feat(golf/messages): premium mobile rebuild — the thread was sunken, not flat - #1833
Open
njrini99-code wants to merge 36 commits into
Open
feat(golf/messages): premium mobile rebuild — the thread was sunken, not flat#1833njrini99-code wants to merge 36 commits into
njrini99-code wants to merge 36 commits into
Conversation
…ts in a row, and an auth hop nobody needed Two causes, and the visible one was not the one I kept assuming. **The rail's loading state did not look like the rail.** Its `loading` branch hardcoded `padding="md"` and the "Conversations" bezel regardless of viewport, while the loaded rail is flat and header-less below `md` and leads with a search field. So entering Messages on a phone drew THREE different layouts in sequence: the route skeleton (flat, search, rows), then this bezel card with no search, then the real rail (flat, search, rows). Two full visible reconstructions on the way to a screen whose shape was known the entire time. That is the "it hot loads when you click Messages" report, and it is a mismatch between three files rather than a rendering fault. The branch now uses the same `isDesktop` gating as the loaded rail and reserves the search field's height, so all three states are geometrically identical and the transitions are invisible. A loading state that does not match the thing it stands in for is worse than none: it manufactures precisely the layout jump it exists to prevent. **And the fetch could not start until an auth round trip finished.** `useGolfConversations` called `auth.getUser()` on mount and gated everything on the result — so the tab paid a full network hop of skeleton before its first query was even sent, on top of the route skeleton Next had already shown. The page already knew who the user was: `useGolfUser()` carries a server-resolved id from the dashboard layout, available synchronously on the first client render. It is now passed in, and the auth call is skipped entirely when we have it. The fallback remains for callers with no context, so nothing else changes behaviour. Deliberately NOT the full server-render of the conversation list. That is a ~400-line refactor across the RPC, active-team scoping, a team-chat merge and batched metadata, and getting the scoping wrong would show multi-team coaches the wrong conversations. It remains the right end state; this removes the two visible reconstructions and one round trip without betting the rail's correctness on a rushed rewrite. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
…und, and add reactions Four changes, and the first is the one that made every previous "redesign" invisible. **The bubbles were darker than the surface they sat on.** The thread well was `bg-surface` — oklch(0.984), the single LIGHTEST colour in the system — and the incoming bubbles were `bg-surface-sunken` at 0.963. Measured in a browser, that is a 0.021 lightness step in the wrong direction: every message was pressed INTO the page rather than resting on it. No amount of radius or spacing work was ever going to fix that, which is why four rounds of it changed nothing anyone could see. `design-tokens.css` already had the answer written down, at --fw-color-canvas: "cards now clearly LIFT off it … the ~0.03 gap is the premium card/page separation the flat-cream look lacked." A conversation is a stack of cards. The well is now `canvas` (0.953) and incoming bubbles are `surface` (0.984) with `shadow-soft` — a 0.031 step, in the right direction. Verified by computed style in a headless render, not by eye. **Metadata was louder than the messages.** Body text was `text-body-sm` = 13px against a spec asking for 15-16. Every group carried a timestamp in Fragment Mono at 0.06em tracking, which is why "9:52 PM" read like a build log. Day separators were all-caps at 0.1em — the exact "all-caps microcopy with huge letter spacing" the spec lists under things to reduce. And "Read" was stamped under every outgoing group, so one screen showed it three times. Now: 15px body, timestamps in the sans face at caption size, sentence-case day markers, and exactly ONE read receipt — under the newest thing you sent, which is all the information there ever was. Radius drops 28px -> 20px, because at 28 a two-letter reply has a radius bigger than its own half-height and renders as a lozenge. **Reactions (spec §18).** New `golf_message_reactions`, one row per (message, user, emoji), with the unique constraint holding the toggle invariant in the DATABASE rather than trusting the client not to double-insert. RLS follows the message: readable and writable exactly when its conversation is, via a new SECURITY DEFINER `golf_conversation_has_me` — which covers DMs, where the existing `golf_conversation_on_my_team` cannot help because it requires a team_id. Insert additionally pins `user_id = auth.uid()`, without which a participant could forge a teammate's reaction. anon has no grant on either. The emoji vocabulary is CLOSED (👍 ❤️ 😂 👀 ✅) and validated server-side. An open set would make this a client-controlled text column on a shared production table; the DB caps its length, the action caps its vocabulary. Long-press now works on ANY message, not just your own. Reacting to what someone ELSE said is the entire point, and gating the gesture on ownership put the feature exactly where it could not be reached. **No more email per message**, by owner instruction. Email was the wrong channel and the loudest one we had: a ten-message thread mailed every participant ten times. Push and the in-app bell already carry the same payload to the app where a reply can actually happen. Deleted rather than flag-gated — a dormant flag on a fanout path is a thing that gets switched back on by accident. Migration applied to production and verified before this commit: RLS on, three policies, anon privileges NONE, helper granted to authenticated only, table in the realtime publication. One test changed rather than deleted. `keyboard-inset` pinned the literal string `[.keyboard-open_&]:pb-4` and failed on a composer that still had the exact behaviour the test is named for — only the number moved. It now asserts the invariant (resting pad reserves the home indicator; the keyboard-open override must NOT re-add the safe-area inset), so it still fails if the override is removed, and no longer fails on a spacing change. NOT done, and not pretended: "Unknown User" in the header is participant resolution returning no profile for that QA thread, not styling. Reply/quote and mentions remain open. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
…en is the conversation The previous commit changed colour, radius and type. Those are ingredients. This changes the composition. **The thread was a panel laid on a page; now it is the page.** Below `md` the InstrumentPanel drops its card tone and its `min-h-[40vh]` floor, so the conversation owns the viewport instead of floating on one. The elevation and the two-pane framing survive from `md` up, where they are actually true. **A short thread now sits at the bottom, where a conversation lives.** It used to hug the top of the well with several hundred pixels of empty ground beneath it — the newest message as far from the composer as the layout could put it. `mt-auto` on the content inside a flex column fixes it with no positioning tricks and no scroll maths: it pushes a short thread down and does exactly nothing once the content overflows, so long threads, prepend and anchor restoration are untouched. **The date landmark stopped being the loudest thing on screen.** Two hairlines spanning the full width made "TODAY" a stronger horizontal element than any message in the thread. It is a centred label now. The unread landmark keeps its rules deliberately — that one is meant to be found, and the contrast between them is what makes it findable. **"Unknown User" is gone.** It is a DEBUG string and it was rendering as ordinary UI for a real DM. The participant is built from a coach/player lookup, so an unresolved one is not unknown: they have no current roster row — they left, or the account is gone. It says "Former team member" now, in the thread header and the conversation list. Truthful beats a placeholder that reads like a fault. (The underlying resolution is unchanged; this fixes what the user is told, which is what was actually wrong.) **Composer.** 20px radius rather than 28 — it is a field, not a sheet. ONE focus treatment: it used to draw a colour-shifted border AND a 2px ring around a field already inside a bordered track, which is three concentric rounded rectangles. And the send button's hit target and its visible control are now different sizes — 44px target, 36px circle — where before they were one 44px tile, which is why send read as a slab instead of a button. Message text 15px -> 17px (spec §10 asks 16-17), with `overflow-wrap: anywhere` so a pasted URL cannot push the thread sideways. Three tests changed, none weakened: - the action-count tripwire moves 447 -> 449 for the two real new reaction actions, annotated like every prior entry in that ledger - the fanout test's email assertion is INVERTED rather than deleted, so re-adding a per-message email fails there instead of shipping quietly - the scroll test stubs the reactions hook, because that file tests scroll position and reactions drag a Supabase client into a jsdom environment that has none. The seam is at the feature boundary, not in Supabase's constructor. Verified by headless render against the real tokens, not by eye: a short thread now resolves at the bottom of the region rather than the top. Still open from the spec, and not claimed: reply/quote (§30), mentions, conversation details sheet (§37), swipe-to-reply (§28), the navigation push transition (§20), structured Helm objects (§41). Physical-device keyboard work (§25/§26) needs an iPhone and cannot be verified here. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
…cator was reserved twice
§7 of the spec, and it was a double-count, not a padding value.
The open-thread surface set its height to
`100dvh - env(safe-area-inset-bottom) - …`, so its box already ended above the
home indicator. The composer inside it then added
`pb-[calc(0.5rem + env(safe-area-inset-bottom))]`. Both reserved the same
~34px, and between the composer's bottom edge and the physical screen edge sat
a strip of the PAGE's colour, not the composer's — which is exactly the
"60-120px of unexplained space" the spec describes, and why it read as a band
rather than as padding.
The inset is now applied exactly once, by the composer, because that is the
half that should own it: a bottom bar's background should run to the physical
edge and pad its CONTENT above the indicator. The surface takes the full
viewport and gives back only the keyboard.
The keyboard term simplifies as a consequence. It was
`max(0, keyboard - safe-area-bottom)` — netting against an inset the surface
was itself subtracting. With nothing to net against it is just
`max(0, keyboard)`.
That simplification broke a sibling gate, and the gate was wrong rather than
the change. It counted occurrences of the literal string
`max(0px,calc(var(--keyboard-height,0px)` and expected exactly 2, so it failed
on arithmetic that still satisfies the property it is named for. It now asserts
the property: both mobile CONVERSATION budgets must reference
`--keyboard-height`. The two budgets are allowed to differ; neither is allowed
to ignore the keyboard.
Getting the scoping right took three attempts and each failure is worth
recording, because two of them would have shipped as a passing test:
- filtering by `4rem-env(safe-area-inset-top` also excluded the LIST branch
- matching class strings with /'[^']*'/ mis-pairs on the prose apostrophes
in this file's comments and matched NOTHING — a gate green by being empty,
which is the exact failure mode `quality-gates.md` exists to name
It now identifies a conversation surface by what follows its height
(`flex-col overflow-hidden`, the scrolling column) rather than by shape or
position, which correctly exempts the no-team EmptyState — centred content,
no composer, no keyboard to yield to.
Falsified before trusting it: reverting the thread surface to a bare
`100dvh` fails the gate with a message naming the surface, and restoring it
passes. A gate that cannot fail is not a gate.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
…ge reactions Shape was verified when the migration was applied (RLS on, three policies, anon privileges NONE, helper granted to authenticated only, table in the realtime publication). Shape is not behaviour, so the policies were then EXERCISED on production in rolled-back transactions. The control matters as much as the attacks: a policy that denied everything would block all three attacks and look identical in the log. It inserts, so the denials below mean the policy discriminates rather than simply refusing. control participant reacts as SELF ................. INSERTED attack A participant forges a TEAMMATE's reaction ... BLOCKED 42501 attack B non-participant reacts to the message ...... BLOCKED 42501 attack C non-participant reads the reactions ........ 0 rows table row count afterwards .......................... 0 Attack A is the `user_id = auth.uid()` conjunct in the INSERT policy. Without it, any participant could attribute a reaction to anybody else in the thread. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
The inbox was a dashboard page with a list on it. Six changes, all §16. **The giant green "New message" pill is gone.** A solid accent lozenge on its own row was the largest, highest-contrast object on the Messages screen — louder than any conversation in the list beneath it — and it spent a full band of vertical space saying what a 44px glyph says. Compose is a pencil-square icon button now, team broadcast beside it, both 44px with real `aria-label`s, so nothing is lost for VoiceOver. **Every group rendered the same generic two-person glyph.** Three conversations meant three identical icons and an avatar column carrying no information at all. A group now gets its own initials the way a person does — "Team Updates" reads TU, "Demo University Golf" reads DU — on the accent wash that still separates it from a DM at a glance. 48px, up from 40. **The preview was `text-eyebrow`: 11px at 0.06em tracking.** That is the type role built for ALL-CAPS labels, applied to a sentence somebody actually said. It is 13px normal sans now, in secondary rather than tertiary ink, and the name above it moved 13px -> 15px so the row finally has a hierarchy instead of two lines of identical type. **TODAY / EARLIER** were `font-fw-display` uppercase at 0.14em. A recency landmark orients; it should not out-shout the conversations under it. Sentence case, caption size. **Timestamps** leave Fragment Mono, same as the thread — a 0.06em-tracked monospace "9:12 AM" is a log line, not a time. They also stop turning accent green on unread: §17 says avoid six simultaneous green treatments, and the bolded name plus the count badge already carry it. **Press physics (§19)** — rows get `active:bg-surface-sunken`, so the acknowledgement paints on touch-down instead of waiting for navigation. Search takes the sunken track as a pill rather than a bordered rounded rectangle, which had it reading as a second card stacked above the rows. Verified by headless render against the real tokens. Row padding went to `py-3` in BOTH row components — the conversation row and the search-result row — so the two lists keep the same rhythm. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
…v bar ## Reply / quote (§30) `golf_messages.reply_to_id`, a self-reference — NOT a join table. A reply IS a message, so a column keeps every existing read, RLS policy, realtime subscription and pagination path working untouched; a join table would have needed policies duplicating the ones golf_messages already has. `ON DELETE SET NULL`, deliberately not CASCADE. Deleting a message somebody replied TO must not delete their reply — those are their words, and cascading would let anyone erase other people's messages by deleting their own. The quote degrades; the reply survives. The column confers NO read access. The quoted row is resolved through golf_messages RLS like any other message, so a pointer into another conversation renders as "Not in this part of the conversation" rather than leaking anything. Quotes resolve from the LOADED window instead of fetching: a per-bubble fetch would open one request per quote for a feature most messages do not use. `reply_to_id` rides the optimistic row too, so the quote paints in the same frame as the bubble rather than appearing a round trip later — the layout jump the arrival motion exists to prevent. The pending reply is owned by the PAGE and stamped with its conversation id. The composer would have been the obvious place, and it is the wrong one: a reply target surviving a thread switch is the composer-draft misdirection bug in a new costume. It is also read and cleared BEFORE the send await, because the composer clears optimistically and a preview left standing during the round trip points at a message already sent. ## Conversation details (§37) Tapping the thread identity opened nothing — a dead control, and the one place a group's membership is obviously "behind". It is a Fairway Sheet now: identity at a size that actually reads, member roster, and search. Scope is what EXISTS. §37 also lists shared media, pinned messages and per-conversation notification settings; none has a backing store, and rendering them as inert rows would be a menu of things that do not work. The member count renders only when the roster actually loaded — "0 members" for a group whose participants are still in flight is a number stating something false. ## The nav bar (§4) The header was `flex` with back, avatar and name packed left, so the title sat at whatever x the back label happened to end at — a different position per conversation — and the right side was empty. It is a three-column grid now: back, centred identity, overflow. A fixed centre is what makes a bar read as chrome instead of content that happens to be at the top. The "Messages" label is gone from the back control. It was defended in this file as naming the destination, but a chevron out of a conversation has one meaning, and the label cost enough left-column width to shove the title off centre. The destination is still named — by `aria-label`, where that belongs. ## Two things that were wasting a whole row each The mobile action band is GONE. It held a giant green pill, then briefly two icons; either way a full row of vertical space on the smallest screen, above the list it was pushing down. Compose and team-broadcast now sit on the rail's search row: one row that says "find one, or start one". The composer is an attached BAR, not a bordered box. A hairline alone left it reading as the last block of the page; an UPWARD shadow is what makes a bottom bar look like it is in front of the thread scrolling under it. Migration applied to production and the column verified before this commit. Verified visually by headless render against the real tokens. NOT done, and not claimed: mentions (§20) and conversation-list reorder motion (§15). Reorder motion is skipped on the spec's own instruction — §14 says do not run layout animation without profiling it, and I cannot profile on a device from here. Shipping unprofiled layout animation across a live list is the failure that section exists to prevent. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
Four chat-app references, and the thing all four had that Helm did not was a filter row under the search. Ours is keyed to what Helm actually knows rather than to a generic All/Favorites: a golf program's inbox is a mix of one-to-one coaching and team channels, and "just the team channels" is a real thing a coach wants at 6am on a travel day. **All · Unread n · Teams**, on the existing Fairway `Segmented` — no new control, and `size="lg"` for the 44px target. The count on the Unread chip comes from the same `unread_count` the rows render, so the chip cannot claim a number the list below it does not show. **Each option only appears if it would change what you see.** `showFilters` requires that some, but not all, conversations match — an inbox of three conversations that are all team channels does not get a "Teams" chip filtering to the same three. A control that cannot change anything is chrome, and this row sits directly above the content it would otherwise be pushing down. **The rows have one right rail now.** The timestamp used to sit on the name's line and the unread badge on the preview's, which gave every row FOUR alignment edges and put the badge in direct competition with the message text for the same horizontal space — a long preview shoved it around. Time over count, one edge, and the preview gets the full width it needs. Two things I checked before building rather than after: - **No presence data exists** anywhere in the messaging hooks or golf types. Three of the four references show "Active Now" / online dots, and Helm's Avatar even has a `status` prop ready for it. Rendering that without a source would be inventing a fact about whether a teammate is at their phone, so it is not here. - **ESLint caught a real bug**, not a style nit: my `visible` list started as a `useMemo` placed BELOW the loading/error/empty early returns — a hook called conditionally, whose call order genuinely changes between renders. It is a plain filter now, which is also what the two lines under it already do over the same array. Also repairs two gates that broke on formatting rather than behaviour, both made to assert intent instead of a literal: - `send-integrity` pinned the whole string `'const sendMessage = async (content: string)'` and broke when §30 added a second parameter — a signature change with nothing to do with the retry/client-id invariant it guards. It matches the function, not its parameter list. - `keyboard-inset` required whitespace before the composer's `pb-[calc(...)]`; that class moved into a `cn()` call and is now preceded by a quote. The class never changed — only what sits to its left. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
… that exist Looking at the four reference inboxes again, the thing I kept missing is what they DON'T have. **None of them has a date-section header.** Not one. They order by recency and let the timestamp column carry it, which is why they read as a single calm surface. Mine cut a five-row list into three labelled blocks — Unread, Today, Earlier — so the eye had to re-acquire the rhythm twice on the way down. The sections are gone, along with `groupConversationsByTime` and `GROUP_ORDER`, which existed only to label them. Floating unread to the top went with them. That was answering "show me what I haven't read", and the Unread CHIP now answers it explicitly and on demand, instead of permanently reordering the list under the reader. Recency is the one ordering that never surprises. Unread still reads instantly — semibold name, primary-ink preview, solid count. Three signals, no section. **The filter is loose pills, not a Segmented track.** Segmented is this repo's view-SWITCHER: a bordered sunken well with a sliding pill, sized to fill its row. Stretched over three options it reads as a squeezed tab bar and draws two more boxes onto a screen trying to be a list. Every reference uses free-standing pills, left-aligned, and they're right — a filter is a set of optional narrowings, not a segmented view of one thing. `SelectablePill shape="round"` is the primitive that already means this. It scrolls rather than wraps, so a fourth chip can't silently become a second row. **The unread badge was a wash.** `tone="accent"` is accent-50 with accent-700 text: at 12px on a cream row it was the quietest thing in the right rail, competing with a timestamp and losing. It's a solid fill now. It's the one element on this screen that should be loud. **Half the avatar column was invisible**, and it was the bubble defect again: the shared Avatar falls back to `bg-surface-sunken` (0.963) and this row sits on canvas (0.953) — a one percent difference. Overridden at the call site, not in the primitive, because every other Avatar in the app sits on a 0.984 card where the default is correct. DMs get a lifted cream disc with primary ink; groups get accent-100/800 rather than 50/700 — an initials circle has to earn from tone the presence a photograph gets for free. Inset hairlines between rows, positioned rather than bordered so they start at the text column and never cut through an avatar (§47). No UI kit installed. It would not have fixed any of the above — every one of these was a value/contrast or information-architecture decision, not a missing component — and it would have brought a second visual vocabulary onto a screen whose problem was already too many competing shapes. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
|
ⓘ Qodo reviews are paused because your trial has ended. Ask your workspace admin to add credits to resume reviews. Manage billing |
|
The latest updates on your projects. Learn more about Vercel for GitHub. 1 Skipped Deployment
|
|
Important Review skippedAuto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Team Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
|
Updates to Preview Branch (agent/messages-instant-entry) ↗︎
Tasks are run on every commit but only new migration files are pushed.
View logs for this Workflow Run ↗︎. |
…tion, real faces
## The composer was a form
`setSending(true)` -> `await onSend(...)` -> *then* `setMessage('')`. The text a
player had already committed sat in the field for the ENTIRE round trip while
the send button ran a bouncing-dot loader. That is form-submission feedback, and
it is the single reason sending felt like a web app: the user had said the
thing, and the interface was still holding it.
It clears in the same tick as the commit now, with `fwHaptic('light')`, and the
field collapses to one line in the same frame — a multi-line send used to leave
a hole where the text had been. The `sending` state is gone entirely; the
optimistic bubble is the feedback, and sending/sent/read/failed belong to it.
`canSend` is deliberately no longer gated on an in-flight send: a gate makes two
quick messages impossible, and clearing on commit already prevents a
double-tap from double-sending.
Failure does not eat the draft. The text goes back — but ONLY if the composer is
still empty, because by then the user may be typing the next thought, and
overwriting that is worse than losing it. Same rule for pending attachments.
## Chat motion is not control motion
`src/lib/golf/chat-motion.ts`. The shared Fairway control transition is a
deliberately slow ~180ms cinematic curve — right for a segmented control, wrong
for a bubble. A message arriving on it reads as the list re-laying out.
bubble arrival spring 420/34/0.75, y 8->0, scale .985->1 (was a flat 200ms tween)
Sent -> Read 140ms crossfade
reaction chip 120ms, scale .92->1, y -2->0
long-press bubble lifts 1.015 / -1px BEFORE the menu
row reorder layout="position", reorder spring
One module because the alternative is the same gesture at 140ms in one file and
220ms in another, which reads as inconsistency long before anyone can name it.
The read receipt is the subtle one: `Sent` and `Read` are different widths, so
swapping them shifted the timestamp beside it. Both states now occupy one
`inline-grid` cell, which reserves the wider label — opacity changes, geometry
does not.
Row reorder is `layout="position"`, not `layout`. Full layout animation scales
the row's box, which distorts text and remounts the avatar image mid-flight —
the "mushy" result the spec warns about, and a guaranteed photo flash.
## The row is not a button
`PressTarget`, not `Button`. The generic Button was being fought with eleven
overrides — `h-auto min-h-0 border-0 font-normal justify-start text-left` — to
stop it being a green capsule, and it brought a coupled click haptic and that
same 180ms curve. PressTarget is unstyled on purpose: real button semantics,
focus and disabled handling, no opinion about shape.
## Real faces
I said AvatarGroup was blocked on data. It was not — the photos were in the app
the whole time, and the resolver already existed. It was just SCOPED to the one
open conversation, so the inbox never asked.
`useGolfGroupAvatars` batches that same resolution across the whole list: three
queries, never one per row, because a per-conversation fetch is an N+1 on the
first screen a coach sees every morning. Participants paginate at 1000 —
PostgREST truncates silently rather than erroring, and a program with many team
channels crosses that. Identities resolve once for the UNION of user ids, since
the same coach is in most of them. Members with a photo sort first, because the
stack shows three and a real face beats initials in a slot that holds one.
Fails soft and silent-empty: any error returns an empty map and rows keep their
initials. A missing face is cosmetic; a throw takes the inbox with it. A group
with one resolved member stays a plain avatar — a stack of one is an avatar
wearing a ring it does not need.
## Also
The reactions migration's `COMMENT` prose tripped the Review Gate's
SECURITY-DEFINER rule, which scans for the phrase and could not see that the
function pins `SET search_path` eleven lines above. Reworded, not suppressed —
a `nosemgrep` there would silence the real case for whoever adds the next
definer function to that file.
Not built, and not claimed: `layoutId` avatar continuity from inbox to thread
(§35) needs both trees alive during the transition, which this master-detail
does not guarantee; NumberFlow unread counts; mentions; slash actions.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
…issing messages
## The worst behaviour in the feature
On a failed send the hook ran `setMessages(prev => prev.filter(...))`. The
player watched their words appear and then VANISH, with a toast as the only
trace of what they had written. That is the single most destructive thing this
surface did, and it did it quietly.
The message stays now. It dims, and its metadata line reads
`Not delivered · Retry` (§29/§30). Retry re-sends under the message's ORIGINAL
id — `golf_messages.id` is the primary key, so a retry that races a send which
actually committed collides on 23505 and the server reports it back as the
success it is. Pressing Retry twice cannot post twice. Same property the
transport retry already relied on.
The bubble softens rather than turning red: an undelivered message is not an
error, it just has not landed. Colour is never the only channel either — the
words are directly beneath it.
The composer no longer restores the draft on failure, because that would put
the same text in two places and the copy in the thread is the one being looked
at. Attachments are the exception: they cannot ride an optimistic bubble, so a
failed photo does come back to the composer.
## You stop silently missing messages
Arrivals while the reader is scrolled up were being counted by nobody — the
thread simply did not move (correct) and gave no indication anything had
happened (not correct). A floating `N new messages` control now appears above
the composer.
Only messages from SOMEBODY ELSE count. Your own send scrolls you down by
design, and counting it would offer to take you to your own message. Reaching
the bottom yourself clears it, because scrolling down IS catching up and a
button offering to take you where you already are is noise. It floats rather
than occupying a row: a permanent slot would cost every conversation height for
a state most never enter.
## Two small ones
Corners settle when a burst continues (§22). Worth recording WHY this is a CSS
transition and not Motion: last -> middle is a `border-radius` change, and
layout animation only interpolates transforms — it cannot smooth this at all.
A transition on the exact property is both correct and far cheaper than putting
every bubble under layout animation to chase it.
Unread counts roll 1 -> 2 with NumberFlow instead of swapping (§52), capped at
9 because past that the badge reads "9+" and there is no number to animate.
## One I caught before it shipped
I wrote `text-danger-600` for the failure text. Typecheck passed — Tailwind
class names are strings, and nothing in the type system knows that token does
not exist. It would have shipped as default-coloured text on the one line whose
whole job is to look wrong. The real token is `fw-danger-ink`.
## Not done, and it is a data limit
§74's group typing copy ("Owen and Cole are typing"). The typing broadcast
carries NO sender identity — the hook exposes a boolean. Naming somebody would
be a fabricated attribution in exactly the surface where getting it wrong
matters. It needs the broadcast payload extended first.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
…at the auto-fixer broke
The Review Gate's `sqlfluff (ratchet)` caught 60 new violations in my two
migration files (LT02 indent +41, LT01 spacing +11, LT05 long lines +7,
CP02 capitalization +1). The ratchet only permits the count to go DOWN, which
is correct: raising a baseline converts a caught regression into permanent debt.
`sqlfluff fix` resolved 62 of them. It also SILENTLY BROKE THE SQL.
Rule RF03 ("unqualified reference in single-table select") rewrote
WHERE m.id = message_id
to
WHERE m.id = m.message_id
in both the SELECT and INSERT policies. But `message_id` there is
`golf_message_reactions.message_id` — the OUTER row the policy is evaluating —
not a column of `golf_messages m`. `golf_messages` has no such column. The
subquery was reaching out of its own scope by design, and the fixer read it as
a mistake.
ERROR: 42703: column m.message_id does not exist
That is a formatter changing the meaning of a SECURITY PREDICATE. Every clean
rebuild would have failed at this file, and the only reason it did not ship is
that I executed the rewritten SQL rather than trusting the linter's exit code.
Both references are now qualified explicitly as
`golf_message_reactions.message_id`, which satisfies RF03 and says what was
always meant.
The fixer also lowercased `FROM PUBLIC` to `from public`. That one IS safe —
unquoted `public` in a GRANT/REVOKE is the PUBLIC pseudo-role, not the schema —
and it is verified below rather than assumed.
VERIFIED, not linted-and-hoped:
- both files execute against production in a rolled-back transaction (201)
- `golf_conversation_has_me` still shows EXECUTE for authenticated and
service_role only, no PUBLIC, no anon
- sqlfluff: 0 violations across both files
- `npm run sql:ratchet`: OK — 7666 violations, no regressions
The remaining 3 long lines were wrapped by hand rather than left for the fixer.
Also recorded both migrations in `supabase_migrations.schema_migrations`. They
were applied through the Management API, which executes SQL directly and does
NOT write the ledger — so production had the table, column, policies and
indexes while the ledger showed nothing. Applied-but-unrecorded drift: a future
`db push` would have re-run them (harmless, they are idempotent) and drift
tooling would have reported a gap that was not real.
SEPARATE, PRE-EXISTING, NOT FIXED HERE: `Supabase Preview` fails on
20260903220000_feature_health_heartbeat_indexes.sql (landed on main yesterday
in #1816) with "CREATE INDEX CONCURRENTLY cannot be executed within a pipeline"
— the CLI wraps each migration in a transaction. The index exists in production,
so it arrived by some other path, but a clean rebuild from migrations dies
there. That breaks `supabase db reset` and recovery-from-migrations for
everyone, and it means the preview validator cannot verify anyone's DDL. It is
not messaging and not mine to decide: dropping CONCURRENTLY takes an ACCESS
EXCLUSIVE lock on that table.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
Caught by rendering it, not by any gate — this is pure geometry, and neither
typecheck nor lint has an opinion about whether a flex child fits its parent.
A conversation row's avatar column is 48px, and the text beside it starts at a
fixed x. The group stack has to live inside that or half the list is misaligned.
Three attempts, and the two failures are worth recording because both looked
reasonable in code:
three `md` (40px) faces + AvatarGroup's "+N" chip ~130px hung out of the row
two `sm` (32px) faces at a 16px overlap 48px back face reduced to
a sliver of half a
letter — reads as a
rendering bug, not
as layering
two `xs` (24px) faces at the default 6px overlap 42px fits, both legible
The "+N" chip is gone with `max`: it is another full-size element, so including
it puts the stack straight back outside the column. Nothing is lost — the row
already names the conversation and the details sheet lists every member.
MAX_FACES drops 3 -> 2 in the hook for the same reason, so it stops fetching
members that could never be rendered.
Gates: lint 0, 1561 test files / 15200 tests pass.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
…ocked ## I was wrong about "blocked on data" I reported group typing copy as blocked because the broadcast carried no sender identity. It does. The handler destructured `userId` and then threw it away to set a boolean. Nothing needed a backend change; I had read my own code as a wall. It tracks WHO now, with ONE EXPIRY TIMER PER TYPIST. The single shared timer was a real bug: a second person starting to type reset the first person's countdown, so in a busy group somebody could show as typing indefinitely after they had stopped. Names resolve locally from the participant map — never from anything a sender puts on the wire — and an unresolved id is DROPPED rather than rendered as "Someone", because a line naming the wrong count is worse than plain dots. Two names, then a count, so five typists cannot grow the line and push the thread. `isOtherTyping` is now DERIVED from that list rather than stored beside it. Two sources for "is anyone typing" is two things that can disagree, and the old boolean could stay true after the last typist was removed. ## Visual finish (§13-23) - **Bubble elevation cut by roughly two thirds**, and outgoing has NONE. `shadow-soft` is the raised-CARD elevation; at that weight a thread reads as a stack of floating Surfaces. Colour already separates cream-on-champagne and green-on-champagne, so elevation only has to hint. A saturated green fill is already the highest contrast step on screen — elevation on top of it is the green-competition problem in §31. - **Incoming avatar in DMs**, 28px, scale-only entrance. I had removed it to save 40px of width, arguing the header already names the person. That was an argument about information; §12's is about PRESENCE, and it is the better one — a thread with a face in it feels like somebody is there. - **Reply quote** drops to medium weight and lower contrast. At 17px response over 12px quote, the response has to be what the eye lands on. - **`Read · 7:43 PM`** as one tertiary line instead of an accent chip beside the time, which had made a routine status the second-greenest thing in the thread. - **`⚠ Not delivered · Tap to retry`** as one PressTarget. Two controls side by side read as web links, and the whole metadata region is now a much bigger target than the word "Retry" was. - **New-messages button is chrome**, not a primary CTA — it is a way back to the bottom, not an action, and a solid green pill competed with the outgoing bubbles it sits among. - **Filters 32px and shadowless**; dividers at half opacity. Together these were putting the CONTROLS above the PEOPLE in the hierarchy (§3) — the eye reached "All / Unread / Teams" before it reached a single name. ## Two real bugs lint caught An arbitrary `text-[11px]` (the token rule exists for exactly this), and reading `ref.current` inside an effect cleanup — which can see a different object than the effect installed timers into, clear the wrong map, and silently leak the real ones. Both captured as locals now. Gates: lint 0, 1561 files / 15200 tests, knowledge 0, docs 0. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
… poll, travel The P2 differentiator. A message stops being only text. ## Schema One migration, because these are one feature and splitting them would create an ordering dependency between files that all touch golf_messages. golf_messages.kind + payload text | system | practice | event | rsvp | poll | travel golf_message_responses one row per person per message golf_message_mentions stored, not re-parsed golf_messages.pinned_at/_by golf_conversation_participants.notification_level / muted_until **A structured message IS a message.** It rides `golf_messages`, so every existing read, RLS policy, realtime subscription, reply, reaction, search and pagination path works on it unchanged. A parallel "team objects" table would have needed its own copy of all of that, and would have put half a conversation outside the conversation. The constraints EARN their place, and I checked rather than assumed: `kind='rsvp'` with a null payload is rejected with 23514 on production. Without that pair of CHECKs a client could post a structured kind with nothing in it and every reader would render an empty card. `golf_message_responses` is unique on (message_id, user_id), which is what makes changing your mind an UPDATE rather than a second vote — held by the database, not by client discipline. Its INSERT and UPDATE policies both pin `user_id = auth.uid()`; without that a participant could answer a poll as somebody else. Mentions are STORED rather than parsed from content at delivery time, because "mentions only" has to be answerable by a query, not by running a regex over every message. Insert is restricted to the message's own author — otherwise anyone could write mention rows against a message they did not send and manufacture a notification. `notification_level` lives on the PARTICIPANT, not the conversation: muting is a property of your membership, and a coach muting a thread must not mute it for the team. `muted_until` expires on READ rather than by a job, because a cron that has not run yet leaves somebody silently muted — the failure you never find out about. ## Rendering `parseStructuredPayload` narrows an untrusted jsonb blob and returns NULL rather than throwing. Postgres will not check the shape of a payload, so a malformed one has to degrade to an ordinary message — a throw inside a message list unmounts the conversation. Every branch requires the fields the renderer actually reads, so a half-written payload cannot produce a card with blank headings. A `Surface` is right here and deliberately wrong for ordinary bubbles: a structured message represents an object with state and actions, which is what a card means. A sentence somebody typed is not. System messages are narration, not speech — centred, quiet, no bubble, no author, no avatar (§40). Times are stored as ISO instants and formatted in the READER's timezone. A coach in Eastern posting "3:30" to a player who has travelled is the classic version of that bug. "View in Calendar" renders only when the payload actually carries an eventId. A button that opens nothing is worse than no button. Poll share is guarded against 0/0 — a NaN width silently renders a full bar. ## One caught by rendering it The unpicked poll bar was `bg-surface` on a `bg-surface-sunken` track: 0.984 against 0.963, a two percent difference, which renders as no bar at all. Same defect as the message bubbles and the inbox avatars, third surface. Both states use the accent ramp now. Two tests updated: the action tripwire 449 -> 451 for the two new actions, and the scroll test stubs the responses hook — that file tests scroll position, and structured answers drag a Supabase client into a jsdom environment with none. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
…hidden it `Button` wraps ALL of its children in a single <span>. Passing a text node and an <svg> as children therefore makes them one inline flow, so `justify-between` had exactly one thing to distribute and the arrow wrapped onto a second line. It goes through the component's own `rightIcon` slot now. The reason this shipped past three green gate runs is worth recording for whoever picks this up. Every screenshot in this branch until now was HAND-WRITTEN HTML approximating the components, not the components. My version had drawn the label and the arrow as two flex items, so the button looked correct in every render I produced — and the emoji I used as icon stand-ins are what prompted the owner to ask why the icons looked wrong. An approximation that drifts from the code is the same failure class as a gate that passes by being empty: it reports success about something it never examined. Screenshots are now produced by rendering the ACTUAL component through react-dom/server with Tailwind compiled against the real source files. The first run of that found this bug immediately. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9
Owner
Author
Latest messaging material + identity passCommit
Verified locally: focused messaging tests (23/23), targeted ESLint, |
njrini99-code
added a commit
that referenced
this pull request
Sep 4, 2026
`open-pr-residue` requires the key set here to equal the live open-PR set, so opening a PR without a row makes the check fail — but a row for THIS PR cannot be on main while the PR is in flight. The file's own schema resolves that: a row absent from main but present at the PR's own head is IN FLIGHT, not unclassified. This is that row, at that head. HUMAN_TEST_PENDING rather than ACTIVE, and the distinction is doing work: the change is finished and green, and what it is waiting for is a thumb. A 450ms hold with 10px of slop cannot be verified by a test that synthesises its own pointer events. It is also waiting on a merge-order call against #1833. worktree_policy PARK_IF_REPRODUCIBLE releases the checkout — it exists only to have cherry-picked and pushed this branch, and parking keeps the branch. Both gates have to permit, so the workspace marker is being released in the same step; neither one implies the other. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019euu9jtJM6WvMVj17coqZ1
njrini99-code
added a commit
that referenced
this pull request
Sep 5, 2026
…e has ended `open-pr-residue` requires the key set to equal the open-PR set, and a row for an in-flight PR lives at that PR's own head rather than on main — so #1836's row rides its own branch. Verified both directions: without it the check names #1836 as unclassified; with it, it does not. Also deletes #1725 and #1738, both MERGED, both past the transitional grace this file's own `$comment` defines — the tool's output says in as many words to delete them "in the PR you are already opening", and that is this one. Leaving them is the exact failure the comment describes: a current-state registry asserting things that stopped being true. **`open-pr-residue` still FAILS, for reasons this PR does not own.** Twelve open PRs carry no row at all — #1834, #1833, #1832, #1831, #1829, #1827, #1759 and five dependabot PRs. That was already true before this commit and is not mine to reconcile; a row is a statement about someone else's work. Recording it here so the red is legible rather than mysterious. Note for merge order: #1835 also edits this file, so whichever lands second resolves that conflict. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019euu9jtJM6WvMVj17coqZ1
…ale assertions
Every failing step reported `conclusion: success` in the UI, because each check
runs `continue-on-error: true` and the job's aggregate reads `outcome`. So the
PR looked like it had no failing steps and a mysteriously failing aggregate.
The aggregate's own log names them: `FAILED ratchet_lint` and
`FAILED world_model`.
**ratchet_lint — a real defect, fixed rather than baselined.**
`@typescript-eslint/no-unused-vars` went 25 -> 26. `npm run lint` never saw it:
that lints `src/**`, and the ratchet lints `src` AND `scripts`. The variable is
`viteOutput` in the QA capture harness, which accumulated every line Vite wrote
to stdout and stderr and then dropped them — while the `finally` below it
asserted the run "has already surfaced any Vite startup or browser failure
above". It had not: a config or port failure surfaced as a bare
`waitForServer` timeout with the actual reason discarded. Now the catch prints
it. The variable is used because it should always have been, not renamed `_`.
**world_model — a stale generated artifact.** `WORLD_MODEL.json`/`.md` no longer
matched their sources after this branch's registry and code changes. Regenerated.
**The five unit failures were assertions pinning markup, not contracts.**
Every underlying contract this branch was accused of breaking still holds:
12-col proportional grid held (gap-6 -> gap-0 + md:border-y)
rail spans 5/12 md, 4/12 lg held (md:border-r inserted between them)
thread spans 7/12 md, 8/12 lg held
720px readable cap held
masthead never on phone height held MORE strongly — ViewHeader is gone
"official team channel" naming held — moved into an aria-label
Each assertion matched a whole class string or one capitalisation, so additive
styling and an icon-ified entry point read as regressions. They now pin the
tokens a real regression would remove: `grid-cols-12` with `md:grid`, each span
pair with anything permitted between them, no `<ViewHeader>` at all, and the
channel named in either register. A flexbox rewrite or an unhidden masthead
still fails them.
Verified locally at this head: lint 0, lint:ratchet 0 (62 warnings, no
regressions), full unit suite 1566 files / 15230 passed, docs:check 0,
knowledge:check 0, world-model check 0.
NOT addressed here: `Snapshot Testing` (a Sentry preprod check, external to
this repo's gates), and the production-schema-ahead-of-main reconciliation this
PR's description documents — that is what merging it does, and it is the
owner's call, not a CI fix.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019euu9jtJM6WvMVj17coqZ1
…te kept #1827 ("honour the Messages notification toggles") landed on main while this branch was open, and both changes answer the SAME report: a player sent 33 notifications in a day whose toggle was read by nothing. They answer it differently, and the conflict is semantic rather than textual: #1827 gate email behind `email_messages`, a preference defaulting ON #1833 remove the email channel, by owner instruction — "stop sending emails every time there's a message. It should just be the app notification." Removal wins: it is the stronger answer to the same complaint and the owner asked for it in those words. But taking either side whole breaks the file — `prefsFor` is DEFINED inside the region #1833 deletes and USED below it by the push gate, which is #1827's durable half. So the resolution keeps the removal AND the gate: no email at all, push still honouring `push_messages`, the in-app bell still ungated because it is how the message is discovered at all. Three of #1827's tests asserted email IS sent. Each was re-pointed at the contract underneath it rather than deleted: - "emails every recipient even though RLS returns none" was never about email; it pins the ADMIN FALLBACK resolving a recipient RLS cannot see. Now asserted on the bell. - "does NOT email when email_messages is off" would still pass if the channel returned for everyone who never touched the toggle — `email_messages: true` is the DEFAULT. Re-pinned with the preference ON, which actually catches the regression. - quiet-exemption still matters; it now asserts on push, with `push_messages` set explicitly because its documented default is OFF and the test would otherwise have passed for the wrong reason. Verified on the merge: typecheck, lint --max-warnings 0, lint:ratchet 0, world-model regenerated and matching, full unit suite 1568 files / 15252 passed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019euu9jtJM6WvMVj17coqZ1
njrini99-code
added a commit
that referenced
this pull request
Sep 5, 2026
…wasn't the sheet Two defects the approved-design pass left standing. Both are invisible to every check we had, which is why they survived a screenshot review. 1. Long-press cancelled on the FIRST pointermove. That reads as correct — they moved, so they meant to scroll — and it passes in jsdom, which emits no spurious moves, and in a screenshot, which cannot photograph a gesture. On a real digitizer a resting finger emits a continuous dribble of sub-pixel moves for the whole press, so the deliberate stationary hold the feature exists for was the one gesture guaranteed to be cancelled. Now a DISTANCE threshold: 10px, the platform convention on both sides (Android getScaledTouchSlop() ~= 8dp, iOS allowableMovement = 10pt), extracted as exceedsLongPressSlop so a test pins the number rather than the prose. A zero threshold fails MessageThreadPane.longPress.test.ts. 2. The actions opened as an inline icon-only row at the MESSAGE's position. The approved design is a bottom sheet — grabber, dimmed thread, labelled Copy / Edit / Delete with Delete below a divider — and the difference is not decorative: an inline row beside a message near the top of the screen is out of thumb reach, and four unlabelled icons make Delete a guess. Now the Fairway Sheet (vaul), which also supplies the scrim, focus trap, Escape and drag-to-dismiss that were hand-rolled listeners before, so this deletes more code than it adds. One sheet per thread resolved from mobileActionsId, not one per bubble; a stale id resolves to null and closes rather than acting on a row that no longer exists. Reply is deliberately absent here — it needs the reply_to column that lives on #1833, and adding a fourth row now would collide with that branch in exactly this region. Verified: tsc --noEmit clean; eslint --max-warnings 0 clean on both changed files (the first draft used a raw <button> and h-13, a class this Tailwind config does not define — both caught and fixed, not shipped); 83 tests across 15 files pass over messages + overlays; markdown ratchet reports no regressions. Not verified locally: the gesture itself on hardware, which is the thing item 1 is about. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019euu9jtJM6WvMVj17coqZ1
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Mobile messaging rebuild. 16 commits. Not merged, not deployed.
The bug that made four earlier visual passes invisible
The thread well was
bg-surface— oklch(0.984), the lightest colour in the system — and incoming bubbles werebg-surface-sunkenat 0.963. Every bubble was 0.021 darker than the surface under it. Pressed into the page, not resting on it.design-tokens.cssalready had the answer at--fw-color-canvas: "the ~0.03 gap is the premium card/page separation the flat-cream look lacked."This defect recurred three more times in the same session — the rail's avatars (
surface-sunkenoncanvas, 1%), the loading skeleton bars, and the unpicked poll bar (surfaceonsurface-sunken, 2%). This design system has four neutrals inside a 3% lightness band. Picking two of them for figure-and-ground produces something invisible, every time. Check any new fill against its actual ground.What's in here
Thread — the screen is the conversation below
md; three-column nav bar; short threads bottom-anchor viamt-auto; 17px text; spring arrival (420/34/0.75);Sent → Readcrossfade in reserved geometry; reactions; reply/quote; details sheet; long-press lift; failed sends that stay put with Retry; new-messages button; incoming avatars; group typing by name.Inbox — real member faces (
AvatarGroup);PressTargetinstead of an over-riddenButton; one continuous list (no date sections); loose filter pills; one right rail per row; solid unread badge with NumberFlow;layout="position"reorder; compose on the search row.Composer — clears in the same tick as commit (was
await onSend(...)then clear, with a bouncing-dot spinner);fwHaptic('light')at commit; nosendingstate.Structured Helm objects — practice, event, RSVP, poll, travel, system narration. A structured message is a message: it rides
golf_messages, so every existing RLS policy, realtime subscription, reply, reaction and search path works on it unchanged.Removed — per-message email fanout (a ten-message thread mailed every participant ten times).
Verification — read this before trusting any screenshot in the history
Screenshots earlier in this branch were hand-written HTML approximating the components, not the components. That approximation drifted and hid at least one real bug (a wrapped "View in Calendar" button, fixed in the last commit).
The working method is in the last commit: render the actual component via
react-dom/server, compile Tailwind against real source files, screenshot that. It found the bug on its first run.Nothing here has run on a device. The composer's safe-area arithmetic changed, springs were added, keyboard geometry moved — the three most device-sensitive things in the spec. §55's physical-iPhone pass is exactly what a harness cannot do.
NOT DONE
Built schema, no UI: mentions (tables + policies exist; no
@autocomplete), pins (pinned_at/pinned_byexist; no affordance), mute (notification_level/muted_untilexist; no sheet).Not started: slash actions (
/practice,/poll…), in-app notification banner, offline/reconnecting status, search-hit highlight, scroll-edge material, swipe-to-reply,layoutIdavatar continuity (§32 — needs both trees alive across the master-detail transition; investigate before building).Verification tier — none of it: observability instrumentation, render profiling, the 320–430px overflow lab, screenshot regression pack, physical-iPhone QA, latency measurement.
Traps worth knowing
sqlfluff fixwill break your SQL. Rule RF03 rewroteWHERE m.id = message_idtom.id = m.message_idin two RLS policies —message_idbelonged to the outer row.ERROR: 42703. Caught only by executing the rewritten SQL. Always run migrations after auto-fixing them.text-danger-600, which doesn't exist; the real token isfw-danger-ink. Typecheck and lint both passed.ghin the Bash sandbox works for one or two calls, then fails withx509: OSStatus -26276. Run bursts outside the sandbox.Supabase Previewis red for a PRE-EXISTING reason —20260903220000_feature_health_heartbeat_indexes.sql(fix(bridge): five production fixes found by reading the live Bridge, in one PR #1816) usesCREATE INDEX CONCURRENTLY, which cannot run inside the CLI's transaction. The index exists in production, so it arrived another way, but a clean rebuild from migrations dies there. Not caused by this branch, and it means the migration validator can't verify anyone's DDL right now.RLS was exercised, not just shaped
For reactions, in rolled-back transactions on production, with a control that passes so the denials aren't vacuous:
golf_message_responsesandgolf_message_mentionsfollow the same policy shape but were not exercised this way. Worth doing before shipping.Local gates green at every commit: lint 0, 1561 test files / 15,200 tests, knowledge 0, docs 0.
🤖 Generated with Claude Code
https://claude.ai/code/session_01WUAKe2pAbPcQ6cH1s4D3P9