elizaos--eliza/packages/docs/ongoing-development/mvp/MVP.md

LifeOps Personal Assistant — MVP

The first doc to read. It defines what the MVP is, who it serves, what is in and out of scope, an honest assessment of where the code stands, and the bar each of the nine workstreams must clear to be called done. It synthesizes the nine workstream research docs under research/; each section links back to the doc that owns the detail.

Live task state: project board 15. Coordination and evidence mechanics: ../README.md.

Product definition

The MVP is chat, onboarding, the current views, and centrally LifeOps — all scheduling, calendar, coordination, reminders, goals, todos, and tasks, as one personal assistant. Chat is the product; views are read surfaces the agent (or the user, via chat) drives.

It must serve business users, children, adults with ADHD/ADD/Asperger's/ autism, neurotypical adults, and elderly people — through the same single interface, with no therapy language and no special rails. There are no persona "modes" and no audience-segmented flows. The product meets each audience by covering their real life against the same machinery: brushing teeth twice a day, workouts, getting the report done on time for class, life goals, night-owl and rotating-shift sleep rhythms, medication-adjacent routines with no medical framing, bills, appointments, and elderly check-ins.

"No special rails" is already the architecture, not an aspiration. Persona differences are owner facts + structural knobs — reminder intensity, anchors, no-reply policy, quiet hours — consumed by one runner (plugins/plugin-scheduling/src/scheduled-task/runner.ts), which matches on structural fields and never on promptInstructions text. The MVP work is proving each persona's knob-set produces the right behavior, not building persona machinery (research/01).

The guiding constraint is minimize additional scope: turn what exists into an MVP by fixing, testing, verifying, and validating the important stuff. Prefer deleting and simplifying over adding. Every workstream doc independently reached the same conclusion — the feature surface is already sufficient; the missing thing is verification.

Personas & journeys

Eight personas, condensed from research/01. Where a bench persona already exists (packages/benchmarks/lifeops-bench/.../scenarios/_personas.py), the name is reused for continuity. Each maps to a pack in the persona scenario corpus.

P1 — Casey, 29, product designer with ADHD (Telegram; packs A1/A2). Opens the app mid-chaos and takes fast-start — zero questionnaire tolerance. First win: a rambling multi-topic dump where the one buried commitment ("call pharmacy before 5") is captured and fires on time. Week 2: the follow-through ladder chases twice then shrinks the ask; a missed step gets a replan, never a lecture. Loves: capture-from-mess with no forms, non-shaming retries, the two-minute-step task-initiation help. Struggles: trust — only 13 of 52 ADHD-pack scenarios are verified, and one misbehaving capture or escalation gets the app deleted. Intuitive fix: nothing new — verify A1/A2 to the bar the other packs already meet.

P2 — Sam, 12, middle-schooler (in-app + push only; no email, no finance). A parent provisions the device; Sam answers "What should I call you?" and accepts brush-teeth from the offered habits. First win: a morning brush-teeth nudge in kid-plain words, marked done from chat. Week 2: "i have a book report friday and i haven't started" → captured todo, back-planned, done on time. Loves: it talks like a person, not a chore chart. Struggles: the customize questions are adult-worded and multi-select-in-chat is hard at 12; zero scenarios exist in a child's voice. Intuitive fix: prove fast-start is genuinely sufficient with child-voice scenarios — don't reword anything.

P3 — Priya, 21, college student, night owl (Discord; pack B1). Onboards at 2am; states morning = 11:00–14:00 and it is accepted without judgment. First win: her day anchors to her actual wake, not 7am. Week 2: a term-paper deadline back-planned across her real nocturnal work windows, sleep protection holding. Loves: relative_to_anchor scheduling — the assistant that doesn't assume mornings. Struggles: the student "report due for class" flow has zero scenarios; it's exercised today only by executive-flavored flows. Intuitive fix: student-voice deadline scenarios asserting the ladder fires inside her active windows.

P4 — Ray, 47, rotating-shift nurse (WhatsApp/SMS; pack B2). Onboards between shifts; fast-start, then tells the agent the roster. First win: "nights starting Monday" re-anchors gm/gn and protects the new sleep window. Week 2: a shift swap re-anchors routines and holds low-priority nudges during the moved sleep window. Loves: the only assistant whose "morning" moves with the roster; medication-adjacent routines stay plain reminders, no medical framing. Struggles: B2 is the least-verified pack (4/22) — re-anchoring is exactly the behavior most likely to be silently wrong, and a wrong anchor wakes a night-shift nurse at 7am. Intuitive fix: verify B2.

P5 — Marcus, 33, autistic software engineer (desktop; Discord; packs E1/F1). Deliberately chooses the customize path — he wants the explicit questions and explicit defaults. First win: plans stated literally and executed exactly as stated, with two-phase preview-then-confirm before any write. Week 2: predictable check-in cadence, no surprise tone shifts, no unsolicited scaffolding. Loves: structural predictability — same trigger, same wording, same time; the runner's structural-fields-only contract makes this a guarantee rather than a vibe. Struggles: the launcher merges ~30 builtin tabs (including trajectories/runtime/database) — overwhelming for a fresh install if the viewKind gate leaks. Intuitive fix: verify the fresh-user view set is the curated MVP set.

P6 — Dot, 74, retired teacher (iPad; iMessage; daughter set it up). The daughter completes setup; Dot talks to it like a person. First win: dentist Thursday 3pm captured from a rambling paragraph + a day-before reminder. Week 2: a daily check-in she reliably answers, the quiet-user-watcher catching the day she doesn't, bill reminders in plain words. Loves: no menus, no tech vocabulary required. Struggles: one verified scenario covers her entire life; the check-in → missed → gentle follow-up loop — her family's actual reason for installing it — is covered only generically, and its no-reply leg is unverified in her voice. Intuitive fix: verify the elderly week-1 loop end to end including the no-reply path.

P7 — Elena, 41, small-business owner (Gmail + Google Calendar; packs D1/F1 + executive-assistant). Customize path; enables inbox triage + follow-ups. First win: a morning brief with the three things that matter. Week 2: a VIP breakthrough during quiet hours, a bill approval that fails closed on silence. Loves: comms-flood triage (D1 24/26 verified) and the approval queue's conservative no-reply defaults. Struggles: nothing structural — she is the best-covered persona; her risk is regression while the neurodivergent packs get attention. Intuitive fix: keep F1 green.

P8 — Jordan, 39, separated co-parent (iMessage + Calendar; planned pack J1). Fast-start, then connects calendar and messages when custody logistics become the immediate pain. First win: recurring exchange reminders and a neutral calendar view of the custody rhythm. Week 2: a last-minute swap request becomes an owner-approved, civil logistics draft without exposing child details or giving legal or therapy advice. Loves: factual language and coordination across both households. Struggles: the cross-domain handoff between scheduling, money, messaging, and child privacy has a planned ledger but no uniformly verified scenario pack. Intuitive fix: verify J1 end to end rather than add a co-parent-specific product mode.

Real-life scenario matrix

The owner's real-life list → the feature that covers it → scenario-test status, from research/01. ✓ covered and verified · ◐ partial / thin · ✗ no coverage.

Real-life scenario	Feature that covers it	Test status
Brush teeth 2×/day	habit-starters pack + reminder ladder	✓ 15 brush-teeth-* incl. 10 languages, night-owl, cancel/retry
Workouts	reminders + blockers	✓ workout-blocker-basic, workout-spanish, stretch-breaks
Report done on time for class	OWNER_TODOS/CALENDAR + reminder escalation	✗ no student/child deadline scenario
Life goals	plugin-goals + OWNER_GOALS + grounding loop	◐ sleep-only (goal-sleep-{basic,spanish,french-formal})
Night owl / irregular sleep	anchors + relative_to_anchor triggers	✓ B1 pack, 19/24 verified
Rotating shifts	re-anchoring + sleep-window protection	◐ B2 pack, 4/22 verified
Medication-adjacent, no medical framing	plain reminders	✓ vitamins-* ×3, adhd-medication-refill-fuzzy-date-capture
Bills	OWNER_FINANCES + approval queue	✓ bill-approval-and-payment, quarterly-tax-payment-runbook
Appointments	CALENDAR + day-before reminder	✓ persona-* dentist family (elderly, ESL, typo, voice)
Elderly check-ins	daily-rhythm check-in + quiet-user-watcher	◐ single persona-elderly-nontechnical scenario
Child as user	same single interface	✗ none
Onboarding → first win	first-run service + seeded packs	✗ no live scenario

Scope: in / out

In scope. Chat as the primary surface (continuous overlay, gestures, scrolling, search, mic). Cloud-only in-chat onboarding + chat-native tutorial + help-as-knowledge. The current views as read surfaces the agent drives. LifeOps: scheduling, calendar, reminders, goals, todos, tasks, coordination — one scheduler, structural behavior. The in-chat widget vocabulary (form, choice, followups, task, secret/oauth, [CONFIG]). A sparse resting home (time/weather + notifications + wallet). Bidirectional voice on every surface. On-device Android + iOS fleet e2e. Doc-driven coordination + inline evidence. The unifying constraint: fix / test / verify, don't add.

Out of scope (aggregated from every workstream's non-goals):

• New default packs, persona "modes", child/elderly UI variants, or any therapy-adjacent framing or rails. Onboarding question rewrites, new onboarding steps/wizards, name pickers, provider marketplaces (01, 04).
• A new widget framework, registry, or renderer; per-plugin marker kinds; a hosted generic /forms/:id page; a widget builder or user-authored widgets (02).
• New home widgets, a widget-settings/customization surface, launcher-tile or rail-gesture changes, a second notification surface, new server-side wallet/price plumbing (03).
• Re-adding a Help view or searchable FAQ; redesigning the tutorial; live investment in the flagged local-model onboarding path beyond flag-boundary tests (04).
• Thread virtualization, framer-motion replacement, sheet-animation rework; cross-conversation UI beyond search-jump; semantic/vector message search; feature work on the kiosk/detached chat paths (05).
• New views, tabs, or layouts; write actions for read-only diagnostic surfaces; semantic actions for third-party plugin views; the voice implementation itself (06).
• On-device Whisper (retired, stays retired); new TTS voices / voice cloning; cloud Kokoro streaming synthesis; any new bespoke voice test harness (07).
• Greening the CI x86_64 emulator; per-appex provisioning-profile minting; any device-farm / remote-brokering service; store-release / TestFlight automation (08).
• Any new CI gate; GitHub issue forms; automated board sync / status dashboards; a CI lane for the live-only persona corpus; publishing this folder on the docs site (01, 09).

What we have / what's weak / what's missing

The honest read, aggregated from every workstream's current-state section. The pattern is consistent: the architecture is built and mostly right; the gap is verification and a handful of concrete correctness bugs.

What we have (built and working). One scheduler with structural, data-driven persona behavior. Seven default packs with conservative no-reply defaults. Cloud-only in-chat onboarding, chat-native tutorial, and help-as-knowledge — all shipped (#13377, PRs #13393/#13521/#13394). The plugin-extensible inline widget registry, the typed connector-agnostic interaction protocol in @elizaos/core, and the out-of-band sensitive-request pipeline with a real hosted page. The chat gesture/detent system (pull-to- maximize landed #13531, CI-gated). A view:interact protocol + generic agent-surface bridge with ~40 views opted in. On-device Kokoro TTS on all three native surfaces and measured fused ASR on desktop (WER 0.008, RTF 0.262). Per-platform e2e orchestrators, a physical-iPhone deploy pipeline, a content-derived renderer build stamp. A working PR-evidence gate that already accepts inline GitHub attachments.

What's weak. The persona corpus is 134/212 verified, and the shortfall concentrates exactly in the make-or-break neurodivergent packs — ADHD capture 8/28, ADHD follow-through 5/24, shift-rotation 4/22. The widget round-trips have zero live-LLM scenario coverage. The home carries seven non-MVP widgets and a wallet that contradicts its own doctrine (hides with no holdings, prices never refresh); an AppRunsWidget polls every 5s ungated. Cloud voice now has a committed Railway benchmark and a scheduled live contract lane, the client cloud-ASR path is wired, the STT model is deploy-configurable, and reproducible Kokoro/Whisper Railway service definitions live in-repo. However, the scheduled lane still lacks a green owner-provisioned Railway run, and no browser test proves real mic audio reaches Railway Whisper and returns an audible live-agent reply. The chat transcript can silently become horizontally scrollable (overflow-y-auto without overflow-x-hidden), and the perf gate carries a known-red CLS 0.80 failure. Several view mutations (settings sections, documents, memories, files) have only the brittle synthetic-DOM path, no semantic action — a dead end for voice.

What's missing. Two persona voices with zero coverage (child-as-user, student report-deadline) and no live onboarding-journey scenario at all; goals coverage is sleep-only. History pagination in chat (a hard 80-row cap over a 200-message window, no before cursor) and message search reachable from the primary surface (it exists end-to-end but is mounted only in the desktop detached shell). The onboarding verification of record is dead — app-live-e2e.yml, the only lane proving a real cloud account + real LLM work after onboarding, has had no green run since ≤2026-06-25 (failure/cancelled every scheduled run for 10+ days), and no device lane covers the production default cloud sign-in path. The Android runner runs stale builds by default (install-if-missing skips fresh APKs — the Capacitor bake-in footgun, live in the runner). Railway ownership and live deploy evidence remain unproven even though the service definitions and scheduled contract lane are in-repo.

Workstreams

The nine research docs, one line each. Full detail behind each link.

1. MVP, personas & UX journeys — verify the persona corpus to a uniform bar; author the two zero-coverage voices + the onboarding scenario; scenarios only, no features.
2. In-chat widget system — fix the dead connector form-link, add date/time field types, and prove the widget + settings-in-chat + hosted-secret round-trips with live-LLM scenarios.
3. Launcher widgets — keep the resting home sparse; respec the wallet to doctrine; remove seven non-MVP widgets; prove the sparse state with an e2e fixture.
4. Onboarding, tutorial & help — resurrect the live e2e lane, automate the default cloud path on devices via the SIWE wallet, and add a liveness contract to every onboarding path.
5. Chat window UX — lock the transcript to vertical scroll, add history pagination, wire search into the overlay, pulse the mic; protect the already-built gesture system.
6. View↔chat integration — close the gap list: give every view mutation a semantic chat action so the view is a pure state renderer and voice-first works for free.
7. Voice pipeline — benchmark cloud vs on-device per platform, wire web cloud ASR for real, un-hardcode the STT model, make the Railway services first-class, prove bidirectional voice e2e.
8. Device testing pipeline — fix the Android stale-install bug, ship a fleet devices:status build-stamp check, one triage bundle per run, one-command physical-iPhone lane.
9. Doc-driven development — the process this folder encodes: discussion → doc → board issues → PR with inline evidence → doc updated; retire .github/issue-evidence/.

Acceptance bar

MVP done is evidence-driven, per AGENTS.md: a reviewer confirms each change works without reading the code, from artifacts attached inline to the PR. "Tests pass" is not proof; "CI is green" is not proof. Per workstream, done means:

1. Personas — the coverage gate (check-lifeops-persona-catalog-coverage.mjs) reads verified = target for A1/A2/B2/F1; child-voice, student-deadline, onboarding-journey, elderly week-1, and non-sleep-goal scenarios exist and pass with hand-read live-model trajectories and catalog status notes recording model + judge score.
2. Chat widgets — the connector form-link is gone (free-text fallback); date/time field types render; live-LLM trajectories show FORM emit→submit→use, CHOICE pick, [CONFIG] emission, and the settings + hosted-secret round-trips working, reviewed by hand.
3. Launcher — an e2e fixture asserts the healthy-quiet home shows exactly time/weather + wallet (+ notifications when present), each keeper self-hides, and the wallet shows BTC/SOL/ETH by default / top-3 held, refreshing.
4. Onboarding — a green scheduled app-live-e2e.yml run with reviewed artifacts (real cloud login + real-LLM reply); iOS-sim + Android-emu lanes that onboard through the real UI via the SIWE wallet and record MP4 + JPG; every onboarding path ends with one real chat turn asserting a non-stub reply.
5. Chat UX — a wheel/deltaX regression e2e proves vertical-only scroll; scroll-up pagination verified against the perf gate; search reachable on the overlay with a working jump; mic pulses on the primary surface; stale chat-full-clear/-maximize specs deleted.
6. View↔chat — a real-LLM scenario changes each targeted view's state ("turn on voice", "delete that document", "reset my background") without the model emitting a raw selector; a ratchet asserts every builtin view mutation maps to a registered action.
7. Voice — the cloud-vs-local benchmark table (TTFB, RTT, WER, sizes) is published inline with the web-TTS default decision recorded; one lane proves mic → STT → live agent → TTS → audible reply on web (Railway) and desktop (local) with an MP4-with-audio, including mic-denied / silence / network-drop failure paths.
8. Device testing — the Android runner reinstalls on stamp mismatch; devices:status reports installed buildId/commit per connected device vs develop HEAD; each device-e2e run emits one triage bundle (MP4 + JPG + logs + summary.json + junit.xml); the physical-iPhone lane is one command.
9. Process — this folder's README + MVP doc exist (the markdown-links lane is green), the PR-evidence gate no longer accepts retired issue-evidence paths, the discussions are open, and the board carries every MVP issue.