Three guppies. One TV. Two MacBooks. A phone number. Decisions get made — yes or yes.

A user calls a Twilio number, talks to a snarky "fish" voice agent, two options appear on a TV behind a fish tank, the guppies vote (majority of which side they're on), and the winning option is then executed by a remote MacBook in real time.

caller ── Twilio ──► bridge (Fly.io) ──► Deepgram (STT)
                          ▲                    │
                          │                    ▼
                    ElevenLabs (TTS)     Claude Haiku 4.5
                          ▲                    │
                          └──── tool calls ────┤
                                               ▼
                          web (Vercel) ── Redis ── Pusher Channels
                                               │
                            ┌──────────────────┼─────────────────┐
                            ▼                  ▼                 ▼
                        /display TV       vision (Mac mini)   remote-agent
                       (A/B + countdown)  (OpenCV @ 30 Hz)    (Playwright + osascript)

1. Caller dials the Twilio number; audio streams into the bridge.
2. Bridge transcribes with Deepgram, feeds the running transcript to Claude, and speaks Claude's replies back via ElevenLabs.
3. Claude listens for whatever mess the caller is in, frames it as a binary choice, and calls present_options(stage, A, B). The bridge POSTs that to the web orchestration API, which writes Redis state and publishes to Pusher.
4. The TV display subscribes to options and renders A vs B with a countdown. Vision publishes per-frame fish counts on fish-pos at 30 Hz; the display tallies which side has the majority over a 1-second rolling window.
5. When the countdown ends, /display publishes a decisions event with {stage, chosen, text, vote}. The bridge's wait_for_decision tool unblocks, Claude announces the winner, then calls dispatch_action(stage, chosen, text).
6. The remote-agent is subscribed to agent-tasks, picks up the dispatch, and carries the winning option out on a second MacBook — composing an iMessage, setting a Reminder, drafting an Outlook email, posting to LinkedIn, whatever the option calls for. Status updates stream back through Pusher to the TV.

The persona is built around real-life "should I, or should I really" moments — breakups, missed meetings, getting laid off, and so on. Claude turns the caller's situation into a pair of options (one defensible, one regrettable), the council picks, and the remote agent makes it happen.

The option set isn't fixed — Claude composes options live from the call, and new outcomes can be wired into the remote agent without touching the bridge.

web/            Next.js + Vercel — TV display + orchestration HTTP API + Redis state
bridge/         Node service on Fly.io — Twilio call audio ↔ Deepgram ↔ Claude ↔ ElevenLabs
vision/         Python on Mac mini — DJI Osmo Pocket 3 capture, OpenCV multi-fish detection,
                live MJPEG preview server on :8765
remote-agent/   Node on remote MacBook — Playwright + osascript runner that executes the chosen option
scripts/        Mock event publisher for end-to-end smoke tests
docs/           Per-component docs + hour-by-hour workflow + demo script

1. docs/00-overview.md — system map and glossary.
2. docs/03-bridge.md — call flow, tool calls, persona prompt.
3. docs/05-remote-agent.md — how the remote agent picks up and runs an option.
4. docs/workflow.md — hour-by-hour build order.
5. Each runtime has its own README inside its folder.

• Twilio (phone) → Fly.io bridge → Deepgram (STT) → Claude Haiku 4.5 (brain + tool use) → ElevenLabs (TTS)
• Vercel hosts the Next.js display + orchestration API
• Pusher Channels for cross-runtime realtime pub/sub: fish-pos, options, decisions, agent-tasks, agent-status
• Upstash / Redis Cloud (via Vercel Marketplace) for per-call state
• Playwright (headed Chromium) + osascript drive the remote MacBook
• OpenCV + DJI Osmo Pocket 3 for fish position detection

./setup.sh                                  # installs deps for all four runtimes
cd web && npm run dev
open http://localhost:3000/display/dev      # click-through preview of every TV state

cd scripts && npm run mock -- demo          # full ig-swipe stage runs on /display

YES or YES — because when you put a goldfish in charge of your life, every option is the right one.