PlayMusic Backlog Snapshot

Sep 19 — release #1495 and the landing split at /

The big one landed at 04:34 (merge commit — a squash would have inherited [skip ci] from the wizard_config snapshots and skipped every Action): Drops SICK redesign P-0→P-8, Alchemy character/song/tile fixes and XP parity, /foryou chrome + Game Lists behind ?list=, Sky Rider options + collectibles, ad click-id capture (migration 109), PlayMusic Approved + bulk upload (108), ffmpeg normalisation, meshy-6-lite, real ESLint on both frontends. Migrations 107/108/109 ran 04:36; gateway 100% at 05:24 after a failed first promote. #1503 (05:36) fixed /alchemy serving the retired engine studio — AlchemyRoute fell back to the engine view whenever creation_flows.alchemy was false, which prod has; the route now ignores the flag and only the /create picker consults it. #1508 (19:03) shipped the landing split: direct and organic visitors to / are assigned at the edge to home or /foryou via a 30-day pm_lx cookie with a hop crumb; Googlebot, paid clicks (gclid) and Playgama referrals are exempt, so / stays the indexed SSR page. It also versioned the dead-R2-host repair scripts that had already fixed 2,421 catalog rows on prod. #1510 (19:27) stamped landing_variant on every event to both sinks — the split had been assigning prod visitors for 24 minutes with nothing joining an arm to an outcome.

Sep 22 — release #1527: crash telemetry on for 100% of prod, create expectation gate

Every game load now has a load_attempt_id and emits a stage event per step (open → engine_url_ready → iframe_loaded → engine_first_message → world_building → world_ready → playing), so a failed load is attributable to a stage instead of inferred from an absence (#1519). Crash reporting is narrow and opt-in — four kinds, no tracing or replay, SDK out of the bundle unless VITE_SENTRY_DSN is set — and that variable was set on Sep 21, so it went live with this deploy; #1524 opened connect-src for the ingest host. #1525 records auth failures that block generation; #1529 keeps the E2E/smoke suites out of Sentry (navigator.webdriver) and tags their analytics rows automation. Engine-side, #513/#514 add the hooks the engine never had: window error / unhandledrejection, a guarded Babylon Engine constructor, WebGL context loss and restore, and a render-stall escalation sampled from engine.frameId — fatal only for boot failures and a confirmed stall, so a game that keeps running never lands behind the host's error overlay. Promoted as engine #517. The create expectation gate (#1521–#1526): one screen between a paid click and the wizard, first-visit campaign arrivals only, 50/50 against a holdout, once-only, touches no wizard state. It also fixed two things in the shared Modal: focus containment (Tab walked out of every dialog in the product) and an elevation prop. Plus album-art upload from the CMS song page (#1520, with a worker carve-out), a failed CSS chunk preload now treated as a stale-deploy reload (#1528), and ffmpeg failures logging the tail of stderr — Sentry issue PYTHON-B8 had 4,000+ events that all read "ffmpeg version 5.1.9 … built with gcc" (#1530).

Sep 23 — releases #1542, #1545, #1548: survival picks, /games, Drops consent, and the embed banner

#1532: the config-builder worker's survival preset was applied with a blanket Object.assign, overwriting the collectible and chaser the creator picked in Step 3 — every published Horror Survival game shipped the default heart and Zed_1. The preset now fills only missing fields; games published before the worker shipped need a republish. #1534: /games grouped by the legacy projects.category string that nothing on the Front Page reads; the SPA and the crawler render now both read wizard_config.home_sections, one chip per carousel. Five slugs (drift, shooter, obby, flight, platformer) exist as both a carousel and a legacy category and the carousel now wins, so those indexed hubs shrink to the carousel's picks (staging: drift 45→6, shooter 44→3). #1533/#1535: Alchemy's Share button was createGame() with none of Publish's gating and a second project row per press — removed; pickers stop closing themselves; obby's Iso camera survives the apply. #1540/#1543: Drops gets a CookieYes banner wired to Consent Mode v2 — until now nothing ever granted consent, so every Drops GA4 hit was a cookieless ping and GA4 showed 0 events over 7 days; the script loads async after a stalled cdn-cookieyes.com was holding first paint. #1538: pm_arm and intro captured on first touch and stamped on every event — on Sep 23, 0 of 26,792 prod events carried intro. The embed banner took three releases. Yandex reported the CookieYes banner painting over the game in the PlayHop shell. #1541 (in #1542) pushed gtm.blocklist: ['cvt_KDQSW'] above the snippet on embedded surfaces — works on staging, not on the prod container. #1544 (release #1545, 22:01) gated the CMP per platform profile (features.consent true for web only; playhop gets a profile) with the sandboxedScripts token, verified against the live prod container — and the banner still showed. #1547 (release #1548, 23:05) found the real cause: prod builds carry a Google Ads snippet injected by vite.config.js that pushes three gtag entries before main.jsx runs, and GTM only honours a blocklist that precedes them; staging has no Ads id, so it could neither reproduce nor prove. The blocklist is now unshifted to index 0. Canary verified 23:15 (dataLayer[0] = blocklist, no CookieYes request, GTM + GA4 loading); promoted 23:12–23:20. #1548 also carried a nine-commit Drops batch (SB-13…SB-20 + a prosumer-api song_id::uuid compare so the 4-second status poll uses its index) that reached staging by a direct push — a mis-set upstream, since fixed.

Telegram: its own GA4 property, a For You feed, and per-launch attribution

Telegram analytics moved off the shared main-site GA4 property (where prod co-mingled with playmusic.gg on one stream) onto a dedicated property with four web streams — player/creator × prod/staging, each Measurement ID baked per build mode; the creator stream had been dead (#37). A traffic_type=internal QA flag keeps prod-stream testing filterable. #36: a For You feed screen — CSS scroll-snap with one IntersectionObserver, no gesture code, like the site's — with rankFeed as a fixed pipeline (pinned → curated-first → stable category-affinity sort → a diversity swap so no two neighbours share a song; the catalog has three "Tinfoil Hats"), a GET /api/foryou reader, every event attributed to the launch that started it, and the SDK + fonts served first-party so a filtered network can still boot. Earlier: /play opens the Mini App on the browse grid, and the prod webhook secret is set so the bot answers /start (#33–#35).

Five overnight hotfixes straight to main (Sep 18–19)

PG-85: context menu, selection and drag suppressed on the portal chrome, the consumer embed (the frame the shell cannot reach) and, engine-side, the whole game field with a text-field exemption (#1493, engine #510). Strict-lifecycle portals: the modal sends HOST_CAPS so the engine stops granting free revives there, and the HUD lives display is driven off the revive grant rather than FREE_REVIVES (#1496, engine #511). Back-merged before #1495 so the release didn't conflict in GameModalV2.jsx.

Engine

#509 (Sep 19, pairs with #1495): Drops UI theme preset + data-ui-theme stamp with a Drops-scoped flatten, self-hosted Anton/Spectral; Sky Rider creator-enabled collectibles with picked models, belly-to-earth rest pose, lean HUD, configurable off-beat wording, negativeFeedback config; Audio Flyer values chords like Music Runner, draws the whole chart, a colour collect is a collect; shooter boss gets real HP and opening spawn HP is capped to what the player can fire; Vector Beat Arena's loudest section stopped spawning (#500); coaster honours resultsMode 'continue'; a dead rig is caught instead of freezing the character, with a script that detects skin weights that tear the mesh; GLB sub-mesh transforms applied to vegetation instance matrices. #515–#518 (Sep 22): the aimer renders the backdrop image in Mania and Taiko; the Catch backdrop scrim is configurable.

Drops batch (SB-13 → SB-20)

Back from Style keeps the character pick; the preview names every game type, not 9 of 16; the success screen and the song page share the built game, not the song; the song-level Result card is gone; songs open in the basic view and return to Studio; GTM events no longer inherit the previous event's params; success-screen GA events name their song; creator's name casing kept in the greeting (SB-12); Rename-song allows editing the artist (SB-9); iOS fixes for monochrome glyphs, fixed-sheet shift and the marquee loop (SB-5/6/7).

Smaller

Alchemy on Sep 19: Play Preview shows the scene before the overlay covers it; Play swallowed the first press then went black then trapped; Tile Mode is Drop-only and Publish keeps it; Platformer stops spawning obstacles it doesn't have; the preview camera resets; the four slots stay on screen on mobile. Thumbnail backfills authenticate their show_all read; 122 catalog has_thumbnail rows repaired on prod Sep 22. Modal reads elevation consistently. Dev container for F5-debugging one API (#1419) and a write-jira-task skill (#1537/#1539) — no runtime effect.

Nearly drained

#1549: Alchemy's Obby builds a real course rather than a level strip; no human in the car modes; a picked ship stays picked. Open against staging: #1546 (the one lint error the SB-20 change left), #1536 (homepage promo popups on the /foryou landing), #1518 (song-pipeline revision), #1490 (tracks upload as file + JSON metadata).

Promote the config the code is waiting on

Much of #1495 is config-driven and dark on prod until wizard_config staging → prod is promoted through the CMS: Sky Rider options and entity slots, Alchemy definitions, feed lists, promo popup, contest CTA. Prod has no wingsuit/coaster rows at all. Then D1 0025_add_song_genre by hand + Sync to Edge in Game Library, or the music-genre list filters match nothing. Then republish the Horror Survival games published before the #1532 worker shipped. The coasterSkyRiderDefinitions test has been red since Sep 19 because #1483's staging D1 edit was never made — a CMS edit of the staging wingsuit row, not a code change.

Make the intro-gate experiment readable, then fix what the phone screenshot showed

In this order: (1) emit the assignment event for the control arm in creationIntroGate.js so a holdout leaves a record — a handful of lines, and until it lands the experiment returns nothing; (2) collapse the song banner on the game-mode step (h-[40vh] in StepSongPreview.jsx) — CSS only, four or five modes above the fold instead of one and a half; (3) geo-target the cookie banner in GTM; (4) re-measure the mode step alone. Keep the ads-side intro override out so the product's 50/50 stays clean.

Experiment plumbing before spend

The CMS split links (split_destinations) have never carried prod traffic; Experiment 1's Wizard-vs-Alchemy comparison needs one live. Confirm client_fp is populated on prod rows — still no fix in any release this window. Source maps for Sentry: sourcemap is off and nothing runs sentry-cli, so every crash report points at minified offsets.

Consent follow-ups the PRs named

A durable GTM trigger exception on pm_surface equals embed, then drop the blocklist line. CookieYes dashboard: mark the SuperTokens cookies Necessary, keep Support GCM on. Ask the Yandex reporter to re-check the draft. A playhop partner-registry key.

Still owed from Sep 15

Reconcile the 23 SuperTokens accounts orphaned by the Jul 27 → Sep 15 /me regression — no commit or script for it in this window.

Landing split at / — live and readable since Sep 19 ~19:56

Direct + organic arrivals at / are split between home and /foryou; landing_variant on every event in GA4 and analytics_events.event_params. Visitors assigned between 19:03 and the #1510 promote have a cookie but no stamped events — unreadable, not backfillable. Paid, Googlebot and Playgama traffic is exempt by design, so this reads organic behaviour only.

Create expectation gate — live since Sep 22, and not yet readable as an experiment

The banner is the treatment, not an obstacle: the question is conversion with it against without, with CTA click-through as the intent signal. Early paid-cohort numbers at the game-mode step (Sep 23 pulls, DB access has since been blocked): never saw it 79 → 43 chose a mode (54%), 23 made a game; saw it 29 → 12 (41%), 3 made a game. Dismissed 17 → 8 (47%); did not dismiss 12 → 4 (33%). Of 16 shown the banner, 4 clicked the CTA (25%), 8 dismissed it, 4 did neither and left. Directionally lower with the banner; not significant (p ≈ 0.23) and not to be acted on. Three things stop this being a real read, two fixable this week: (1) the control arm does not record itself — describeIntroAssignment is only called from CreationExpectationGate.jsx, which only renders when the banner shows, so a holdout visitor emits nothing; the only holdout rows are four the ads override forced, and the 79 "never saw" are holdouts, returning visitors and segment exclusions mixed together. (2) the ads-side intro parameter overrode the product's assignment for two days, forcing the banner onto people assigned to control — removed Sep 23, so the 50/50 runs clean from then. (3) 29 vs 79 is too small either way; ~400 per arm to detect a 10-point gap. The fix for (1) is a handful of lines in creationIntroGate.js — emit the assignment event for the control arm at the moment it is assigned. Until it lands the experiment consumes traffic and returns nothing. #1538's intro stamp and scripts/analytics/intro_experiment.py still stand for the treatment side.

Mobile wizard: the song banner owns 40% of the screen at the moment we ask for a choice

Screenshotted on a 664-px phone viewport: song banner 266 px (40%), "Choose Game Mode" heading ~80 px (12%), yellow tutorial callout ~150 px (23%), game modes ~170 px (25%) — one and a half of twelve modes visible. The banner is h-[40vh] sm:h-[45vh] in StepSongPreview.jsx, so it claims the same share of every phone regardless of size, and it was still an empty black box eight seconds after load. The cookie banner then lands at 3.2 s and covers the bottom 57%, starting 20 px below the song banner; for anyone who hasn't dismissed it the picker is effectively gone — which is what the strong geography correlation was measuring. That is the second problem, not the first, and the earlier read-out had them the wrong way round. Collapsing the banner on the mode step is a CSS change and would put four or five modes above the fold; geo-targeting the cookie banner is a tag-manager setting; then re-measure that single step — what exists is a strong correlation across three markets, not proof. The rest of the readout is unchanged: the cost movement, the loop journey, and the stable bad news that about half of games are never played.

Crash + load-attempt telemetry — first week of data

Sentry receives from 100% of prod since Sep 22; automation excluded and tagged. Read with docs/load-attempt-telemetry.md's SQL. Caveat: a missing terminal stage is "unknown", not a crash, and stacks are minified until source maps ship.

Consent defaults differ between prod and staging

The prod GTM container's CookieYes template has all six consent signals defaulted to granted for all regions; staging's table is empty so the template denies by default. On playmusic.gg, GA4 gets storage before any banner interaction; staging is opt-in. Flagged, unchanged — an owner decision, not a code fix. Partner embeds run GA4 with cookies and no consent mode; ruled as designed (the partner owns the consent relationship).

Correction to the Sep 10 read-out

The hp-vs-lp label was six campaigns, not one A/B — three search, three YouTube Demand Gen — and YouTube DG was 91% of tagged sessions with almost no human signal. Any homepage-vs-feed comparison has to drop it first; pooled, it made the homepage look ahead. Search alone is too small to call in either direction. Re-read is in progress; the Sep 10 tab's per-channel conversion rates stand, its landing-page rows do not.

Hotfix: new-account creation had been broken since 27 Jul

jsonb_build_array is variadic, so under server-side binding Postgres could not infer parameter $6 and rejected the whole INSERT on /me. OAuth succeeded, SuperTokens created the account, /me 500'd, no users row was written, the client fell back to guest and people retried with another provider. Introduced by #924 on Jul 26; the last non-guest row was 2026-07-27 11:43. 23 real accounts exist in the auth core and are invisible to the app. Fixed in prod (23eb6b02, same as #1407 on staging); verified the INSERT prepares and completes for email and null-email providers. Follow-ups: reconcile the 23 orphaned ids; re-read every sign-up number since 27 Jul.

Experiment 1 ("Create a game with [song]") is instrumented end-to-end on prod

?song=<id> opens Wizard Step 1 with the advertised song selected (b3b6a398); &mode=<id>[&submode=] advances past the game-type step, and a hand pick disarms the auto-advance for good (#1383 + Aikido follow-ups). 50/50 split short-links draw an arm server-side, pin it in a 30-day per-code cookie, append pm_arm, forward the ad's own query, and emit campaign_split_assigned for human deduped clicks only; the CMS Campaign Builder gets an A/B section with per-arm arrivals (6c6f9e46, migration 106). Plain /c/ links now forward gclid and utm_* too (9ec8fc90, 8ebd7aba). One creation-funnel vocabulary — creation_deeplink_arrive → start → step → preview_ready → publish / error / abandon, tagged with tool and arm, arm persisted across the sign-in wall — is fired by both the Wizard and Alchemy, which previously had no tracking at all (e5312ebb). creation_created is the terminal event; creation_publish means listed (6a2612d4). Google Ads receives Game Created and Mode Selected as conversions with the real labels (b48a0f13, 0a05de94). Staging E2E covers deep-link + funnel (7e0e2de1).

/foryou gets a way to Create without leaving the feed

The Ads landing surface is a fixed full-screen pager with no BottomNav, so the only way off it was the back chevron. A slim TikTok-style tab bar now sits over the pager with the four BottomNav destinations and Create promoted to a centre + (cf4ea30d); a new feed_create_click event carries card index and visit depth so we can tell whether the feed produced the creator. The registered XP pill is a real tap target and opens the missions drawer instead of /profile (57185907, a4c0a68c); bottom-nav labels are readable over light content (3fa62bb2).

Telegram mini-app: three production releases

A 32-commit staging→main promote (28ecc73) fixed the "stuck on Finding the Match" chain (leave stranding the next match, engine-spoke gating, deferred picker, engine addressed at contentWindow) and the room-config gameMode bug that made guests refuse the host's room; the same gameMode fix was ported to the TikTok activity. GA4 via gtag.js + LiveOps portal funnel landed, then four analytics defects were fixed in the production funnel (c6d55cd): every prod event was labelled environment=staging (the *.workers.dev hostname matched an includes("dev") test — LiveOps filtered to production showed no Telegram traffic at all); the #tgWebAppData launch fragment with user ids and the signed auth hash was reaching Google Analytics on every hit; scoreless runs dropped game_play_end/song_completed. Two more releases followed: a KV-backed leaderboard fallback with the global board derived from raw scores, and brand-2026 alignment (slate palette, site mark, card design, preview clips). CI: Telegram auto-deploys staging on push, production stays a deliberate dispatch.

Alchemy hardening (#1351–#1364)

White screen from a TDZ in the mode-options callback (#1364); mode cards get thumbnails and only the options that belong to the mode (#1362); Obby was unselectable until a song was picked (#1358); the slot panel confirms a selection, honours skip flags and gets the brand pass (#1359); characters come from the catalog the Wizard reads and FX is hidden on course modes (#1356); the platformer course no longer rebuilds on every parent render (#1351).

Run-end and retention on mobile

Play again is a reachable target on mobile and stands the run-end UI down when pressed (1d48e720, c68164d8); the progression flag resolves when the guest read times out, and the test checks the flag field rather than the config object (92e1fe5d, f8f96209); modal analytics lifecycle and teardown delivery hotfixed (18976f16).

Prod plumbing that was quietly failing

CSP now allows CookieYes hosts alongside Cookiebot — the prod cookie banner had been blocked since the migration (f0221895). The CDN answered 206 to plain GETs on /playmusic-pipeline/*, so card-image generation rejected its own reference frames (aafb02cb). Prod portal-cms was missing MEDIA_SERVICE_URL and GAME_PIPELINE_URL, so screenshot capture was unavailable (14c99e92). The Drops Doppler overlay was repointing consumer storage — prod NoSuchBucket (3e3ff5a7).

Smaller

Asset Library gains display-name editing (090e3db4), rename returns 404 for an unknown id (#1368). Dead Discord invite in header and footer and the "Music Music Games" typo hotfixed (fff5f8cd). Survival's env picker ignored its own exclusions (b668fe31). Engine main: the clicker/merge Play card vanished one frame after the mode showed it (2d222c65); no lane-switch line on a side-camera platformer (engine #487).

Game Lists — CMS-managed discovery feeds, now on /foryou

Editors build named lists (manual picks or rules over game genre / music genre / artist / song / game id) with a default sort and the filters players may change; resolved client-side with zero per-filter requests (60170e67). Built on the existing generator as a new wizard_config type game_feeds (D1 edge, Promote flow) rather than a new R2 pipeline; D1 0025 adds song_genre to projects. Hard rule: a game with no preview clip is never a list member. Product moved the player surface from /browse to /foryou and restored /browse byte-for-byte (05a4eaff): with no visible list the feed is exactly the previous one; with ≥1 list a second chrome row (list ▾ · Filters · sort ▾ · count) drives the pager and the view lives in the URL.

Alchemy reaches parity with the Wizard (#1399–#1428)

A game made in Alchemy now earns what one made in the Wizard — 150 XP, missions ticked, celebration toast, first-creation mark for guests (#1426). Publish is gated, the game screen's chrome is restored and Shuffle can no longer pick art that does not exist (#1422); the preview stops before Publish creates the game (#1423) and the published game has a way out (#1424). Mobile picker takeover, song-art parity, slot labels and publish identity (#1411); the world picker offers what the Wizard offers and a world you're given is one you can pick (#1427); the preview names the mode it is previewing (#1428); the engine's menu is hidden while building, play is full-screen, mobile gets an exit (#1406); tutorial held, auto-picks shown, finished game marked (b50421d9).

Worlds cleanup (#1429–#1433)

Skyfall is Sky Rider's world only; Boujee Drift is not a world and leaves the pickers; Boujee Night is withheld from modes that render a character; Skyfall, Backrooms and Boujee Night tiles resolve to new art. The Wizard decides the world step from the mode definition rather than mode-id lists (7dfc932c) and shows a World confirmation tile only where a world was applied silently.

Songs: PlayMusic Approved flag + drag-and-drop bulk upload

Informational "PlayMusic Approved" flag (migration 108) on single and bulk forms; new POST /api/tracks/batch-upload takes multiple MP3s with per-file error isolation, then a per-file table → shared fields → cost estimate → generation → live batch status (291e1dc4). Fixed along the way: the "artist required" check ran before ID3 extraction, so a blank artist could never be backfilled from the file. Per-song metadata is editable after a bulk upload.

Share attribution — shares were invisible, not absent

Every share copied a raw /play/<id> URL and never called the share API: share_links last gained a row on 15 Jan, and all 22 September share events carried short_code = null, so a recipient looked exactly like direct traffic. The resolver was fine (verified on prod: 307 with ?ref=, bad codes handled, click counted); only the caller was missing. Tracked short links are now minted on menu-open (a1747d16).

Also on staging

Portal config serialises each game's multiplayer flag (ccb373e2); GTM event-trigger regex widened to feed/identity/retention/session events (ea0a5dd5); migration 107 wired into both runners; wizard-config sync accepts game_type_order so it writes anything at all (b3e95314). Engine: Vector Beat Arena's loudest section stopped spawning, plus a song indicator (engine #500); wingsuit lean HUD by default with configurable off-beat wording; coaster honours resultsMode 'continue'; shooter boss gets real HP and opening spawn HP is capped to what the player can fire; a dead rig is caught instead of rendering a frozen character, with a script to detect skin weights that tear the mesh; audio-flyer values chords like Music Runner, draws the whole chart, and a colour collect is a collect not a detonation; MODE_CAPS coverage pin 24 (engine #481). Integrations: TikTok deploy falls back to staging not production; the MP lobby reads the engine's lobbyState; guest toast wording matches the host side.

Launch Experiments 1 and 2

Per the eight-week plan: "Create a game with [song]" → Wizard Step 1 vs Alchemy through one split link (~$500 / $500), and "Play rhythm games" → one curated page (~$1,000). Verify ad → /c/ → correct song → creation → publish, and ad → page → play, with attribution intact, before releasing full spend. The curated rhythm page is what Game Lists is for — it needs the next release, or a feed_order-pinned interim.

Staging → prod release

Game Lists on /foryou, Alchemy #1399–#1428, worlds cleanup, bulk song upload, share attribution, GTM triggers; migrations 107 and 108 plus D1 0025. Then an engine promote for wingsuit / coaster / shooter / audio-flyer / VBA and a wizard_config promote behind it.

Reconcile and re-baseline accounts

Fold the 23 orphaned SuperTokens ids into users; re-read every sign-up and is_guest figure since 27 Jul; restart the sign-up series from Sep 15.

Telegram

Deliberate Deploy Telegram production dispatch for anything still on staging; confirm prod events now carry environment=production and that no tgWebAppData reaches GA.

Sign-ups: series invalid 27 Jul → 15 Sep

0 non-guest users were created in that window because creation was broken, not because nobody tried — 23 accounts made it into the auth core. Last week's Ads read-out ("0 sign-ups, 5 guests") cannot be read as an ads result. Re-baseline from Sep 15; treat any account-based KPI in the Aug 6 → Sep 10 tabs as unmeasured.

Shares: 22 events in September, 0 attributable

Instrumentation gap (no short code minted), fixed on staging only. Keep "share → play" at n/a until the release ships; do not read the September zero as "sharing doesn't work".

Telegram traffic was hidden from the production LiveOps view

Until c6d55cd every prod Telegram event was labelled staging. Any Telegram number filtered to production before Sep 16 is zero by mislabelling; re-pull after the fix.

Experiment 1 readiness

Arm assignment is server-side per short code with a 30-day cookie and a forged-pm_arm guard, so the pill-vs-missions failure mode (no config key, split drifting 24–97% hourly) does not apply. Both tools fire the same funnel. Still to confirm before spend: client_fp populated on prod rows (the Sep 3–7 blind spot — no fix for it appears in this week's history), and that the split link survives the sign-in wall end-to-end on prod, not just staging E2E.

Production release #1325 (Sep 9) — pipeline scale, catalog security, guest retention

Staged 7 days of work to prod: pipeline asset-ingest scaling, catalog visibility hardening (auth + search perms), Alchemy brought into the live creator flow (feature-gated, safe to iterate), guest identity self-healing (#1350) so half-states don't mint duplicate guests, and analytics event QA (tracking backlog cleared).

Rollercoaster & Sky Rider become first-class game types

Both modes unlocked in the creator wizard as standalone game types, not variants. Engine now supplies correct defaults per mode; wizard overrides that forced wrong models got reverted. Rollercoaster got dark-mode support (#1335). Sky Rider dark-mode variant now available too.

Alchemy editor completed across 13 PRs

Shipped in pieces Sep 3–10: CMS controls for Alchemy modes (#1337), asset slots UI (#1342), fill-the-blank mode messages (#1344), True Platformer course (#1346), theme toggle (#1348), publish payload (#1347), song sections UI (#1343), camera & FX controls (#1339), world pins & player slot (#1338), and environment names (#1349). Platformer reference-error fixed (#1341). Isometric view dropped from wizard config (#1340).

Guest identity self-healing (retention win)

Laggy publishes + dead previews + creator flips left guests with session cookies but missing localStorage tokens, triggering the create gate to mint a second guest identity on next visit. Now detects and merges half-states into one identity (#1350). Tests locked in.

Beat Dance preview two-character body fixed

Preview showed engine's runner character + selected dancer simultaneously. Scoped runner suppression to real games only; song preview now correct from start. Dance mode-init now seeds the selected character before apply, so preview and game show the same model (#1332).

Search converts 59× better than YouTube (but tracking gaps inflated both numbers)

Search: 64 of 770 sessions started a game (8.31%, 95% CI 6.6%–10.5%). YouTube: 11 of 7,756 (0.14%, 95% CI 0.08%–0.25%). The gap is real (p < 0.0001). But YouTube's tagged-session count is ~11× too high because the app's browser wipes its memory each session, so every load looks like a new person; search inflated less. YouTube's 0.58% "human activity rate" (tap, scroll, or play) vs search's 14% is the finding: YouTube delivered almost no people, not YouTube people didn't play.

Landing-page question (homepage vs /foryou feed) unanswered

Search showed feed → 9.4% conversion vs homepage → 6.9%, but needed ~900 tagged sessions per page to tell if real; test had 469 feed, 291 homepage. YouTube showed 0.16% feed vs 0.08% homepage, also too small. Statistically, both differences are noise (p > 0.23). Takeaway: need ~1,800 total search sessions per test repeat to answer this.

Pill vs missions A/B test didn't run

Configuration key to randomise pill vs missions display never shipped to production. Split swung 24%–97% hour-by-hour (a real coin flip would stay ~50%), so no experiment took place. Pill rates: search 10.5%, YouTube 0.22%. Missions: search 8.3%, YouTube 0.08%. Differences are not trustworthy.

Sign-ups and game creation stayed flat

0 sign-ups (5 accounts created were all guest). 7 games created (0 published). 19 of 361 games were played at least once. Campaign generated 7–10× traffic increase but no creator lift.

Key unknowns: device fingerprinting, landing-page tagging

Device fingerprint (client_fp) was empty on every production record during the test—the one field that would answer "is YouTube real people or app-auto-fetches" outright. Tagging was inferred after the fact by landing-page assignment; a proper tag in the ad link would settle homepage vs feed. Both fixes are done on staging and ready to ship. Next test should run with both live.

Production release — three weeks live on playmusic.gg

Everything from Jul 12→21 — SEO/LLM discovery, the performance pass, portal LiveOps & retention, catalog thumbnails, PlayHop — reached production in one staging→main push (#947). It bundled an auth reconciliation: a migration (094) that backfills linked account IDs and merges the duplicate accounts that formed when the same person signed in across staging and prod.

The creator wizard becomes a game-mode toolbox

Horror Survival is now its own mode (with a Horror Runner submode), Spy Hunter joined as a vehicular-combat play style, the shooter gained UA Mode and Zen Flow styles, and Merge became a selectable type. New environments (boujee night/drift, backrooms) and skydomes surfaced too. Creators now pick from a genuine menu of genres instead of a single runner.

CMS-defined asset slots per game type

The CMS now decides which asset slots — traffic, enemies, obstacles, collectibles, characters, environments — a given game type exposes (#934, #942), and players can opt into extra slots (#926). Content, not code, shapes what each mode lets you customize.

Wizard-config snapshot / restore + auto-sync

Wizard configuration now syncs additively from code on every deploy (#931) and can be snapshotted and restored from D1 (#932), so a bad edit or a missed seed can no longer silently wipe modes or environments.

Home create surface polish

featuredCreate modes moved onto the Home page (out of the /create picker) with a readable "New" badge, and the create banner now sits under the Featured Drop hero.

Contests & Promotions promote-to-prod

Contests and promo campaigns can now be promoted staging→prod through the same lineage-tracked flow as games (#923); contest migrations wired into both runners.

Creation funnel visibility

New LiveOps view into the create→publish funnel plus per-song completion attribution (#937).

Play counts stop jumping

The homepage/modal play count now reads a live D1 edge counter (#941), so it no longer flips between a stale replica and the live number.

PlayHop distribution live

PlayMusic games now run inside Yandex's PlayHop portal via its SDK (#825) — a third external surface alongside Discord and TikTok — including the Yandex SDK integration and its auth/player quirks.

SEO/LLM discovery pages

The discovery layer went from hubs to a real graph: directory pages (#849), entity-graph IDs linking songs/artists/creators/games (#848), LLM-oriented hub enrichment (#850), an is-indexable gate wired via migration (#839), and a sitemap/robots audit (#834) that cleaned up what search engines and AI crawlers see.

Performance pass — bundle + images

The main JS bundle dropped from ~809 kB to ~625 kB (SuperTokens UI out of boot, path-based chunking) (#871), the mobile LCP chain was broken up with hero preload (#870), and wizard/catalog images now route through CDN resize — roughly 11–20× smaller (#917). Noticeably faster first load, especially on mobile.

Cross-environment auth hardening

Signing in with a second provider now merges into the existing account instead of creating a duplicate (#886); the OAuth connect flow was locked to same-origin with CSRF nonces (#885); stale cross-env session cookies get cleared (#879); and cached API responses vary on Origin to stop cross-origin CORS bleed (#877).

Contests feature

A Contest Hub with a homepage entry point, landing pages, and CMS authoring (#914), a brand-styled leaderboard with podium + player position (#915), and contest-rules links in promo popups (#913).

Draft-first player games

Player-created games now start as drafts; publishing is a PlayMusic-gated step (#916), plus a post-run "Create your own game" prompt on playmusic.gg.

Portal LiveOps analytics

Retention split by acquisition source × destination — playmusic.gg vs portals side-by-side (#861, #863) — and play-count consistency fixes (#857).

Engine — new play styles

A full shooter pass (walls, power-ups, boss, HP display, weapons), a config-bootable Horror Survival, a Spy Hunter vehicular-combat mode, and a new 2D platformer package with double-jump, gravity-flip, and on-beat scoring.

Create → Publish shipped to production

The new two-step draft→publish flow went live (#732, #738): players create a draft, then publish as a distinct, gated step. This was the release's highest-risk item and cleared prod after a staging publish-route fix.

Catalog thumbnails end-to-end

Song/character and asset thumbnails now render across the CMS asset library and game cards, backed by a has_thumbnail flag (#718), ingest-time rendering (#796), and cache-busting on rotation — no more initials-only placeholders.

Playgama ads hardened

Ads now pause the whole game (not just audio) on overlay (PG-35/37), respect multiplayer eligibility (PG-39), carry legal/noindex correctness (PG-44), and route share correctly on TikTok/Discord (PG-40) — the fixes that make monetized play feel intentional rather than jarring.

Faster embeds

Embedded games got edge-cached config (#788), engine preloading (#789), and deep-link slug resolution (#790), cutting the time from click to playable when a game is embedded elsewhere.

Search & Social report live

The Soundcharts + Google Trends CMS report (admin-only) went to production with its migration and scheduler.

Engine — tutorials & first shooter

A full audit/rewrite of every game and sub-mode's instruction copy (#310), plus the first UA-Mode shooter with a generic projectile/weapon pipeline (Pistol).

Engine — flight & platformer fixes

Per-mode flyer orientation, dynamic floor, collision anchoring, and a single config-driven platformer health/lives model.

SEO & social plumbing

Organization sameAs socials (#807), a Discord social footer (#805), and robots rules that disallow internal routes (#809).

Gameplay-video sharing suite

Players can record a run and share it with rich social links (#696), a video library (#697), and an in-player share (#698) — turning a play session into shareable content across social surfaces.

Personalized home

My Games and Followed Creators shelves now inject over the CMS home layout (#687), plus a "create your own" card (#684) and the brand-2026 polish — Featured Drop, posters, badges (#702).

Content takedowns + enforcement

A global takedown system that hides taken-down content everywhere and blocks republish, with coverage logging (#659, #661, #675) — the compliance backbone for a UGC catalog.

Engine — Alchemy asset picker & Match3

The Alchemy slot panel became a "fill-the-blank" UX driven by live catalog data (#280), a Match3 mode family shipped with four forks, and a SIDEKICK avatar builder PoC (customize + animate) opened up.

Search & Social report + Director NLE

First cut of the Soundcharts / Google-Trends CMS report (#703) and a Director prompt/NLE mode with fuller mode coverage (#704).

Creator wizard batch (#706–#717)

Multiplayer defaults on in-card, unpublished-song characters surfaced, flight preview live-swap, model-less song handling, and a raft of scroll/preview fixes.

Share Hub / Invite Players

A refreshed share hub with an Invite Players component and QR codes (#719).

Ingest & catalog correctness

Vehicle category (#713), idempotent incremental sync (#721), and is_public preserved so CMS "hide" toggles survive re-ingests (#725).

Prosumer — first end-to-end vertical slice is live

The independent-artist self-serve track went from architecture doc to running code: a full slice spanning the API, a dedicated frontend, database migration, and a staging deploy (#455). We then worked through its first real-world bugs — sign-in was 500-ing and song upload was 502-ing, both traced to the frontend's own nginx proxy (SuperTokens headers not forwarded; API resolver + Host header; a reserved PORT env blocking the gateway route). Artists can now stand up on the prosumer frontend on staging.

SEO/LLM — full indexing + landing-page builder (P1–P5)

Last week's AEO foundation grew into the real thing: SEO/LLM indexing with a block-based CMS landing-page builder, image/video content blocks with upload, and the ability to link songs/artists on a game page and manage Song/Artist hubs in the CMS. Migrations wired, human-preview sign-in gate fixed. This is the machinery that turns curated song × genre combos into publishable, discoverable pages.

TikTok Mini Game — Discord parity + multiplayer (on staging)

A third distribution surface came together this week: the TikTok Mini Game landed on staging with full feature parity to the Discord Activity — games browser, scores, XP, song progress, likes/achievements, game-proxy, and multiplayer signaling — plus a two-tab dev harness so multiplayer can be exercised without TikTok partner access (one browser acts as streamer + viewer against the real engine). A cross-distribution platform-profiles framework was also drafted, with TikTok as profile #1. Open PRs to staging (integrations #16, PMH #585), not yet merged to production.

Preview capture — real gameplay, finally

The preview-video pipeline landed on the approach that actually works: capture real gameplay via page screenshots at speed, starting games with a DOM click (Playwright's normal click was hanging on Cloud Run) instead of frozen frames or engine-side exports. Preview videos now show the game genuinely being played — the missing piece for rich, trustworthy game cards.

Home deep-link control

A global home-deep-link default plus a per-game tri-state override in the CMS — fine-grained control over whether a home-screen click routes straight into a game or to its detail page.

Genre & Song/Artist hubs

Game pages now link their song and artist; hubs show all categories with a published-fallback; game-page hub links wired end-to-end — part of the same discovery push as SEO.

Asset rotation & CDN correctness

Model load + rotation fixes across multi-storage backends, credential handling, and a CDN purge on rotation so the immutable-cached viewer reloads the new model instead of a stale one.

Wizard camera & game-mode fixes

Camera reset-to-standard and the top/flight/obby view fixes (partial APPLY_CONFIG no longer stomps a picked Top variant); flight generated-ship-on-floor and drop/obby fixes in the engine; a 4-player game-over guard and leaver-forfeit result for multiplayer.

Discord multiplayer — reconnection fixed + single-socket gateway

Discord Activity multiplayer had two structural problems, both now fixed. Rooms died on an idle drop — redundant reconnect attempts were tearing down healthy connections and never recovering, so you had to relaunch the activity. And you couldn't start a fresh lobby after a match — Discord only allows one live connection per activity, so opening a second one just hung.

The fix is a new single-socket gateway: one persistent connection for the whole session, owned by the outer activity frame (which survives Discord's menu navigation), with rooms switched over control messages instead of new connections. It's behind a flag with a one-line rollback. We also tidied the lobby/spectator experience: watch-instead choice on invites, switching to spectator mid-match, spectators see the results screen, and a host "force-start" to unblock stuck lobbies.

Playgama PoC — rewarded ads through the vendor layer

We wired Playgama in as a first-class ad vendor with a new ad manager that routes rewarded-ad requests with CMS-driven vendor selection — proving the end-to-end monetisation path on a partner channel. Sits on top of last fortnight's session-tracking fixes, so Playgama sessions now report clean length data alongside the ad path. The vendor abstraction means the next partner plugs in without a rewrite.

SEO / LLM discovery pages — AEO foundation (in flight)

A new workstream to make curated games discoverable by search engines and LLMs: llms.txt, server-rendered game-detail pages with structured data (JSON-LD), AI-bot crawl rules, and per-game SEO overrides in the CMS. The model is deliberately curated, not automatic — we hand-pick the song × genre combinations worth owning, gated by a single editorial toggle that defaults to no-index. Architecture is merged; the first feature slice (P1a) is on a branch, not yet to production.

Game previews — big push

Manual preview-video upload and preview controls in the CMS Game Library, hover-to-play preview videos on the consumer feed, and a "has preview video" filter. Capture got much more robust across every mode — dismissing the start screen reliably, capturing actual gameplay rather than the menu, and not mistaking rhythm-mode audio for "song complete." Plus an in-engine 3D model preview in the CMS that shows real in-game orientation.

Wizard & vehicle-picker fixes

Camera perspective only resets on an actual game-type change; the mode accordion centers when it opens; the character/vehicle picker got a right-edge fade so hidden cars are discoverable; going Back from Step 3 keeps your picked vehicle; song-generated vehicles now reuse the catalog thumbnail.

Multiplayer correctness — drift & flight

Resolved the phantom 3rd car in drift and the oversized opponent ghost; fixed flight-ghost orientation, grid offset, flicker and glide smoothness — all derived locally with zero extra bandwidth. Drift traffic is now deterministic across screens (host-seeded, fixed-step). Results screens support 3–4 players and abandoned matches resolve to the survivor.

Skydomes & camera (preview parity)

Picked skydomes now render correctly in the wizard preview across aimer and obby — resilient texture loading with retry and CDN fallback, fixed an upside-down skydome bug, and the camera re-applies on perspective changes. Also: the aimer wizard preview now autoplays a muted demo so you can see real gameplay before generating.

Relay repo split complete

The multiplayer relay is now its own repository with CI/CD at parity with the engine — branch-driven deploys, secrets via Doppler, version drift-checking, and contributor docs. The relay side of the single-socket gateway shipped here too.

Auth account-linking — true merge + reconciliation

Signing in with a second provider now resolves to your existing account instead of creating a duplicate (#886), backed by a new user_identities model and a backfill (migrations 091–094) that merges accounts already duplicated across staging and prod. The OAuth connect flow was locked down — same-origin, CSRF nonce, no unauth fallback (#885). This was the release's highest-risk item: it was reverted once on Jul 21 (#896) and re-landed after hardening.

Creator wizard — new game modes & CMS asset slots

Horror Survival (now its own mode) plus a Horror Runner submode, UA Mode & Zen Flow shooter styles, Spy Hunter on Drift Drive, and new environments (boujee night/drift, backrooms) + YSTO skydomes + a Merge type. The CMS now defines which asset slots each game type exposes (#934, #942), players can add extra slots (#926), and settings show/hide via CMS-authored rules (#944).

Contests

A Contest Hub with a homepage entry point, landing pages, and CMS authoring (#914), a brand-styled leaderboard with podium + player position (#915), contest-rules links in promo popups (#913), and Contests + Promotions added to Promote-to-Prod (#923). Migrations 0011/0012 (Postgres) + 0024 (D1).

Wizard-config safety net

Wizard configuration now syncs additively from code on every deploy (#931), with no-store reads and D1 snapshot/restore workflows (#932) — a bad edit or missed seed can no longer silently wipe modes or environments in production.

Draft-first player games

Player-created games now start as drafts; publishing is a PlayMusic-gated step (#916), plus a post-run "Create your own game" prompt on playmusic.gg (#908).

Front-end performance pass

Main JS bundle cut from ~809 kB to ~625 kB (#882), the mobile LCP chain broken up, wizard/catalog images routed through CDN resize (~11–20× smaller) (#917), and a boot-smoke gate to catch dead bundles.

LiveOps & play counts

Creation-funnel visibility + per-song completion attribution (#937), and the D1 edge play_count kept live so homepage/modal counts stop jumping (#941, #907).

Home / modal polish & worker CSP

Featured Drop banner placement with readable "New" badges, aspect-safe hero media, share unblocked on desktop/iframe surfaces (#905), and engine-iframe frame-ancestors + CORS fixes (#900–#903, #877).

The two-verb spine

Press play. Can't find it? Make it. — Press vs make: playing is the default, making is the fallback. Keep that contrast intact wherever both actions appear.

Reference lines (calibration set — pass all eight personas)

• Play games to the music you love.
• Press play. Can't find it? Make it in under a minute.
• Your music. Any game.
The test: if a new line would make the Explorer, the Curator, or the Philosopher wince, rewrite it.

Tier 1 — Homepage (broad, hand-written)

H1: Play games to the music you love
Subhead: Any song, any genre — from flappy flight to endless runners. Can't find your combo? Make it in under a minute.
CTA: Press play · Can't find it? Make it →

LLM-quotable definition (the line an LLM quotes back)

PlayMusic lets you play games set to any song or artist — across genres like flappy, runner, drift, obby, and rhythm — and create your own in under a minute when the combo you want doesn't exist.

Tier 2 — Genre /play/{genre} · Tier 3 — Song /play/song/{artist}-{title}

Both fill {variables} from the same wizardMeta.gameTypeDefinitions the wizard renders from, plus the song catalog. Add a genre in the CMS → its landing page exists with correct copy, no second edit. SEO surfaces lean plainer still — shaped like the query, genre/artist leads the line.

We can finally see load times

Until this week, partner complaints about "45-second load times" were a guessing game — we had no data. The game now reports back exactly how long each step took (initial download, configuration, game assets, audio) for every play session, on every partner.

New "Portals" dashboard in CMS

Anyone in the business can now open the CMS, go to LiveOps → Portals, and see per-partner load times, trends over time, and how many players are coming back to replay. Sparklines on each partner card give an at-a-glance health check.

Drop-out tracking

We also capture when a user bails before the game finishes loading — historically these players were invisible. Now we know how many we're losing and where.

Smart asset loading

The game now adapts how aggressively it downloads assets based on the player's connection and device. Fast connections on good devices get more in parallel; slow phones get protected from getting overwhelmed. This was originally scheduled later in the roadmap but came along for free.

Hover-to-prefetch on portals

When a player hovers over a game card for 200ms, we start loading the game in the background. By the time they click, it's already partway loaded — feels instant.

Discord embeds report in

Discord-embedded games now route through the same analytics pipeline as everywhere else, so we'll see Discord performance alongside Playgama, Y8, etc.

The new load-time tracking works inside our own site (playmusic.gg). It doesn't yet report from games embedded directly on partner sites like Playgama or Y8, or from our portal shells. So tomorrow's prod dashboard will show consumer-site data first; partner-embed coverage is the next follow-up.

Y8 partner attribution fixed

Y8 traffic wasn't being properly credited in our funnel — every session looked like it came from "unknown". Fixed: Y8 sessions now show up correctly in partner reporting, so we can have a real conversation with them about volume and payouts.

Discord Activity foundation

The "game-inside-Discord" launch surface is now mostly in place — Discord Activity UI redesigned, share + mobile UX polished, analytics flowing through the same pipeline as our other partners. Featured Artist Experience landed with STYNGR admin controls for curating which artist gets the prime slot.

Playgama feedback addressed

Items Playgama flagged during partner review have been closed out and shipped to prod.

One-click "promote to prod" for partner content

Previously, getting a new song or partner-portal update from staging to production required manual coordination across systems. Now it's one click from cms-staging — for the team running partner relationships, this is daily friction gone.

New game mode: OSU (rhythm tap)

Full rhythm-tap mode landed in production, with a polished wizard flow for picking it during game creation. Expands the catalogue beyond runner/drift/flight into the rhythm-game category.

Wizard rebuilt around real data

The "create a game" wizard used to have a lot of hardcoded options. It's now driven by actual config — so when content/engineering add a mode, it shows up in the wizard automatically. Less drift, less "why isn't X showing up?".

CMS song delete + maintenance polish

CMS got a clean song-delete flow (previously a support ticket) and maintenance-page polish so partner-visible downtime looks professional.

LiveOps got drill-down views, music data quality, and trend visibility

Five separate LiveOps improvements shipped in two weeks. The internal ops tool keeps getting tighter — drill down from a top-line metric into the underlying sessions, see music data quality at a glance, catch errors that used to go silent. Less time spent asking engineers "can you look this up for me?".

Better partner reporting infrastructure

D1 + Worker Redis runbooks finalised; observability fixes shipped. When something looks weird in the LiveOps dashboard, we now have actual playbooks for triaging it instead of inventing one each time.

Two same-day production hotfixes

A bug made collectibles invisible in some game modes (May 4); fixed same day. A separate bug caused phantom "MISS" popups in Cyber Run / Cosmo Run (May 5); also fixed same day. Both were caught before they hit large player numbers — operational maturity getting better.

Flight-mode ship revive UX

Smoother revive flow when players crash in flight mode — small UX nicety, but the kind of polish that compounds.

Mobile UI polish on Discord/partner shells

Mobile drawers, auto-hiding bars, share-link flow refined — the embedded-game experience feels closer to a native app.

Production load tolerance improved

Database connection pooling (PgBouncer) extended to production, memory increased on the busiest services, CPU throttling tuned. Net result: we can take more concurrent players without falling over, and OOM crashes that used to happen during traffic spikes are gone.

Engine release process formalised

The game engine now has a proper "canary cutover" promotion system — new versions go to a small slice of traffic first, with one-click flip to full traffic once we're confident. Riskiest part of any release (game engine itself) is now the safest.

Director Studio — full NLE timeline + AI cue generation

The engine-side of Director Studio matured into a real non-linear editor: multi-lane timeline, drag-edit cues, scrubbable ruler (SEEK_TO_TIME to scrub the running game), and seamless live mode-switching (APPLY_MODE) with stem crossfade between levels. Switch game mode / map / VFX / song / camera mid-song without a restart.

New & expanded game modes

Drift Runner (music runner with a steerable car), Crash Course (grouped as a Drift sub-mode), the Aimer family (OSU / Taiko / Mania / Catch) with configurable Builder/Breaker stem game-over and Drop integration + fever danger, Obby multiplayer, and a Horror Survival preset (backrooms popout zombies). FlyTap collectibles made reachable; Shooter HUD + enemy preload fixed.

Multiplayer — 4-player, mid-join, spectators

Engine threaded explicit 4-player + spectator caps, a mid-game join flow, and ghost fixes across Drift/Flight/Obby (orientation, phantom-vehicle dedup, ghost height/animation, input gating). Mid-join spectators now start rendering, not just players.

Template kit — Lego-core + env capability bounds

Phase 1–2 of Lego-core: standalone registries with parity-tested canCombine() and a harness-v2 lego route — the seam for mix-&-match templates and mechanics. Plus an env capability-bounds system, Synty Apocalypse traffic pool, and drift/crash-course driving extended to ground+road environments.

SIDEKICK avatar builder (PoC)

Proved live face/head + hair swap on a shared body via slot-agnostic equipPartGLB, natural arms via A-pose re-bind, and head-only hair binding. Foundation for a customizable character pack.

Alchemy mixer redesign

Slot-panel "fill-the-blank" UX wired to live catalog data, More Songs/Games dropdowns, and stable character alternates.

Discord Activity — live channel

The Discord Activity surface went from foundation to a working channel: redesigned activity UI with portal data + brand styling, mobile drawers / auto-hiding bars / share-link flow, Featured Artist Experience with STYNGR admin controls, and the game iframe tagged ?portal=discord. Activity analytics now route through the /game-api proxy and carry portal slug + user-agent in the envelope.

TikTok auth + new embedded channels

TikTok added as an auth provider (SuperTokens authorize-URL fix). Embedded-channel scope now spans Discord, TikTok, and partner portals — with Reddit / Telegram flagged as next.

Partner portals — faster, cache-correct, measured

Portal cold-start fix Phase 2 (edge-cache /api/portal/config/* in KV); scoped edge-cache purge on portal create / delete / config-save; FRVR dev harness + reliable FRVR ad triggering with correct revives; CMS "Test Portal" button + automated suite; per-portal Clear Cache. Partner portals: Y8, PlayGama, FRVR, PlayHop, CrazyGames (leaderboards disabled).

Campaign builder + UTM shortener

CMS Campaign Builder with a general /c/ short-link shortener and full UTM ingestion; environment-specific short/destination URLs; gateway routes /api/c/ to player-api for campaign links. Closes the loop from promo card → tracked session.

Attribution survives the whole funnel

First-touch UTM/referrer persisted across OAuth; GA4 ↔ Playgama join keys upgraded; shared pm_session_id/pm_user_id stamped into the GA4 dataLayer; CF geo (CF-IPCountry → X-Country) forwarded end-to-end so analytics country isn't NULL. Abandoned sessions closed via pagehide beacon + hourly reaper.

Game wizard — accordion picker, persistence, funnel tracking

Accordion-style game-mode picker; persistent wizard state so reload/back recall the engine preview in sync; loading screen on re-sync; full end-to-end funnel instrumentation (wizard + Director Studio). Real game-mode labels on Profile + CMS cards (was always "Runner").

Preview videos (P1)

Capture pipeline for hover-preview videos landed — the missing piece for richer game cards (alongside file upload, in-game capture, and AI generation).

CMS — themes, songs, portals

Operator-editable display names for consumer themes; ISRC/UPC/label/parent-label persisted on song upload; per-tab delete for custom portals; cron layer for portal-cms operationalized (never ran in prod before).

Security & reliability hardening

ADMIN auth enforced on CMS portal/game management; SSRF guard on catalog fetch; staging/canary hostnames blocked from indexing; warm instances kept on user-facing services; remaining console.logs DEV-gated. Multiple Aikido review passes closed out.

Live Babylon engine + per-game config.json everywhere

Finished migrating every game-load path to use the live Babylon engine + per-game config.json. Removed the last legacy "baked HTML export" code paths from the consumer frontend, CMS, and the play API.

Portal import populates engine config

Portal imports now correctly populate game titles and song metadata in the engine config.

AI chart review hardening

Added robustness to the AI-driven chart review pipeline: handles truncated Gemini responses, configurable model, 16K token limit, robust JSON parsing for double-encoded responses.

Portal attribution end-to-end

Every game session played from a partner portal (Playgama, Discord, Y8, etc.) now records which portal it came from. Spans the engine, consumer frontend, and backend; backed by a new game_sessions.portal_id column. Critical for partner reporting and revenue attribution.

Wizard song analytics

When a user picks a song, navigates the wizard, and creates a game, we now track impressions, clicks, and "song builds" — feeding the LiveOps Songs leaderboard for editorial decisions ("which songs should we refresh?").

Share metrics fix

Share funnel had been silently broken since March — frontend was firing one event name and backend was looking for another. Fixed; share metrics will populate from now on.

Tween-friendly share menu

Dropped X/Twitter, added Messages (iMessage/SMS), reordered for tween-relevance. Native Discord share works for connected users.

Rich link previews

Share links pasted into Discord/Slack/iMessage/Twitter now unfurl as rich cards with the game thumbnail, title, and description. Previously rendered as a generic placeholder.

Music integration semantics fix

A pipeline-vs-engine vocabulary mismatch ("drop" / "break" / "breaker" weren't valid engine values) was causing audio to drop out and never recover on some songs. Both the publish path and the engine's runtime are now defensive.

Admin security hardening

All /api/admin/* endpoints now require ADMIN-role authentication at the router level. Removed two shared-secret backdoor endpoints that were guarded only by a hardcoded string in the codebase.

Playgama partner experience

New CDN-cached path for engine assets (immutable headers); portal D1 sync respects curated CMS data instead of overwriting; tab-pause/resume on the game modal; vehicle-URL safety fix to prevent cross-mode contamination.

Quick Builder cross-link

CMS sidebar's Quick Builder now opens the consumer-app /create flow on the matching environment.

Boujee Baby (Kendrick) — Drift mode

New environment (NYC streetscape, taller buildings, highway signs, custom sky), rigged character riding a Taxi Drift vehicle, organic camera pull-back, custom collectibles (flowers + trash), uniform Office wall with 8-color variety, denser lamps, full song-spaced gameplay tuning. Demonstrates the pipeline producing a polished one-of-one experience for a specific artist.

It Takes A Village (Canyon) — Platformer mode

New canyon environment with 3-height collectibles, double-jump + extended hang-time, "no game-over" forgiving mode, rolling wheels, diamond collectibles, song progress bar, always-on tutorial, results-screen polish. Audio unmute fix shipped alongside. Shows platformer mode now ready to host artist content, not just abstract levels.

Why these matter

Both are end-to-end proof that the song → concept art → 3D → game pipeline can deliver branded artist experiences. They also surfaced and forced fixes on real edge cases (vehicle wheel preservation, character swapping, audio fallback, void-damage modes) that now harden the engine for everyone.

Real-time relay infrastructure

Cloudflare Workers relay with Durable Objects GameRoom per match. WebSocket hibernation recovery so a worker restart doesn't drop in-flight games. 30Hz tick rate with smooth client-side interpolation between state ticks. Production relay URL wired up in the harness.

H2H (head-to-head)

Full ghost renderer with synced jump, slide, and celebration animations; lane sync; vertical-curve following on hills; postMessage integration with parent frame; dedicated MultiplayerHUD; H2H results screen.

Duet co-op mode

Stem-builder co-op where two players cooperatively trigger stems. Synced stem state, host config sharing on room join, dedicated co-op results screen.

Lobby + spectator system

Lobby with countdown, ready button gated on config load, room-state config persistence in Durable Object storage. Spectator camera + spectator lobby — anyone can watch live multiplayer matches without joining.

Customization

Ghost color picker, opacity slider, emote system, room settings panel, host toggles for ghost/score display. Player avatars with speed multiplier. Auto-removal of stale ghosts.

Currently gated

Modes restricted to Music Runner + Duet during testing. Flight and Drift disabled in MP for now — they'll come online as physics-sync mode matures. Mid-May ships the gated set; Q2 expands across all 10 modes.

Frame-ancestors + trusted origins for partner embedding

CSP frame-ancestors updated to allow Y8, Playgama, GamePix, and localhost. Portal-shell can be embedded in y8.com iframes without breaking. Same plumbing extends to Discord Activity iframe at launch.

Portal attribution end-to-end

Game sessions now record portal_id at the engine level — Phase 1 (engine sessions) and Phase 2 (frontend + backend regex matching) both shipped. When Discord Activity goes live, it'll plug into the same attribution pipeline with no additional work.

Embed-mode UX

Login button hidden in /e/ embed routes. Recording.js skipped in portal/embed contexts. Tab-pause/resume on the game modal. Cumulatively this is the embed UX Discord Activity will use.

What's left for Mid-May launch

Discord Activity SDK integration, Discord OAuth scopes, lobby/voice channel awareness, and submission to Discord's developer portal. Underlying iframe + attribution + embed UX is already production-tested via Y8 / Playgama traffic.