[richshaw]

Part of #503, which the owner's design comment narrows to "the check at Publish". Its other checks (merging copies, empty fields, unnamed fields, ungrouped dates) aren't in this PR, so #503 stays open for the owner to close or rewrite. This is PR 4 of 5 for the "Publish and share" redesign in #498 (row I), stacked on #578. Review it against claude/relaxed-dijkstra-cpx7zo-3-publish-button.

The problem

Publish published whatever was stored, including:

• a hand-made draft whose only slot was still empty, which went out as "Untitled slot";
• an AI draft with a slot that still said "TBD".

Once published, a signup can't go back to being a draft. There was also a quieter race: edits on the Build tab save about 800 ms after the last keystroke, so a Publish straight after typing published the version from before the edit.

The fix

Publish and share sends what's waiting to save, then checks.

• When it's pressed, the button first sends any Build edits still waiting to save and waits for them. That includes saves already on their way and those a Build tab left behind when the organizer switched tabs.
• While it waits: aria-busy and aria-disabled are set at once, so a second press is refused and focus stays on the button.
• The busy look: "Checking…" with a spinner shows only if the answer takes longer than about 300 ms. It stays at least 400 ms, unless an answer shows first: it goes in the same render as the check's panel, the share panel or a refusal.
• The width holds: both labels sit in one grid cell, so the header doesn't move. The accessible name is always the visible label.
• The wait is capped at 10 s, then it goes on. If a save is refused, it doesn't publish. It says "A change didn’t save. Fix it, then publish." instead: under the header on a computer, above the button in the bar on a phone.
• The check: the server action reads the draft fresh, policy guard included, and runs checkBeforePublish. It's a pure function, src/lib/publish-check.ts, and returns the first problem:

Problem	Panel	Buttons
No slots	"Nothing to sign up for yet" / "There are no slots yet. Add one, then publish."	Add a slot, Not now
No fields	"Nothing to sign up for yet" / "Add a field that says what people sign up for, then publish."	Not now
Every slot empty	"Nothing to sign up for yet" / "The slot is empty…" or "Every slot is empty. Say what people sign up for, then publish."	Fill it in, Not now
A slot still says TBD, TBC or TBA, as a whole word	"Before people see it" / "Cashier TBD · Sat, Nov 14 still says “TBD”. People will see that, and a published signup can’t go back to being a draft." or "3 slots still say “TBA”, “TBD” or “TBC”."	Name it first, Publish anyway

• How the warning names a slot: by its values, leaving out any field whose value is made only of those words, give or take punctuation, such as "TBA", "(TBD)" or "TBD/TBC". So "Cashier TBD" stays in, and a Who of "TBA" gives "Fruit and water · Sat, Apr 6 still says “TBA”".
• The first three block, and the form can't wave them through. Publish anyway passes only the warning.
• With nothing wrong, it publishes at once and opens After Publish, a share panel to send the link #574's share panel.
• A signup that's no longer a draft skips the check and gets today's "This signup changed since the page loaded. Reload to see the latest."
• The MCP publish_signup tool, POST /api/signups/[id]/publish and the service are unchanged, apart from the service taking an optional acknowledged that only the header's Publish anyway passes (below).

The panel

• It's a popover by the button on a computer, and a sheet over the bar on a phone, with the fix first and 48px buttons.
• It's named by its title and described by its text, and focus starts on the fix.
• Escape, Not now and ✕ return focus to Publish and share. On a computer, pressing Publish and share again closes it.
• The header's live region says nothing for a check.

The fixes

• Fill it in and Name it first open that slot's editor on the Build tab, scrolled into view, with focus in the field that needs it: the TBD's own field, or the first field.
• Add a slot adds one and opens it.
• From the Responses or Settings tab, the request waits in a small client bridge in the layout (BuildBridge) while it navigates to Build. The Build tab takes it once when it registers. It's never in the URL, so Back, Forward, a reload or a bookmark can't replay it.
• A request the organizer walks away from before Build loads is dropped, so a later visit to Build can't add an empty slot. That happens when:
  • the path settles on any tab but Build;
  • they make a plain click on any tab but Build, the one they're on included (a Ctrl/Cmd, Shift, Alt or middle click doesn't navigate this window and keeps it);
  • they go Back or Forward;
  • any form is submitted, such as Settings' Save, which comes back to the same path.
• A fix for a slot this tab hasn't loaded gives focus back to the button.

What's recorded. #525 records signup.publish_refused in the publish service. This check now stops "no slots" before the service, so the check writes the same event itself, through recordPublishCheck in src/services/signups.ts (policy guard, best effort, outside any transaction; a failed read is logged and never fails the publish):

• { code: 'conflict', reason: 'no_slots' | 'no_fields' | 'all_blank' } for a blocking problem. no_slots is the same shape Record link copies, failed AI drafts and refused publishes #525 writes.
• { reason: 'placeholder', warning: true } for the warning. A warning doesn't stop anyone, so docs/telemetry.md says to count real refusals as the rows whose payload has code.
• After a warning, Publish anyway adds acknowledged: 'placeholder' to the signup.published row. A publish after the slots were fixed has no flag. The REST and MCP publishes can't set it.

Choices this makes

• The check runs on the server, on the stored draft, after the flush. The header's own data is the layout's, which doesn't refresh while the organizer edits in Build.
  • The read isn't locked. A concurrent edit, say from an AI assistant, could empty the slots between the check and publishSignup's lock. The service's own "Add a slot before publishing." still covers the case where the last slot is deleted.
• Only placeholder words that stand alone count. "Tbacco" and "ÉTBA" don't. The regex is the one from the closed Point out fixable problems in a draft on its review page #530.
• The check panel is in the first load, not lazy. Lazy-loading it saved 213 B on the signup tabs and added 21 B to every page's webpack runtime, which put / and /s/[slug] over their budgets.
• The busy look follows the owner's tophat (comment below): no flash for a quick answer, and "Checking…" rather than "Publishing…", since the answer may be a check.
• Left for PR 5: marking slots that still say TBD on the Build page, and Done (or Enter) after a fix putting focus back on the button.
• Known limits:
  • After the 10 s cap, the check reads what's stored. A save slower than that can still be on its way, so the panel can say a slot is empty that the organizer has just typed into.
  • A Duplicate whose save takes longer than that 10 s can take focus from the fix's slot when it lands. The root is Put a duplicated slot right after the original and open it #568's code, which opens the copy whenever its save returns, whatever was opened since; main has the same steal when another slot is opened meanwhile. It wants a follow-up there.
  • Back or Forward to Build shows Next's cached Build payload, which Build's own saves don't refresh. That's older than this PR and worth its own issue.

Tests

• Unit:
  • publish-check.test.ts: each kind, the order of precedence, near-misses, the slot label (whole-value and punctuation-only placeholders left out, "Cashier TBD" kept), every distinct word, and the copy for one slot and for several.
  • useBuildState.test.ts: the flush sends pending saves and waits for those in flight and those that aren't debounced (a duplicated slot included), loops until nothing waits, reports a failure, and doesn't send twice on unmount.
  • useDelayedFlag.test.ts: on fake timers, an instant answer never shows, a slow one shows at 300 ms, an early end keeps it to 400 ms, and a press after it hid waits again.
• Component:
  • PublishButton.test.tsx: the flush resolves before the action; the 10 s cap; a refused save stops it; the panel's focus and dismissal; the fixes; a phone sheet. On fake timers: "Checking…" on both buttons only after 300 ms, gone the moment the panel or a refusal shows and not back after Escape; and the stacked labels' names.
  • BuildBridge.test.tsx, with the real TabsNav:
    • a request is taken once, a remount doesn't replay it, and it's kept through a re-render on the same page and a click on Build;
    • it's dropped when the path settles elsewhere, on a plain click on the current or another tab, on Back, and on a form submit, but kept through a Ctrl-, Cmd-, Shift- or Alt-click and a middle click (auxclick);
    • a departed tab's failure stops counting once a new tab registers.
• DB:
  • actions.db.test.ts: each kind blocks or warns and leaves a draft; an acknowledged placeholder publishes; an acknowledged blocking problem still blocks; a viewer is refused; a non-draft gets the reload message.
  • Telemetry: each check kind writes one row with the right payload, by the organizer who pressed it. Publish anyway writes acknowledged: 'placeholder' on signup.published, and a fix-then-publish, or a Publish anyway after the slot was named elsewhere, writes { from, to } only. A viewer writes nothing. A read that throws records nothing and still returns the check. The service's own row still appears when the check is bypassed.
• e2e (organizer.spec.ts, 1280 and 390):
  • an empty slot stops Publish, Escape returns focus, and Fill it in focuses What;
  • a TBD slot warns, published from Responses: Name it first lands on Build with the slot open, and Publish anyway opens the share panel;
  • Back and Forward after Add a slot from another tab write nothing.
• Help walkthrough: the phone run goes through the empty-slot check, the TBD warning and Name it first, and both runs keep Put a duplicated slot right after the original and open it #568's Duplicate steps.
• Brought up to date with One Publish and share button, in a bar on phones; drafts show no public link #578 and, through it, main and the earlier PRs: Record link copies, failed AI drafts and refused publishes #525 and Tell Magic Compose to use one slot with a capacity, not a row per person #526 in 2fd9d88 and 0c9f1c5; Put a duplicated slot right after the original and open it #568 in 5d166f4; and the owner's tophat rounds on Ask before closing a signup, and move Close signup into the ⋯ menu #571 and After Publish, a share panel to send the link #574, with Cut an over-long Magic Compose title or description instead of throwing the draft away #579, in af65c82, daad50c, 84f146e and f95e870.
• Passed:
  • on f95e870: pnpm lint && pnpm typecheck && pnpm knip && pnpm test (1675);
  • on 84f146e: pnpm test:db for the touched files, pnpm build && pnpm budgets, and the chromium e2e for organizer.spec.ts and the publish walkthrough (20);
  • on e8150c0: the full pnpm test:db (562) and the full chromium e2e with CI=1 (72).
• Independent checks: a tophat on 3952591, and code reviews of every round, including the owner's tophat round, are below. Several of them include a running-app check at 1280 and 390. The review summarises the first five rounds. The latest fix (b3316f9, "TBD/TBC" in the label) closes an edge case the latest re-check found; I read it, and it hasn't had a separate check of its own.

Budgets

Every signup tab's first load grows by about 1.7 kB gzip, and Build's by 2.7 kB. The tabs carry:

• the Publish and share button, its check panel (static, as above) and its delayed busy look;
• the bridge to the Build tab, with its drop triggers.

Build also carries the flush that makes Publish wait for its saves, and the code that opens a slot from the check. Each budget goes to the next whole kB above its new size. The owner's tophat round added about 220 B per tab, which still fits.

Page	#578 (21eb67c)	This PR (f95e870)	Change	Budget
/app/signups/[id]/build	155.3 kB	157.98 kB	+2.7	156 → 158
/app/signups/[id]/responses	128.1 kB	129.84 kB	+1.7	129 → 130
/app/signups/[id]/settings	127.7 kB	129.38 kB	+1.7	128 → 130
/app/signups/[id]/edit, /app/signups/[id]	126.7 kB	128.42 kB	+1.7	127 → 129
/, /s/[slug], preview			+6–7 B	unchanged

Docs

• Help: create-and-publish-a-signup has a note under the Publish step covering Nothing to sign up for yet, Fill it in, Before people see it, Name it first and Publish anyway. Those names are in its UI list, and the walkthrough clicks through them. No screenshots change.
• CHANGELOG: a Find fixable problems in a draft and offer one-tap fixes #503 entry under Changed, including what's recorded and the busy look. The A viewer who chooses Publish, Close or Delete gets the error page #390 line no longer lists "Add a slot before publishing." as something Publish shows.
• docs/telemetry.md: the signup.publish_refused row lists:
  • the check's reasons and both of its sources;
  • how to count real refusals;
  • how to read acknowledged: 'placeholder'. A publish with no flag after a warning means the slots were fixed first, or the signup was published through the API or an AI assistant, which don't check.
• Privacy page: already says the log records "a publish … that didn't go through", and the new flag holds nothing personal, so its text and date are unchanged.

The stack

1. Ask before closing a signup, and move Close signup into the ⋯ menu #571: Close signup moves into ⋯ and asks first (One click on "Close signup" closes a signup for good, from the spot where Publish was #457).
2. After Publish, a share panel to send the link #574: the share panel, with copies and shares recorded (After publishing, help the organizer share the link #505).
3. One Publish and share button, in a bar on phones; drafts show no public link #578: one Publish and share button, in a bar on phones; drafts show no public link (Make Publish hard to miss, and say when it fails #499).
4. This PR: the check when Publish and share is pressed (part of Find fixable problems in a draft and offer one-tap fixes #503).
5. After a fix, hand focus back to Publish and share; mark slots that still say TBD #581: focus back on the button after a fix, and slots that still say TBD marked in Build.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HoJtwsyrp6YC4DU9JAnc7H

[richshaw]

Tophat: passed

I ran 3952591 locally with pnpm dev, Postgres 16, and Chromium at 1280×900 and 390×844. An independent agent that didn't write the change did the run.

The head has moved on since. It now also has the catch-up merges with #578 and main, the telemetry for the check's outcomes, and the bridge's expiry (2fd9d88..9946d9d). That round is getting its own independent running-app check, which I'll post below when it's done.

Check	Result	What I saw
Empty slot, computer	✅	A popover by the button, "Nothing to sign up for yet" / "The slot is empty. Say what people sign up for, then publish.", named by its title and described by its text. The button shows aria-expanded=true. Focus starts on Fill it in. Escape and Not now return focus to Publish and share, and pressing it again closes the popover without publishing. Fill it in opens the slot with focus in "What value". The signup stays a draft.
Empty slot, phone	✅	A sheet over the bar with Fill it in first. Both buttons are 358×48. Escape, Not now and ✕ return focus to the bar's button. The field opens on screen, above the bar.
No slots, no fields, and fixes from another tab	✅	On Settings, a draft with no slots: Add a slot lands on Build's own address with a new slot open, after a single POST. Three rounds of Back and Forward, and a reload, wrote nothing and reopened nothing. Every slot empty, from Settings: Fill it in opens the first slot. No fields: only Not now, which takes focus.
TBD warning, computer	✅	One slot: "Fruit and water · Sat, Apr 6 still says “TBA”…". Name it first puts focus in the field that holds the TBA, not in the first field. Several slots: "3 slots still say “TBA”, “TBD” or “TBC”." "Tbacco" and "ÉTBA" are ignored. Publish anyway publishes, opens the share panel and focuses Copy link.
TBD warning, phone, from Responses	✅	Name it first (focused) sits above Publish anyway, both 48px. Name it first lands on /build with the slot open and "Cashier TBD" focused. Back and Forward reopen nothing and write nothing. On the 14th slot of a list, the field scrolls into view above the bar.
Publish sends waiting edits first	✅	Type, then press Publish at once: the slot PATCH goes out before the server action, and the signup publishes with the new value. With the PATCH slowed by 3 s, the button reads "Publishing…", focus stays on it, and the panel opens at 4.2 s. Type "Juice TBD", switch to Responses, then press Publish: it waits for the save, then warns about "TBD".
10 s cap	✅	With the PATCH held for 16 s, the button still reads "Publishing…" at 8.5 s. The panel opens at 10.4 s, checking what's stored.
A refused save	✅	A real 404, for a slot deleted through the API while its editor was open. Publish stops with "A change didn’t save. Fix it, then publish." No panel opens, and the signup stays a draft.
No longer a draft, and can't be published past	✅	Published through the API, then Publish pressed: "This signup changed since the page loaded. Reload to see the latest." Emptying the slot through the API while the TBD warning is open, then pressing Publish anyway, switches the panel to "Nothing to sign up for yet", and the signup stays a draft.
Screen reader	✅	The live region says nothing when a check opens, and "Published. Anyone with the link can sign up now." after a real publish.
Help article	✅	The new paragraphs match the app's wording.
Console and server log	✅	No console errors.

What changes for organizers: Publish and share first saves anything still being typed, then checks the draft:

• Nothing to sign up for: a panel by the button says so and won't publish. Its button opens the slot to fill in, or adds one.
• A slot still says TBD/TBC/TBA: the panel names it and offers Name it first or Publish anyway.
• A refused save: it says so instead of publishing.

Looked off, but not a failure:

• If a save takes longer than 10 s, the check can say a slot is empty that has just been typed into. That's the cap, working as designed.
• After a refused save, the builder shows the stale slot. That's older than this PR.

Screenshots are in the session linked below.

https://claude.ai/code/session_01HoJtwsyrp6YC4DU9JAnc7H

---

Generated by Claude Code

[richshaw]

Review: passed (rounds 1–5)

This PR delivers the narrowed part of #503 within the redesign. Publish and share first sends any edits still being saved, then checks the draft on a fresh read:

• Nothing to sign up for (no slots, no fields, or every slot empty): it won't publish. Fill it in or Add a slot opens the slot in Build, from any tab.
• A slot still says TBD, TBC or TBA: it warns, with Name it first or Publish anyway.
• Nothing wrong: it publishes at once, as before.

The REST and MCP publish paths are unchanged. Rounds 1 to 4 were each an independent agent that didn't write the change: code-review at high effort as a starting list, each item checked by hand, and from round 3 a running-app check on its own dev server at 1280 and 390. Round 5 is three small commits (29703de, 5dcfaac, 9c4281e) that fix round 4's findings. I read those myself, and a final independent pass over the whole stack will look at them again.

Blocking (0). Earlier rounds had four, all fixed:

• Round 1: the fix's open-slot request lived in the address (?slot=new), so Back, Forward or a bookmark replayed it and added another slot. It now sits in the layout's BuildBridge, and the Build tab takes it once. e2e checks Back and Forward.
• Round 3: a pending Add a slot survived when the organizer stayed on the tab they pressed it on (clicked it again, or Settings' Save came back to the same path), so a later Build visit added an empty slot. The bridge now drops the request on a plain click on any tab but Build, on Back and Forward, and on any form submit. Round 4 re-ran the exact scenario with Build's load delayed 5 s, at both sizes: no slot.
• Round 3: docs/telemetry.md said a warning and a Publish anyway could be told apart by whether a signup.published follows, but fixing first then publishing gives the same rows. Publish anyway now writes acknowledged: 'placeholder' on the signup.published row, and the doc says so. Round 4 checked the rows for Publish anyway, a fix then publish, and a slot named in another tab, and that REST and MCP can't set the flag.
• Round 4: the CHANGELOG still gave round 3's old rule. Fixed in 29703de.

Fixed from optional findings (each marked with the round that found it):

• Flush gaps (round 1): a save sent as a Build tab unmounts, saves a departed tab left in flight, and a failed save are all waited for and counted, and a failed save stops Publish with "A change didn’t save. Fix it, then publish." A 10 s cap keeps a stuck save from holding Publish.
• A pending fix never expired (round 2): it's dropped when the path settles anywhere but Build, and on the triggers above.
• Check outcomes recorded (round 3): the check stops before the service, where #525 records refused publishes, so it writes the same signup.publish_refused row itself (recordPublishCheck, outside any transaction, after the write guard). A read failure there is caught and logged, so it never fails the publish. Tests assert the actor on each row. The doc says to count real refusals as rows whose payload has code, since warnings are rows too.
• A Ctrl/Cmd or middle click on a tab (round 4) dropped the fix while this window still went to Build. Only a plain primary click drops it now (5dcfaac, tested).
• Wording: "each slot's first such word", the "Add a slot before publishing." CHANGELOG line, the bridge's comments, and the doc's note that an API publish carries no flag either.

Optional, left as is (3):

• A Duplicate whose save takes longer than Publish's 10 s wait can take focus from a fix's slot when it lands (round 4, reproduced with the save delayed 12 s). The root is #568's code, which opens the copy whenever its save returns, whatever was opened since. main has the same steal when another slot is opened meanwhile. One follow-up should fix it there.
• A departed Build tab's failed save is dropped once a new Build tab registers. The new tab shows the stored values, so nothing wrong is shown, but the failure isn't mentioned again.
• A save slower than 10 s can leave the check looking at the stored values, so it may say a slot is empty that was just typed into. That's the cap working as designed.

The merges: PR 3's catch-ups with main (#525, #526, then #568) are merged in. Round 4 checked the #568 merge: a duplicate made just before Publish is waited for, and the check sees it. Duplicate still opens the copy right after the original with focus in What. A fix's focus wins over a duplicate's that saved within the wait.

Earlier threads: none.

Checks:

• CI is green on this SHA (9c4281e), and was on e8150c0.
• The tophat passed on 3952591 (comment above). Rounds 3 and 4 re-ran its flows on later heads, at 1280 and 390, with no console errors.
• Run locally on e8150c0: pnpm lint && pnpm typecheck && pnpm knip && pnpm test (1595), pnpm test:db (562), pnpm build && pnpm budgets, and the full chromium e2e suite with CI=1 (72). On 9c4281e: lint, typecheck, knip and unit tests (1599), pnpm build && pnpm budgets, and the chromium e2e for organizer.spec.ts and the publish walkthrough (20).

gate=owner: this is UX. It adds a check panel to Publish and share, opens slots from other tabs, and records each check's outcome.

https://claude.ai/code/session_01HoJtwsyrp6YC4DU9JAnc7H

---

Generated by Claude Code

[richshaw]

Final independent check: passed. An agent that didn't write the change checked round 5's three commits, which the review above says I had only read myself:

• 29703de: the CHANGELOG's acknowledged: 'placeholder' sentence matches the action. The flag is passed only when the check found a placeholder and the form said Publish anyway.
• 5dcfaac: the TabsNav guard is right in Chromium:
  • Enter on a focused tab sends a plain click, so a pending fix is dropped.
  • Ctrl+Enter and Shift+Enter, and Ctrl-, Shift- and Alt-clicks, carry their modifier, so the fix is kept.
  • A middle click sends only auxclick, so onClick never runs.
• 9c4281e: the telemetry wording is right. The REST publish uses the cookie session, so its row has no viaClientId, and the AI assistant's tool adds one.

It found three small things, now fixed:

• 9c44144: comments in TabsNav and on BuildBridge still said any tab but Build drops the fix. They now say only a plain primary click does. The middle-click test case fired a click that browsers never send for a middle click; it now fires auxclick, and there's an Alt-click case.
• 62db8e0: docs/telemetry.md's signup.link_copied row said "a draft's link only says the signup isn't ready". Since One Publish and share button, in a bar on phones; drafts show no public link #578 a draft shows no link to copy, so the row now says only older rows can carry draft.

I read both, and on 62db8e0 lint, typecheck, knip and unit tests pass (1600, with 18 in BuildBridge.test.tsx). CI is green on 62db8e0.

The smoke test of the whole stack at the top (#581), at 1280 and 390, covered this PR's part:

• An empty slot opens "Nothing to sign up for yet" with focus on Fill it in, and writes an all_blank row.
• A TBD slot warns, and Name it first puts focus in its field and writes the placeholder row.
• Publishing after the fix writes signup.published { from: 'draft', to: 'open' }, with no flag.
• No console errors.

Left as is, as the review says: a Duplicate still saving after a fix can take focus when it lands. The root is #568's code on main, and it wants a follow-up there.

Nothing else is left to do here; it's waiting on review.

---

Generated by Claude Code

[richshaw]

Tophat: two things in the publish check

I tophatted the stack locally and ran a UX walkthrough of Publish and share, on a computer and a phone. Both of these are on ad9fa96.

1. The spinner flashes before "Nothing to sign up for yet" opens

On a draft with nothing in it, pressing Publish and share turns the button into its spinner and "Publishing…", and almost at once the "Nothing to sign up for yet" panel opens. The spinner shows for a split second, so it looks like a glitch.

Why: in PublishButton.tsx, busy (saving || pending) shows <Spinner /> and "Publishing…" the moment the button is pressed. With nothing waiting to save, flushPendingSaves() resolves at once and the check comes back quickly, so busy is true only for a moment.

Suggested: show the spinner only when the answer is slow. Wait about 300 ms before showing it, and once it shows, keep it for at least about 400 ms so it never flickers. Nothing moves when the answer is instant, and a slow publish still gets a spinner.

Also, "Publishing…" isn't right when the answer is a check and nothing gets published. The delayed spinner hides that in the fast case. If a check can be slow, the label could fit both outcomes.

2. When the TBD is in the slot's name, the warning names the slot only by its date

A signup with the rows "Cashier TBD · Sat, Nov 14" and "Grill · Sat, Nov 14". Publish and share says:

Sat, Nov 14 still says “TBD”. People will see that, and a published signup can’t go back to being a draft.

It reads as if the date is TBD. Both rows have that date, so it doesn't say which one. (Name it first does open the right row, and the row's TBD tag marks it.)

Why: checkBeforePublish in src/lib/publish-check.ts names the slot by its other values (slotLabel), so the panel can quote the word itself. That works for the PR's example, "Fruit and water · Sat, Apr 6 still says “TBA”", where the TBA is a field's whole value. It doesn't work when the word is part of a longer value, such as the slot's name.

Suggested: leave a field out of slotLabel only when its whole value is the placeholder word. Then this case reads "Cashier TBD · Sat, Nov 14 still says “TBD”", and the PR's example doesn't change. Minor.

---

Generated by Claude Code

[richshaw]

Both are done.

1. The spinner flash

In 666ffd0, then 233333f and 4f0c7a0 from the checks.

When the spinner shows:

• After about 300 ms, through a small useDelayedFlag hook. Once shown, it stays at least 400 ms, unless an answer shows first: the spinner goes the moment the check's panel, the share panel or a refusal appears, so it never sits beside an open panel.
• The label is "Checking…" while it shows, true whether it then publishes or stops at a check.
• A second press is still refused at once (aria-busy and aria-disabled as before), even before the spinner shows.

Keeping the header still: both labels sit in one grid cell, so the button keeps its width and Preview doesn't move. Its accessible name is always the visible label.

Checked: an independent agent timed it frame by frame at 1280 and 390:

• An empty draft opened its check in about 120–250 ms, with no spinner.
• With the answer slowed to 400 ms, "Checking…" showed from about 300 ms until the panel's frame.
• The button stayed 159.9px wide in every frame, 358px in the phone bar.
• No frame showed both the spinner and the panel.
• After Escape, Not now, a refused save, or three quick presses (which sent one request), it neither stuck nor came back.

2. The TBD in the slot's name

In 6962a50, then 5d619eb and b3316f9 from the checks.

The rule: a field is left out of the label only when its whole value is the placeholder word, give or take punctuation, or only such words.

• "Cashier TBD · Sat, Nov 14" now reads "Cashier TBD · Sat, Nov 14 still says “TBD”."
• "(TBD)", "TBD.", "TBD/TBC" and " tbd " are still left out, so "Fruit and water · Sat, Apr 6 still says “TBA”" doesn't change.
• The tests cover each.

Checks:

• Lint, typecheck, knip and unit tests pass (1675 on f95e870).
• actions.db.test.ts and the e2e for organizer.spec.ts and the publish walkthrough pass.
• Budgets: each signup tab grows about 220 bytes (the delayed flag, the answer tracking and the stacked label). Build is 157.98 of 158 kB, with no budget raised.
• CI is green on 84f146e, and running on the head.

---

Generated by Claude Code