Demo walkthrough

A repeatable, end-to-end walkthrough that proves the MVP works from a user’s seat — not unit tests, the real app. Run it on WhatsApp Web (the channel with the most to show: voice in/out, group gating, proactive push) and tick the checklist at the end.

Scope: the six pillars — memory, voice, schedules, web/browser, multi-agent missions, approval gates. Known gaps are listed at the bottom; don’t demo those.

Setup (once)

1. Configure ~/.gaia/gaia.yaml:

   connectors:
     whatsapp:
       enabled: true
   voice:
     enabled: true            # whisper speech-to-text in
     reply_with_voice: true   # edge-tts text-to-speech out (voice-in → voice-out)
   memory:
     enabled: true
     auto_ingest: true
   missions:
     approval_classes: [spend, book, send_as_me, destructive]
     max_tasks: 20            # drop to ~4 only for the runaway-cap demo (#7)
   cron:
     deliver:                 # where scheduled results land when fired off-chat
       channel: whatsapp
       chat: "<your-whatsapp-jid>"

2. Install deps: uv sync --all-extras --all-groups (whisper, edge-tts, browser). For the browser: either bun on PATH (default playwright-mcp backend) or uv run playwright install chromium (native). edge-tts needs network, no espeak-ng.
3. Auth a model: uv run gaia model (or put a Gemini key in ~/.gaia/.env).
4. Start the daemon and pair: uv run gaia start, scan the QR with WhatsApp once. Confirm health: uv run gaia status and uv run gaia doctor.
5. Watch while you demo (separate terminals):
   • uv run gaia logs --follow --events
   • uv run gaia task list (re-run to see the board move)

Scenarios

Run in order. Each is prompt → expect → pass. Prompts are sent as WhatsApp messages to your paired number.

1. Memory — “it knows me”

• Say: “Remember my girlfriend’s name is Grace and I’m cutting on an A/B gym split.”
• Then /reset, then: “What’s my girlfriend’s name and my training split?”
• Pass: recalls Grace + the A/B split; /memory lists both.

2. Voice — “talk to it, it talks back”

• Send a voice note: “give me a one-paragraph summary of today’s AI news.”
• Pass: it transcribes (whisper, see logs) and replies as a voice note (edge-tts). A typed question still gets a typed reply.

3. Schedule — “it reaches out on its own”

• Say: “In 2 minutes, send me a one-line hello.” (or “every day at 8am send me a 3-bullet AI brief”).
• Pass: uv run gaia cron list shows the job; at fire time the message is pushed proactively to the chat (you didn’t message first).

4. Web + browser — “it can act on the live web”

• Say: “Find a well-reviewed ramen spot near Marina Bay Singapore and screenshot its page.”
• Pass: web_search/web_fetch fire (logs) and a screenshot image arrives in chat.

5. Missions (the headline) — “real work done, with brakes”

• Say: “Build a one-page website for my gym A/B plan — design the actual program first, then build the site from it.”
• Pass: the dispatcher forges a trainer soul + a frontend soul; a soul consults an expert or files a subtask and yields, and the parent re-runs with the result (running → blocked → … → done). The site is presented once (screenshot + summary); internal steps are not pushed. Check uv run gaia task list (the tree) and ~/.gaia/agents/<soul>/workspace/ (the files). Events: task_blocked_on_children, consult_soul.

6. Approval gate — “it pauses for me on risky actions”

• Say: “Research a TLV→NYC flight next Tuesday and book the cheapest one.”
• Pass: research runs; the booking task parks and pushes “⏸ … needs approval (spend) … /task approve <id>”. Reply /task approve <id> → it proceeds; /task reject <id> → it fails and you get a notice.
• Restart proof: while it’s parked, uv run gaia restart, then /task — still awaiting_approval; approve → it runs. (The board survives reboots.)

7. (optional) Runaway cap — set missions.max_tasks: 4 first

• Say: “Plan and execute a full product launch: landing page, blog, email sequence, social posts, press release.”
• Pass: the mission pauses after the cap: “⏸ Mission … paused — task cap reached …”.

Ready checklist

• 1 — memory recall across /reset
• 2 — voice-in → voice-out round trip
• 3 — proactive scheduled push (no message from you)
• 4 — screenshot/media delivered from a live page
• 5 — multi-soul mission: forge + consult/subtask + re-run + single deliverable push
• 6 — approval park → approve/reject → runs; survives a restart
• no crashes in ~/.gaia/logs/errors.log; no stuck running tasks after a restart

One grep to confirm each pillar’s event fired:

grep -E "memory_updated|media_out|task_awaiting_approval|task_blocked_on_children|consult_soul|mission_paused" \
  ~/.gaia/logs/events.jsonl

Green = all six ticked, errors.log clean, restart leaves no stuck running tasks → MVP-ready. A failing scenario is a concrete bug to file (label it by pillar), not a vague “not ready”.

Known gaps (don’t demo these)

• WhatsApp Business (Cloud API) inbound webhook isn’t wired (#3) — use WhatsApp Web.
• No /schedules chat command yet — manage schedules with the cron tool in chat or gaia cron … on the CLI.
• Mission planner / re-plan-on-stall is P4 (#130): Gaia decomposes via task_plan, but doesn’t yet auto-re-plan a stalled mission.
• A2A external agents is P5 (#131); token/cost budget (#133) and session-resume (#132) are follow-ups — long missions are bounded by max_tasks/max_hours, not tokens.
• Souls run on the configured model, so a deep mission spends real tokens — keep the caps modest for demos.