Demo — Hina cold-open

The instant the server starts, Hina texts first with a fixed opener, then hands off to the live agent.

4 min

On this page

The scripted opener
What happens, start to finish
How to run it
Knobs
Why hardcoded (vs. the model opener)
During the demo
What I changed

The scripted opener

Three bubbles, sent one at a time with the normal human typing choreography (dots → "type" → send → beat → next). Modeled on the north-star anna cold-open.

hina

hey

i heard you want to learn a language

which one?

japanese

hina

へえ　本当？？

what made you want to learn japanese?

dashed = hardcoded · everything after is the live model picking up the thread.

The exact bubbles live in one constant, DEMO_BUBBLES in agent/core/demo-opener.mjs. Drop the 'hey' line for exactly the two you asked for. I added it because the north-star opens with a standalone greeting before the question — it reads warmer — but it's a one-line edit either way.

What happens, start to finish

You run npm start (+ npm run messages).

The watchdog launches Messages.app with the dylib injected; the helper dials into the server.

On first helper connect, the opener fires once.

Guarded by a latch so a watchdog relaunch can't re-fire it. Waits ~1.5s after connect so it doesn't fire the literal instant Messages finishes booting.

The demo learner is reset to a clean slate.

Wipes +15109358199 back to fresh onboarding (keeps the chatGuid) so the demo is identical every run — the live agent behaves like a first conversation, not someone already onboarded.

The three bubbles send with typing dots.

Same sendBubbles path + sending latch as a real reply, so it looks human and can't collide with an inbound turn or the proactive sweep.

The opener is seeded into history as Hina's turn.

So when you reply "japanese", the model continues from the opener — it won't re-greet or re-ask "which one?". It drops Japanese immediately and runs the rest of onboarding live.

How to run it

Two terminals, from agent/. DEMO=1 is already set in .env.

# terminal 1 — launch Messages with the helper injected
npm run messages

# terminal 2 — start the server (fires the cold-open on connect)
npm start

Watch the demo phone: hey → i heard you want to learn a language → which one? arrive within a couple seconds. Reply japanese and the live agent takes over. To re-run a clean demo, just restart the server — the reset makes it repeatable.

Heads up: DEMO=1 means every npm start now cold-opens and wipes that learner's memory. Set DEMO=0 (or delete the line) to go back to normal operation.

Knobs

Env / location	Default	What it does
`DEMO`	`1` (in .env)	Master switch. `0/off` disarms the cold-open entirely.
`DEMO_HANDLE`	first `ALLOW_FROM`	Who to open on. Unset → `+15109358199`.
`DEMO_RESET`	`1`	Reset the learner to fresh onboarding each run. `0` = demo against existing memory.
`DEMO_BUBBLES` (code)	3 bubbles	The scripted lines, in `core/demo-opener.mjs`. Edit to change the words.

Why hardcoded (vs. the model opener)

The normal opener

Model-generated, reply-only.

"hey, it's hina" came from the onboarding prompt in agent.mjs — it varies run to run and only fires after the learner texts in. Proactive can re-poke a quiet thread but won't script exact words or cold-open a fresh one.

The demo opener

Fixed, fires first, deterministic.

Same bubbles every run, on a clean slate, the second Messages connects. No model call for the opener — just the scripted lines, then the live agent owns everything after. Exactly what a live demo needs.

During the demo

Turn two is the hook. The moment you say "japanese", Hina drops Japanese with no announcement (へえ本当？？) — that's the product, not a tutor.
It always ends on a question. React warmly, then ask — so you keep producing the language.
Ambient info-gathering. It learns your reason / level / interests through the chat and saves them via remember — never a five-question form.
Proactive is still live. If the thread goes quiet, Hina texts first again later — but for the demo the cold-open is the star.

What I changed

File		Change
`agent/core/demo-opener.mjs`	new	The cold-open: scripted bubbles, fire-once latch, clean-slate reset, history seed.
`agent/core/memory.mjs`	edit	Added `resetLearner(handle)` — fresh onboarding, keeps chatGuid.
`agent/config.mjs`	edit	Added the `demo` config block (enabled / handle / reset).
`agent/server.mjs`	edit	Wired `startDemoOpener(...)` into `main()` after the proactive sweep.
`agent/.env`	edit	`DEMO=1` to arm it.

Smoke-tested: config parses, modules import, the opener registers its connect handler without sending. The real send needs Messages.app connected — that's the demo itself.