Attention and focus

Two opposing forces, one policy: DreamLayer must be able to interrupt you out loud when a moment genuinely demands it, and must be able to shut up completely when you demand that. The attention policy decides the first; Focus mode enforces the second; the proactive-cue picker tunes everything in between.

The hark — "Listen!" and "Watch out!"

A hark is the system tapping you on the shoulder: one line, one ring, an earcon, a haptic.

orchestrator/attention.py: AttentionPolicy.evaluate(ctx, commitments) scans the same live Context the anticipation engine sees and produces at most a handful of ranked alerts:

The discipline that keeps it from nagging:

• One hark at a time — attention_tick speaks only the single most important fresh alert, watch-outs ranked first.
• A 30-minute per-key cooldown — the same alert cannot repeat inside half an hour, and a key is only marked consumed if the hark actually spoke (a veil- or focus-suppressed hark does not burn the alert).
• The hark call itself adds a second 120-second cooldown across all harks.
• Normal harks are held during Focus; urgent watch-outs pierce it. Everything is silenced by the Veil.
• set_attention(False) (the phone's "Proactive alerts" toggle, or "Hey Juno, stop keeping watch") mutes the policy entirely.

The heartbeat: pulse(context) — or the background start_pulse(context_fn, interval=15.0) — runs anticipation and attention together on one context snapshot. Seam: the live context feed (place, people in view, clock) that a device build supplies to the pulse.

Focus mode

set_focus(minutes) — default 25 by voice ("Hey Juno, focus mode") — turns the interruptions down while capture keeps running. That second half is the difference from Incognito, which pauses capture itself.

Held while Focus is active: anticipation cards, live caption display, message pop-ups, fact-check cards, delivery reads, answer-ahead, commitment-capture confirmations, and normal harks. Still running underneath: the ledger, the user model, commitment tracking, and recall on demand. Still allowed through: urgent watch-outs, and anything you explicitly ask for.

clear_focus() ends it early; focus_active() reports it; turning Focus on from the phone also unlocks the Saga's Deep Focus badge.

The proactive-cue picker

Finer than on/off: set_cue(kind, on) mutes any of the three anticipatory kinds — event, person, place — before the engine's ranking pass, so you can keep "leave in 8 minutes" while silencing arrival reminders. The phone nests these three under its Proactive cards toggle. cue_kinds() reports current state.

The deviation nudge

Related but distinct: the Tell engine (tell_check) compares fresh transcript against your prior commitments and raises a DeviationAlertCard when the new words contradict the old plan — before-versus-now across a dashed divider, with a severity dot.

The Glance Arbiter — which lens owns a look

The hark decides when to speak; the Glance Arbiter (orchestrator/glance.py) decides which lens a look belongs to — without a mode picker, because a menu is friction on glasses. On a look, glance(frame) classifies what's in view and every candidate lens bids; the arbiter fires the clear winner, offers a one-tap chooser when it's genuinely ambiguous, or does nothing.

It reuses the Object Lens provider-registry shape, lifted up a level: there, providers declare matches(sighting) and the registry merges rows into a panel; here, candidates declare bid(reading, ctx) and the registry ranks them into one decision.

• Two-tier read. A free coarse pass (classify_coarse over cheap on-device cues — a face flag, a text-density estimate, a form grid, a question mark, a language guess) runs first; only when it can't tell a form from a question from prose (is_ambiguous) does the hub spend the Brain's vision tier for a fine read. The big model is used when it changes the answer, not on every glance.
• Fire vs offer. The top bid fires outright when it beats the runner-up by the gap (or a spoken intent forced it, or it's the only bidder); otherwise a GlanceChoiceCard offers the close contenders ("Answer it · Fill it in · Translate"). A pick runs that lens and teaches the arbiter.
• It learns you. Per-scene priors (GlancePriors) reinforce the lens you keep choosing for a kind of scene, so tomorrow's ambiguous look leans your way. They persist as a small JSON on the hub beside the vault (glancepriors.json, the same pattern as the user model) — read once at start, rewritten on each pick, in-memory only when there's no vault. The local file stays the source of truth so a glance never waits on the network; the dict is still serialisable, so a Mac Brain can later mirror it across hubs.
• Spoken steer. A recent "what's the answer / how do I fill this / explain this" biases the very next look to that lens for a few seconds — say it, then look.
• Calm. Hysteresis holds a fresh decision through a debounce window, so a glance wandering across a page doesn't flip lenses. Veiled ⇒ nothing.

The candidate lenses today: Person (a face → Social Lens), Scholar (answer / form / plain-words), TasteLens (a shelf or menu → the pick), Rosetta (foreign text → translate), and Juno (an object, or a weak fallback to name what's here). The chooser, when a look is genuinely ambiguous:

Who gets to interrupt — the summary