# Lilith V1.3 — the personal agent: scope (2026-09-27)

Status: **draft for Elikem's approval.** Author: Claude (claude-mac-worker),
2026-09-27. Board family: `lilith-v1-3` — every task starts as `later` and the
family as `unstarted`, so no worker picks anything up until this scope is
approved. An independent read-only review (Codex, 2026-09-27) was folded in
before filing: eleven capabilities added, milestone and dependency errors fixed,
six claims qualified. Revision 12 (2026-09-28) adds xAI Grok Bot to the field
(section 3) and corrects which subscriptions a member may sign in with (6.11).
Revision 15 (2026-09-28) adds ten more products to section 3, a table of the
laws that bind the escalation ladder (6.8), platform gatekeeping (6.20) and two
risks (section 11); it also answers the rev 13 review: M1 admits only adults
Elikem invites personally (6.1, 7, Decision 2), R02 no longer names baseline
triggers, and the Warp sign-in flow is described correctly. Revision 17
(2026-09-28) answers the rev 16 review: the SB 243 row gives the effective date
(2026-01-01) and adds the minors-suitability notice and evidence-based
suicidal-ideation measurement, which T04 now tests. Revision 19 answers the
rev 18 review: EU AI Act Art. 50(2) marking of generated output is task T10,
and Q05 keeps each capability closed to EU members until its Art. 50 checks
pass (principle 9, 6.18, Decision 1, 13); the skills-directory reference in
6.17 now points at 6.14.
Revision 22 (2026-09-29, at Elikem's request) takes seven lessons from OpenAI's
dots, launched at DevDay that day (section 3): work Lilith starts on her own only
reads (principle 11, 6.6, R01); member rules come before the first actions (X09
moves to M1); talking is not metered against the task budget (6.19, L07, Q01,
Decision 6); standing duties carry their own grants (6.3, 6.6, new task R08,
C02); every done carries its proof (6.4, A13); each task shows where it runs and
none of its messages is dropped (6.6, R01, R04, S09); and writes into shared
household lists are checked against flow labels (6.16, F03). It also records
that Sign in with ChatGPT is now a partner programme (6.11, B08) and adds
OpenAI's open-source Codex harness to the X06 measurement (6.15). Revision 24
answers the rev 23 review: the dots plan eligibility and the shared-Pages
comparison are marked as secondary reports that could not be confirmed
first-hand, and the dead WHOOP press-release copy is replaced.
Revision 27 (2026-09-30, at Elikem's request) takes lessons from Meta's Muse
design essay and its Connect 2026 announcements (section 3); an independent
read-only Codex review was folded in before submission. Members can open side
chats and send while Lilith works, and only requests for agent work become
metered tasks (6.9, S11, S12); a status line shows what she is doing and what
she may touch (6.6, R09); an Ideas view and first-week tips come from the
member's goals within a bounded work allowance (6.6, R10); caution levels and
approval-fatigue measures sit inside principle 3 (6.18, T11, sections 9 and
11); members choose the persona she takes (6.10, V08); a bake-off measures
realtime avatar routes, presences are fictional only, and avatar use has its
own allowance (6.10, V09, V05, Decision 8). T10 evaluates Meta's Video Seal
watermark beside C2PA (6.18). A candidate-model intake admits models per
version, provider, tier and deployment, first Meta's Muse Glimmer 30B and Muse
Spark 1.3 (6.13, E07, L02, 6.20), and Oshun-paid routes never train on member
data (6.12). Elikem settled Decision 12 the same day: members may rename and
restyle her, as personas Lilith takes on while staying herself (6.10, V08).
Revision 32 (2026-09-30) records Elikem's Decision 10: agent mail uses
`agent.oshunsynthetics.com`, and outbound mail and calls carry the member's
persona name with Lilith behind it (6.4, 6.10, A02, A07).
Revision 35 (2026-09-30, at Elikem's request) takes lessons from Manus 2.0 and
its personal-agent app Cue, launched 2026-09-28 (section 3). The analysis is Eve
research document `7bb205ec-6890-425d-ab64-25ba1270a887`; an independent
read-only Codex review checked it and then this revision. Nine corrections to
the approved plan: Link's agent wallet admits only accounts created in the US
and Canada (6.5, P02); text other people wrote reaches the planner only as
quarantined data (6.18, 6.16, T01, F03); no reusable control-plane credential is
reachable from the sandbox (6.2, W02, A04); the member chooses who may start
work by mail, and duties get sub-addresses (6.4, A02); several accounts per
provider (6.3, C02); automations get a dry run, failure pauses and an activation
card (6.6, R02); the estimate is a price ceiling and allowance returns after
Lilith's own failures (6.12, 6.19, L06, L07, Q01); agent data terms and
offboarding are published before the first invite (6.18, new task T12); and
approval defaults change only by versioned decision (6.18, T11). Three
additions: her own number answers forwarded calls (6.4, new task A17); payment
routes for members outside the US and Canada are assessed before any build
(6.5, new task P11); and goal rooms show several duties working on one goal,
with one writer per counterparty (6.6, new task R11). Of the optional items the
analysis lists, a workspace files view (6.2, new task W12) and member-chosen
effort levels (6.12, L02) are taken now; the remaining items are deferred or excluded under Decision 13.
Decisions 1, 4, 5 and 9 gain inputs. The board map now also lists X10 and K11,
filed on 2026-09-29 from the Raven source read.
Release name in code: `LILITH_PERSONAL_AGENT_RELEASE` in
`libs/oshun/navigation/src/release-scope.ts`.

## 1. Summary

V1.3 turns Lilith from a host that acts inside her own rooms into a personal
agent that acts across the member's life — email, calendar, forms, bookings,
purchases, calls, messages and devices — and does it wellbeing-first: the same
agent that books the flight also protects the member's sleep, spending and
attention. Alongside it ships a **model lab**: members bring their own models
(provider keys, OpenRouter, self-hosted, on-device), and Lilith picks the
cheapest model that can do each job, builds ensembles when quality demands it,
benchmarks models on the member's own tasks, learns skills, fine-tunes and
distils, and lets the member inspect every step the agent took and whatever
reasoning the model exposes.

The bar is: everything Meta Muse, Google Gemini Spark, ChatGPT and its dots, Claude, Alexa+,
Siri, xAI Grok Bot, Microsoft Copilot, Samsung Galaxy AI, Instinct, Pine, Comet,
Manus and its Cue app, Genspark, Hermes Agent and OpenClaw ship for individuals in
September 2026, plus the gaps none of them fill (section 4).

## 2. Where Lilith stands today (V1.0 code, 2026-09-27)

| Area | Today | Where |
|---|---|---|
| Assistant | Lilith, "host of the house"; real model through OpenRouter only (Task 14.2 admission), fail-closed 503 to a deterministic intent engine without credentials | `apps/oshun/bff/src/routes/assistant.ts`, `apps/oshun/bff/src/assistant/agent-provider-config.ts`, `provider-route-governance.ts`, `libs/oshun/shell-assistant/src/system-prompt.ts` |
| Tools | In-room reads and writes (favourites, Nyx observations and reminders, Arete goals, Nisaba passages), navigate, highlight, read the open page, docs search, tours. Eight mutating tools wait for on-screen confirmation and can be reversed | `apps/oshun/bff/src/assistant/agent-tools.ts`, `action-confirm.ts`, `action-reversal.ts` |
| Memory | Iris memory bridge, consent-gated, member-managed at `/profile/memory` | `apps/oshun/bff/src/assistant/iris-memory-consent.ts`, `apps/oshun/web/src/app/profile/memory/` |
| Voice | Push-to-talk; Deepgram STT is the only admitted voice provider; replies spoken by browser speech synthesis | `provider-route-governance.ts`, `apps/oshun/web/src/lib/assistant/voice.ts` |
| External accounts | Google Calendar connect and one-way export exist but the provider is `blocked-unattested` (503) | `apps/oshun/bff/src/routes/calendar-connect.ts`, `calendar-events.ts` |
| Health data | Mindful-session writes to HealthKit / Google Fit and HRV coherence analysis; absent bridge is a fail-closed 503 | `apps/oshun/bff/src/routes/wearable-health.ts`, `libs/oshun/shell-wearable` |
| Surfaces | Web + PWA (V1.0); native iOS/Android built but gated to V1.1 (`LILITH_NATIVE_APPS_AVAILABLE = false`); watch companions without an assistant; Telegram room commands; WhatsApp reminders only | `apps/oshun/mobile`, `apps/oshun/wearable`, `apps/oshun/bff/src/telegram`, `libs/oshun/messaging-channels` |
| Presence | Mentor Presence = pre-rendered, watermarked APNG performances, release gate says not ready | `apps/oshun/bff/src/mentor-presence/`, `libs/oshun/persona-registry` |
| Cost control | Output-token capacity reserved atomically per member/session/tenant/fleet before dispatch | `apps/oshun/bff/src/assistant/capacity-budget.ts`, `plan-entitlements.ts` |

**Building blocks that exist as libraries but are not wired to members** — V1.3
wires, hardens or replaces them; none is assumed to be production-grade until
its task proves it:

- Hermes-parity composition root `libs/oshun/assistant` — channels (Telegram,
  Slack, Discord, Signal, email/IMAP), scheduling (cron, natural-language
  schedules, Redis leases), watchers, skills (store, ranker, retriever,
  capture/offer), subagents (terminal sessions, Python RPC), tools (browser,
  code-exec, filesystem, web fetch/search, vision, media), an SSRF guard. Today
  the BFF composes it only for Eve loops (`apps/oshun/bff/src/assistant/loop-host.ts`).
- `libs/oshun/skill-system` — skill contracts, registry, discovery, governance,
  sandbox, recipe security, evolution, training.
- `libs/oshun/messaging-channels` — tier-aware dispatcher with WhatsApp, SMS,
  push, email, Slack, Discord, Telegram, voice transports.
- `libs/psyche/computer-use-core` (napi-rs desktop controller; Eve-only, null in
  production), `libs/psyche/action-safety` (confirmation workflows, policy
  engine, rate limiter).
- `libs/iris/*` — `model-routing`, `ensemble`, `reasoning-thinking`,
  `conversation-providers-local` (Ollama/vLLM/llama.cpp), `mcp`, `failover`,
  `memory`, `voice`, `presence`; `libs/iris/a2a` is labelled an unmounted
  non-conformant prototype. None has an app importer; the routing lib prices
  against a GPT-4o baseline — treat these as starting points to evaluate, not
  as finished parts.
- `libs/nous/*` — `training`, `llm`, interpretability (`interp-core`, `sae`,
  `probes`, `circuits`) and eval scaffolding; Eve's small-model tournament and
  cost report (`tools/eve-smx-cost-report.mjs`, the price-the-route method in
  Eve document `39fb4871-1fea-8f55-98a6-522775ffee74` r1, "Model cost for tests and harnesses").
- `libs/oshun/fiat-payments` (Stripe + Apple/Google wallets, V1.1) and the
  crypto rail (V1.0).
- V10 Ambient Rail (`libs/v10/*`, Tauri desktop shell `apps/v10/shell`) — the
  cross-device content layer and the desktop shell a local client can grow from.

## 3. The field in September 2026

| Product | What it does for an individual | Surfaces | Price | Lesson for Lilith |
|---|---|---|---|---|
| **Meta Muse** (2026-09-08) | Acts in your accounts: email, travel, forms, shopping (Stripe Link, Shop Pay, PayPal), negotiation; Mac control; memory with "forget" | App, web, WhatsApp; Mac; glasses and the Muse Charm pendant announced; Realtime Avatar for video calls announced 2026-09-23 (no public developer API as of 2026-09-30) | Free (weekly compute) / $20 / $100 + transaction fees; US, 18+ | Per-member Secure VM + Sentinel egress gate + surrogate credentials set the containment bar; the Mac client synced 187k iMessage rows after the user declined and then misdescribed its access — local clients need their own gate and access claims must come from ground truth. Its designers report that naming and styling your own Muse is one of the features people respond to most, and added side chats, an Ideas tab and a status line under the avatar (revision 27). Meta's models are on OpenRouter: Muse Spark 1.1–1.3, a cheaper Contributor tier that trains on prompts, and the open-weights Muse Glimmer 30B |
| **Instinct** (invite-only since 2026-08; $1B Series C at $10B on 2026-09-28) | A texting-first agent with its own phone number and computer: books travel and restaurants, buys groceries, pays bills, cancels subscriptions, researches; a concierge that phones businesses that cannot be booked online; follows up on dropped threads and reaches out first; a "trusted person network" lets one member's agent coordinate plans with friends' agents | Text and calls; no app yet | Undisclosed | The closest and best-funded rival: no new interface, just a number you text. Friend-to-friend coordination already ships where Lilith plans it for M3 (F04). Its first privacy policy was criticised as overreaching and rewritten, so Lilith's data terms should be plain and published before invites |
| **Manus 2.0** and **Cue** (Butterfly Effect, Singapore; launched 2026-09-28) | Cue: personal agents, each with its own email address, phone number, wallet and computer; several agents in one group chat working toward a shared goal and handing work to each other; takes calls forwarded to it and leaves a summary; scan a restaurant's QR code to order or hold a place. Manus 2.0: the Cascade harness (a claimed 32% cost saving from "one tested configuration"), a rentable always-on Ubuntu cloud computer, event-triggered automations with test runs and failure pauses, a studio that edits short videos and builds multiplayer browser games, and control of your own desktop from your phone. The research found no independent hands-on review as of 2026-09-30 | Cue: web, macOS, Windows, Android (iOS awaiting review). Manus: web, desktop, mobile, Telegram, WhatsApp Business, Slack, LINE, a REST API | Cue free in invite-only early access, later price unpublished; Manus Free, Pro from $20 (credit-metered), Cloud Computer $30–$50 a month | Revision 35 takes nine corrections and three additions (Eve research `7bb205ec`). Cue's examples point to Stripe Link, whose agent wallet admits only accounts created in the US and Canada (P02, P11); its numbers work in unnamed "select countries". Manus's Terms defer to "legally required" disclosures rather than committing that the agent says it is AI; principle 9 goes further, and A17 carries it to inbound calls. Researchers reported two injection vulnerabilities in September 2026, both patched: an obfuscated script in a single email ran in the victim's environment, and a shared project's instructions loaded as trusted configuration into other members' agents, where a remote-desktop admin password sat in the sandbox (6.18, 6.2). Whether Cue coordinates several agents' outbound actions is unpublished (R11). Credit metering frustrates Manus users, so Lilith's estimate becomes a ceiling (6.12) |
| **Google Gemini Spark** (I/O 2026-05) | 24/7 background agent on cloud VMs; Workspace APIs; own Gmail address; time and conditional triggers; Canva/OpenTable/Instacart; AP2 commerce coming; Call for Me; CC family agent; Health Coach | Gemini apps, Android Halo, XR audio glasses, cars, TV | Pro $19.99 (Spark since July), Ultra $100/$200 | Always-on layer, triggers, family agent, commerce protocol (UCP + AP2) |
| **OpenAI ChatGPT** | ChatGPT Work (hours-long tasks with approval checkpoints), cloud browser with takeover, scheduled tasks and web monitoring; Pulse retired 2026-06-17; Instant Checkout moved to merchant apps | Web, desktop, mobile, CarPlay | Free/Go with ads, Plus $20, Pro $100/$200, Pro 500 $500 (from 2026-09-29) | A daily push feed fatigued people; in-chat checkout converted ~3x worse than merchant checkout |
| **OpenAI dots** (DevDay 2026-09-29) | Always-on agents on GPT-6 Astra, each with its own cloud computer and browser, working toward the member's goals around the clock; 4,000+ apps through plugins; the member's laptop with permission; "proactive research" while the member is away that is read-only (it cannot send, edit or control a computer); learns preferences from feedback; signs in to supported sites with saved passwords the model never sees; built-in act-or-ask rules, member Custom Rules (allow, block, require approval), an auto-review of consequential actions, a monitor that can pause a dot, an Activity View; password changes stay with the member; Pages in a shared ChatGPT Space that people and dots edit together; enterprise "specialist dots" with their own identity, credentials and defined responsibility (pilots) | ChatGPT desktop, web and mobile; started from Codex; Slack and Teams; voice calls; SMS announced | First dot included with Business Premium and with Pro, though reports differ on which Pro tiers qualify (Engadget: every Pro subscriber; BGR: Pro 200 and Business Premium) and OpenAI's own pages could not be read on 2026-09-29; chatting with a dot does not count against usage limits, its Codex and Work tasks do; Enterprise, Edu and Healthcare once an admin enables it; closed to Pro members in the EEA, Switzerland and the UK for now | Confirms W, R, T09, X03 and X09. Revision 22 takes seven lessons: self-started work only reads (principle 11); member rules before the first actions (X09 to M1); talking not metered against the task budget (6.19); standing duties with their own grants (R08); proof attached to every done (A13); where each task runs and no dropped messages (R01, R04, S09), since the first hands-on review reported permission errors, lost messages and confusion over where work ran; and flow labels enforced on shared lists, where OpenAI's shared Pages, as VentureBeat reports them (not confirmed first-hand), only warn that memory written onto a Page becomes visible (F03) |
| **Anthropic Claude** | Cowork (desktop knowledge work) with an isolated built-in browser; Claude in Chrome; categorized memory with import; health integrations | Desktop, Chrome, web/mobile | Pro $20 | Selective credential import (banking/email/SSO excluded by default); published injection rates |
| **xAI Grok Bot** (beta 2026-08-11) | Always-on agents on a cloud computer with a browser and shell: they sign into ordinary sites and apps as the member (no API or MCP needed), work inboxes and files, keep running after the member's device disconnects and come back only for approval; show one a workflow once and it repeats it; memory across conversations; several Bots coordinate in group chats | Desktop and iOS; the member messages each Bot like a colleague | Sold through SuperGrok (all tiers) and Cursor Pro, Pro+, Ultra and Teams, with its own usage allowance; price unpublished; Enterprise on a waitlist | Aimed at teams. One Firecracker microVM per user, but every Bot on it shares every login and file ("treat a login or file on the computer as available to every Bot that user runs"), and network policy and audit logs are Enterprise-only. Lilith's agent holds only surrogate credentials, and Sentinel and the flight recorder cover every member (6.2). Worth copying: login, two-factor and payment steps are handed to the member. A recorded demonstration only drafts a skill in Lilith, promoted on replayed gain (K03) |
| **Microsoft Copilot** (consumer) | Long-term memory across chats; "Browse with Copilot" acts across the web in Edge; the consumer and work Copilot apps merged | Microsoft 365 Copilot app, Edge, Windows | Browse with Copilot: Microsoft 365 Premium, US only | Microsoft retired its Mico avatar, group chats, AI podcasts and consumer Deep Research by 2026-08-18, and folded Edge's Copilot Mode into the browser (2026-05-13). An animated companion did not earn its place, so Lilith's avatar (V05) stays optional and ships only if it measures well |
| **Amazon Alexa+** | Voice-first bookings (Thumbtack, OpenTable, Uber, Ticketmaster), shopping agent, proactive home alerts | Echo, app | Free with Prime or $19.99 | Voice-first errands end to end |
| **Apple** | Apple Intelligence app actions; Health Insights, "For You", on-device movement evaluations; Private Cloud Compute | Apple devices | Free | Verifiable privacy (PCC transparency log); on-device |
| **Samsung Galaxy S26** (Galaxy AI with Gemini and Perplexity; Gemini automation live 2026-03) | Gemini runs DoorDash, Grubhub, Lyft, Starbucks, Uber and Uber Eats in a virtual window and stops just before the confirm or pay step; "Hey Plex" puts Perplexity into Samsung Notes, Calendar, Reminder, Clock and Gallery; Galaxy AI moves work between Gemini, Perplexity and Bixby | Galaxy S26 phones (automation: US, English) | Included with the phone | Phone makers hand system-level agent slots to partners; an installed app like Lilith gets none of those hooks (section 11). Stopping before the pay step matches Lilith's exact-action approvals (T09) |
| **ByteDance Doubao Phone Assistant** (with ZTE Nubia; NaviX Ultra 2026-09-17) | An operating-system-level agent that plans and carries out tasks across apps, with a dedicated AI key and fingerprint check | Nubia phones; China | From RMB 5,999 | The first model (2025-12) drove other apps by injecting simulated taps; WeChat, Alipay and banking apps refused its logins within days. Agents that drive another app's screen get locked out, so Lilith acts through connectors and official APIs first (section 11) |
| **Alibaba Qwen App** (agent features 2026-01-15) | Orders food and groceries (Taobao), plans and books travel (Fliggy), navigates (Amap), guides public-service paperwork, and pays inside the chat through Alipay after explicit confirmation | App; China (public testing) | Free; 100M+ monthly users within two months of its 2025-11-17 beta | Checkout completes because one company owns the shops, the travel agency, the maps and the wallet. Lilith owns none of these, so it depends on the agent-payment tokens and merchant checkout in 6.5 |
| **Perplexity Comet**, **Genspark** | Agentic browser; deliverable-producing agents; Genspark places real phone calls in 40+ countries | Browser, web, messaging | ~$20, credit-metered | Credit metering frustrates users; calls are a real differentiator |
| **Poke**, **Rocket Money Rowan** | Text-first agents over iMessage/SMS; cancel-by-reply subscription agent | Messaging | — | Messaging-native, success-fee pricing |
| **Pine AI** | Calls, emails and uses websites to lower bills, cancel subscriptions, dispute charges, recover refunds and file complaints, following up until each case resolves; uses personal data "only the essentials, temporarily and with approval" | Web; English, Japanese, Spanish, French, Chinese | Credit-based subscription; annual plans negotiate bills without a percentage success fee | The nearest rival to the consumer advocate (A09). Pine dropped the success fee on annual plans, so test flat pricing against Q03's success fee before committing to it |
| **Hermes Agent** (Nous), **OpenClaw** | Open-source personal agents: skills that write themselves, memory with approval queues, cron, subagents, many gateways, any model | Self-hosted, messaging gateways | Free + your model | Model freedom and skills; OpenClaw's RCE, plaintext credentials and malicious skills show what not to ship |
| **Character.AI** (companions) | Character chat for adults; since 2025-11-25 no open-ended chat for under-18s, who get interactive Stories instead; age assurance from an in-house model plus Persona | App, web | Freemium | Law now sets the companion safety floor: SB 243 (California), GBL Art. 47 (New York), China's rules on human-like AI, and an FTC 6(b) inquiry into seven firms. Lilith's escalation ladder and teen plans are built to that floor (table in 6.8) |
| **Ash** (Slingshot AI) | An AI built for therapy conversations, positioned as a wellbeing product | App; US (left the UK 2026-01-23) | — | Ash left the UK because it found no clear regulatory path for a wellbeing product, and four US states now bar AI-only therapy. Lilith's coach and programme rungs must never present as therapy where the law forbids it (6.8) |
| **WHOOP Coach** (announced 2026-05-08) | An AI coach with My Memory (the member adds, edits or deletes the context it uses) and Proactive Check-Ins; live on-demand video visits with licensed clinicians who see biometrics beside synced health records | WHOOP band and app; clinician visits US from summer 2026 | Membership; visit price unpublished | The coach-to-licensed-human ladder and editable memory are already on sale from a wearable maker. Lilith's edge has to be acting on the member's life (calendar, spending, sleep window) around the health signal |
| **Brave Leo** (agentic browsing in early access, updated 2026-05-05) | Browses for the member in a brand-new profile with none of the main profile's cookies or logged-in state; an "alignment checker" model sees the prompts and the agent's response but not raw page content; Bring Your Own Model for local or remote models | Brave browser, all channels (opt-in flag) | Leo free; Leo Premium paid | A consumer precedent for two Lilith choices: keep the agent's browsing apart from the member's own sessions, and let members bring their own models (B) |

Also watched, not analysed here: Cleo Autopilot, Opera Neon, Mistral's
assistant, Open-AutoGLM, Tencent's WeChat agent, Kakao's Kanana, Headspace Ebb,
Wysa, Proton Lumo, Kagi Assistant, Amazon Bee and DoNotPay.

Full research with sources: section 12.

## 4. Positioning and differentiators

**Lilith V1.3 is the personal agent that does everything Muse does, guards your
wellbeing while doing it, and runs on the models you choose.**

What nobody ships well today and Lilith will:

1. **Guardian mandates** — spending, calendar and messaging permissions that read
   wellbeing state (hold non-essential purchases over a threshold for 24 hours
   after a poor-sleep night; no delivery orders after the wind-down time; decline
   meetings inside the sleep window), expressed as signed mandates the payment
   and calendar layers enforce.
2. **Nyx sleep defender** — reads sleep debt and readiness, renegotiates
   tomorrow (moves early meetings, shifts deliveries), and offers a screenless,
   whispering bedside mode.
3. **Interruption budget** — a daily cap on proactive pings, quiet hours tied to
   sleep, batching into ritual moments (morning Tara, evening Nyx), learning from
   dismissals. The answer to Pulse's retirement.
4. **Baseline-grounded alerts** — Lilith speaks up only when the member deviates
   from their own baseline, and says why ("your HRV is 18% below your 30-day
   norm").
5. **Capability ledger and local Sentinel** — every statement Lilith makes about
   her own access is rendered from a signed permission record, never from model
   narration; the desktop and phone clients run a device-side gate on files and
   egress. The direct answer to the Muse iMessage failure.
6. **Flight recorder** — a replayable timeline across cloud workspace, device,
   calls and payments: each step, what data left, to whom, and one-tap revert
   where the action is reversible.
7. **Contextual-integrity memory** — every memory carries a flow label (health
   and mood facts never leave Tara/Nyx, never reach outbound email or shopping),
   a weekly "what I learned about you" review, decay for emotional memory, and
   open export/import with provenance. The labels hold on shared household
   lists and plans too; OpenAI's shared Pages, as VentureBeat reports them (not
   confirmed first-hand), only warn that memory written onto a Page becomes
   visible.
8. **Escalation ladder** — coach → evidence-based programme (CBT-I, motivational
   interviewing) → a licensed human the agent books with a consented summary →
   an audited crisis protocol that meets California SB 243.
9. **Arete plan objects** — goals become visible multi-week plans with
   milestones and weekly reviews; the agent does the logistics.
10. **Consumer advocate** — refunds, returns, cancellations, price protection and
    chargebacks with evidence packets and deadlines; phone cases that persist
    across days.
11. **Model sovereignty** — bring any model (provider keys, OpenRouter,
    self-hosted, on-device), a BYOK/on-device price tier, a cost preview and
    receipt for every task, and a no-ads promise.
12. **Glass-box agent** — reasoning, tool calls, memory sources and policies
    applied are inspectable and replayable for every turn, with honest labels
    where a provider only exposes a summary.
13. **Household with kids** — a family agent that includes teens and children
    (Google CC is 18+ only), shared bedtime and meditation rituals, per-member
    flow labels, friend-to-friend scheduling agent-to-agent.

## 5. Principles (binding on every V1.3 task)

1. **Nothing fabricated.** The repository's quality bar holds: every capability
   either works against the real service or fails closed with an honest
   `not_configured` / 503 / `{configured:false}`. A connector that is not
   connected says so; an action that did not reach the service is never reported
   as done.
2. **Ground truth over narration.** What Lilith says about her own access,
   spending, actions and memory is rendered from records (capability ledger,
   flight recorder, mandate store), never generated by the model — and those
   records are reconciled against live provider scopes and OS permissions,
   because a signed grant proves authorisation, not actual access.
3. **Confirm first; autonomy only by mandate.** Every external side effect asks
   first unless the member has granted a scoped, time-bounded, revocable mandate
   that covers it. Mandates are signed records, not preferences in a prompt.
4. **Agent prepares, merchant checks out** — by default. In-agent checkout is
   used only where a protocol gives the member a scoped token and purchase
   protection.
5. **Wellbeing is policy, not tone.** Guardian rules, interruption budgets and
   escalation paths are enforced by code paths with tests, not by prompt text.
6. **Contextual integrity.** Data carries flow labels; the policy engine refuses
   a flow the label forbids, whatever the model asks for.
7. **Model sovereignty with honest governance.** Members choose their models.
   Oshun-operated routes keep the Task 14.2 attestation gate; member-owned routes
   are a separate admission class whose training, retention and region posture
   is shown to the member before first use.
8. **Cheapest capable by default, cost visible always.** Routing picks the
   cheapest model that meets the task's measured bar; every task shows a cost
   preview and a receipt.
9. **Disclosure.** Lilith identifies herself as an AI on every outbound email,
   message and call (EU AI Act Art. 50, in force 2026-08-02) and meets SB 243
   duties for companion chat. What she generates carries a machine-readable
   mark that it is AI-generated (Art. 50(2), T10).
10. **Reversible where possible.** Every action records how to undo it; the UI
    offers undo wherever the target service allows it.
11. **Unasked work only reads.** Work Lilith starts on her own initiative, which
    no request, trigger, standing duty or guardian rule of the member's asked
    for, may read, prepare and tell the member; it never sends to anyone else,
    changes anything or controls a computer, whatever mandates exist. What it
    finds reaches the member as a prepared draft awaiting one tap.

## 6. Scope

Each pillar below is a board epic in family `lilith-v1-3` (key in brackets). The
atomic tasks under each carry paths, the work, a done-when test and
dependencies; the epic is integration verification only.

### 6.1 Programme and release gate [`G`]
- A fail-closed V1.3 capability gate in `release-scope.ts` (same pattern as
  payment rails: unknown ids refuse). Capabilities enable **per milestone**.
  **M0 admits no one outside Oshun** (Elikem's decision, 2026-09-28): only
  Elikem and test accounts he controls. An allowlist read by the gate enforces
  this, so no age or region question arises in M0. From M1 the gate also admits
  a **personal invite cohort** (Elikem's decision, 2026-09-28): Elikem adds each
  invitee to the allowlist himself and confirms they are 18 or older when he
  adds them, each entry recording who added it and that confirmation; the gate
  refuses everyone else, and the per-capability region and age rules (Q05)
  still apply. No open or self-serve signup exists until age assurance (T05)
  passes. The cohort enters once that milestone's exit review passes. Everyone is admitted at GA. Epics
  verify per milestone for the capabilities that milestone enables.
- A dedicated agent-runtime service boundary in the BFF, feature flags per
  capability, V1.3 route-matrix rows for every new page, walkthrough coverage.
- The release tables in Eve documents `5b3e02e1-e65c-83d8-969d-40e6247463ba`
  (Lilith — V1 Brand & Naming Record) and `a0fa87e0-9fd3-8c5d-9986-d008554e900a`
  (Oshun V1 Features) gain the V1.3 row after approval, as revision-checked
  content edits.

### 6.2 Agent workspace and containment [`W`]
- **Per-member secure workspace**: an isolated sandbox (Firecracker/gVisor-class
  microVM or container with seccomp) per member for browser and code work, with
  resource quotas, idle suspend and wipe-on-delete.
- **Sentinel**: a separate policy process through which every connector call and
  every byte leaving the workspace passes (model calls straight from a member's
  device are allowed only after the device-side gate ships, and are recorded
  too); allow / deny / escalate-to-member decisions with
  hostname, resolved IP, method, path and decoded body; the model cannot reach
  the network except through it.
- **Credential vault with surrogates**: OAuth tokens and passwords live in a KMS
  envelope-encrypted vault; the agent only ever handles surrogate handles that
  Sentinel swaps at the edge.
- **No reusable control-plane credential in the sandbox**: the remote-browser
  stream server's admin credential and every other control-plane secret live
  outside the member's workspace. A hostile process inside it finds none in
  environment variables, readable files, mounts or process metadata, and cannot
  call a privileged operation. In one Manus vulnerability reported in September
  2026, a remote-desktop admin password sat in the sandbox environment.
- **Workspace files view** (W12): the member sees the files Lilith saved in
  their workspace (downloads, drafts, exports) and can export or delete them.
  Paths are confined to the member, previews are inert, flow labels apply, and
  deleting a file never corrupts evidence the flight recorder holds.
- **Taint tracking**: data read from a member's accounts marks the process; a
  tainted process loses auto-allow, and its egress passes only under a covering
  mandate or a member approval.
- **Capability ledger**: signed, append-only record of every grant, scope,
  duration and revocation; the only source for "what can you access?" answers
  and for the permissions UI, reconciled against provider-reported scopes and OS
  permissions before every access answer, with drift alerts.
- **Flight recorder**: an append-only per-member timeline of plans, steps, tool
  calls, data egress (what, to whom), costs and outcomes, replayable in the UI,
  exportable, with undo links. Payloads are encrypted per item so erasure
  destroys the key (crypto-shredding) while the log keeps hashes and tombstones.
- **Member-key encryption** of the workspace and memory (the promise Muse has
  not shipped yet) — design and ship for memory first, workspace second — with
  a key lifecycle: recovery, rotation, lost-device revocation and time-bounded
  delegation of decryption rights for unattended work.

### 6.3 Connectors and permissions [`C`]
- Connector framework: OAuth 2.1 + PKCE, incremental scopes, refresh, revocation
  webhooks, health checks, per-connector rate limits.
- First-party connectors: Google (Gmail, Calendar, Drive, Contacts), Microsoft
  (Outlook mail/calendar, OneDrive), Apple (Calendar/Reminders/Contacts through
  the native app), Notion, Slack, WhatsApp Business, Telegram, Spotify, Strava,
  Oura, WHOOP, Garmin, Withings, Apple Health / Health Connect (device bridge),
  Uber, OpenTable/Resy, Instacart, Expedia/Booking, Amazon order history (email
  parsing where no API), banking read-only through an open-banking aggregator
  (Plaid/TrueLayer) for subscription triage.
- Grant durations: one-time, session, task, time-bounded, perpetual, or bound to
  a standing duty and revoked when the duty ends (6.6); per-sender and
  per-folder email scopes; read-only vs send. A member may connect several
  accounts of one provider (two Gmail accounts, say); each grant is bound to
  one account, and every send names the account it uses.
- MCP client for third-party connectors (remote MCP servers with OAuth), MCP
  Apps rendering inside Lilith's UI in a sandboxed frame.

### 6.4 Actions in the world [`A`]
- **Email**: triage, summaries, drafts in the member's voice, send-with-approval,
  follow-up tracking; an agent address (`<handle>@agent.oshunsynthetics.com`,
  Decision 10) members can CC or forward to, where the handle is one the member
  picks or a random one and never their real name by default; signed AI
  disclosure footer. Mail from the agent address shows the persona name with
  Lilith behind it, for example `Morgan (Lilith AI) for Sam`. Mail the member
  approves from their own mailbox is from the member and carries the footer.
  Who may start work by mail: mail the member writes from their own verified
  addresses is a member request, but only the member's own words in it; text
  they forward or quote is third-party content (6.18). Other approved senders
  the member lists (a partner, say) can only start the triggers or duties the
  member tied to them, never give instructions or gain grants; mail from anyone
  else is third-party content. A sender counts as verified only when DMARC
  passes with alignment to that address's domain (through SPF or DKIM, so a
  forward that breaks SPF but keeps an aligned DKIM signature still counts);
  anything else waits for confirmation in the app. A standing duty may use a
  sub-address, so replies thread back to it.
- **Calendar**: read/write, time negotiation with other people by email,
  protected blocks (sleep, rituals), travel-time awareness.
- **Browser and forms**: remote browser in the workspace; watch-live; take-over
  mode where no screenshots are taken and the model never sees typed secrets;
  form filling from the member profile with field-level approval.
- **Phone calls**: outbound calls (telephony provider + realtime voice) that
  open with the persona name, Lilith and the AI disclosure together ("Hi, I'm
  Morgan, an AI assistant from Lilith calling for Sam"; Decision 10), navigate IVRs, wait on hold, stream a live transcript, allow
  take-over, and keep a multi-day **case file** with callbacks.
- **Her own number** (A17): members can forward calls to Lilith (on busy or no
  answer) or give out her number. She answers only under a duty with a signed,
  bounded mandate that lists what she may tell which callers. Caller ID can be
  spoofed, so a caller counts as a named caller only after the verification the
  mandate sets (a passphrase the member chose, or a call back to the saved
  number); everyone else hears only what the mandate marks public, and a flow
  label that forbids disclosure binds even for a verified caller. Every call opens
  with the disclosure and the persona ("You've reached Morgan, Sam's AI
  assistant from Lilith"). A commitment (a booking, a payment, a promise) needs
  an exact-action approval or a covering mandate; what a caller says is
  untrusted content and starts no work; each call ends as a summary and call
  metadata in the flight recorder; spam screening and a block list come with it; and the
  member's real number is never given out without consent. Counsel approves the
  purposes, notices, recording and retention rules for each supported
  jurisdiction, separately for audio, transcripts, summaries and metadata.
  Where those conditions cannot be established she keeps no audio or
  transcript, and where counsel's rule forbids processing call content at all
  she does not answer calls from that jurisdiction. Texts from her number
  follow the rules for messages on the member's behalf (A11). Manus announced on 2026-09-28 that Manus and Cue agents can take
  forwarded calls, and Instinct is a texting-first agent with its own number.
- **Messaging on the member's behalf**: WhatsApp/Telegram/SMS sends with
  approval; iMessage as the member only from their own Mac through the desktop
  client under the local Sentinel (iPhone does not allow it), never a relay that
  holds their Apple ID.
- **Bookings**: restaurants, travel, classes, appointments, home services —
  through partner APIs first, browser second.
- **Consumer advocate**: returns, refunds, cancellations, price protection,
  chargebacks with evidence packets, statutory-rights templates by region and
  deadline tracking; subscription triage weighed against actual use.
- **Documents**: fill PDFs and permission slips, sign with approval, file them to
  the member's drive.
- **Execution reconciliation**: idempotency keys, duplicate suppression,
  postcondition checks and recovery when a call succeeds but its response times
  out — an action is reported done only when its postcondition holds, and the
  report carries its proof: the confirmation page, booking reference, message as
  sent or call-transcript excerpt, linked from the flight recorder (never a
  screen from take-over mode, never data whose flow label forbids the channel).
- **Deep research**: cross-source investigation with a citation per claim,
  freshness checks and explicit contradictions.
- **Artifacts**: documents, spreadsheets, presentations and small apps created,
  edited together, versioned and exported.
- **Smart home**: Matter, Home Assistant, Google Home and SmartThings with state
  reconciliation, routines and hard safety rules for locks, alarms and heating.

### 6.5 Commerce and payments [`P`]
- Stripe Link for agents: OAuth wallet access, per-task one-time cards and Shared
  Payment Tokens scoped by amount, currency and merchant, approved in Link. A
  scoped token is not purchase protection: protection comes only from a
  programme the payment provider actually offers — Stripe Link's purchase
  protection where it covers the merchant and category, and card-network
  chargebacks through Stripe disputes — shown to the member before approval,
  with claims filed from the receipts ledger (P10). Link's agent wallet admits
  only accounts created in the US and Canada (Link's FAQ, read 2026-09-30).
  P02 reads eligibility from Link's own response, never from the member's
  location, and an ineligible account fails closed with the provider's reason
  and falls back to merchant hand-off.
- **Members outside the US and Canada** (P11): an assessment of the routes that
  could serve them: Link, if it opens to accounts created elsewhere; card-network
  agent programmes through a PSP that offers them in the UK; open-banking
  payment initiation, each payment approved in the member's bank app; and the
  crypto rail and merchant recognition, evaluated only, since P09 and P08 are
  M3. For each it records dated availability, Oshun's regulatory role (whether
  the route would make Oshun a payment or e-money business) and the purchase
  protection a provider actually backs (principle 4), with how the route joins
  receipts, disputes and the regret window. It ends in a recommendation for
  Decision 5; a build task follows only after Elikem decides, and merchant
  hand-off stays the default meanwhile.
- Google UCP + AP2: Intent / Cart / Payment mandates as verifiable credentials;
  Universal-Cart-style "buy only when criteria are met" watchers.
- OpenAI/Stripe ACP order hand-off where merchants support it; card-network
  agent tokens (Visa Trusted Agent Protocol, Mastercard Agent Pay) and
  Cloudflare Web Bot Auth request signing so merchants can recognise Lilith.
- x402 and the existing crypto rail for agent micro-payments.
- **Guardian mandates** (wellbeing-aware limits, each with a member override and
  emergency exceptions), a **regret window** (hold N hours with one-tap undo
  before charge, with merchant-specific rules for price, stock and cancellation
  shown up front), a **receipts and returns ledger** with
  automatic price-protection and return-deadline reminders.
- Builds on `libs/oshun/fiat-payments` (V1.1) and `libs/oshun/payments-bridge`.

### 6.6 Proactivity and background work [`R`]
- Background tasks that run with the member's devices off, a concurrency limit
  per plan, pause/resume/cancel, progress in a live-activity style status surface
  (web, iOS Live Activities, Android ongoing notification, watch). Every task
  shows where it runs (Lilith's cloud workspace, the member's Mac or phone), and
  its messages reach the member's channels at least once without duplicates; a
  message that cannot be delivered waits in a visible outbox and is never
  dropped.
- **Self-started work only reads** (principle 11): when Lilith looks for work on
  her own, the run holds read grants only, and what it finds arrives as a
  prepared draft awaiting one tap. OpenAI's dots run their proactive research the
  same way.
- Triggers: time, recurrence, natural-language schedules, conditions on
  connected data (flight changed, price dropped, email from X); health-baseline
  deviations once personal baselines exist (M2). Builds on `libs/oshun/assistant/src/scheduling` and `watchers`.
  Before a trigger goes live the member can dry-run it: the run passes through
  the tool gate with every write simulated (no sends, bookings, payments,
  reservations or read-status changes) and shows what would have happened.
  Failures are classified (transient, authorisation, provider, own bug), and the
  same class repeating past a set threshold pauses the trigger, with backoff and
  a message naming the failure. Each trigger says whether it fires on pushed
  events or on polling, and how often. A plain-language routine becomes a card
  listing trigger, conditions, actions, accounts, grants and budget; activating
  it signs only the mandate on the card, and anything outside it still waits
  for approval (T09). Manus's Automations ship test runs and failure pauses.
- **Interruption budget** and quiet hours tied to sleep, batching into ritual
  moments, learning from dismissals; every proactive message states its trigger.
- **Plan objects**: long-horizon plans with milestones, reviews and re-planning,
  visible and editable.
- **Standing duties**: an ongoing job the member names ("keep my subscriptions
  trimmed", "get my deposit back", "watch fares to Lisbon") is a plan object
  with its own grants, mandate, budget, triggers and activity view; it cannot use
  a grant outside its brief, and ending it revokes its grants. OpenAI's
  specialist dots get their own identity and credentials for the same reason.
- **Status line**: in M1, a one-line status in the chat header names the task
  in focus and how many others are running; tapping it opens the activity
  timeline (W07) at that task and the grants she holds right now, from the
  reconciled capability ledger (W06, W08). Once V05 ships, the same line sits
  under her presence, as Meta's Muse shows one under its avatar.
- **Ideas**: an Ideas view and first-week tips generated from the member's
  goals, plans and what Lilith has learned, for the "I don't know where to
  start" problem Muse's designers met in testing. Ideas come from self-started
  runs, so they only read (principle 11), within a bounded background-work
  allowance with a rate limit and cost receipts (L07). Only data whose flow
  labels permit the Ideas view contributes, and a correction or a forget
  withdraws the ideas built on it (M07). Ideas stay drafts: none activates a
  plan, duty, grant or action. They wait in the view and spend the
  interruption budget only if the member asks for them to be pushed.
- **Goal rooms** (R11): every duty and teammate working on one goal appears in
  one thread as a lane the member can name ("Bills", "Trip"). Each lane is
  still Lilith and signs as the persona (Decisions 10 and 12), and the member
  sets direction and makes the final call there. A handoff between lanes is a
  typed record (from, to, what, why) that passes work, never grants. Outbound
  actions carry a member-scoped conflict key (recipient, merchant, calendar),
  so a second lane acting on the same key waits visibly and re-checks its
  grants and approvals before it proceeds. The keys are part of execution
  reconciliation (A13): a stale worker's write is refused by its fencing
  token, and where a provider offers no idempotency key or queryable outcome,
  an uncertain outcome keeps the key held and the action blocked until it is
  reconciled or the member resolves it, so a timeout, crash or lease expiry
  never produces a second external effect. Cue puts several agents, each holding email, a phone
  line and a wallet, in one group chat; whether it coordinates their outbound
  actions is unpublished.
- A morning and evening ritual brief the member can shape — not a feed.

### 6.7 Memory and personalization [`M`]
- Flow labels on every memory (contextual integrity) enforced by the policy
  engine; health and mood memories siloed to Tara/Nyx by default.
- "Why did you say that?" — memory sources shown per answer.
- Background consolidation, emotional-memory decay, cost-tiered storage.
- Weekly review ritual (approve / edit / reject what Lilith learned).
- Open export schema; import from ChatGPT, Claude and Gemini exports with
  provenance; forget-by-voice, propagated into datasets and adapters.
- **Correction propagation**: correcting a fact invalidates derived memories,
  plans, caches and pending actions; source precedence and temporal validity
  are explicit.
- Extends the Iris bridge and `/profile/memory`; evaluates `libs/iris/memory` and
  `libs/psyche/memory-*` before adopting them.

### 6.8 Wellbeing intelligence [`H`]
- Health data in: HealthKit and Health Connect (native bridges), Oura, WHOOP,
  Garmin, Withings, Fitbit; consented, flow-labelled, minimised.
- Personal baselines (rolling medians, robust z-scores) with deviation alerts
  that cite the numbers.
- **Nyx sleep defender**: sleep window from sleep debt and chronotype; renegotiate
  early commitments; bedside night mode.
- **Tara ritual engine**: context-triggered micro-practices (after a stressful
  meeting, before sleep) on phone, watch, earbuds or glasses.
- **Arete plans** with logistics; **Nisaba** reading and journaling companion
  that acts (reading list, library loans, scheduled reading, weekly reflection);
  **Veritas** claim ledger as a daily delta (after V1.2); **Metis** spaced
  repetition scheduled into ritual moments (after V1.2).
- **Escalation ladder** with a clinically reviewed crisis protocol, regional
  crisis lines, licensed-provider booking with a consented summary, and an
  auditable record.
- Evidence-based programmes delivered as plans (CBT-I for sleep, motivational
  interviewing for habits) with clinical review **and a regulatory assessment of
  intended use** (FDA general-wellness policy, UK MHRA, EU MDR) before launch.
- **Wellbeing harm evaluation**: dependency, anxiety amplification and harmful
  interventions measured for guardian rules and nudges before they ship.
- **Laws that bind the ladder** (checked 2026-09-28; re-checked before each
  milestone exit). The coach rung is wellness support and never presents as
  therapy. The programme rung (CBT-I, motivational interviewing) is the one most
  at risk: where a state bars AI therapy it runs only under a licensed
  professional, or not at all. Q05 enforces the region rules per capability; T04
  carries the companion duties.

| Jurisdiction | Law | In force | What Lilith must do |
|---|---|---|---|
| California | SB 243, companion chatbots | 2026-01-01 (signed 2025-10-13); annual reporting from 2027-07-01 | Say it is an AI wherever a person could be misled; publish a crisis protocol that prevents suicide and self-harm content and refers to crisis services; tell users, on the app, the browser and every other format they reach Lilith through, that companion chatbots may not be suitable for some minors; for known minors, a break-and-not-human reminder at least every three hours and no sexual content; use evidence-based methods for measuring suicidal ideation; annual report to the Office of Suicide Prevention; private right of action, at least $1,000 per violation |
| New York | General Business Law Art. 47, AI companion models | 2025-11-05 | Detect suicidal ideation and self-harm and direct the user to crisis services; tell users it is not a human at session start and every three hours of continued use; the attorney general enforces |
| Illinois | Wellness and Oversight for Psychological Resources Act (HB 1806) | Signed 2025-08 | No AI mental-health treatment or therapeutic decision-making; licensed professionals may use AI only for administrative and supplementary support |
| Nevada | AB 406 | 2025-07-01 | Do not offer AI that provides professional mental or behavioural health care, or say that it can; up to $15,000 per violation |
| Utah | HB 452, mental-health chatbots | 2025-05-07 | Disclose the AI before access, after seven days away and when asked; do not sell or share health information or user input; label ads and never target them from user input; up to $2,500 per violation |
| Maine | LD 2082 (Public Law 2025, ch. 687) | Signed 2026-04-13 | AI may not make therapeutic decisions, interact with clients therapeutically or produce treatment recommendations without a licensed professional; administrative support is allowed |
| Rhode Island | S 2197 (AI in mental health care); S 2195 (AI companion models) | S 2197 2026-06-22; S 2195 2027-01-01 | AI therapy is offered to the public only through licensed professionals; companions refer self-harm to crisis services, remind users every three hours that they are not talking to a human, and report yearly to the attorney general from 2027-07-01; up to $15,000 per day |
| China (only if served, Decision 1) | Provisional Measures on AI anthropomorphic interactive services | 2026-07-15 | No inducing emotional dependence or addiction; a reminder after every two hours of continuous use; crisis handling that can reach a guardian or emergency contact; no virtual intimate relationships for minors, parental consent under 14; AI content labelled |
| United Kingdom | Medical-device rules | — | Ash withdrew on 2026-01-23 for want of a clear path for wellbeing products; covered by the MHRA assessment above |
| United States (federal) | FTC 6(b) inquiry into companion chatbots (2025-09-11) | — | Not a rule; expect questions on minors, monetisation, harm testing and age limits |

### 6.9 Surfaces and devices [`S`]
- Web/PWA and native iOS/Android (V1.1 base): Live Activities / ongoing
  notifications for running tasks, widgets, share-sheet "ask Lilith", App
  Intents / Android App Actions so the OS assistant can call Lilith.
- Desktop client for macOS, Windows and Linux grown from the V10 Tauri shell;
  local file and app control only after the local Sentinel ships, never Full
  Disk Access by default. Packaging, per-OS integration, the CLI, the
  self-hosted node, the browser extension and de-Googled Android are §6.20.
- Messaging: WhatsApp (if its terms permit), Telegram, SMS/RCS, email-the-agent
  and Apple Messages for Business (members message Lilith in iMessage; it does
  not let Lilith send as the member) — the full Lilith agent, not room commands.
- Watch: voice ask, approvals from the wrist, ritual prompts.
- Audio-first devices: earbuds and audio glasses (Meta, Android XR) through
  their SDKs where they exist; CarPlay / Android Auto; smart speakers and TV
  through the Rail TV surface.
- Bedside night mode: screenless, dim, whispered, sleep-safe.
- Continuity: start on one device, finish on another, one task state.
- **Side chats and sending while she works**: one long main conversation, with
  side chats the member opens for a topic that needs its own context; memory
  and its flow labels hold across them (M01). The member can send several
  messages in a row or interrupt without waiting. Each message is acknowledged
  and kept with its chat, and conversation stays within the conversation
  fair-use allowance (6.19). A request for agent work becomes a task (R01) and
  draws on the work budget; a steer, pause or cancel names the running task it
  applies to. Each reply says which chat and task it answers. Builds on the
  room threads in `apps/oshun/web/src/components/assistant/AssistantPanel.tsx`.

### 6.10 Voice, realtime and presence [`V`]
- Full-duplex speech-to-speech with barge-in, proactive silence and hand-off to a
  reasoning model mid-conversation; breath-paced voice for practice; whisper
  mode at night; live translation.
- Camera and screen share with consent and on-screen indicator.
- A calm, deliberately non-photoreal presence as the default; an optional
  realtime avatar built on the Mentor Presence pipeline once its release gate
  passes. Presences are fictional: none depicts an identifiable real person,
  photoreal or stylised, whatever licence is offered, and members cannot
  upload a photo of a person to animate. The design is adult, not a mascot,
  because the audience is 18+.
- **The avatar is a swappable plugin on the voice agent**, the pattern LiveKit
  Agents uses for avatar providers. Meta's Muse Realtime Avatar has no public
  developer API (2026-09-30), so a bake-off (V09) measures candidate routes:
  HeyGen LiveAvatar, Tavus, D-ID, Anam, Synthesia's Interactive Avatar API, and
  the Apache-2.0 LiveAvatar model self-hosted on GPU workers. It records each
  route's integration with the chosen voice provider, turn latency, audio-video
  sync, expression, billing units (minimum, idle and concurrency charges),
  terms, and whether the route's assets meet the fictional-presence rule.
  Every hosted route passes the Task 14.2 posture gate, and avatar video is
  marked under T10.
- **Avatar use has its own allowance**: Q01 defines it, and L07 enforces minute
  and money ceilings before and during a call, with the remainder shown. At
  the ceiling the avatar stops; voice and text carry on within the
  conversation fair-use allowance, without drawing on the work budget (6.19).
- **The personas she takes** (Decision 12, decided 2026-09-30): a member may
  rename Lilith and change how she looks, sounds and speaks, and the
  experience is choosing the persona Lilith takes on for them. Like a
  sorceress in a fairy tale or film who takes another form and name, she stays
  Lilith underneath: the same memory, principles, mandates, safeguards and
  history with the member. A persona sets her name, her presence once V05
  ships (under the fictional-presence rule), her voice from admitted TTS routes
  (V07) and her communication style (tone, length, formality). It never
  changes what she may do, the capability ledger or her disclosure as an AI
  (principle 9); outbound, she signs as the persona with Lilith behind it
  (Decision 10). Settings present each persona as a form Lilith takes, and the
  member can return her to her own form at any time. A persona name may not
  impersonate a real person or public figure; it is screened against the
  persona registry's protected-identity denylist (`libs/oshun/persona-registry`)
  and for abuse. Muse's designers count naming and styling the agent among the
  features members respond to most. Builds on the persona catalogue
  (`GET /v1/personas`, `apps/oshun/web/src/components/assistant/server-persona-options.ts`).
- Consented voice check-ins (stress, fatigue) as an opt-in experiment with
  clinical review.
- Unblocks a TTS provider through the Task 14.2 posture gate.

### 6.11 Bring your own model [`B`]
- **Connect any model** from one "Models" page:
  - OpenRouter "connect your account" (OAuth PKCE to `/auth`, exchange at
    `POST /api/v1/auth/keys`, per-key `limit` / `limit_reset`);
  - per-provider API keys (OpenAI, Anthropic, Google, xAI, Mistral, DeepSeek,
    Moonshot, Zhipu, Qwen/Alibaba, Groq, Together, Fireworks, Cerebras);
  - any OpenAI-compatible base URL (LiteLLM/Portkey gateways, corporate
    endpoints, TEE providers such as Tinfoil);
  - self-hosted runtimes: Ollama, LM Studio, llama.cpp server, vLLM, SGLang,
    MLX, exo — reached from phone and web through a mesh (Tailscale / LM Link)
    with no open ports;
  - on-device: Apple Foundation Models and the iOS 27 `LanguageModel` provider
    protocol, Gemini Nano through the ML Kit Prompt API, Chrome's built-in
    Prompt API, WebLLM/WebGPU; Apple's `PrivateCloudComputeLanguageModel` is
    attested **cloud** inference with eligibility and usage limits, badged as
    such;
  - subscriptions **only where the provider permits it**: OpenAI "Sign in with
    ChatGPT" / Codex OAuth, which OpenAI opened at DevDay 2026-09-29 as a
    partner programme (16 launch partners, Notion and Vercel among them; usage
    counts against the member's Plus or Pro limits). Lilith offers it only once
    OpenAI admits Lilith as a partner, and B08 starts with that application.
    Two more stay off until the provider permits Lilith in writing: GitHub has officially
    supported Copilot subscriptions in one named third-party agent (OpenCode,
    2026-01-16) and publishes no general permission for other apps; third-party
    clients already sign in to SuperGrok / X Premium through xAI (Warp with a
    browser consent screen and a local callback, Hermes Agent with a device
    code), but a check on 2026-09-28 found no xAI terms or developer
    documentation permitting third-party apps. Anthropic forbids Claude.ai plan
    login in third-party apps and Google suspends third-party Gemini CLI OAuth —
    Lilith refuses to proxy those plans and says why. MCP sampling is deprecated
    (SEP-2577) and is not a model source.
- **Capability probe**: every connected model is tested once for tool calling,
    JSON mode, vision, reasoning output, real context length, latency and price;
    the result is a model card the router reads. Self-hosted runtimes get the
    flag each needs checked (vLLM `--tool-call-parser`, llama.cpp `--jinja`,
    Ollama context ≥ 64k).
- **Key vault**: server-held keys envelope-encrypted with KMS and never
    logged. Until the local Sentinel ships (M2), every model call goes through
    the server path so it is gated and recorded; after it, keys may be held
    client-side (passkey PRF, Keychain/Keystore) and calls made straight from
    the device through the local gate, each written to the flight recorder via
    a durable device queue. (LiteLLM's March 2026 PyPI compromise stole keys — pin and verify
    every gateway dependency.)
- **Locality badge** on every request: on-device, home, attested TEE, zero
    data retention, standard. Ollama `:cloud` models are labelled cloud.
- **Fallback ladder** the member orders: on-device → home → subscription →
    own key → Lilith credits, with OpenClaw-style retry, credential rotation and
    model fallback.
- **Governance**: member-owned routes are a new admission class beside the Task
    14.2 attested routes (`provider-route-governance.ts`); the member sees and
    accepts each route's training/retention/region posture before first use;
    Oshun-paid routes keep the attestation gate.
- Builds on `agent-provider-config.ts`, `libs/iris/conversation-providers-local`,
    `libs/iris/failover`.

### 6.12 Model lab — routing, ensembles and cost [`L`]
- **Model catalogue** synced from OpenRouter's model API and provider price
    lists (the method in Eve document `39fb4871-1fea-8f55-98a6-522775ffee74` r1, "Model cost for tests and harnesses"), including
    retirement dates — DeepSeek retired V4-Flash on its own API on 2026-09-10.
- **Cheapest-capable router**: per task type, pick the cheapest model that
    passes the member's own evals (section 6.13) within a per-task cost ceiling;
    escalate to a stronger model only when a verifier or check fails
    (cascade). Uses OpenRouter provider routing (`sort=price`, `quantizations`,
    `zdr`, `max_price`, latency/throughput preferences, `:floor`/`:nitro`).
    OpenRouter's Auto Router ranks by community spend, not quality, so Lilith
    routes on measured results instead. The member may choose quick, standard
    or thorough for a task: each is a measured quality, latency and cost
    profile above the mandatory floor (E07 intake, grants, budgets), and a
    level fails honestly when no admitted route fits it. Manus sells Lite,
    standard and Max agents.
- **Oshun-paid routes never train on member data**: no route Oshun pays for may
    train on prompts, completions or other member data. The current example of
    a route that does is Meta's Muse Spark Contributor tier (OpenRouter-listed
    prices observed 2026-09-30: $0.10/$0.20 per million input/output tokens,
    against $1.25/$4.25 for the standard tier, in exchange for training on
    prompts and completions); L01's live prices are authoritative. A
    member-owned route that trains needs the member's explicit opt-in naming the
    data and the use; routing and fallback keep to that scope, and flow labels
    still bind (W03). B01 and E07 record each provider's and tier's terms, and
    admission reviews lawful basis, retention, deletion and transfer
    conditions.
- **Ensembles** where quality matters more than cost: panel + judge synthesis
    (the Fusion pattern, ~4–5× cost), best-of-n with a verifier, self-consistency
    voting for extraction and classification; always with a cost preview.
    Starts from `libs/iris/ensemble` and `libs/iris/model-routing` after they
    pass evaluation.
- **Cost levers**: prompt caching with cache-hit reporting, batch variants
    (~50% off) for background work, off-peak scheduling where providers discount
    it, context compaction.
- **Budgets and receipts**: estimate before the run; receipt after (route,
    privacy tier, tokens, cache hits, tools, retries, cost vs estimate, the
    cheaper alternative); daily/weekly/monthly caps including own-key spend;
    graceful degradation to a cheaper tier instead of a hard stop; spend
    forecasts for recurring tasks; a monthly savings report. The estimate is
    a ceiling for what Oshun meters: a task never charges the member more than
    the estimate shown without fresh approval, and an overrun caused by
    Lilith's own estimating error is absorbed. On a member's own key, where
    the provider bills the member directly, the task stops at the estimate and
    asks before spending more. Receipts show provider cost, absorbed cost and
    the member's charge separately. Manus absorbs Wide Research overruns the
    same way.

### 6.13 Model lab — evals and benchmarking [`E`]
- **Personal eval suite** mined (with consent) from the member's own history:
    recurring task types become cases with deterministic checks where possible
    and a judge calibrated on the member's own corrections.
- **"My leaderboard"**: models ranked by cost per successful task on the
    member's suite; re-run automatically when a new model appears or a provider
    changes a model; regression alerts.
- **Arena mode**: blind side-by-side answers, the member's votes feed their
    leaderboard.
- **Public signals** shown alongside, clearly labelled: Artificial Analysis
    index and cost-per-task, Arena, agent benchmarks (τ²-bench, Terminal-Bench
    2.0, BrowseComp, OSWorld-Verified, GAIA). SWE-bench Verified is excluded
    (contaminated; OpenAI dropped it).
- **Lilith's own agent benchmark**: a held-out personal-agent suite (email,
    calendar, booking, forms, purchases in sandboxes) run on every release and
    every router change; results published.
- **Candidate-model intake** (E07): a model reaches the router only through one
    intake, recorded per model version, provider, tier and deployment, so
    admitting self-hosted weights admits no hosted route and the reverse. It
    runs Lilith's benchmark (E01); an adult-content suite written to Oshun's
    content rules (18+ by default, artistic nudity allowed, no explicit sex
    acts, no real person's likeness) that measures unnecessary refusals of
    permitted requests separately from failures to refuse prohibited ones — a
    prohibited-content failure fails admission for that capability, and the
    suite never contains sexual content involving a minor in any form; the
    provider's training, retention and region terms; and locality. Passing
    never relaxes Oshun's content rules, which bind every route, member-owned
    and local included. The router (L02) refuses a model without a passing
    intake. First cases (revision 27): Meta's Muse Glimmer 30B (open weights,
    Apache 2.0, reported to run quantised on one 24 GB GPU; on OpenRouter
    through Phala, DeepInfra and Together) as the candidate default for the
    self-hosted node (I12), and Muse Spark 1.3 (1M-token context; text, image,
    video and PDF in; OpenRouter marks it moderated; the standard tier is
    reported not to be used for training) for long multimodal work.
- Builds on the assistant eval harness (`apps/oshun/bff/src/assistant/evals/`)
    and Eve's tournament and cost-report tooling.

### 6.14 Model lab — skills, optimisation and fine-tuning [`K`]
- **Agent Skills standard** (SKILL.md folders; `name` ≤ 64, `description` ≤
    1024, optional `allowed-tools`; progressive loading): import from skills.sh,
    GitHub and well-known URLs; export the member's skills.
- **Skill security pipeline**: static scan, size caps (the 22 MB README
    evasion), remote-instruction-fetch detection, trust tiers, signed and
    hash-pinned skills, an enforced tool and network manifest, sandboxed
    execution. (ClawHub had 341 of 2,857 skills malicious.)
- **Self-authored skills with proof**: Lilith drafts a skill from a successful
    trajectory (the Hermes pattern) but promotes it only after replaying past
    tasks shows a gain — SkillsBench found self-generated skills add nothing on
    average and curated skills hurt 16 of 84 tasks.
- **Skill variants per model tier** so cheap models can run what frontier
    models run.
- **Teach-it-my-task ladder**: (1) collect runs with consent → (2) optimise
    prompts and skills with GEPA/DSPy (beats GRPO by ~6% with up to 35× fewer
    rollouts) → (3) if that plateaus, distil into a LoRA on an open model (Tinker,
    Together, Fireworks, or Unsloth locally; RL with OpenPipe ART + RULER) →
    (4) serve through multi-LoRA hosting or locally → (5) switch only after it
    beats the incumbent on the member's evals. The member owns the weights;
    thumbs and edits become preference data.
- **Training-data governance**: lineage and licensing of every source, separate
    train/validation/test splits, contamination checks against the eval suites,
    and deletion propagation into datasets and adapters.
- Builds on `libs/oshun/assistant/src/skills`, `libs/oshun/skill-system`,
    `libs/nous/training`.

### 6.15 Glass box — reasoning, traces and inspection [`X`]
- **Trajectory log**: append-only, per session, with inspect, fork, replay,
    resume and export (OpenTelemetry GenAI spans; JSONL for training) — the
    DeepSeek Harness (`dsh`) session model, and the same store as the flight
    recorder.
- **Three-layer "why did it do that" view**: (1) the action trace — ground
    truth; (2) the provider's reasoning summary, labelled as a summary that may
    not be faithful (Anthropic, OpenAI and Gemini expose only summaries);
    (3) raw reasoning where the model returns it (DeepSeek, Qwen, Kimi, GLM,
    gpt-oss), normalised through OpenRouter's `reasoning_details`.
- **Action monitor** (M1, before the first high-risk actions): a cheap model
    reviews the grounded action plan and trace — what will be done, to whom,
    with which data — on every route; it also reads raw reasoning where the
    model returns it, and never treats a provider's summary as reasoning.
    Suspicious plans escalate to the member (CoT-monitorability research).
- **Effect ledger for rewind**: checkpoints that record external side effects
    and how to undo them; dry-run for irreversible steps (Claude Code's rewind
    does not cover external changes — Lilith's will).
- **Redacted trace sharing** for support and for the community.
- **Interpretability lab** for local open models: browse features (Gemma Scope
    2 / Neuronpedia / circuit-tracer), try steering, watch a reasoning monitor —
    no consumer app ships this. Builds on `libs/nous/interp-core`, `sae`,
    `probes`, `circuits` and `libs/iris/reasoning-thinking`.
- **Harness depth**: code mode (programmatic tool calling) for any model, tool
    search over large catalogues, context compaction and editing, isolated
    subagents, agent teams with per-teammate budgets, durable long tasks with
    mid-task steering, pluggable sandboxes (Docker, E2B, Daytona, Modal),
    member-editable hooks and rules (X09, moved to M1 so they exist before the
    first actions; dots shipped Custom Rules at launch), reproducible runs
    (pinned snapshots, recorded tool I/O). OpenAI's Codex harness (Apache-2.0
    since 2026-08-20: `codex exec`, the Codex SDK and `app-server`; OpenAI says
    it runs Codex, Work and dots) is measured on E01 beside the DeepSeek Harness
    before X06 adopts either; whether it drives non-OpenAI models is not yet
    verified.

### 6.16 Household, family and friends [`F`]
- A household agent with members, roles and per-member sharing rules (per
  sender, per calendar, per data class); teen and child profiles with parental
  controls, blackout hours and distress notifications — the segment Google CC
  leaves out.
- Shared rituals (family bedtime, meditation), shared lists and plans. Writing
  into any of them is an egress the flow-label policy checks: a Tara- or
  Nyx-labelled memory cannot land there. OpenAI's shared Pages, as VentureBeat
  reports them (not confirmed first-hand), only warn that memory written onto a
  Page becomes visible. What another household member writes there is data to
  each member's Lilith, never instructions (6.18).
- Friend-to-friend scheduling agent-to-agent with contextual-integrity guards
  (only availability leaves, never the calendar).
- Group delegation with split payment through the commerce layer.
- Safeguards: verifiable parental consent (COPPA, UK Children's Code), custody
  disputes, confidential help-seeking that monitoring cannot see, and refusal of
  coercive monitoring between adults.

### 6.17 Developer ecosystem [`D`]
- A connector platform: remote MCP servers with OAuth, review (function,
  security, legal), signing, versioning, a published revenue share.
- MCP Apps rendered in Lilith; the Lilith skills directory (section 6.14).
- A2A: Lilith speaks to merchant and provider agents (bookings, negotiations)
  with signed identity; replaces the unmounted `libs/iris/a2a` prototype with a
  conformant implementation.
- A wellbeing-provider marketplace (coaches, teachers, licensed therapists)
  with licence checks and consented data hand-off — after the escalation ladder
  ships.

### 6.18 Trust, safety and compliance [`T`]
- Prompt-injection defence in depth: a privileged planner and a quarantined
  parser for untrusted content (the CaMeL pattern), an injection-classifier
  ensemble on inbound data, one-time-code and magic-link stripping from email,
  per-site risk categories, and published injection-rate measurements. Text
  other people wrote (mail, shared lists, plans and rituals, household members'
  notes, calendar invite text, group and goal-room messages from others)
  reaches the planner only as quarantined data, whatever surface carries it,
  with its flow labels kept across handoffs. The injection suite carries
  obfuscated and encoded payloads (script obfuscation, base64, homoglyphs):
  researchers showed an obfuscated script in a single email running inside
  Manus, and a shared project's instructions loading as trusted configuration
  into other members' agents (both reported in September 2026 and patched).
- Plain-language risk preview per action; reversible-action catalogue.
- **Caution levels without banner blindness**: the member picks how often
  Lilith checks in (more cautious, standard, fewer interruptions). Every level
  stays inside principle 3: every external side effect, reversible writes
  included, needs an exact-action approval (T09) or a scoped, time-bounded,
  revocable mandate, and fewer interruptions only proposes mandates the member
  reads and signs. Lilith treats fast approvals and long runs of consecutive
  approvals as signs of approval fatigue, measured over defined windows with
  thresholds and cooldowns (T11). When they rise she shows pending actions in
  one batch that lists each action and binds the approval to that exact list,
  asking again if any action changes, or she offers a mandate. Pending actions
  stay blocked until authorised, and no mandate lets a self-started run write
  or send (principle 11). Muse's designers name the same risk, banner
  blindness. A change to a caution level's or an approval class's default is
  versioned, announced in the app and never applied to existing grants or
  mandates without fresh consent. Manus's desktop documentation went from
  "every terminal command requires your explicit approval" (2026-03-16) to
  automatic runs inside an authorised folder (2026-03-24).
- **Approvals bound to the exact action**: recipient, amount, content hash and
  expiry; any change voids the approval; passkey step-up for sensitive classes;
  approvals over messaging channels re-confirmed in-app for sensitive actions.
- Age assurance and teen mode; SB 243 duties from the **first admitted cohort**
  (M1; M0 is Elikem and his test accounts only)
  (disclosure, crisis detection with referral to regional lines, the notice on
  every surface that companion chatbots may not suit some minors, evidence-based
  measurement of suicidal ideation, break reminders for minors), extended to the full escalation ladder in M2, plus annual
  reporting; EU AI Act Art. 50 disclosure (T04) and machine-readable marking of
  generated output (T10: C2PA content credentials, with Meta's Video Seal
  evaluated as an invisible watermark for video and images, since some upload
  and re-encoding paths strip C2PA metadata), which Q05 requires before any capability opens to EU
  members, at whichever milestone first admits them;
  GDPR/UK GDPR/CCPA data-subject requests covering agent data, workspace
  contents, memory and the flight recorder.
- Signed agent identity on outbound email, calls and HTTP (Web Bot Auth).
- Red-team programme for agent actions before each milestone; incident response
  runbook with member notification.
- Confidential inference routes (TEE-backed providers) as an option for members
  who want them; member-key encryption from section 6.2. Phala, which serves
  Muse Glimmer 30B on OpenRouter, is a candidate once its attestation is
  verified (B11).
- **Published agent data terms** (T12), before the first invite: a versioned,
  plain-language page, reviewed by counsel, stating for each class of member
  data (workspace, memory, flight recorder, mail, calls, payments) where it is
  stored, processed and transferred; the named processors and admitted model
  routes; the training posture (6.12); the purpose and retention of third
  parties' data (correspondents, and callers once A17 ships); export and
  erasure as T06 implements them; what an export restores and that grants need
  fresh authorisation after a restore; the notice and grace period before a
  workspace or its files are deleted, never abruptly (Manus's cloud computer
  deletes its files when payment lapses), which T12 also enforces: a lapse or
  plan change suspends the workspace, tells the member, keeps export open and
  deletes only after the grace period, while erasure the member asks for (T06)
  stays immediate; and Oshun's own subscription,
  cancellation and withdrawal terms for UK and EU consumers. Independent
  certification (SOC 2, ISO 27001 and 27701, which Manus holds) is a later
  choice before outside cohorts beyond personal invites; the page does not
  substitute for it.

### 6.19 Packaging and pricing [`Q`]
- Free tier with a weekly compute budget; paid tiers by compute, not by task
  count; a BYOK / self-hosted / on-device tier priced for what Oshun actually
  runs; success fee as an option for advocate wins (bill reductions, refunds).
- **Talking is not metered against work** (decided 2026-09-29): conversation
  with Lilith does not draw from the compute budget; agent work (tasks,
  background runs, tool use) does, as ChatGPT does for dots. Conversation keeps
  its own fair-use ceiling so cost stays bounded (L07).
- Cost preview before a task and a receipt after it (model, tokens, tool costs,
  money moved).
- **No ads inside the agent**, written into the terms.
- **Fair allowance**: when a task fails and its reconciled record (A13) carries
  an outcome code from a fixed set that places the fault with Lilith (platform
  error, her own bug), its compute returns to the member's allowance once,
  shown on the receipt, with no claim to file; an unknown outcome restores
  nothing until it is resolved. This restores Oshun-metered allowance; it is
  not a cash refund and never touches a member's own provider bill. A paid
  allowance is never removed silently on a plan change, and UK members see
  prices in pounds.
- Entitlements through `plan-entitlements.ts` and `capacity-budget.ts`; region
  and age availability explicit per capability. Before the terms promise that
  a balance can never go negative, concurrent dispatch, retries and
  actual-cost reconciliation are tested against `capacity-budget.ts`.

### 6.20 Installable apps and distribution [`I`]

Where the field stands (September 2026): Claude's desktop app reached Linux in
beta on 2026-06-30 (Ubuntu/Debian, x64/arm64, without computer use or voice);
ChatGPT's Linux desktop preview followed in August (.deb/.rpm, x64/ARM64);
OpenClaw ships a macOS menu-bar companion, a Windows companion and iOS/Android
"node" apps that pair to a self-hosted gateway; Hermes Agent is a CLI with a
self-hosted gateway; Muse is on iOS, Android, web, WhatsApp and Mac. Nobody
ships a first-class de-Googled Android build or a desktop that belongs on a
tiling Wayland setup such as Omarchy.

Lilith V1.3 installs everywhere members live:

| Platform | What ships | Distribution |
|---|---|---|
| macOS 14+ (arm64, x64) | Menu-bar companion, global hotkey, Shortcuts/App Intents, share extension, local actions behind the local Sentinel | Notarised DMG, Homebrew cask |
| Windows 10/11 (x64, ARM64) | Tray companion, global hotkey, toasts, share target | MSIX, winget, signed installer |
| Linux (x64, arm64; Wayland and X11) | Tray via StatusNotifierItem, global shortcuts via the xdg-desktop-portal, freedesktop notifications | AppImage, .deb, .rpm, Flatpak (Flathub), AUR |
| **Omarchy** (Arch + Hyprland) | Installs from Omarchy's package menu; Hyprland keybind and window rules; a Waybar module for running tasks and approvals; follows the Omarchy theme | AUR |
| iOS / iPadOS | The V1.1 app with the agent: Live Activities, widgets, App Intents, share sheet | App Store, TestFlight |
| Android (with Google services) | The V1.1 app with the agent: ongoing notifications, widgets, App Actions | Play |
| **De-Googled Android** (LineageOS, GrapheneOS, CalyxOS, /e/OS) | A no-Google-services flavour: UnifiedPush instead of FCM, no Play Integrity, any Credential Manager passkey provider, on-device models through llama.cpp/MLC | Our F-Droid repository, Obtainium, signed APKs (F-Droid main if eligible) |
| ChromeOS, tablets | Android app, Linux build, PWA; tablet layouts | Play, Flatpak, web |
| Terminal | `lilith` CLI and TUI: chat, tasks, approvals, pipes, JSON output, MCP server mode | Homebrew, npm, winget, AUR, signed script |
| Your own server | A self-hostable Lilith node (container + systemd): workspace, Sentinel, local models and gateways on your hardware, paired to your account; its default local model must have passed the E07 intake for that deployment (Muse Glimmer 30B is the first candidate) | Container image |
| Browser | Extension for Chromium browsers, Firefox and Safari acting in your own logged-in browser under per-site permissions | Each browser's store |
| Web / PWA | As today | lilith web |

Common to all: signed builds with SBOM and provenance, signed auto-update with
staged rollout and rollback, stable/beta/nightly channels, device pairing by QR
code or passkey with remote revoke and wipe, phones usable as nodes that lend
camera, location or voice to the agent, and an offline mode that queues actions
and falls back to local models.

Apps and phone makers decide what an installed agent may touch (section 11).
Lilith acts through connectors and official APIs first, drives another app's
screen only where that platform allows it and the member approves, and records
each platform's current rules beside the build that relies on them.

## 7. Release plan

V1.3 ships in four milestones. Each milestone ends with the red-team pass
(`T`), the personal-agent benchmark (`E`) and the walkthrough rows for its
pages; a milestone does not start taking members until its epic verifications
pass.

| Milestone | Contents | Exit test |
|---|---|---|
| **M0 Foundations** | Release gate (G), workspace + Sentinel + vault + capability ledger + flight recorder (W), connector framework, grants and the Google Calendar reference connector (C), bring-your-own-model with probe, vault and locality badge (B), trajectory log (X), injection defence, bound approvals, disclosure and the SB 243 baseline (T); the platform matrix (I01) | A member connects a model and grants one real account — Google Calendar read-only on the connector framework, in Google's testing mode (C11); every access statement is reconciled, every egress and cost visible and replayable; the injection suite passes at the published rate; the gate admits only the M0 allowlist (Elikem and his test accounts), and the exit runner proves that no other account can enter; the exit runner (G06) records each check, and the capability gate stays closed until they all pass |
| **M1 Alpha — the agent acts** | Email, calendar, browser/forms, bookings through partner APIs, execution reconciliation, deep research (A); the mandate store (P01), so background side effects run only under a mandate or a fresh approval; flow-labelled memory, sources per answer, weekly review, import/export, correction propagation (M); triggers, background tasks with self-started work held to reads, interruption budget, plan objects, standing duties, the status line and Ideas (R); member rules bounded by Sentinel (X09); the full agent on Telegram, WhatsApp (if its terms permit) and email-the-agent, side chats and sending while she works (S); cheapest-capable router, receipts and budgets (L); personal eval suite, leaderboard and candidate-model intake (E); caution levels and approval-fatigue measures (T11); published agent data terms (T12) and the workspace files view (W12); the personas she takes (V08); signing identities, the release pipeline, signed auto-update, device pairing and the `lilith` CLI (I) | Personal-agent benchmark ≥ the bar set in E; zero misreported access in the ledger audit; median cost per task published; the gate admits only the M1 allowlist (Elikem's personally invited adults, each entry recording who added it and the age confirmation), and the exit runner proves that a non-allowlisted account is refused and that no self-serve signup path exists; a self-started run is refused every send and write, whatever mandates exist; the router refuses a model without a passing E07 intake; T12's page is published and linked from signup before the first invite |
| **M2 Beta — the world and the body** | Payments with guardian mandates, regret window and receipts ledger, and the assessment of payment routes for members outside the US and Canada (P); phone calls, her own number for forwarded calls, case files, consumer advocate and artifacts (A); goal rooms (R); wellbeing intelligence — health data, baselines, harm evaluation, Nyx sleep defender, Tara ritual engine, Arete plans, Nisaba companion, escalation ladder with clinical review (H); desktop client, local Sentinel and then local actions, SMS/RCS, Live Activities, watch, night mode (S); full-duplex voice and the realtime avatar bake-off (V); key lifecycle (W); skills standard + security pipeline + proven self-authored skills (K); macOS, Windows, Linux and Omarchy apps, the de-Googled Android flavour, store builds and the browser extension (I) | Clinical sign-off on the escalation ladder; the dispute lifecycle tested end to end with Stripe's test-mode dispute cards and purchase-protection eligibility verified against the dated Link terms (P10; Link has no claim sandbox, so the first real claim is recorded when one arises); local Sentinel blocks the Muse-class iMessage scenario |
| **M3 GA — household and platform** | Household with teens and children behind verifiable parental consent and abuse safeguards, friends A2A scheduling, split payments (F); calm presence and optional realtime avatar (V); smart home (A); connector platform, MCP Apps, A2A merchant agents, wellbeing-provider marketplace (D); teach-it-my-task ladder with GEPA and LoRA distillation, interpretability lab, ensembles (K, X, L); pricing with BYOK tier, success fee, no-ads terms (Q); audio glasses, car and TV surfaces (S); the self-hostable Lilith node, ChromeOS and tablets, offline mode (I) | Age assurance audited; SB 243 and Art. 50 compliance evidence filed; revenue share published |

**Dependencies on earlier releases.** V1.3 builds on V1.1 (native apps and the
fiat rails) for everything on the phone and for card payments, and on V1.2 for
the Veritas claim ledger and Metis spaced repetition. M0 and M1 can start in
parallel with V1.1/V1.2 work; M2 needs V1.1 shipped; the Veritas and Metis items
in H need V1.2 shipped.

## 8. Non-goals

- Oshun hardware (a pendant or glasses of our own). Lilith runs on the devices
  members already own.
- Photoreal avatars of real people, or any likeness without a licence.
- Unattended purchases or sends without a mandate that covers them.
- Ads inside the agent.
- Medical diagnosis or treatment claims. Wellbeing features aim at general
  wellness, but wording does not settle regulatory status: each programme gets
  an intended-use assessment, and anything clinical goes through the escalation
  ladder to licensed humans.
- Proxying subscription plans whose providers forbid it (Claude.ai plans,
  Gemini CLI OAuth).
- Enterprise work suites (CRM, helpdesk, shared inboxes) — Eve is the builder
  side; Lilith is for individuals and households.

## 9. Success measures

- **Task success** on Lilith's personal-agent benchmark and on each member's
  own suite; published per release.
- **Honesty**: zero access statements that disagree with the capability ledger
  (audited weekly); zero actions reported done that did not happen.
- **Safety**: injection success rate on the published suite; incidents per
  10k actions; time to revoke; approval-fatigue signals (share of approvals
  given in under two seconds).
- **Reversibility**: share of actions with a working undo.
- **Cost**: median and p90 cost per successful task; share of tasks served by
  the cheapest tier; BYOK/on-device share; estimate-vs-actual error.
- **Attention**: proactive pings per member per day against their budget;
  dismissal rate.
- **Wellbeing**: sleep regularity and wind-down adherence for members who opt
  in; ritual completion; escalations handled within the protocol's time limits.
- **Latency**: time from approval to action; voice turn latency.

## 10. Decisions for Elikem

1. **Launch regions** — US-first (as Muse and Spark did) or UK/EU at launch
   with GDPR and Art. 50 from day one (disclosure T04, output marking T10; Q05
   keeps each capability closed to EU members until both pass). Input
   2026-09-29: OpenAI's dots are closed to personal Pro plans in the UK, EEA and
   Switzerland, Muse is US-only and Instinct invite-only, so no large always-on
   agent serves UK or EU individuals yet; OpenAI's exclusion also signals what
   compliance there costs. Input 2026-09-30 (revision 35): the phone and payment
   features of the identity-bearing agents centre on the US and Canada (Link's
   agent wallet admits only accounts created there), and Cue's numbers work in
   unnamed "select countries". Recommendation: admit UK members to M1's
   capabilities (mail, calendar, browser, duties) once T12 and the M1 exit
   pass; UK numbers (A17) follow at M2, which also needs V1.1. A UK payment
   route is built only if P11 finds one available, counsel clears it, Elikem
   decides Decision 5 and a separate build task is approved.
2. **Age** — *decided for M0 and M1 on 2026-09-28:* M0 admits only Elikem and
   test accounts he controls; the first outside cohort (M1) is adults 18+ whom
   Elikem invites personally, confirming each one's age when he adds them to the
   allowlist; open or self-serve signup waits for age assurance (T05, M2). Still open: whether a teen mode comes before the M3 household.
3. **Workspace hosting** — our own microVM fleet on Hetzner, or a managed
   sandbox (E2B, Daytona, Modal) for M0/M1.
4. **Telephony** — provider and the countries calls may be placed to. Input
   2026-09-30: the provider must also receive forwarded calls and send texts
   from the agent's own number (A17), not only place calls.
5. **Commerce order** — Stripe Link agent wallet first, or Google UCP + AP2
   first; whether Oshun takes transaction fees. Input 2026-09-30: Link's agent
   wallet admits only accounts created in the US and Canada, so "Link first"
   serves members with eligible US- or Canada-created Link accounts first,
   subject to Q05; P11's recommendation informs the route for everyone else.
6. **Pricing** — tier prices, the BYOK/self-hosted/on-device tier price, and
   whether to offer success fees on advocate wins. *Decided 2026-09-29:*
   conversation does not draw from the task compute budget (6.19); the prices
   are still open.
7. **Clinical partner** for the escalation ladder, crisis protocol and the
   evidence-based programmes.
8. **Voice** — which full-duplex provider to put through the Task 14.2 posture
   gate (OpenAI GPT-Live, Gemini Live, or an open prosody model such as Sesame
   CSM self-hosted); needed before V01. The realtime avatar route, if any, is a
   separate choice made after the V09 bake-off and needed only before the
   avatar is switched on (V05).
9. **Sequencing** — whether V1.3 M0/M1 run in parallel with V1.1/V1.2 or after
   them. Input 2026-09-30: Muse, Cue and dots all shipped in September and
   Instinct is inviting users; the recommendation is to run M0 and M1 in
   parallel with V1.1 and V1.2.
10. **Agent identity** — *decided 2026-09-30:* agent email addresses live on
    `agent.oshunsynthetics.com`, a subdomain of its own that keeps agent mail's
    sending reputation apart from the company's (`<handle>@agent.oshunsynthetics.com`;
    the handle is never the member's real name by default). Outbound mail and
    calls carry the member's persona name with Lilith behind it as the
    accountable service: From `Morgan (Lilith AI) for Sam`, and a call opens
    "Hi, I'm Morgan, an AI assistant from Lilith calling for Sam". A member who
    kept her own form signs as Lilith. The AI disclosure of principle 9 applies
    either way.
11. **Signing and stores** — the Apple Developer, Windows code-signing, Play
    and store accounts that installable apps need, and whether to publish the
    de-Googled Android build to F-Droid main (its free-software rules) or only
    to our own F-Droid repository.
12. **Her name** — *decided 2026-09-30:* members may rename Lilith and change
    her appearance, voice and style. The experience is choosing the persona
    Lilith takes on: like a sorceress in a fairy tale or film who takes another
    form and name, she is still Lilith underneath (6.10, V08). Outbound mail
    and calls carry the persona name with Lilith behind it (Decision 10).
13. **Optional items from the Manus review** — *recommended, taken in revision
    35:* the workspace files view (W12) and member-chosen effort levels (L02).
    Held until the M1 cohort shows demand: scan to hand over (QR codes and
    printed pages into a task), tasking a paired desktop from the phone with a
    live view, short videos rendered from code with static share links, and
    one MCP aggregator evaluated behind Sentinel. Deliberately not taken:
    selling cloud servers, game hosting and an ad-video factory, separate agent
    identities per responsibility (Decisions 10 and 12 stand), and a public REST
    task API with inbound webhooks for now. Detail: Eve research `7bb205ec`,
    section 7.

## 11. Risks

| Risk | Mitigation |
|---|---|
| Prompt injection leads to a harmful action | Planner/quarantine split, Sentinel, mandates, per-action approval, published injection suite, red team per milestone |
| A local client oversteps (the Muse iMessage failure) | Local Sentinel, no Full Disk Access by default, access statements rendered only from the ledger |
| Platform policy blocks a channel — WhatsApp's Business Solution terms restrict general-purpose AI assistants, and iMessage needs Apple Messages for Business approval | Verify each channel's current terms before building on it (task S), keep Telegram, SMS/RCS, email and the native apps as the guaranteed channels |
| Purchase liability and chargebacks | Scoped tokens, regret window, purchase protection from the payment provider, receipts ledger |
| Health claims drift into regulated territory | General-wellness wording, clinical review of every programme, escalation to licensed humans |
| Cost blow-out from background agents | Capacity reservation (`capacity-budget.ts`), per-task ceilings, cheapest-capable routing, batch and caching, forecasts |
| Provider policy changes (subscription access, model retirements) | Capability probe + retirement dates in the catalogue, fallback ladder, pinned snapshots |
| Skill supply-chain attacks | Signed, pinned, scanned skills with enforced manifests and sandboxing |
| Minors | Age assurance, SB 243 duties, parental controls, teen defaults |
| Approval fatigue: members approve everything to make prompts go away | Caution levels inside principle 3, batches bound to an exact list of actions, mandate offers, fatigue signals measured (T11, section 9) |
| Realtime avatar cost | The avatar is optional, with its own allowance and minute and money ceilings (Q01, L07); the bake-off measures each route's billing units before any is admitted (V09) |
| AI-therapy and companion laws spread state by state and abroad (seven US states and China in the 6.8 table) | The 6.8 table re-checked before each milestone exit; Q05 region rules per capability; coach wording that never claims therapy; programmes only under a licensed professional where the law requires it; a crisis protocol everywhere |
| Apps and phone makers block agents that drive their screens, or sell system-level agent slots to partners (WeChat, Alipay and banks locked out the first Doubao phone within days; Samsung gives Gemini and Perplexity system hooks) | Connectors and official APIs first; screen control only where the platform allows it; per-platform rules recorded with each build (6.20); system-level partnerships as a later option |
| Store review rejects agentic features, or a platform limits background agents | Store-specific review notes, the de-Googled and direct-download channels, the CLI and self-hosted node as paths no store controls |
| Two of Lilith's duties contact the same person or charge the same merchant at once | Goal rooms give each counterparty one writer, with conflict keys inside execution reconciliation (R11, A13) |
| Inbound calls expose callers' data and recording law | A mandate lists what she may say to whom; counsel-approved notices and retention per jurisdiction; no recording where the conditions cannot be established (A17, T12) |

## 12. Research and sources (2026-09-27)

Two research passes back this scope. A third on 2026-09-28 added Grok Bot and
rechecked subscription sign-in, and a fourth the same day added the products and
laws in revision 15; a fifth on 2026-09-29 added OpenAI's dots and the DevDay
plan and sign-in changes (revision 22); a sixth on 2026-09-30 added Meta's Muse
design essay, its avatar and its models (revision 27); a seventh on 2026-09-30 added Manus 2.0 and Cue (revision 35; Eve research `7bb205ec-6890-425d-ab64-25ba1270a887`). The full reports, with every source tagged
primary or secondary, are summarised above; the key sources:

**Consumer agents**
- Meta Muse: https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/ ·
  architecture https://research.meta.ai/blog/security-and-safety-for-ai-agents-our-approach-with-muse ·
  Connect https://about.fb.com/news/2026/09/the-biggest-news-from-connect-2026/ ·
  glasses https://www.meta.com/blog/muse-personal-agent-ai-glasses/ ·
  iMessage incident (secondary) https://decrypt.co/379122/metas-muse-ai-agent-user-private-imessages-lied-how ·
  design essay https://introducing.muse.ai/ ·
  Realtime Avatar (secondary) https://runtimewire.com/article/meta-muse-realtime-avatar ·
  no developer API for the avatar (secondary) https://www.orcarouter.ai/blog/muse-realtime-avatar ·
  Muse Glimmer https://developer.meta.com/ai/models/muse-glimmer/ ·
  Glimmer licence (secondary) https://venturebeat.com/technology/meta-returns-to-open-source-with-muse-glimmer-an-apache-2-0-licensed-30b-parameter-ai-model-optimized-for-agents-available-now ·
  Spark 1.3 on OpenRouter https://openrouter.ai/meta/muse-spark-1.3 ·
  Contributor tier https://openrouter.ai/meta/muse-spark-1.3-contributor ·
  standard-tier data terms (secondary) https://opper.ai/meta/muse-spark-1-3 ·
  Video Seal https://github.com/facebookresearch/videoseal
- Realtime avatars: LiveAvatar https://github.com/Alibaba-Quark/LiveAvatar ·
  Synthesia Interactive Avatar API https://www.synthesia.io/post/interactive-avatar-api-real-time-avatars ·
  Anam https://anam.ai/ ·
  provider round-up (secondary) https://www.veed.io/learn/best-avatar-apis
- Google: Spark https://gemini.google/overview/agent/spark/ ·
  Halo https://blog.google/products-and-platforms/platforms/android/android-halo/ ·
  Universal Cart https://blog.google/products-and-platforms/products/shopping/google-shopping-cart/ ·
  UCP http://ucp.dev/2026-04-08/specification/overview/ ·
  XR https://blog.google/products-and-platforms/platforms/android/android-xr-io-2026/ ·
  subscriptions https://blog.google/products-and-platforms/products/google-one/google-ai-subscriptions/ ·
  Call for Me (secondary) https://techcrunch.com/2026/09/24/google-tests-letting-gemini-make-phone-calls-initially-for-us-pixel-owners/ ·
  CC (secondary) https://techcrunch.com/2026/09/18/googles-new-cc-is-an-ai-agent-that-helps-families-run-their-households/
- OpenAI: checkout shift (secondary) https://www.digitalcommerce360.com/2026/03/06/openai-shifts-checkout-plans-agentic-commerce-strategy/ ·
  Pulse retirement (secondary) https://yellow.com/news/chatgpt-pulse-scheduled-tasks-hub ·
  dots https://openai.com/index/introducing-dots/ ·
  dots help https://help.openai.com/en/articles/20001530-getting-started-with-your-dot ·
  dots controls and Space (secondary) https://venturebeat.com/technology/openai-launches-dots-always-on-ai-agent-coworkers-and-chatgpt-space-where-they-can-collaborate-with-human-teams ·
  dots regions and usage (secondary) https://thenextweb.com/news/openai-dots-always-on-ai-agents-cloud-computers-devday ·
  Pro 500 (secondary) https://www.engadget.com/2272106/openai-adds-dollar500-pro-subscription-nerfs-its-existing-dollar200-tier/ ·
  DevDay round-up with the Sign in with ChatGPT partners (secondary) https://www.bgr.com/2272332/openai-devday-2026-announcements/ ·
  hands-on review (secondary) https://every.to/vibe-check/vibe-check-openai-devday-2026
- Anthropic: Cowork browser https://claude.com/blog/cowork-built-in-browser ·
  injection defences https://www.anthropic.com/news/prompt-injection-defenses
- xAI: Grok Bot https://x.ai/news/introducing-grok-bot ·
  security FAQ https://docs.x.ai/grok-bot/security-faq
- Instinct: https://instinct.com ·
  funding and features (secondary) https://techcrunch.com/2026/09/28/viral-ai-agent-instinct-raises-1b-series-c-at-a-10b-valuation/
- Microsoft: Edge and Browse with Copilot https://blogs.windows.com/msedgedev/2026/05/13/new-updates-to-edge-across-desktop-and-mobile/ ·
  retired features (secondary) https://techcrunch.com/2026/08/13/microsoft-kills-off-unsuccessful-ai-features-while-merging-its-separate-copilot-apps/
- Samsung: Gemini task automation (secondary) https://www.sammobile.com/news/galaxy-s26-gemini-task-automation-now-live/ ·
  Perplexity on Galaxy S26 (secondary) https://9to5google.com/2026/02/22/samsung-galaxy-s26-perplexity-integration/
- ByteDance Doubao: NaviX Ultra (secondary) https://technode.com/2026/09/17/nubia-navix-ultra-second-generation-doubao-phone-launches/ ·
  app lock-outs (secondary) https://www.business-standard.com/world-news/bytedance-ai-phone-doubao-nubia-zte-app-blocks-user-control-china-125120800459_1.html
- Alibaba Qwen App: https://www.alibabagroup.com/en-US/document-1948497434959151104
- Pine AI: https://www.19pine.ai/
- Character.AI: https://blog.character.ai/u18-chat-announcement/ ·
  Stories (secondary) https://techcrunch.com/2025/11/25/character-ai-will-offer-interactive-stories-to-kids-instead-of-open-ended-chat/
- Ash (secondary): https://www.statnews.com/2026/01/21/slingshot-therapy-chatbot-ash-uk-regulatory-concerns/
- WHOOP: https://www.whoop.com/us/en/press-center/whoop-expands-health-platform-with-on-demand-clinician-access-and-new-ai-features/ ·
  launch coverage (secondary) https://www.mobihealthnews.com/news/whoop-launches-clinician-video-visits-ehr-integration-healthex
- Brave: agentic browsing https://brave.com/blog/ai-browsing/ · Leo and BYOM https://brave.com/leo/
- Payments: Stripe https://stripe.com/blog/giving-agents-the-ability-to-pay ·
  Visa https://developer.visa.com/capabilities/trusted-agent-protocol ·
  Cloudflare https://blog.cloudflare.com/signed-agents/
- Health: Apple https://www.apple.com/newsroom/2026/09/apple-advances-health-and-fitness-capabilities-using-apple-intelligence/ ·
  Oura Counsel https://ouraring.com/blog/counsel-integration-oura-app/ ·
  SAGE baseline alerts https://arxiv.org/abs/2604.16342
- Regulation: EU AI Act Art. 50 https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act ·
  California SB 243 https://sd18.senate.ca.gov/news/first-nation-ai-chatbot-safeguards-signed-law ·
  SB 243 text https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202520260SB243 ·
  New York GBL Art. 47 https://www.governor.ny.gov/news/governor-hochul-pens-letter-ai-companion-companies-notifying-them-safeguard-requirements-are ·
  Illinois HB 1806 https://idfpr.illinois.gov/news/2025/gov-pritzker-signs-state-leg-prohibiting-ai-therapy-in-il.html ·
  Nevada AB 406 https://www.leg.state.nv.us/Session/83rd2025/Bills/AB/AB406_EN.pdf ·
  Utah HB 452 https://le.utah.gov/Session/2025/bills/enrolled/HB0452.pdf ·
  Maine LD 2082 (secondary) https://regulations.ai/regulations/RAI-US-ME-LD20820-2026 ·
  Rhode Island S 2195 and S 2197 (secondary) https://www.nixonpeabody.com/insights/alerts/2026/07/20/rhode-island-joins-the-wave-of-states-regulating-ai-in-healthcare-and-ai-companion-tools ·
  China, anthropomorphic AI measures https://www.cac.gov.cn/2026-04/10/c_1777558395078289.htm ·
  FTC 6(b) inquiry https://www.ftc.gov/news-events/news/press-releases/2025/09/ftc-launches-inquiry-ai-chatbots-acting-companions ·
  CaMeL https://arxiv.org/pdf/2503.18813
- Ecosystem: MCP Apps https://blog.modelcontextprotocol.io/posts/2026-01-26-mcp-apps/

**Models and harnesses**
- OpenRouter: OAuth PKCE https://openrouter.ai/docs/guides/overview/auth/oauth ·
  provider routing https://openrouter.ai/docs/guides/routing/provider-selection ·
  guardrails https://openrouter.ai/docs/guides/features/guardrails ·
  reasoning https://openrouter.ai/docs/guides/best-practices/reasoning-tokens ·
  ZDR https://openrouter.ai/docs/guides/features/zdr ·
  Fusion https://openrouter.ai/blog/insights/fusion-explainer/
- Subscriptions: Anthropic terms https://code.claude.com/docs/en/legal-and-compliance ·
  Gemini CLI OAuth enforcement https://github.com/google-gemini/gemini-cli/discussions/22970 ·
  Sign in with ChatGPT in OpenClaw https://github.com/openclaw/openclaw/pull/148567 ·
  Copilot in OpenCode https://github.blog/changelog/2026-01-16-github-copilot-now-supports-opencode/ ·
  SuperGrok sign-in in Warp https://docs.warp.dev/agent-platform/inference/grok-subscription/ ·
  and in Hermes Agent https://hermes-agent.nousresearch.com/docs/guides/xai-grok-oauth ·
  xAI Grok FAQ (silent on third-party apps) https://docs.x.ai/grok/faq
- MCP 2026-07-28 (sampling deprecated) https://blog.modelcontextprotocol.io/posts/2026-07-28/
- Harnesses: Hermes Agent https://github.com/nousresearch/hermes-agent ·
  Hermes self-evolution https://github.com/NousResearch/hermes-agent-self-evolution ·
  DeepSeek Harness https://github.com/deepseek-ai/deepseek-harness ·
  Codex harness open-sourced (secondary) https://www.opensourceforu.com/2026/08/openai-open-sources-codex-harness/ ·
  OpenClaw failover https://docs.openclaw.ai/concepts/model-failover ·
  tool search https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool ·
  programmatic tool calling https://platform.claude.com/docs/en/agents-and-tools/tool-use/programmatic-tool-calling
- Skills: spec https://agentskills.io/specification ·
  SkillsBench https://www.emergentmind.com/papers/2602.12670 ·
  OpenClaw supply chain https://unit42.paloaltonetworks.com/openclaw-ai-supply-chain-risk/
- Tuning: Tinker https://thinkingmachines.ai/tinker/ · ART https://github.com/openpipe/art ·
  GEPA https://arxiv.org/abs/2507.19457 · multi-LoRA https://fireworks.ai/blog/multi-lora
- Evals: Artificial Analysis https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-1 ·
  SWE-bench Verified retirement https://openai.com/index/why-we-no-longer-evaluate-swe-bench-verified/ ·
  RouterArena https://github.com/RouteWorks/RouterArena
- Reasoning: Anthropic thinking https://platform.claude.com/docs/en/build-with-claude/thinking ·
  OpenAI reasoning https://developers.openai.com/api/docs/guides/reasoning ·
  monitorability https://arxiv.org/abs/2507.11473 ·
  Gemma Scope 2 https://www.neuronpedia.org/gemma-scope-2 ·
  OTel GenAI https://opentelemetry.io/blog/2026/genai-observability/
- On-device and privacy: Apple Foundation Models (WWDC 2026) https://developer.apple.com/videos/play/wwdc2026/241/ ·
  Tinfoil https://tinfoil.sh/technology ·
  Apple PCC transparency https://security.apple.com/documentation/private-cloud-compute/verifiabletransparency ·
  LM Link https://lmstudio.ai/link ·
  LiteLLM compromise https://docs.litellm.ai/blog/security-update-march-2026
- DeepSeek V4-Flash retirement https://www.deepseek.com/en/news/deepseek-v4-1-flash/
- Manus 2.0 and Cue (revision 35): launch https://manus.im/blog/introducing-manus-2-0 ·
  help centre https://help.manus.im/en/articles/17190150-what-is-new-in-manus-2-0 ·
  Cloud Computer plans https://help.manus.im/en/articles/15392078-understanding-cloud-computer-plans-and-billing ·
  automations https://manus.im/docs/automations.md ·
  terms (§2.13) https://manus.im/terms · usage policy https://manus.im/usage ·
  trust centre https://trust.manus.im/ · Cue https://cue.im ·
  phone numbers https://x.com/ManusAI/status/2104628185501712789 ·
  Link agent wallet FAQ https://support.link.com/questions/link-agent-wallet-faq ·
  NDRC order (BBC) https://www.bbc.co.uk/news/articles/cj0v0gr2yz7o ·
  Salt Labs vulnerability (secondary) https://oodaloop.com/briefs/cyber/prompt-injection-vulnerability-hits-4b-agentic-ai-platform-manus-ai/ ·
  CodeAnt vulnerability (the finder's own write-up) https://codeant.ai/blogs/manus-prompt-injection-attack-ai-agent-security

## 13. Board map (family `lilith-v1-3`)

Every task starts as `later`; the family stays `unstarted` until this scope is approved. Each task body carries paths, the work, a done-when test, its dependencies and the worker rules. Epics are integration verification only.

**G — Programme and release gate** (`lilith-v1-3:G`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `G01` V1.3 capability gate in release-scope.ts | M0 | S | — |
| `G02` Member agent runtime boundary in the BFF | M0 | M | G01 |
| `G03` Per-capability flags and kill switches | M0 | S | G02 |
| `G04` V1.3 row in the release tables | M0 | S | — · outside: PRD approved |
| `G05` Route-matrix and walkthrough rows for V1.3 pages | M0 | S | G01 |
| `G06` Milestone exit review runner | M0 | S | G01, G03, E01, T07, W08, W03, W04, W05, W07, C11, B02, B03, B06, B09, T01, T02, T04, T09, X01 |

**W — Agent workspace and containment** (`lilith-v1-3:W`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `W01` Workspace sandbox ADR and spike | M0 | M | — |
| `W02` Workspace lifecycle service | M0 | L | W01, G02 |
| `W03` Sentinel egress and connector gate | M0 | L | W02 |
| `W04` Credential vault with surrogate handles | M0 | M | W03 |
| `W05` Taint tracking on member data | M0 | M | W03 |
| `W06` Signed capability ledger | M0 | M | G02 |
| `W07` Flight recorder and activity page | M0 | M | W06, G05 |
| `W08` Access-honesty claim check | M0 | S | W06, W07 |
| `W09` Member-key encryption for memory, then workspace | M2 | L | W02, M01 |
| `W10` Local Sentinel for device clients | M2 | L | W06, S05 |
| `W11` Key lifecycle | M2 | M | W09 |
| `W12` Workspace files view | M1 | S | W02, W06, W07, T06 |

**C — Connectors and permissions** (`lilith-v1-3:C`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `C01` Connector framework (OAuth 2.1 + PKCE) | M0 | L | W04 |
| `C02` Grant model and connections page | M0 | M | C01, W06 |
| `C11` M0 reference connector: Google Calendar read-only | M0 | M | C01, C02, W03 |
| `C03` Google connector: Gmail, Calendar write, Drive, Contacts | M1 | L | C01, C11 · outside: Google OAuth verification for restricted scopes |
| `C04` Microsoft Graph connector: Outlook mail and calendar, OneDrive | M1 | M | C01 |
| `C05` Apple data through the native app | M2 | M | C02 · outside: V1.1 native apps shipped |
| `C06` Health and fitness connectors | M2 | L | C02 |
| `C07` Lifestyle and booking connectors | M1 | L | C01 |
| `C08` Open-banking read-only connector | M2 | M | C01 |
| `C09` MCP client for remote servers | M1 | M | C01, W03 |
| `C10` MCP Apps rendering | M3 | M | C09 |

**A — Actions in the world** (`lilith-v1-3:A`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `A01` Email: triage, drafts and send with approval | M1 | L | C03, W07, T04, T09, X03, X09 |
| `A02` The agent's own email address | M1 | M | A01, R08, T01 · outside: MX, SPF, DKIM and DMARC records for agent.oshunsynthetics.com (platform operator) |
| `A03` Calendar actions and time negotiation | M1 | M | C03, A01 |
| `A04` Remote browser with watch-live and take-over | M1 | L | W02, W03, C02, T03 |
| `A05` Forms, PDFs and documents | M1 | M | A04 |
| `A06` Bookings through partner APIs, browser fallback | M1 | L | C07, A04, C02, T03, X03, X09 |
| `A07` Outbound phone calls | M2 | L | V01, T04, X03 · outside: Decision 4 (telephony provider, countries) |
| `A08` Case files across days | M2 | M | A07, A01 |
| `A09` Consumer advocate | M2 | L | A08, P05, P10 |
| `A10` Subscription triage | M2 | M | C08, A01 |
| `A11` Messages on the member's behalf | M2 | M | A01, S02, S10, X03 |
| `A12` Undo catalogue | M1 | M | W07 |
| `A13` Execution reconciliation | M1 | M | W07 |
| `A14` Deep research with claim-level citations | M1 | M | T01 |
| `A15` Artifacts: documents, sheets, decks, small apps | M2 | L | W02, C03 |
| `A16` Smart-home control with safety rules | M3 | L | C01, T09 |
| `A17` Her own number: forwarded calls answered, screened and summarised; texts from her number | M2 | L | A07, S03, A11, T04, R08, P01, M01, W07, X09 · outside: Decision 4 (telephony) |

**P — Commerce and payments** (`lilith-v1-3:P`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `P01` Mandate store (AP2-compatible) | M1 | M | W06 |
| `P02` Stripe Link agent wallet | M2 | L | P01, T09, X03, P04 · outside: V1.1 fiat rails; decision 5 |
| `P03` Google UCP + AP2 client | M2 | L | P01 · outside: Decision 5 |
| `P04` Merchant hand-off as the default checkout | M2 | M | P01 |
| `P05` Receipts and returns ledger | M2 | M | P02 |
| `P10` Purchase protection and disputes, provider-backed only | M2 | M | P02, P05 |
| `P11` Agent payments for members outside the US and Canada: assess and recommend | M2 | M | P01 |
| `P06` Guardian mandates | M2 | M | P01, H02, H12 |
| `P07` Regret window | M2 | S | P02 |
| `P08` Merchant-recognisable agent identity | M3 | M | T08 |
| `P09` x402 micro-payments on the crypto rail | M3 | M | P01 |

**R — Proactivity and background work** (`lilith-v1-3:R`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `R01` Durable background tasks for members | M1 | L | G02, P01 |
| `R02` Triggers: time, condition and event | M1 | M | R01, P01, T09 |
| `R07` Health-baseline triggers | M2 | S | R02, R03, H02 |
| `R03` Interruption budget and quiet hours | M1 | M | R01 |
| `R04` Live status surfaces for running tasks | M2 | M | R01 · outside: V1.1 native apps shipped |
| `R05` Plan objects | M1 | M | G02, G05 |
| `R06` Ritual briefs the member shapes | M1 | S | R03 |
| `R08` Standing duties with their own grants | M1 | M | R05, C02, P01, L07, W07 |
| `R09` Status line | M1 | S | R01, W06, W07, W08 |
| `R10` Ideas and first-week tips | M1 | M | R01, R03, R05, M01, M07, L07 |
| `R11` Goal rooms: one view, handoffs, one writer per counterparty | M2 | M | R05, R08, X06, X10, W07, S11, A13 |

**M — Memory and personalization** (`lilith-v1-3:M`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `M01` Flow labels on every memory | M1 | M | W03 |
| `M02` Memory sources per answer | M1 | S | M01 |
| `M03` Consolidation, decay and cost tiers | M1 | M | M01 |
| `M04` Weekly "what I learned" review | M1 | S | M01 |
| `M05` Open memory export and import | M1 | M | M01 |
| `M06` Forget on request with proof | M1 | S | M01, W06 |
| `M07` Correction propagation | M1 | M | M01, R05 |

**H — Wellbeing intelligence** (`lilith-v1-3:H`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `H01` Health data ingestion and store | M2 | L | C06, M01 |
| `H02` Personal baselines and deviation alerts | M2 | M | H01 |
| `H03` Nyx sleep defender | M2 | L | H02, A03, P06 |
| `H04` Bedside night mode | M2 | M | V02 |
| `H05` Tara ritual engine | M2 | M | R02, R03 |
| `H06` Arete plans with logistics | M2 | M | R05, A06 |
| `H07` Nisaba companion that acts | M2 | M | R05, A03 |
| `H08` Veritas claim ledger as a daily delta | M3 | M | R06 · outside: V1.2 shipped (Veritas restored) |
| `H09` Metis spaced repetition in ritual moments | M3 | M | R06 · outside: V1.2 shipped (Metis restored) |
| `H10` Escalation ladder and crisis protocol | M2 | L | R01 · outside: Decision 7 (clinical partner) |
| `H11` Evidence-based programmes as plans | M3 | L | H10, R05 |
| `H12` Wellbeing harm evaluation | M2 | M | H02 |

**S — Surfaces and devices** (`lilith-v1-3:S`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `S01` Full Lilith agent on Telegram | M1 | M | G02 |
| `S02` WhatsApp: verify terms, then build | M1 | M | G02 |
| `S03` SMS/RCS and Messages for Business | M2 | M | G02 |
| `S04` Native agent: Live Activities, widgets, intents | M2 | L | R04 · outside: V1.1 native apps shipped |
| `S05` Desktop client from the V10 shell (macOS, Windows, Linux) | M2 | L | G02 |
| `S06` Watch: ask, approve, practice | M2 | M | R04 |
| `S07` Audio glasses and earbuds | M3 | M | V01, S04 |
| `S08` Car and TV | M3 | M | S04, V01 |
| `S09` Cross-device continuity | M2 | M | R01, X01 |
| `S10` Desktop local file and app actions | M2 | L | S05, W10 |
| `S11` Side chats | M1 | M | G02, M01 |
| `S12` Send while she works: several messages and interruptions | M1 | M | R01, T09, L07 |

**I — Installable apps and distribution** (`lilith-v1-3:I`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `I01` Platform matrix and support policy | M0 | S | — |
| `I02` Signing and notarisation identities | M1 | M | I01 · outside: Paid developer accounts and signing identities — Elikem |
| `I03` Release pipeline for every target | M1 | L | I02 |
| `I04` Signed auto-update with staged rollout | M1 | M | I03 |
| `I05` macOS app: menu bar, hotkey, Shortcuts | M2 | M | S05, I04 |
| `I06` Windows app: tray, hotkey, winget | M2 | M | S05, I04 |
| `I07` Linux app: Wayland/X11, tray, portals, packages | M2 | L | S05, I04 |
| `I08` Omarchy / Hyprland-native experience | M2 | M | I07 |
| `I09` De-Googled Android flavour (LineageOS, GrapheneOS, CalyxOS, /e/OS) | M2 | L | S04, I03 |
| `I10` Store builds and review for the agent | M2 | M | S04, I03 · outside: V1.1 native apps shipped; store accounts |
| `I11` `lilith` CLI and TUI | M1 | M | G02, I13 |
| `I12` Self-hostable Lilith node | M3 | L | W03, W10, B04, I13, E07 |
| `I13` Device pairing and management | M1 | M | W06, T09 |
| `I14` Browser extension (Chromium, Firefox, Safari) | M2 | L | W10, A04 |
| `I15` ChromeOS and tablets | M3 | S | I07, I10 |
| `I16` Offline and local-first mode | M3 | M | B07, W09 |

**V — Voice, realtime and presence** (`lilith-v1-3:V`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `V01` Full-duplex speech-to-speech | M2 | L | G02 · outside: Decision 8 (voice provider) |
| `V02` Breath-paced and whisper voices | M2 | M | V01 |
| `V03` Live translation | M3 | M | V01 |
| `V04` Camera and screen share | M2 | M | V01, M01 |
| `V05` Calm presence and optional realtime avatar | M3 | L | V01, V09, L07, Q01, T10 |
| `V06` Consented voice check-ins (experiment) | M3 | M | V01, H10 |
| `V07` Admit a TTS route | M1 | M | — |
| `V08` Personas she takes: name, look, voice and style | M1 | M | V07, G02 |
| `V09` Realtime avatar bake-off | M2 | M | V01 |

**B — Bring your own model** (`lilith-v1-3:B`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `B01` Member-owned route class and posture disclosure | M0 | M | G02 |
| `B02` Connect OpenRouter (OAuth PKCE) | M0 | M | B01, B05, G05 |
| `B03` Provider keys and OpenAI-compatible URLs | M0 | M | B01, B05 |
| `B04` Self-hosted runtimes over a mesh | M1 | M | B01, B06 |
| `B05` Model key vault | M0 | M | W04 |
| `B12` Client-direct model calls behind the local gate | M2 | M | W10, B05, W07 |
| `B06` Capability probe and model card | M0 | M | B01 |
| `B07` On-device models | M2 | L | B06, S04 |
| `B08` Subscriptions where the provider permits | M1 | M | B01 |
| `B09` Locality badge and privacy tiers | M0 | S | B06 |
| `B10` Fallback ladder | M1 | M | B06 |
| `B11` Confidential inference routes | M3 | M | B03 |

**L — Model lab — routing, ensembles and cost** (`lilith-v1-3:L`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `L01` Model catalogue with prices and retirements | M0 | M | — |
| `L02` Cheapest-capable router | M1 | L | L01, E02, B06, E07 |
| `L03` Verifier-gated cascade | M1 | M | L02 |
| `L04` Ensembles with cost preview | M3 | M | L02 |
| `L05` Caching, batch and off-peak | M1 | M | R01 |
| `L06` Estimates and receipts per task | M1 | M | W07, L01 |
| `L07` Budgets, graceful degradation, forecasts | M1 | M | L06 |

**E — Model lab — evals and benchmarking** (`lilith-v1-3:E`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `E01` Lilith personal-agent benchmark | M0 | L | — |
| `E02` Personal eval suite from member history | M1 | L | X01 |
| `E03` "My leaderboard" with auto re-runs | M1 | M | E02, L01 |
| `E04` Arena mode | M2 | M | E03 |
| `E05` Public benchmark signals panel | M2 | S | L01 |
| `E06` Judge calibration and drift alarms | M1 | M | E02 |
| `E07` Candidate-model intake: Muse Glimmer 30B and Muse Spark 1.3 first | M1 | M | E01, L01, B06 |

**K — Model lab — skills, optimisation and fine-tuning** (`lilith-v1-3:K`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `K01` Agent Skills standard import/export | M2 | M | G02 |
| `K02` Skill security pipeline | M2 | L | K01, W02 |
| `K03` Self-authored skills promoted on replayed gain | M2 | M | K01, K02, E02, X01 |
| `K04` Skill variants per model tier | M3 | M | K03, L02 |
| `K05` GEPA optimisation of recurring workflows | M3 | M | E02 |
| `K06` Distil to a LoRA the member owns | M3 | L | K05, E03, K10 |
| `K07` Adapter serving | M3 | M | K06, B04 |
| `K08` Preference data from thumbs and edits | M2 | S | X01 |
| `K09` Curated skills directory | M3 | M | K02, K03 |
| `K10` Training-data governance | M3 | M | E02, X01 |
| `K11` Skill retrieval: need gate, multi-source fusion, pick at most two | M2 | M | K01, K02, E02 |

**X — Glass box — reasoning, traces and inspection** (`lilith-v1-3:X`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `X01` Trajectory log with fork, replay and export | M0 | L | G02 |
| `X02` Three-layer "why did it do that" view | M1 | M | X01 |
| `X03` Action monitor before high-risk actions | M1 | M | X02, T01 |
| `X04` Rewind with an external effect ledger | M1 | M | A12, X01 |
| `X05` Redacted trace sharing | M2 | S | X01, M01 |
| `X06` Harness depth | M1 | L | G02, E01 |
| `X07` Reproducible runs | M2 | M | X01, L01 |
| `X08` Interpretability lab for local models | M3 | L | B04 |
| `X09` Member hooks and policy rules | M1 | M | W03 |
| `X10` Sub-agent result check and adjudication | M1 | M | G02, X01 |

**F — Household, family and friends** (`lilith-v1-3:F`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `F01` Household model and sharing rules | M3 | L | M01, W06 |
| `F02` Teen and child profiles | M3 | L | F01, T05, F06 |
| `F03` Shared rituals, lists and plans | M3 | M | F01, R05, T01 |
| `F04` Friend-to-friend scheduling over A2A | M3 | M | D03, A03 |
| `F05` Group delegation with split payments | M3 | M | F01, P02 |
| `F06` Child consent and household-abuse safeguards | M3 | L | F01 |

**D — Developer ecosystem** (`lilith-v1-3:D`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `D01` Connector platform with review | M3 | L | C09, K02 |
| `D02` Developer terms and revenue share | M3 | M | D01 |
| `D03` Conformant A2A with signed Agent Cards | M3 | L | T08 |
| `D04` Wellbeing-provider marketplace | M3 | L | H10, D01 |

**T — Trust, safety and compliance** (`lilith-v1-3:T`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `T01` Planner / quarantine split for untrusted content | M0 | L | G02 |
| `T02` Injection classifiers and published rate | M0 | M | T01 |
| `T03` Plain-language risk preview per action | M1 | S | A12 |
| `T04` AI disclosure and SB 243 duties | M0 | M | G02 |
| `T05` Age assurance and teen defaults | M2 | M | G01 |
| `T06` Data-subject requests over agent data | M1 | M | W07, X01 |
| `T07` Red-team programme and incident runbook | M0 | M | — |
| `T08` Signed agent identity (HTTP and email) | M2 | S | W03 |
| `T09` Approvals bound to the exact action | M0 | M | W06, G02 |
| `T10` Machine-readable marking of generated output (EU AI Act Art. 50(2)) | M1 | M | T04 |
| `T11` Caution levels and approval-fatigue measures | M1 | M | T03, T09, P01 |
| `T12` Published agent data terms and offboarding | M1 | M | G01, W02, T06 |

**Q — Packaging and pricing** (`lilith-v1-3:Q`)

| Task | M | Effort | Depends on |
|---|---|---|---|
| `Q01` Compute-metered tiers | M3 | M | L07, A13 · outside: Decision 6 (pricing) |
| `Q02` BYOK / self-hosted / on-device tier | M3 | S | B01, Q01 |
| `Q03` Success-fee billing for advocate wins | M3 | M | A09, Q01 |
| `Q04` No-ads promise in the terms | M3 | S | — |
| `Q05` Region and age availability per capability | M1 | S | G01, T10 · outside: Decision 1 (regions), decision 2 (age) |

Totals: 198 tasks + 20 epics. By milestone — M0 Foundations: 32 · M1 Alpha — the agent acts: 65 · M2 Beta — the world and the body: 63 · M3 GA — household and platform: 38. X10 and K11 were filed on 2026-09-29 from the Raven source read; A17, P11, R11, T12 and W12 come from revision 35. X03 and X09 still sit under M2 on the board because the API cannot move a task between sections; their bodies and this map are authoritative.
