UPDATEE:: — stats, brave stt actually works, hydration fix

tl;dr

• deepgram stt works on brave now (and edge) — words dissolve, wpm moves, no more fake “google speech blocked” errors
• root cause: brave kills the vercel http audio bridge + we fell back to browser speech (which shields also blocks)
• fix: brave/edge → render websocket proxy (wss://monkeyspeak.onrender.com/...); chrome/firefox still use vercel bridge
• fixed utterance_end_ms=400 deepgram 400 spam (minimum is 1000 on live listen)
• fixed bridge sending keepalive json into the pcm stream (corrupted audio, empty transcripts)
• stats page + profile hub — streaks, achievements, session history, same minimal desk ui as home
• live: https://monkeyspeak-delta.vercel.app
• repo: https://github.com/nothariharan/monkeyspeak (commit might differ by the time u read — i keep pushing)

---

CHANGES::

deepgram / stt (the painful part)

routing by browser

browser path brave / edge websocket → render proxy → deepgram chrome / firefox (prod) same-origin POST /api/deepgram/live bridge on vercel localhost ws://localhost:8080 when backend is up, else bridge

🐜🪲🐞 that made it look dead

1. utterance_end_ms=400 — deepgram returns http 400 on live ws below 1000ms. was baked into client urls. clamped to 1000 everywhere (client, bridge, render backend, server.js)
2. keepalive in pcm body — http bridge was writing {"type":"KeepAlive"} into the upload stream every 3s. deepgram got garbage bytes, transcripts stayed empty, no obvious error in ui
3. brave + duplex fetch — bridge connects, mic on, waveform moving, but ndjson response never streams back. zero words dissolved
4. bad fallback — deepgram fail → auto webspeech → brave shields block google → ui says google is blocked even though user picked deepgram. now we don’t fallback on brave/edge when deepgram is preferred

other STT polish

• bridge waits for BridgeReady before marking session ok
• server-side keepalive via deepgram sdk on the bridge route (not in pcm stream)
• removed verbose [STT:deepgram] console spam from prod — debug only with NEXT_PUBLIC_DEBUG_STT=true
• gsap scale not eligible for reset warning fixed on momentum meter

stats + profile

• new /stats dashboard — totals, streak, recent runs, wpm trend charts
• profile hub drawer — nickname, activity heatmap, achievement badges
• shared stats-* / desk css so stats doesn’t look like a different app
• achievements + streak logic with unit tests (lib/achievements.ts, lib/stats/streak.ts)

UI / polish

• home + results + settings aligned to same minimal desk style
• daily challenge card, error/not-found pages match the rest
• monkey favicon + readme/docs polish

Infra

• GET /api/deepgram/proxy-health — same-origin render health check (browser never hits render directly, no cors scream)
• render monkeyspeak-dg-proxy redeployed via render cli

what actually happened on brave

1. user picks deepgram → app tries vercel http bridge
2. brave never delivers streaming ndjson reliably
3. fallback kicks webspeech → shields block google
4. ui shows “brave shields may be blocking google’s speech service” — confusing af because user didn’t pick browser mode

fix was routing + stop lying via fallback. not more deepgram dashboard staring.

---

what’s next (maybe)

• GHOST RACE — new mode where u speak against ur pb (most wanted)
• signed run tokens so leaderboard posts tie to a finished test
• shared rate limit (redis/kv) if leaderboard spam shows up
• vendor ort wasm properly so vad worker stops whining in console

---

ok that was a lot of browser-specific nonsense for a monkey that just wants you to read words out loud 🐒🐒🍌

---