The banquet menu (18 dishes) and extras/supplies/delivery items (16)
in the training catalog were seeded with photo:'', which the catalog
tiles render as a plain red placeholder square. Sourced a real photo
for each item (mostly Pexels, free license, picked and visually
reviewed one by one to avoid mismatches -- several first search hits
were wrong: a live chicken for "chicken caesar salad", a person in
frame, branded bottles/cups) and two AI-generated (banquet-caprese,
banquet-roastbeef). Converted to square 1024x1024 WebP with sharp.
Wires the paths through the actual source of truth: ops/demo/trial-
banquet-data.mjs and trial-extras-data.mjs now set photo to the real
path, then ops/demo/build-trial-*.mjs regenerated public/demo/*.json
and the matching Supabase seed-function migrations, so the client's
static JSON fetch and the server-side company-provisioning function
stay in sync.
Bumped the demo-catalog fetch's own cache-busting query string in
training-catalog.js (and its script-tag/precache version), since it's
fetched by the client with a separate version from the outer script
tags -- the same stale-cache class of bug fixed earlier this session.
Verified: all 34 referenced files resolve (200) and decode as real
1024x1024 images; confirmed via the actual rendered catalog tiles
(#tiles img[src]) that the extras category -- where the user's
red-square screenshot was taken -- now serves the real photo paths.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The deployed PHP reverse proxy at api.caterium.ru only forwarded/allowed
a fixed set of CORS request headers, missing x-supabase-api-version which
supabase-js v2.112.4's auth client sends on every request. That made the
browser reject the preflight and fail the actual login call client-side
with a generic "Failed to fetch" (not a server error, so it never showed
up in server logs) - every login was broken since the proxy went live.
Fixed on the live server and committed the previously SCP-only script
here so future edits go through git instead of being SSH-only.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
production had diverged from main with 20 unreviewed direct-push
commits never merged back (account center feature, owner-only employee
roles, and a login-screen redesign - the exact "cream login" work that
replaced the old dark table-photo screen). Neither QA nor the
audit fixes on main had ever seen this code.
Conflict resolution:
- service-worker.js: kept production's newer cache-refresh mechanism
(CRITICAL_FRESH, forceFresh, withAccountCenter, v81 cache name) and
combined both sides' CORE asset lists (account-center-v1780.js +
login-signature-v1776.js from production, auth-security-v1774.js +
order-enhancements-v1775.js from main).
- deploy-timeweb.yml: kept main's version, which already independently
verifies service-worker.js's sha256 alongside the login/logo files -
strictly more thorough than production's version of the same check.
Also fixes fallout from production's commits never having been
QA-tested before landing: package.json was bumped to 17.8.0 with
nothing else in the codebase updated to match (reverted to 17.7.3,
matching package-lock.json/release-manifest.json/app-runtime.js, since
no other release artifact actually changed), and three tests
(static-security.mjs, edge-security-v1774.mjs, release-check.mjs) had
hardcoded strings (old PWA cache name, old employee role list) that no
longer matched the code they were checking.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
sun_employee_prepare_v28 and sun_employee_finalize_v28 - the RPCs
caterium-create-employee's whole authorization model rests on - were
applied directly to the production database and were never committed,
so the actual authorization logic wasn't auditable from the repo.
Recorded verbatim via pg_get_functiondef() against the live database
on 2026-09-12 (read-only; nothing was re-applied). Confirmed both
match what the security audit inferred from the Edge Function's
error-message handling: caller identity + workspace-owner/platform-admin
check + plan/feature gates + member limits, all enforced here rather
than in the Edge Function itself.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Full production-ready docker-compose stack (db, kong, auth, rest,
realtime, storage, imgproxy, meta, functions, studio) targeting
api.caterium.ru, plus bootstrap script for a fresh Cloud Server,
Caddy reverse-proxy config (HTTPS, WebSocket, upload limits), and
dump/restore/verify/storage-sync scripts for moving off the managed
Supabase project (cksuehzcimitsxmeloes). Does not touch public/ or
any live runtime behavior — frontend cutover is documented separately
in frontend-cutover.md and only applied after Etap 8 verification.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
caterium.ru and app.caterium.ru are separate Timeweb sites with separate
document roots. The apex domain's root is reserved for an unrelated future
site, so auto-detection must never be able to land there. Drop the generic
$HOME/public_html and unverified app.caterium.ru/public_html guesses in
favor of the confirmed real path ($HOME/caterium-app/public_html), and
require CATERIUM_WEB_ROOT explicitly in the documented cron command.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The documented cron command redirected to $HOME/.caterium-deploy/cron.log
before the script could create that directory, so the very first run
failed before executing. Document mkdir -p as a one-time setup step,
add $HOME/public_html to web-root auto-detection (the confirmed real
document root for this account), and note the CATERIUM_REPO_URL override
needed when the account's SSH config uses a non-default Host alias.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Introduce the first data layer, move order auto-completion to server cron with legacy/cloud compatibility, clean error telemetry, update PWA/versioning, and add regression coverage.
