# Tech Debt Running list of known shortcuts, deferred hardening, and "good enough for now" decisions that need follow-up before they bite us in production. Format: `[date]` short title, then enough context for someone (or future-you) to act on it without re-deriving the discussion. --- ## Backend ### `[2026-06-01]` Bookkeeping INSERT sits in the pairing critical path **File:** `backend/src/services/pairing.service.js` (`acceptPairingRequest`, ~line 506) **What happened:** the `INSERT INTO customer_transactions` runs *after* the session is flipped to `ACTIVE` but *before* `startSessionTimer`, `startSessionListener`, the customer `PAIRED` WS notify, and the other-mitra dismiss fan-out. A `varchar(20)` overflow on `type = 'first_session_discount'` (22 chars) threw there, so every first-session-discount pairing half-completed: no transaction row, no server-side timer, no PAIRED push (customer recovered via polling), and a 500 returned to the mitra so its app never opened the chat. **Fixed now:** column widened to `VARCHAR(128)` (migrate.js), so the INSERT no longer throws. **Why it's still debt:** a *bookkeeping* write can still abort *critical* pairing steps if it ever fails again (constraint change, DB hiccup, future longer enum). Hardening: either move the `customer_transactions` INSERT to the end of `acceptPairingRequest`, or wrap it in a `try/catch` that logs-but-doesn't-throw, so transaction recording can never again half-complete a pairing. Same applies to the equivalent INSERT in `extension.service.js`. --- ### `[2026-05-11]` Public `GET /api/public/bestie/available` needs rate limiting before prod **File:** `backend/src/routes/public/public.bestie-availability.routes.js` **Decision:** The endpoint was made unauthenticated by business requirement — SHome1st renders before any JWT exists, and the CTA must reflect global mitra availability so users see whether bestie is online before committing to onboarding. Response is intentionally a single boolean (no count, no IDs). **Why it's debt:** No auth + no rate limit. The 10s in-memory cache bounds DB load, but a single attacker can still hammer the endpoint to: - run sustained traffic against the public listener (DoS surface) - scrape `available` over time to infer mitra online/offline patterns (weak information leak — only "is anyone online", but still a signal) **Mitigation before prod:** - Per-IP rate limit (suggested: ~30 req/min/IP, headroom over the legitimate 5s client poll cadence = 12 req/min/IP). - Implement via `@fastify/rate-limit` plugin so other public endpoints can share the policy as we add them under `/api/public/*`. - Verify Cloud Run / NLB preserves real client IP and that `request.ip` reflects it (Fastify already has `trustProxy: true`). **Not required:** auth, captcha, or removing the count from `/api/client/mitra-availability` (that route stays authed for CC/debug). --- ## Client app ### `[2026-05-11]` Social-login (Google / Apple) has no entry point after S3a rewrite **Files:** `client_app/lib/features/auth/screens/register_screen.dart` (no longer renders them); `client_app/lib/core/auth/auth_providers_provider.dart` (still wired). **Decision:** `RegisterScreen` was rewritten to match Figma `S3Phone` 1:1 (step-dots + name greeting + privacy card + tanpa-verif ghost link). Figma S3a shows no Google/Apple buttons, so they were removed from this screen. **Why it's debt:** Google/Apple buttons used to render here when the `authProvidersProvider` flags were enabled. Today both flags are `false` (creds pending — see `Phase 3.4 Status` memory), so nothing visible is missing. But the moment `/api/shared/auth-providers` flips either flag, the buttons have nowhere to live. **Fix-when-creds-arrive:** - Decide where Google/Apple buttons belong (likely a dedicated login screen reachable from the SHome1st "masuk →" banner), or whether to bring them back to S3a as Figma-friendly tiles above the phone input. - `loginGoogle` / `loginApple` on `authProvider` are still intact, so the wiring is one button widget away. ### `[2026-05-12]` Stage 10 — Bestie Offline Popup variant not wired on BestieHistoryList **File:** `client_app/lib/features/home/screens/bestie_history_list_screen.dart` **Decision:** Stage 10 follow-up restored `BestieHistoryList` as a separate picker screen (per mermaid §4) and made offline rows un-tappable (dimmed). Mermaid §4 actually calls for a **Bestie Offline Popup (returning variant)** to surface when the user picks an offline bestie — with options "cari bestie lain" and "tanya admin". **Why it's debt:** Today the offline row is just disabled. The user gets no explicit prompt to redirect them into the blast flow or to contact admin. **Fix:** wire `BestieOfflinePopup` with `variant='returning'` on offline-row tap. The popup widget already exists from Stage 8 (Tanya Admin sheet ships with the wiring); just needs to be triggered here. ### `[2026-05-12]` S5 ESP screen retired from spec — code still ships it **Files:** `client_app/lib/features/onboarding/` (S5ESP screen + nav wiring); `screens/onboarding.jsx::S5ESP` (Figma reference still in handoff); any `espSelectionProvider` / `espSkippedProvider` Riverpod state. **Decision:** Business removed the ESP multi-select step from the customer flow on 2026-05-12. Both verified and anonymous branches now go from `VerifChoiceSheet` straight to the `usp_seen?` gate. See `requirement/flow_customer.mermaid.md` §2. **Why it's debt:** Stage 2 (commit `2645bcd`) shipped the ESP screen and its state providers. The screen is still reachable in the current build. The mermaid spec is the source of truth — the code has drifted behind by one business decision. **Fix:** - Delete the ESP screen widget and its route registration. - Remove `espSelectionProvider` / `espSkippedProvider` and any nav step that routes through ESP. - Wire `VerifChoiceSheet → USPGate → (USP screen | skip → next)` directly. - Drop the "ESP is decorative only" memory (it's now superseded by removal). - Keep `screens/onboarding.jsx::S5ESP` in the Figma handoff folder — it's history, not active design.