Cue

Helps you say the right thing.

A multi-mode conversation coach for Even Realities G2 smart glasses. Listens to the conversation, surfaces 2-3 suggested responses on the display in real time. Pick a mode (Date / Argue calm / Sales close / Sting / Listen well / Custom) to shape the suggestions. The app never speaks for you — it offers cues you say in your own voice.

Status: v0.4.3 (per-session transcript persistence — review past conversations on phone)

If you've deployed the personal Worker (see worker-template/README.md) and pasted its URL + bearer token in phone settings, Cue streams audio over chunked HTTP → Deepgram for transcription, and POSTs your rolling transcript to the Worker's /suggest endpoint for LLM suggestions. If those settings are blank or the Worker is unreachable, Cue falls back to the v0.1.0 timer-driven mock suggestions so the app stays demonstrable.

Version	What's in it
v0.1.0	Scaffold, mode picker, privacy opt-in, mic toggle, glasses UI, mock suggestion driver
v0.2.0	Worker template (Deepgram + LLM bridge), real audio capture via audioControl, transport layer, live captions, debounced LLM suggestions. Mock fallback preserved.
v0.2.5	Test infrastructure: chunked HTTP transport, JSDOM tests, worker integration tests, app.json lint, KNOWN_QUIRKS, WebKit harness for iOS WKWebView parity.
v0.3.0	End-of-utterance detection (sentence-final punctuation + silence-gap + max-wait), sentence-aware transcript trimming, battery glyph in glasses header, idle auto-pause after 5 min, word-boundary line wrap on suggestions, per-mode bullet glyphs, first-word emphasis.
v0.3.1	Phone-side idle-auto-pause-min setting — was a 5-min hardcode. 0 disables.
v0.3.2	Battery glyph now appears on idle screen (was only visible mid-session). Mock-bridge JSDOM coverage for v0.3 state machine (auto-pause, mode cycle, foreground re-paint). Stale wss:// whitelist entry removed (chunked HTTP supplanted WebSocket in v0.2.5).
v0.3.3 (current)	Mock-mode adds a longer-suggestion entry that exercises the v0.3 word-wrap path + custom-mode coverage (no more silent fallback to date-mode). lint-app-json.mjs extended to auto-detect whitelist gaps + stale entries by greping src/. Regression script gains a countStateLogs render-loop liveness assertion (would catch a silent main-loop death).
v0.4.0 (planned)	Worker-side dedupe of repeated suggestions, retry/backoff on rate-limit, partial-transcript pulses if Deepgram streaming becomes available.

How it works (current v0.2.0)

1. One-time — deploy the personal Cloudflare Worker (see worker-template/README.md). You get a https://<sub>.workers.dev URL and a SHARED_SECRET bearer.
2. Wire it to Cue — paste both into phone-side settings. (Skip this step and Cue runs in mock mode.)
3. Open Cue from the Even Hub launcher.
4. Privacy notice appears on first launch — read and accept (or decline) before the mic can be enabled.
5. Pick a mode in the phone-side settings page.
6. Put on the glasses, open Cue. The idle screen shows ◉ live if a Worker is configured, ◌ mock otherwise.
7. Tap glasses to start a session. With a Worker, audio streams to Deepgram and you'll see live transcript captions on glasses; suggestions arrive ~6s after each transcript update. Without a Worker, the timer-driven mock fires.
8. Glasses double-tap when not micced = cycle mode. Ring double-tap during a session = "fresh topics" prompt (date / custom modes).
9. Glasses double-tap during a session = exit (also stops mic).

Privacy is a real feature, not boilerplate

• Mic OFF by default every session. No exceptions.
• Explicit opt-in required on first launch via a modal.
• Mic indicator always visible when listening — never hidden.
• No persistence of audio — when real STT lands in v0.2, audio streams through and is dropped. Transcripts buffered ≤3 min in Worker memory.
• No analytics that include conversation content.
• You are responsible for ensuring it's legal where you are. Recording someone without their knowledge violates two-party-consent laws in CA, FL, IL, MD, MA, MT, NH, PA, WA, and many countries.

Modes

Glyph	Mode	Use it for
★	Date	Curious, warm. Suggests questions and follow-ups. Ring-tap for fresh topics when stuck.
◇	Argue calm	Validating, deescalating. For tense conversations.
▶	Sales close	Listens for objections, suggests handlers.
⚡	Sting	Sharp witty comebacks. Banter mode.
●	Listen well	Reflective listening prompts ("what I hear is…", "tell me more").
◆	Custom	Use your own system prompt (write it in phone settings).

Glasses gestures

Gesture	Action
Single tap (mic off)	Start mic session
Single tap (mic on)	Stop mic session
Double tap (mic off)	Cycle to next mode
Ring double tap (mic on)	Request fresh topics (proactive — date / custom modes)
Glasses double tap (mic on)	Exit app (also stops mic)

Development

npm install
npm run dev          # Vite dev server on :5176
npm run build        # tsc + vite build
npm run pack         # evenhub pack → cue.ehpk
npm run deploy       # build + pack
npm test             # Vitest unit tests
npm run test:watch   # Vitest watch mode

Test on real glasses via QR:

npx evenhub qr --url http://<your-mac-lan-ip>:5176

Test in simulator:

npx evenhub-simulator --glow http://localhost:5176

Source files

File	Purpose
src/main.ts	Entry, state machine, phone settings UI, glasses render
src/even.ts	Glasses bridge wrapper (text container, input routing, mic capture, battery)
src/transport.ts	Worker transport — chunked HTTP POST for audio + REST for suggestions
src/modes.ts	Mode registry — id, label, glyph, system prompt, behavior flags
src/utterance.ts	Pure heuristics — end-of-utterance trigger, sentence-aware trim, word wrap, battery glyph
src/mock.ts	Mock-mode timer-driven canned suggestions (fallback when Worker unset)
src/storage.ts	Native setLocalStorage wrapper for mode + privacy + Worker config
worker-template/	Cloudflare Worker source — Deepgram + /suggest LLM bridge
tests/*.test.ts	Vitest unit tests (54 passing)
scripts/regression.mjs	Simulator-driven e2e flow check (mock-fallback path, 4/4)

Roadmap

Full plan in ~/Documents/Pulse/ROADMAP.md § "Plan: Cue". Remaining for v0.4+:

• Worker-side dedupe of suggestions repeating the same advice
• Retry/backoff on Deepgram or LLM rate-limit
• Partial-transcript pulses if streaming-Deepgram path becomes viable on WKWebView
• Phone-side IDLE_AUTO_PAUSE_MS setting (currently a 5-min hardcode)

Packaging note

The Worker URL is per-user (each deployer gets their own *.workers.dev subdomain). Before running npm run pack, replace https://your-cue-worker.example.workers.dev in app.json's permissions[].whitelist with your own Worker URL — otherwise the WebView in the packaged build will block the request.