What this pins

Anatomy order, as shipped (MatterDetailView.swift, verified 2026-07-30): title (21 bold, in the body — the nav bar carries none) → Why this matter → Owed pill when the board projection has items → With (cast chips, scrolling, age pinned trailing) → Now (ONE card; fronts divided by hairlines) → Timeline + “N threads · N messages” → the docked Door. Headers are 15pt semibold sentence case. Guidelines are agent-internal — held and enforced at review/send/incoming, never a displayed section.
The ··· menu is real, and holds exactly one item: Mark done (ToolbarItem(placement: .primaryAction) → a Menu whose single Button is Mark done with a checkmark). So the verb ships — it just lives in the overflow, not under a Done-when quote.
Designed but NOT built — do not read these off the drawing as shipped: a Done-when section with an editable quote and a Mark-done that cites it; a bordered zone-variant Status card; the inline Send / Change ▾ pair; a side-ask row. The earlier version of this file drew all four as though they existed. What the screen actually offers for a prepared reply is Review the draft reply →, which opens the composer through openSuggestedDraft. The Done proposal in Scene ⑤ is in the same category: server-side real (#577), client-side unrendered.
DECISION 2026-07-30 — Timeline v3. Both changes REVERSE the 2026-07-21 redesign (pingbo docs/adr/matter-detail-redesign/plan.md §3, §4/§10), which is why they are written down rather than just drawn:
① Each thread carries an up-to-date summary under its title. The redesign called this “the rejected digest” on the grounds that a thread’s what it’s about is already the Gmail subject titling the section. That reasoning holds for what it’s about and misses where it stands — which is what a reader of a 15-message thread actually needs, and what the fold hides. Source (v0): the newest message’s summary, projected at read time — no new column, no new model call, and up-to-date by construction because the newest message defines it. Falls back to the body’s opening where that summary is absent (it is on 550 of 853 Matter-homed rows today). Not v0: a distilled thread-state sentence — that needs its own model call per arriving message AND its own freshness watermark, the exact machinery #569 had to build for L1; worth doing only if the projected line reads thin in dogfood.
② The message rows get their left rail back — node, connector, fixed mono time column, then the initial avatar. The redesign removed it (“no vertical rail”) reasoning that a thread is a conversation read forward, not a reply-pile, and that the section header already groups. True — but the rail is not doing grouping here, it is doing time: with a summary line now sitting above the rows, the rows need a scannable spine. Ported from that ADR’s own timeline-ab-mock.html variant A rather than reinvented, so the two drawings stay one language.
③ Who and when are ONE column, and the row is vertically centred. The initial icon sits directly over its own datetime instead of standing beside a separate time column — they are one fact about a message, so they read as one block, and it returns ~42px of width to the summary on a 376pt screen. The dot, the identity block and the summary share one centre line. This reverses an earlier top-aligned version whose stated worry was that a two-line summary would push the icon off the rail’s axis; centring the rail itself is what answers that — the connector became a full-height element behind the dot (clipped at the first and last dot), so the dot sits at the row’s middle however tall the summary grows, which top-alignment never gave. Narrows matter-activity-projection’s revived rule twice over: that ADR had the time in an “adjacent fixed column” with the avatar in the content row, and centred the node on the date/time block alone rather than on the whole row.
④ Every thread group collapses. Real disclosure, drawn in both states above (the second group is closed). What stays visible when closed is the point: title, date and the thread summary — collapsing is only safe BECAUSE ① exists, since a closed thread that showed only a subject would hide where it stands. Hidden: the message rows and the fold. Not built on the client: the shipped Timeline renders every section’s rows with no disclosure.
⑤ Every Now front says WHY it is on you, or why it is waiting — and this one needs no new AI field. Both members of the Thread Now union already carry evidenceMessageIds (it is in threadNowBaseShape), so the reason renders as the citation that produced the Now: sender · date · the line itself, with the trust-ladder ↗. Unciteable by construction — a reason with no evidence cannot be drawn. needs-you cites the message that put the ball on you; waiting cites what YOU last did, why the other side owes a reply, and — free from nudgeAt, which only the waiting variant carries — when Pingbo will chase. The waiting front previously said nothing beyond its summary; a lane that asserts “waiting on Allister” while showing no basis is the same unevidenced claim #564 is about, one level down.
⑥ The row’s timestamp is ONE token from MailTimestamp, never a time-over-date stack. That formatter (PingboShared/MailTimestamp.swift) escalates by age and returns exactly one string: 1m under an hour (floored — a just-arrived item never reads “0m”) → localized time same-day (jm) → Yesterday → the weekday within the week (EEE) → MMM d same-year → MMM d, yyyy older — with a separate Chinese branch (N 分鐘 / 昨天). An earlier draft of this file drew 10:02 over Jun 24, a shape the app cannot produce: a Jun 24 message in the same year renders Jun 24 and no time at all. Consequence for layout: the column must fit the longest rung in the reader’s locale, not a fixed HH:mm. This supersedes matter-activity-projection’s revived rule that the time column always shows “localized system short time” sized for Jul 2, 2025 — the shipped formatter does not always show a time.
⑦ The rail spans exactly first-dot → last-dot — never above the oldest message or below the newest. It cannot be one element: the dots’ positions depend on how tall each summary wraps, which CSS cannot read. So each dot draws its own two segments (up to its row’s top edge, down to its bottom edge) and consecutive rows meet at the boundary; the first dot drops its upward segment and the last drops its downward one. The fold carries a full-height segment of its own, since it sits between two dots and would otherwise leave a gap where hidden messages are.
⑧ Spacing rules the render forced, each with a reason: the seam between the identity block and the summary is 15px while the seam inside it (dot→icon) is 9 — a block reads as one unit only when its internal spacing is tighter than its external. The datetime is anchored to the icon (absolutely, inside the centred row) rather than stacked after it: the icon is centred in a row as tall as the summary, so a row-relative datetime drifted further from the icon the more the summary wrapped. And the Why→Owed gap tightens from 24px to 12 — 24 is what ships (MatterDetailView’s .padding(.bottom, 24)) and it left the pill floating between two blocks; with no pill the 24 stands, because that IS a section break.
⑨ The section chevron is a hanging gutter. It sits in the group’s left padding, so the thread title, the thread summary and the message rail share one left edge. Inline, it indented the title alone and the header stopped reading as one column.
Unchanged by this decision: sections stay current-first, rows stay oldest→newest within a thread, the fold stays after the opening message, and there is still no CURRENT chip.
Timeline is grouped by thread, and its rows are messages only: a tappable subject + last-activity header, then avatar · one-sentence summary · timestamp per message, with View N earlier replies folded in after the OPENING message (the first message is what everything after it answers). Agent moves are not interleaved as rows today — an earlier version of this legend claimed they were. The row’s sentence is message.summary when the wire carries one and the body’s opening when it does not, which is visibly worse and known.
A prepared reply is a link, not a Send button (today): when the server already wrote a draft for a front’s thread, the front shows Review the draft reply → and the composer opens with reply context loaded — that is where reply-all recipients come from. The merged verdict brief with inline Send / Change ▾ remains the design (FEAT-verdict-brief §21), unbuilt on this screen.
Multi-front: a Matter with several open strands renders every front in the one Now card, separated by hairline dividers — not a critical-path card plus a “side-ask” row, which is what this legend used to claim. Each front carries its own state line and its own Open thread ↗, and maps to exactly one Timeline section by threadId.
Multi-thread: a line links messages, never threads — threads are derived groupings. The Timeline is thread-blind (a row never names its thread); Open thread ↗ zooms to that row's thread, and the zoom view's title is the thread subject with the Matter title in the sub row. The scope new thread · same matter (and forward) is how a Matter grows its second thread.
Colour tokens do NOT port across by name — match on hex. This file’s --accent is the lime (#CCEF33 light / #D4F038 dark), which is PingboColor.lime — lime500/lime450. But PingboColor.accent in the app is a legacy alias for primary, i.e. forest #1C251F — the token file says so outright: “equals primary (forest), NOT the lime pop.” Same word, opposite colour. The lime Send drawn in the gallery below IS correct — PingboActionButton(“Send →”, variant: .lime) (Composer.swift:187) fills with PingboColor.lime and labels with onLime = forest800 #162219, which is this file’s --on-accent exactly.
Trust ladder = three zoom levels, a concept — NOT labels in the UI: distilled (the Timeline row you read) → single original (that one raw email) → whole raw thread (its whole thread, WKWebView). The app never prints the words "distilled/original"; you just tap deeper. No “Open in Gmail” door in daily UI (north star); durable export/delete live in You.
The Door (orb + placeholder) is the ask/tell input, docked at the floor of every case-file.