Mallory Design Archive

Three independent polish items, none touching the backend — two ship on a Pages deploy alone, one needs a small native inset bridge.

Continue used to open the keyboard against the app's voice-first priority — now it starts listening immediately and carries thread context through a dismissable chip.

Notifications now forward everything except a small seeded blocklist — the old allowlist was silently missing the user's own email app.

One quiet visual register — hairline rows, serif titles, outline pills — replaces three surfaces that had each drifted their own way.

The bottom nav becomes the single, indicator-bearing switcher across Intel/Loops/R&D — and a drain-on-connect fix means a delegation made while Q is offline is never lost.

A single per-theme --focus token fixes a contrast fail that no single colour could — oxblood too dark on dark, gold too light on light.

Q's task chat had drifted onto its own code path for text turns and its own TTS entrypoint — this pass unifies the parts that actually have a live consumer and explicitly drops the parts that don't.

The vault pushes digests to the edge every 15 minutes; the chat turn never waits on the Hub to know what's true.

"On it. I'll dig in and ping you." — the turn returns instantly, Opus does the real work on the VM, and the answer rejoins the same thread.

One tray under the composer, one segmented sheet, one native report viewer — digs finally have a home in the PWA.

Flip built: true once in one file, and the persona's "what I can't do" list stops lying on its own.

The Briefing modal dies; Mallory just tells you, in the feed, in his own voice — and it can't drift from what he actually can do.

One dot, four systems — tap the ring and see whether the brain, the memory, or the Hub itself is actually behaving.

Two lifecycle gaps closed with a hybrid model: digs get manual Dismiss, loops keep their auto-hide — now both retrievable from one Archive view.

A continuous listen-reply-listen loop, default ON, with barge-in, auto-follow for long replies, and a fail-safe mic — one coherent pass over the voice view.

Markdown, tables, images, choice chips and safe links land in the feed, all stripped back to clean prose before TTS ever touches them.

Detected calendar changes and aging @mentions finally reach Alex's phone outbound — gated by quiet hours, dedup and a daily cap so the safe failure direction is "too quiet," not "too noisy."

The Hub pushes a vault-note index to D1 on its existing cadence, so a single FTS5 query can rank conversations and notes together — no Worker→Hub call on the hot path, no new infra.

Working · Maths Circle or Off · Personal: nine possible moments gate nudges, shape briefings, and softly bias recall — with manual override always winning, and no per-mode persona rewriting.

One silent card per reply — Mallory points, the Worker fetches, SSRF never gets a foothold.

Four new silent blocks — diagrams, widgets, launchers, charts — all client-side, all sanitised, none of them touch the Worker.

The LLM gets one job — a structured pull — everything after it is jq, forever.

Every teaser gets a Worker-side Haiku pass — the app only needs to know how to ask for the right app.

Presence lives in his voice pointing you there; detail lives in Updates, exactly once.

Three PRs, one review: stop the confident-wrong count, stop the stacking pushes, let handled items die.

The reminders table already existed and had zero callers — this slice gives it its first.

One action — fire a nudge — reused so hard that mode-routing, dedupe, and quiet hours come for free.

Not a to-do app. A curated state-of-your-world where draining is the point, not accumulation — and most of it drains on its own.

v1 optimised for the steady-state glance. It had no answer for a day you couldn't look — and the judgement layer had never once run.

130 of 142 items resolved themselves and none were ever nudged — so measure first, and let two weeks of misses decide whether the rest gets built at all.

The mic on home, a Loops task, an Intel dig, or R&D all open the *same* canvas — scoped to context, and deep enough for a sustained working session.

Chat-to-Q was visibly broken (wrong-axis alignment, no auto-scroll, no persona colour) — the tactical fix ships first, then a shared primitive stops it recurring.

Persona is *derived* from mode, so showing both a persona chip and a status pill was showing the same state twice — and pushing the menu off-screen.

Mallory's personality stays invariant — modes no longer bend *his* character, they hand the desk to a different persona whose character suits the moment.

Recall today only works *in-turn* — there's no way to search directly, no browsable history, and no memory of facts about the operator.

The edge KV schema has nowhere to put a coordinate. The phone's whole job is collapsing raw sensor data into {activity?, hat?} on-device and shipping only the verdict.

Weather & news as in-turn tools — hardcoded hosts, no SSRF, no secrets, and a turn never fails because a lookup did.

"Done that" ticks a step; "stop here" parks the whole session; reopening gets a spoken recap of where you left off.

A dedicated Ubuntu VM on the Tower's GPU box, reachable over Cloudflare Access SSH, so Claude Code can drive Hermes as durably as it drives the Hub.

vault_guard_path is the keystone invariant: Hermes writes strictly under Hermes/, so its rebase-and-retry pushes can never content-conflict with the Hub's.

The Tower pushes; the Worker never pulls — hermes_events' dedupe ledger means Hermes can re-emit its whole recent window every cron run for free.

The Tower dials *out* to a Durable Object WebSocket hub — no inbound port on the Tower, no port-forward — so the edge-never-calls-Tower invariant survives making Q live.

"Added 16 items to the queue" — and wrote nothing. The only path that touched q_tasks was the human delegate composer; Talk-to-Q had no tool at all.

Asked "what's going on," Q returned Alex's day-job tasks — because SOUL.md/USER.md were empty, the backlog was empty, and it was still running the weaker default model.

Reusing the existing qsocket instead of a new edge→Tower path means /chat never opens a Tower connection — the edge-local invariant holds by construction, not by discipline.

The Tower runs hermes -z <text> one-shot with no session — folding Q into the home feed just made painfully visible that Q has always answered blind.

npm run eval puts Mallory through his paces — persona, recall, capability honesty, reasoning — scored against a committed baseline, on-demand only, never in CI.

Same battery, six candidate models, one question: which task classes can downgrade to a cheaper model without hurting Mallory's answers?

Thrifty, Balanced, Sharp, or Custom — a segmented slider that turns model routing from a code edit into an operator choice, with the quality bar drawn straight from the eval battery.

The harness exists; the baseline doesn't. Commit it once, and every push and every week gets checked against it automatically.

Time Travel covers corruption while the account is intact; this covers everything else — off-platform artifact, tested rebuild, account loss.

The unguarded failure mode isn't a hard crash — it's a job that quietly stops running and never pages.

A bug, a loop, or a heavy day currently has no brake and no signal — this stops a runaway the same day it starts.

Not a ping per message — one coalesced countdown for the whole outstanding set, and a celebration when nothing's waiting.

Two Slacks should never cancel each other's resolution — a connection registry makes that structurally impossible.

Four self-reported surfaces, four zeros. So the fifth is not a form — it is Mallory noticing the correction inside the conversation already happening.

The board is still up because closing a successor is not the same as closing a predecessor. This writes the second sentence.

The map of what Mallory + Hermes are, the six pieces they run on, and the invariants that hold the whole estate together.

The forward-looking intent behind Mallory: what he's for, why he feels the way he does, and why three personas still add up to one Service.

The concrete shape of the estate — components and responsibilities, data stores, the trust/auth model, and how the edge-local invariant is actually enforced in code.

The concrete request/response sequences behind chat, recall, nudges, briefings, digs, Hermes as dev-driver, digest sync, and mode resolution.

Where the estate has been, the phased direction a council review set, and the open questions still steering what gets built next.

Sixteen numbered decisions — context, rationale, and traded-away alternatives — for the calls that define how the estate is built and run.

The runbook implementing D14 — what's backed up, where, how to rebuild at three severity tiers, and the recovery order when everything is gone.

How a Q-research artifact (or any persona deliverable) gets from an isolated session to the right delivery channel without Alex manually copy-pasting.

Four new capability surfaces landed on main — but the build got ahead of the landing.