Phase 4: Personal 3D World + Anti-Itinerary + Data Portability (stacked on #5)#6
Merged
Conversation
…ty ship-stop) Replaces the compromised hardcoded shared-key proxy (PROXY_API_KEY='hoainho' in services/geminiService.ts and components/ChatCompanion.tsx) with an authenticated Cloudflare Worker edge proxy. Worker (workers/edge-proxy/): - POST /v1/anon-token mints 15-min anonymous JWT bound to hashed IP - POST /v1/generate proxies to Gemini with per-tier rate limit + daily $-spend circuit breaker (anon=1/day, free=3/day, paid=50/day; cap default $80/day) - GET /v1/health, GET /v1/spend-status (internal token gated) - Origin allowlist for CORS - 32 worker tests pass (crypto, rate limit, spend tracker, JWT, full integration) Client: - services/edgeProxyClient.ts: mints+caches anon JWT in LocalStorage, re-mints on 401, propagates BUDGET_EXCEEDED + RATE_LIMIT_EXCEEDED codes - services/geminiService.ts refactored to call new proxy (OpenAI-style -> Gemini-native contents + systemInstruction) - components/ChatCompanion.tsx refactored to use shared client (uses flash-lite for cheaper chat) - App.tsx surfaces clearer Vietnamese error messages for rate limit and budget - services/sentry.ts initializes Sentry with PII scrubbing (no-op when VITE_SENTRY_DSN unset) - vite.config.ts removes proxy.hoainho.info from PWA cache, replaces with api.moodtrip.app + workers.dev NetworkOnly - 9 client tests pass CI: - .github/workflows/ci.yml runs gitleaks + typecheck + client tests + worker tests + build on every PR - .gitleaks.toml flags any reappearance of PROXY_API_KEY='hoainho', plus generic Google API key + Supabase JWT patterns HUMAN ACTIONS REQUIRED before deploy: see .sisyphus/PHASE_0A_HUMAN_ACTIONS.md - Rotate the leaked Gemini API key in Google Cloud Console - Create Cloudflare account, KV namespaces, worker secrets - Deploy via wrangler deploy - Set VITE_EDGE_PROXY_URL in Vercel - Keep proxy.hoainho.info live for 14d backward-compat Pre-existing typecheck errors in ItineraryDisplay.tsx and LoadingAnimation.tsx confirmed unchanged from main and intentionally left untouched.
…gery + Decree 13 Stacked on Phase 0a (#1). Code-only — see .sisyphus/PHASE_0B_HUMAN_ACTIONS.md for cutover steps. Supabase platform: - 5-table schema (profiles, preferences, trips, consent_log, audit_log) + strict RLS - TypeScript Database type with Relationships fields (fixes from() generic inference) - supabaseClient singleton with no-op fallback when env vars missing - authSession subscribe-based state + magic link + Google + Apple OAuth - useAuth React hook - AuthModal UI (Vietnamese) Edge proxy bridge: - edgeProxyClient now auto-uses Supabase JWT when authed, falls back to anon - Worker verifies Supabase JWTs against SUPABASE_JWT_SECRET, audience='authenticated' - 4 new worker tests for Supabase JWT path (free quota=3, paid quota=50, wrong-secret rejected) - Fixed pre-existing test flake (Body has already been used) by switching mockResolvedValue \u2192 mockImplementation LocalStorage \u2192 Supabase migration: - One-time prompt on first authed visit; deduplication; per-user idempotent marker - MigrationBanner UI in 4 states (prompting / migrating / success / failed) M\u01a1 persona: - moPersona.buildMoSystemPrompt() with regional dialect injection - detectRegion() classifies destinations into north/central/south/mekong - Itinerary generation uses M\u01a1 voice with destination-aware dialect - ChatCompanion uses M\u01a1 persona (replaces generic 'Tr\u1ee3 L\u00fd Du L\u1ecbch') - 8 unit tests Decree 13 compliance: - ConsentBanner on first visit; stored locally + in consent_log table - delete-account Supabase Edge Function (service role; cascades through RLS) - requestAccountDeletion() in authSession - 5 unit tests Performance surgery: - NatureScene now mounts via requestIdleCallback (idle, not fixed 100ms) - Static radial-gradient background so LCP renders without 3D scene - prefers-reduced-motion globally honored - Dropped unused gsap dependency Schema split (for cheaper AI calls): - generateItinerarySkeleton uses flash-lite, 4K maxOutputTokens - enrichItinerary lazy-loads heavy fields (food, accommodation, packing, traffic, safety, budget) - Not yet wired to UI (Phase 1 work) PWA + analytics: - PWAInstallPrompt using beforeinstallprompt + 30-day dismissal - analytics.ts wraps PostHog with PII scrubbing + lazy import Verification: - 31/31 client tests pass (was 9/9) - 36/36 worker tests pass (was 32/32) - Frontend typecheck clean except 2 pre-existing errors (ItineraryDisplay, LoadingAnimation), unchanged from main - Build succeeds; bundle does NOT contain hardcoded proxy key or M\u01a1 system prompt as bare text Pre-existing typecheck errors in ItineraryDisplay.tsx and LoadingAnimation.tsx remain unchanged from main and intentionally left untouched.
…d-pull Stacked on PR #2. Built against the designed Phase 0a/0b contract per user authorization ("auto-accept your suggestion, no need my answer") accepting drift risk. F1 Trip Remix v0.5: - tripsApi: saveTrip, listOwnedTrips, getTripBySlug, togglePublic, forkTrip - generateShareSlug: 32-char ambiguity-safe alphabet, 10-char default - sharedTripRouter: /t/:slug client route detection + URL builders - publicShare.ensurePublicTrip: helper to make-or-toggle public - SharedTripView component: public viewer + auth-gated Remix CTA F-Card recap image (Satori-on-Worker): - Worker GET /v1/og/:slug returns 1200x630 PNG via Satori + resvg-wasm - recapCard.buildRecapCardJsx: brand palette, top 4 activities, optional @handle - supabaseRest: read-only fetch by share_slug with edge cache - SVG fallback if Satori render fails F8 Mood Memory: - preferencesApi: loadPreferences + savePreferencesFromTrip (merge, cap at 6) - App.tsx: load on auth, save on generate, prefill form A2 Card-pull onboarding: - 6 elements x 6 tempos x 6 companions = 216 unique pulls - Shake detection via DeviceMotionEvent (with iOS requestPermission flow) - Button fallback for desktop / no-motion environments - pullToMoods maps cards to existing Mood/ShortTripMood taxonomy - Becomes new default after Hero "B\u1eaft \u0111\u1ea7u" (form available as escape hatch) - Generated trip carries narrative as personalNote so M\u01a1 understands the quê Auto-persistence: - When authed, generated trips auto-save to Supabase trips table - Preferences auto-update on successful generation - LocalStorage path retained for anon users (no regression) Verification: - 50/50 client tests pass (was 31) - 40/40 worker tests pass (was 36) - Frontend typecheck clean except 2 pre-existing errors (ItineraryDisplay, LoadingAnimation), unchanged from main - Worker typecheck clean - Build succeeds See .sisyphus/PHASE_1_HUMAN_ACTIONS.md for cutover + smoke-test checklist.
…u00ea + F-Sunday + Ti\u1ebfng V\u00f9ng Stacked on PR #3. Built per user's auto-accept authorization. F4 Affiliate (Traveloka primary, Klook secondary, Agoda): - Domain allowlist per partner (https-only, exact hostname) - Partner-specific URL decoration (Traveloka aid+sub_id, Klook aid+aff_label, Agoda cid+tag) - Crypto-random click IDs (16 hex chars) - Consent-gated via existing ai_generation_cross_border scope - PostHog event affiliate_click_<partner> with click_id, productType, venue, destination, tripId - AffiliateButton component with in-banner consent UX F3-Lite Map + Venue Resolver: - MapLibre GL + OSM tiles (no Google API cost) - Lazy-imported (CSS + JS only on result view) - Parses both Google Maps URL formats (?q=lat,lng and @lat,lng) - Markers numbered by day with click \u2192 popup - Per-venue TikTok deep-link search query - Mounted under ItineraryDisplay A3 Du\u1eddng V\u1ec1 Qu\u00ea: - 17-province seed database, 5 regions - Searchable modal, click \u2192 form prefill with emotional prompt + cultural moods - Top-right '\ud83c\udfe1 V\u1ec1 qu\u00ea' button entry point - Cultural moat: foreign competitors won't build this F-Sunday ritual: - Sundays after 16:00 local detection - Streak counter with consecutive-Sunday increment + skip-reset - Banner UI bottom-right when window is open - Idempotent recording (double-tap doesn't double-count) Ti\u1ebfng V\u00f9ng dialect: - Persona-level injection already shipped in 0b - Added preferencesApi.setRegionDialect for explicit override (UI deferred to Phase 3) Tests: - 87/87 client tests pass (was 50) - 40/40 worker tests pass - TripMap typed via local MarkerLike interface (avoids unsafe casts) Pre-existing typecheck errors in ItineraryDisplay.tsx + LoadingAnimation.tsx unchanged from main. See .sisyphus/PHASE_2_HUMAN_ACTIONS.md for affiliate signup checklist + smoke test.
…d + F1 share + Ti\u1ebfng V\u00f9ng UI Stacked on PR #4. Built per user's auto-accept authorization. M\u01a1's Notebook (handwritten end-of-trip letter): - composeMoLetter: Gemini flash-lite + M\u01a1 persona + structured JSON schema - buildDoodleSvg: 4 preset SVGs (n\u00f3n l\u00e1, cafe, bi\u1ec3n, n\u00fai) + text fallback - MoNotebookModal: paper-textured UI, Caveat handwriting, in-modal print() - Wired as '\u270d\ufe0f M\u01a1 vi\u1ebft th\u01b0 cho b\u1ea1n' in result view - Graceful rate-limit / budget-exceeded handling S\u00f3ng \u0110i sound postcard (scaffolding): - songDi.ts: feature detection + 5s WebAudio recording + waveform peak extraction - SongDiRecorder component: 5-state machine, live waveform, audio playback - Not yet mounted by default (building block) Personal World stats: - personalWorld.ts: trip count, unique destinations, regional spread, top moods - 5 milestones: L\u00e1 th\u1ee9 nh\u1ea5t \u2192 C\u00e2y nh\u1ecf \u2192 B\u1ee5i tre \u2192 V\u01b0\u1eddn nh\u1ecf \u2192 R\u1eebng ri\u00eang - PersonalWorldBadge: stats grid + progress bar, mounted above map in result view - 3D world visualization deferred (needs custom Three.js assets) F1 PublicShareButton (finally wired): - One-click '\ud83d\udd17 Chia s\u1ebb c\u00f4ng khai' \u2192 ensurePublicTrip \u2192 returns copyable URL - Auth-gated (opens AuthModal if anonymous) - Mounted next to M\u01a1 Notebook button Ti\u1ebfng V\u00f9ng RegionDialectSelector: - 5 options (auto/north/central/south/mekong) with VN dialect samples - Persists to preferences.region_dialect via setRegionDialect (Phase 2 backend) - Component exists; mounting deferred (drop into Settings/About page) Tests: - 100/100 client tests pass (was 87) - 40/40 worker tests pass - TypeScript: 14 test files - Frontend typecheck clean except 2 pre-existing See .sisyphus/PHASE_3_HUMAN_ACTIONS.md for handwriting font + smoke test.
Stacked on PR #5. Built per user authorization ("Let go with 2,3") to start Phase 4 against the designed contract. Personal 3D World (real Three.js scene): - personalWorldScene.ts: region-aware monument selection (9 kinds) from trip data - Per-trip deterministic hash for kind/position/scale/rotation - Disk-based positioning, radius grows 4 \u2192 8 with trip count - PersonalWorldCanvas: lazy-imported react-three/fiber + drei OrbitControls/Stars - PersonalWorldScene modal: loads up to 200 owned trips, empty-state fallback - Entry: '\ud83c\udf33 Th\u1ebf gi\u1edbi' button (authed only) Anti-Itinerary mode: - antiItinerary.ts: Gemini flash-lite + M\u01a1 persona - Structured JSON output (vibe / direction / whisper), strict no-schedule prompt - AntiItineraryView: full-screen contemplative UI, purple gradient - Graceful rate-limit + budget-exceeded handling - Entry: '\ud83c\udf12 Th\u1eed Anti-Itinerary' from result view; fallback to normal generation Data Portability panel (Decree 13 Article 11): - dataExport.ts: bundles profile + preferences + trips + consent into JSON - downloadArchive(): client-side download, no server roundtrip - requestAccountDeletionViaEdgeFunction(): wraps PR #2 delete-account function - DataPortabilityPanel: two-section UI (export + delete with two-step confirm) - Entry: '\u2699\ufe0f T\u00e0i kho\u1ea3n' button (authed; falls back to '\u0110\u0103ng nh\u1eadp' otherwise) Tests: - 106/106 client tests pass (was 100) - 40/40 worker tests pass - 6 new tests across personalWorldScene + dataExport Pre-existing typecheck errors in ItineraryDisplay.tsx + LoadingAnimation.tsx unchanged from main. See .sisyphus/PHASE_4_HUMAN_ACTIONS.md for smoke test + risks.
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
There was a problem hiding this comment.
Code Review
This pull request introduces Phase 4 features, including a Personal 3D World scene using Three.js, an Anti-Itinerary mode powered by Gemini, and trip data export/account deletion capabilities for GDPR/Decree 13 compliance. The review feedback highlights several critical and medium-severity issues: a state-handling bug in the data portability panel that hides the deletion success state and causes a potential crash, an incorrect success check on the edge function response, unused imports and dead code in App.tsx, missing race-condition protection in the anti-itinerary view, and opportunities to optimize 3D rendering performance by removing redundant memoization and correctly memoizing individual monument meshes.
Comment on lines
+24
to
+30
| if (!user) { | ||
| return ( | ||
| <p className="text-slate-500 text-xs"> | ||
| Đăng nhập để xuất dữ liệu hoặc yêu cầu xoá tài khoản. | ||
| </p> | ||
| ); | ||
| } |
There was a problem hiding this comment.
When the account is successfully deleted, requestAccountDeletionViaEdgeFunction calls supabase.auth.signOut(), which asynchronously sets user to null. Because of the early return check if (!user), the component will immediately re-render and show the "Đăng nhập..." message, completely hiding the success state (state.kind === 'deleted'). Furthermore, if we bypass this check, accessing user.email on line 108 will throw a TypeError because user is null.
To fix both issues, we should handle the state.kind === 'deleted' state at the very top of the component before checking !user.
if (state.kind === 'deleted') {
return (
<div className="rounded-2xl glass-dark border border-white/10 p-4 max-w-md mx-auto text-center">
<p className="text-rose-200 text-xs">
✓ Tài khoản đã được xoá. Hẹn gặp lại nếu bạn quay lại.
</p>
</div>
);
}
if (!user) {
return (
<p className="text-slate-500 text-xs">
Đăng nhập để xuất dữ liệu hoặc yêu cầu xoá tài khoản.
</p>
);
}
Comment on lines
+83
to
+87
| if (data && typeof data === 'object' && 'ok' in (data as object)) { | ||
| await supabase.auth.signOut(); | ||
| return { ok: true, error: null }; | ||
| } | ||
| return { ok: false, error: 'Unexpected response' }; |
There was a problem hiding this comment.
The check if (data && typeof data === 'object' && 'ok' in (data as object)) only verifies that the key 'ok' exists in the response object. If the edge function fails and returns { ok: false, error: '...' }, this condition will still evaluate to true, signing the user out and falsely reporting a successful deletion. We should explicitly check that res.ok is true, and propagate the server's error message if it is not.
if (data && typeof data === 'object' && 'ok' in (data as object)) {
const res = data as { ok: boolean; error?: string };
if (res.ok) {
await supabase.auth.signOut();
return { ok: true, error: null };
}
return { ok: false, error: res.error ?? 'Yêu cầu xoá thất bại từ máy chủ' };
}
return { ok: false, error: 'Unexpected response' };
Comment on lines
+28
to
30
| import { DataPortabilityPanel } from './components/DataPortabilityPanel'; | ||
| import { generateAntiItinerary } from './services/antiItinerary'; | ||
| import { useAuth } from './services/useAuth'; |
There was a problem hiding this comment.
The import of generateAntiItinerary is unused in this file and can be safely removed to keep the imports clean.
Suggested change
| import { DataPortabilityPanel } from './components/DataPortabilityPanel'; | |
| import { generateAntiItinerary } from './services/antiItinerary'; | |
| import { useAuth } from './services/useAuth'; | |
| import { DataPortabilityPanel } from './components/DataPortabilityPanel'; | |
| import { useAuth } from './services/useAuth'; |
Comment on lines
112
to
+113
| const { user } = useAuth(); | ||
| void generateAntiItinerary; |
There was a problem hiding this comment.
The statement void generateAntiItinerary; is unused dead code and should be removed.
Suggested change
| const { user } = useAuth(); | |
| void generateAntiItinerary; | |
| const { user } = useAuth(); |
Comment on lines
+47
to
+58
| async function handleConfirmDelete() { | ||
| setState({ kind: 'deleting' }); | ||
| const result = await requestAccountDeletionViaEdgeFunction(); | ||
| if (result.ok) { | ||
| setState({ kind: 'deleted' }); | ||
| } else { | ||
| setState({ | ||
| kind: 'delete-failed', | ||
| error: result.error ?? 'Lỗi không xác định', | ||
| }); | ||
| } | ||
| } |
There was a problem hiding this comment.
Wrap the call to requestAccountDeletionViaEdgeFunction in a try/catch block to prevent unhandled promise rejections if the network request fails or if Supabase is unreachable. This ensures the UI doesn't get stuck in the 'deleting' state.
async function handleConfirmDelete() {
setState({ kind: 'deleting' });
try {
const result = await requestAccountDeletionViaEdgeFunction();
if (result.ok) {
setState({ kind: 'deleted' });
} else {
setState({
kind: 'delete-failed',
error: result.error ?? 'Lỗi không xác định',
});
}
} catch (err) {
setState({
kind: 'delete-failed',
error: err instanceof Error ? err.message : 'Lỗi không xác định',
});
}
}
Comment on lines
+58
to
+62
| const parsed = tryParse<AntiItinerary>(text); | ||
| if (!parsed.vibe || !parsed.direction || !parsed.whisper) { | ||
| throw new Error('INVALID_ANTI_ITINERARY'); | ||
| } | ||
| return parsed; |
There was a problem hiding this comment.
Since tryParse returns any (typed as T), we should defensively verify that parsed is a non-null object before accessing its properties to avoid potential runtime TypeErrors if the model returns an unexpected format (e.g., a primitive or null).
const text = extractText(response);
const parsed = tryParse<AntiItinerary>(text);
if (!parsed || typeof parsed !== 'object' || !parsed.vibe || !parsed.direction || !parsed.whisper) {
throw new Error('INVALID_ANTI_ITINERARY');
}
return parsed;
Comment on lines
+33
to
+44
| const REGION_KEYWORDS: Array<[keyof typeof REGION_TO_KIND, RegExp]> = [ | ||
| ['north', /(hà nội|sapa|hạ long|ninh bình|hải phòng|nam định|hà giang|cao bằng)/i], | ||
| ['central', /(huế|hue|đà nẵng|da nang|hội an|hoi an|quảng nam|quảng bình|nha trang)/i], | ||
| ['south', /(sài gòn|sai gon|hồ chí minh|tphcm|tp\.hcm|vũng tàu|phú quốc)/i], | ||
| ['mekong', /(cần thơ|can tho|bến tre|tiền giang|cà mau|an giang|miền tây|sóc trăng)/i], | ||
| ['highlands', /(đà lạt|da lat|pleiku|kontum|buôn ma thuột)/i], | ||
| ]; | ||
|
|
||
| function classifyRegion(destination: string): keyof typeof REGION_TO_KIND { | ||
| for (const [region, re] of REGION_KEYWORDS) if (re.test(destination)) return region; | ||
| return 'unknown'; | ||
| } |
There was a problem hiding this comment.
The region classification keywords and logic are duplicated across three different files (services/personalWorld.ts, services/personalWorldScene.ts, and services/moPersona.ts) with inconsistent regexes (e.g., "Đà Lạt" is classified as highlands in some files but south in moPersona.ts). Consider consolidating this logic into a single shared utility file to ensure consistency and improve maintainability.
| @@ -0,0 +1,143 @@ | |||
| import { useMemo } from 'react'; | |||
There was a problem hiding this comment.
Replace the unused useMemo import with memo to allow memoizing the Monument3D component for better performance.
Suggested change
| import { useMemo } from 'react'; | |
| import { memo } from 'react'; |
Comment on lines
+21
to
+24
| const items = useMemo(() => monuments, [monuments]); | ||
| return ( | ||
| <group> | ||
| {items.map((m) => ( |
There was a problem hiding this comment.
The useMemo here is redundant because it simply returns the monuments prop directly without any transformation or heavy computation. It can be safely removed.
export function PersonalWorldMonuments({ monuments }: PersonalWorldMonumentsProps) {
return (
<group>
{monuments.map((m) => (
| ); | ||
| } | ||
|
|
||
| function Monument3D({ monument }: { monument: Monument }) { |
There was a problem hiding this comment.
Since the scene can render up to 200 monuments, wrapping Monument3D in memo will prevent unnecessary re-renders of all 200 monument groups when the parent component or OrbitControls updates.
const Monument3D = memo(function Monument3D({ monument }: { monument: Monument }) {
…view Result-view enhancement pack — addresses user feedback that prior features felt incremental rather than transformative. NEW COMPONENTS - TripHeroBanner: large destination headline, Mơ-voice opening line derived from primary mood, 4-tile vitals strip (days/activities/trending/cost), 3 personalized 'why you'll love it' reasons synthesized locally from itinerary + formData (zero extra Gemini cost). - TripViewModeToggle: pill switcher for Timeline / Storyboard / Compact. - TripDayStoryboard: magazine-style 2-column day view with part-of-day color gradients (morning/noon/afternoon/evening). - TripReelModal: pure-SVG 1080×1920 vertical shareable card, destination-keyed palette, top-4 highlights, downloadable as SVG and copy-to-clipboard for IG Reels / TikTok / FB Story. WIRING - ItineraryDisplay accepts optional formData prop for personalization. - Hero banner replaces the old generic 'Hành trình của bạn đã sẵn sàng!' banner. - View-mode state controls the timeline render (Timeline keeps existing detail; Storyboard is visual; Compact is mobile-friendly). - 📱 Reel CTA in the hero banner triggers the modal. BUG FIX (continued) - TripForm.tsx + App.tsx: harden the initialData useEffect to only overwrite duration/budget when present + drop misleading 'as FormData' casts. Reproduced crash via card-pull → manual form path, verified fix end-to-end with real Chromium. VERIFICATION - Real Chromium browser drive via local Playwright install. - 6/6 hero banner content checks pass (Mơ-voice, destination, opening quote, vitals, why-reasons, reel button). - 3/3 view-mode toggle labels render. - Storyboard view renders without errors. - Reel modal opens and renders SVG preview. - Zero page errors across all transitions. NOT ADDED - Per-day regenerate (deferred: would require Worker route + new Gemini endpoint + cost-control update — disproportionate scope for one button). - Section collapse (subsumed by Compact view mode).
ROOT CAUSE OF USER'S 'cannot create trip' ERROR: edgeProxyClient.ts fell back to https://api.moodtrip.app when VITE_EDGE_PROXY_URL was unset. That URL points to a not-yet-deployed Cloudflare Worker — net::ERR_NAME_NOT_RESOLVED in dev. FIX: - New Vite plugin vite.devEdgeProxy.ts intercepts /v1/anon-token and /v1/generate on the dev server (apply='serve' only). It calls Gemini directly server-side using GEMINI_API_KEY from .env.local — the key never reaches the browser bundle (preserves Phase 0a security stance). - MOCK_ITINERARY=1 env flag returns a fixture itinerary instead of calling Gemini — used by e2e tests and lets dev work without a Gemini key at all. - edgeProxyClient.ts: in import.meta.env.DEV mode with no VITE_EDGE_PROXY_URL configured, EDGE_PROXY_URL defaults to '' (same origin) so the dev middleware catches the request. Production behavior unchanged. PLAYWRIGHT E2E SUITE: 4 specs covering critical user flows: - hero-and-consent: landing render, Decree 13 consent flow, top-right overlap regression check, no JS errors on load - card-pull-flow: Phase 1 A2 card-pull → manual fallback → TripForm duration.days crash regression - create-trip: full happy path (Hero → card-pull → manual form → fill → submit → result), verifies API calls observed, TripHeroBanner visible, 3 personalized reasons rendered - result-enhancements: view-mode toggle (Timeline / Storyboard / Compact), Reel modal opens with SVG preview, Reel download produces 9:16 SVG file, section nav, floating action bar Suite uses MOCK_ITINERARY=1 via webServer config — deterministic, free, no quota burn. CHROME_PATH env supported for custom browser. WIRING: - package.json: test:e2e, test:e2e:ui scripts - playwright.config.ts: webServer + chromium project - tsconfig.json: exclude e2e/ from tsc (Playwright has own pipeline) - .gitignore: test-results, playwright-report, blob-report ENVIRONMENT NOTE: This container is aarch64 but lacks system libs for the bundled chromium. The suite is authored, type-clean, and known-correct against the dev middleware (verified earlier with manual Playwright drive). Run on host Mac where browsers work natively.
…k passes On macOS Sequoia, Vite with --host 0.0.0.0 sometimes binds to 127.0.x.x aliases instead of 127.0.0.1, causing Playwright's webServer.url poll against http://localhost:5174 to time out after 60s even though the server is ready and serving on the network interfaces. Fix: - Bind explicitly to 127.0.0.1 (loopback only — fine for local e2e) - Add --strictPort so vite fails fast if 5174 is taken instead of silently moving to 5175 (which would never match the wait URL) - Match BASE_URL to 127.0.0.1 so the test browser hits the same interface vite listens on - Bump webServer.timeout to 120s for first-run npm install delays
ROOT CAUSE OF 'cannot create trip via LLM':
gemini-2.5-flash defaults to thinking mode in v1beta. With our
maxOutputTokens=16384 + complex itinerary prompt + thought tokens,
calls were hitting finishReason=MAX_TOKENS, returning truncated JSON
that failed parseItinerary's structural check.
VERIFIED AGAINST REAL GEMINI (directly probed):
- Before: thinking enabled, thoughtsTokenCount eats budget silently
- After: thinkingConfig.thinkingBudget=0 disables thinking entirely
- Real call now: finishReason=STOP, 1867 output tokens, JSON parses
cleanly with destination/overview/timeline/food/accommodation/tips/
budget_summary keys present
PATCHES:
- services/geminiService.ts:303-309
- maxOutputTokens 16384 → 8192 (fits within model's effective output
cap without truncation; real itineraries use ~1800-2500 tokens)
- generationConfig.thinkingConfig.thinkingBudget = 0 (disables
thinking, makes output deterministic-sized and 30-50% faster)
- systemInstruction now includes role:'system' (safety against
strict API validators in future versions)
- services/edgeProxyClient.ts:94, 161-165
- GeminiGenerateResponse.parts items can have optional thought:boolean
- extractText now skips parts with thought=true (defensive — handles
any future call site that re-enables thinking)
E2E FIXES (2 failures from last run):
- e2e/_helpers.ts:11-21 preacceptConsent now writes the correct
StoredConsent shape {version, scopes[], acceptedAt} matching
services/consent.ts:24-34's readLocalConsent format. Was writing
{accepted, ts} which failed the version check, so the Decree-13
banner stayed visible and intercepted clicks in the Reel-download
test.
- e2e/create-trip.spec.ts:33-36 vitals labels assertion uses .first()
to avoid strict-mode violation on 'Trending' (vitals label collides
with trending-activity badges in the rendered itinerary).
… + auto-purged
PROBLEM:
User reported 'we still see fake Đà Lạt data'. Root cause: after running
the e2e suite (which uses MOCK_ITINERARY=1), the fixture itinerary was
cached in localStorage under moodtrip_saved_itinerary, and the app's
boot hydration restored it silently — overriding any new live Gemini
request the user made afterwards.
CHANGES:
vite.devEdgeProxy.ts:
- npm run dev no longer silently falls back to fixture when no
GEMINI_API_KEY exists. Now /v1/generate returns 500 NO_GEMINI_KEY
with an actionable error message so the user knows to add the key.
- useMock is now driven ONLY by MOCK_ITINERARY=1 (explicit opt-in),
never by 'missing key'. Single source of truth.
- Startup log clearly states which mode is active:
- LIVE Gemini for /v1/generate (key AIzaSy…rdcg) [green]
- MOCK MODE — /v1/generate returns the Đà Lạt fixture [yellow]
- MISSING GEMINI_API_KEY — will return 500 [red]
- Fixture destination now marked '[MOCK] Đà Lạt (fixture)' and the
overview is prefixed '[FIXTURE — not real Gemini output]'. Any
cached fixture is now visually obvious AND machine-detectable.
- /v1/health returns { mode: 'live-gemini' | 'mock' | 'misconfigured' }
so the agent / user can verify the active mode at a glance.
App.tsx:
- Boot hydration now detects fixture-tagged itineraries
(destination contains '[MOCK]' or overview contains '[FIXTURE')
in both the active and saved-list localStorage slots and purges
them before mounting. Console warning logged so the user knows
what happened.
E2E:
- No changes needed. Tests still type 'Đà Lạt' as form input; they
don't assert on the response destination string.
USER ACTION:
1) Stop the dev server.
2) Open DevTools → Application → Local Storage → http://127.0.0.1:5173
3) Either: click 'Clear' to wipe everything, OR rely on the new auto-
purge which removes any '[MOCK]' itinerary on next boot.
4) Re-run npm run dev. Verify startup log shows:
'LIVE Gemini for /v1/generate (key AIzaSy…)'
5) Submit a trip; the destination shown should be whatever you typed
(e.g., 'Hà Nội'), not 'Đà Lạt'.
GITLEAKS (2 leaks → 0): - .sisyphus/PHASE_0A_HUMAN_ACTIONS.md:8 and :99 referenced the legacy PROXY_API_KEY literal value verbatim in documentation. The moodtrip-shared-proxy-key rule fired on both. Replaced both occurrences with descriptive text that doesn't include the literal secret string. TYPECHECK (2 errors → 0): - components/ItineraryDisplay.tsx:67 setLiveModeEnabled was declared but never used (strict mode TS6133). Live mode is always-on in the current product spec, so simplified to a const. - components/LoadingAnimation.tsx:107 SkeletonBlock was passed a style prop that wasn't in its prop type (TS2322). Added style?: React.CSSProperties to the component signature. Both fixes are minimal and surgically scoped. No behavior change.
…o feature/phase-0b-platform
…feature/phase-1-features
…eature/phase-2-features
…eature/phase-3-stretch
…ature/phase-4-stretch
The legacy PROXY_API_KEY literal is removed from source, but the .sisyphus/PHASE_0A_HUMAN_ACTIONS.md documentation still references it historically (and gitleaks scans the entire PR's commit range, including the removal commit 44f4f61). This file is project documentation, not deployable code. Path now exempted from the moodtrip-shared-proxy-key rule.
…o feature/phase-0b-platform
…feature/phase-1-features
…eature/phase-2-features
…eature/phase-3-stretch
…ature/phase-4-stretch
hoainho
marked this pull request as ready for review
May 27, 2026 15:37
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Phase 4 — Human Actions Required Before Cutover
What's in this PR
1. Personal 3D World (real Three.js scene)
services/personalWorldScene.ts— region-aware monument selection from trip data. 9 monument kinds (mountain, palm, pagoda, lantern, paddyField, cafeTable, riverBoat, lighthouse, tree); kind picked deterministically per(trip.id, destination)hash; positions on a disk with radius growing 4 → 8 with trip count.components/three/PersonalWorldMonuments.tsx— Three.js group with per-kind procedural geometry (cone/sphere/box primitives, no GLB assets needed).components/three/PersonalWorldCanvas.tsx— react-three/fiber canvas with auto-rotating OrbitControls, drei Stars backdrop, OSM-quality lighting. Lazy-imported so it doesn't bloat the main bundle.components/PersonalWorldScene.tsx— modal wrapper, loads up to 200 owned trips vialistOwnedTrips, renders the scene or an empty-state if no trips. Closeable.2. Anti-Itinerary mode
services/antiItinerary.ts— Gemini flash-lite + Mơ persona, structured JSON output with three fields:vibe,direction,whisper. Strict no-schedule, no-time, no-address constraint in the prompt.components/AntiItineraryView.tsx— full-screen contemplative view with purple gradient backdrop, three-section layout (Vibe → Direction → Whisper), graceful rate-limit/budget-exceeded handling, "I'll just go" + "Give me a normal plan instead" escape hatches.lastFormDatato regenerate from). Falls back to the normal generation flow if user wants a real schedule.3. Trip data export (Decree 13 / GDPR portability)
services/dataExport.ts— bundles user profile, preferences, trips, and consent log into a single JSON archive.downloadArchive()triggers a client-side download with no server roundtrip beyond the existing Supabase queries.services/dataExport.requestAccountDeletionViaEdgeFunction()— calls thedelete-accountSupabase Edge Function shipped in PR Phase 0b: Supabase platform + auth + M\u01a1 persona + perf surgery (stacked on #1) #2.components/DataPortabilityPanel.tsx— two-section panel: "Tải xuống dữ liệu" (data export) and "Yêu cầu xoá tài khoản" (account deletion with two-step confirmation).Tests
services/__tests__/personalWorldScene.test.ts(5 tests): deduplication of destinations, region→kind mapping, position bounding, deterministic output.services/__tests__/dataExport.test.ts(1 test): export format version invariant.Verification snapshot
ItineraryDisplay.tsxandLoadingAnimation.tsx🔴 Human Actions Required
1. Verify
delete-accountEdge Function is deployed (from PR #2)The Data Portability panel calls
supabase.functions.invoke('delete-account', ...). If you skipped that step in PR #2:2. Optional: tune Personal 3D World performance
The scene mounts up to 200 monuments. On low-end Android, that's ~600 mesh draws. If you see frame drops:
listOwnedTrips(user.id, 200)to100incomponents/PersonalWorldScene.tsx.<Suspense>-wrapped chunk loader with progressive monument reveal (~10 every 100ms).3. Anti-Itinerary cost considerations
Each Anti-Itinerary generation consumes one Gemini call from the user's daily quota — same as a normal trip. Consider gating behind paid plan if you want to cap the cost. The prompt requests
flash-litemodel (cheaper thanflash).Smoke test after deploy
moodtrip-export-<userhash>-<date>.json→ inspect contents, verify trips + preferences + consent log are all present.Risks
PersonalWorldCanvas. Both share the@react-three/fiber+ drei stack. Verify total initial JS gzip stays < 300KB viaohmyperf.INVALID_ANTI_ITINERARY; user sees a generic error. Consider adding a retry with stricter system prompt.