Weekly recap

Sally stopped being a persona

Home Journey was a linear narrative — 14 days of Sally's life with Opal, told as a story. The problem: you'd read about Sally but never feel like you were her. So we killed the narrative frame and rebuilt it as Sally's actual phone. One suggestion card at the top, a few widgets below it. The hook is sleep — "Set your bedtime tonight" — because that's the one thing every parent actually wants to control first.

Better soul. Clearer cracks.

The onboarding had three options: boring form, chat-based intake, or something weirder. We went weird — Sheikah direction, ancient-tech aesthetic, "I was born underground" as the opening line. It makes Opal feel singular. But when we ran Sally through it, the problem was obvious: 8 screens of mythology before any product value. Users don't have that patience. The aesthetic works. The structure needs to front-load utility and weave myth in alongside it, not before it.

Two different products, same name

We ran a Figma audit of both platforms side by side. Expected skin-level differences — colors, spacing, tokens. What we found: onboarding is substantially more complete on iOS, Autofocus barely exists on Android, and settings/rules/profile/stats all have structural mismatches that go beyond design into actual product decisions. Can't talk about "the Opal experience" when the platforms diverge this much.

The model should see the page

Opal Focus classifies content to decide what to block. The v1 approach was metadata-only — URLs, DOM text, heuristics. It works for obvious cases and fails on everything interesting (memes, branded content, context-dependent stuff). The v2 direction: screenshot the page, send it to a vision model, let it decide. Grok 4.20 via OpenRouter on a Cloudflare Worker. Pipeline includes blur preprocessing, meme detection, allowlists, and per-platform rules. Still early, but the direction is right — metadata heuristics will always have a ceiling that vision doesn't.

Boots. Doesn't hold.

OpalOS is the Android side of the vision — a system-level layer that sits on top of your phone and changes what you see based on a mode. Utility mode strips apps to essentials, Entertainment restores feeds, Focus blocks distractions, Sleep dims everything. It's controlled by an agent you chat with. The prototype runs in the Android emulator: four modes work, agent chat connects to OpenRouter with vision, and there's a gem orb UI. But the overlay keeps re-triggering when you reopen apps, the image picker isn't wired to the backend, and Java 17 was a prerequisite that cost a day to sort out. It proves the concept. It's not something you'd hand to someone yet.

Agent in your pocket, no server

Most AI apps are thin clients — they send your prompt to an API and stream back a response. Hermes Sovereign inverts that. The agent runs locally inside the app, powered by Python 3.13 bundled via the PythonAppleSupport xcframework. SwiftTerm provides the terminal UI. The Hermes agent source ships in the app bundle. A bridge script mediates between Swift and Python. The idea: you don't need internet for the agent to work. You don't need our server. The phone is the server.

Open questions

• site-packages-ios-sim/ and site-packages-ios-device/ are empty placeholder dirs
• No iOS-compiled Python packages yet — host validation only
• Bootstrap wiring not connected to ProjectSovereignApp.swift

Safari, but each tab talks to the agent

Sovereign is the ambitious path. Hermes Tabs is the pragmatic one. A native SwiftUI app where each tab is a separate conversation with Hermes. You tap +, get a new chat. Responses stream in via SSE from the public Hermes proxy at opal-agent.decemberclaw.com. State persists locally in Application Support. It's simple — intentionally so. The tradeoff: it depends on the server being up and the API being available. But it's a real iOS app that talks to the agent today, with none of the Python packaging complexity that Sovereign carries.