.cjs — the whole delivery flow against the built
container, over HTTP only, as a black box: a sales expert and an ordinary
customer both sign in with the master code, both place an order, both orders
appear in the admin grid and open there, and the three token boundaries still
hold in the deployed image. 32/32.
Verified: compose e2e 32/32 · jest 54/54 · build 5/5 ·
against the CONTAINER: complete-profile 38/38, report 40/40, expert login 36/36 ·
locally: apps smoke 43/43, admin smoke 43/43, B2B scan 34 ok/0 broken.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The folder-discovery fix in the previous commit surfaced a second problem.
Databases migrated BEFORE that fix recorded the baseline under the name of
the hand-added flat duplicate that used to sit beside it
(prisma/migrations/0001_init.sql). Once the runner started discovering the
real folder (20260910113720_first), it saw a name it had never recorded and
tried to apply the baseline again — against a database that already had it:
Raw query failed. Code: 42710.
ERROR: type "AdminRole" already exists
The two files are identical SQL (verified byte-for-byte after normalising
line endings; they differ only by one trailing blank line), so the recorded
legacy name genuinely means that migration is applied. A LEGACY_NAME_ALIASES
map now treats '0001_init.sql' as an already-applied marker for
'20260910113720_first', so upgrading an existing database is a no-op instead
of a crash.
Verified against the local dev database this actually broke on
(postgresql://localhost:5432/gold_banking, which had recorded only the old
flat name): 'npm run dev' in apps/api now logs 'No pending migrations' and
'Nest application successfully started' instead of dying. That database's
tracking table was also reconciled to record both real migration names, and
it was confirmed to already contain both migrations' schema (AdminRole,
City.percent, user_type, CustomerKind all present) — nothing was re-run
against it. The dockerised database is unaffected: it already records the
real names, so the alias is inert there.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three real bugs found and fixed while wiring up product images end-to-end:
1. src/prisma/migrate.ts only discovered flat prisma/migrations/<name>.sql
files, never Prisma's own folder-per-migration layout
(<timestamp>_<name>/migration.sql). Every migration created the normal
way (via 'prisma migrate dev') was silently skipped; only a hand-added
flat duplicate (prisma/migrations/0001_init.sql, now removed) ever ran.
This meant schema.prisma had drifted well ahead of what any real deploy
database actually had (missing columns on City, Customer, Branch,
Product, several new tables). Fixed to discover both layouts; the
drift is closed with a new 20260914165255_sync_schema_drift migration.
2. Neither nginx.conf (production) nor vite.config.ts (dev) proxied
/uploads/ to the API — only /api/. UploadedFile/ProductImage URLs
(/uploads/<file>, served by main.ts's useStaticAssets OUTSIDE the /api
prefix) 404'd through both. Both now proxy /uploads/ the same way /api/
already was.
3. docker-compose.yml's api service had no volume for storage/uploads, so
every uploaded file — and by extension every ProductImage row pointing
at one — was lost on the next container recreate/redeploy. Added a
named volume.
Also: main.ts's static file serving now sets Content-Type: image/avif
explicitly (serve-static's bundled mime table predates .avif and served it
as application/octet-stream; browsers rendered it anyway by sniffing, but
a stricter client would not have).
Product photos: copied 15 real jewelry photos into the uploads volume,
registered them as UploadedFile rows, and assigned each of the 5 existing
Product rows a randomly picked one (duplicates allowed, per request),
replacing their /img/no-items.png placeholder ProductImage row. Verified
end-to-end in a real browser against the real containerized API+DB: logged
in, opened the admin product list, confirmed all 5 rows render their photo
(naturalWidth/Height populated, network 200s, no console errors).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
mahmood