2026-07-29 — verex history

Source docs: docs/tasks/jul-28-verex.md → docs/tasks/jul-28-verex-design.md (implemented on claude/jul-28-features); staging deploy per docs/runbooks/deploy.md.

Staging deploy: jul-28 branch live on Cloud Run + Sepolia

Deployed claude/jul-28-features to staging per the runbook — no contract redeploy (the existing Sepolia backbone from deployments.json staging was reused; only off-chain code changed) and the existing demo mnemonic was kept. The seed ran idempotently against the live backbone: the 10 existing markets skipped prepare/register and topped up inventory, the 21 new group members were created fresh (~15 min of Sepolia txs, ~0.011 ETH at ~1.1 gwei), MM books posted. 5 new DB migrations applied through the Cloud SQL proxy.

New this deploy: gcloud run services update verex-api --no-cpu-throttling --min-instances 1 (revision verex-api-00007-c6m) — required so the in-process ChainJob worker keeps CPU between requests; settings persist across future revisions.

Verified live on staging: health OK, 3 groups OPEN with quoted books, one real trade (11.11 Yes @ 45¢) filled instantly in the DB and settled on Sepolia via matchOrders (0xba1e5165…e1157), position visible in the portfolio.

• API: https://verex-api-q6qvjcw5ma-du.a.run.app
• Web: https://verex-web-q6qvjcw5ma-du.a.run.app

Market detail logo + edit page (image URL / rules / category)

No task/design doc — implemented directly from jay's chat request (with a Polymarket screenshot as the visual reference), on branch claude/jul-29-market-edit.

The market detail page now shows the market logo (56px, vs 36px on the grid cards) next to the title, preferring market.imageUrl over the seeded picsum fallback — MarketCard got the same imageUrl ?? marketThumbnail fallback that GroupCard already had. The Create page doubles as an edit page via /create?edit=<slug> (linked from a pencil icon on the detail page): image URL, rules, and category are editable through a new PATCH /markets/:slug. Permission split per jay's spec: the operator (#0) may only change the image URL — rules and category are the creator's copy — enforced in both the UI (disabled fields + banner) and the API (403).

Follow-up: market editing flipped to operator-only

jay clarified the intent ("Only Operator can edit"): the first cut let any wallet edit all fields while restricting the operator to the image URL — inverted. Now PATCH /markets/:slug rejects every accountIndex except 0, the edit form disables all fields (with a banner) for non-operator wallets, and the detail page's Edit link renders only for the operator (new EditMarketLink client component, since the page itself is a server component).

Fix: homepage Featured card ignored the edited image URL

jay's edited logo didn't show on the main page — the market he edited was the highest-volume one, which renders as the homepage's Featured card, and that card still called marketThumbnail(slug) directly. The imageUrl-fallback sweep had covered MarketCard/GroupCard/detail but missed this fourth call site. All four now use imageUrl ?? marketThumbnail(slug).

Fix: group detail page had no logo at all

Follow-up from jay: the group detail page (/group/[slug]) never rendered a logo — unlike the market detail page it simply had no <img>. Added the same 56px imageUrl ?? marketThumbnail treatment next to the group title; verified with the Oscars group (custom URL) and the World Series group (picsum fallback).

Fix: group cards get the signature probability bar

jay flagged (with a screenshot) that grid cards were inconsistent: MarketCard has the top gradient bar driven by the Yes price, GroupCard had none. Group cards now render the same bar with width = the leading outcome's probability (members[0].quoteCenter).

Group cards: stacked per-outcome bar; groups get the edit feature

Two more follow-ups from jay. (1) The group card's top bar now stacks one segment per member's Yes probability using the group chart's palette — moved to GROUP_COLORS in lib/utils.ts so chart and bar can't drift; the grey track shows through as the "+N more" remainder. (2) Groups are now editable like markets: PATCH /market-groups/:slug (operator-only; category cascades to member markets, mirroring creation), the edit page takes ?editGroup=<slug>, and the group detail header gets the operator-only Edit link.

Prod deploy: jul-29 branch live on verex.jaylabs.xyz

Deployed claude/jul-28-features (through 4cdb36c) to prod per the runbook day-2 path: SKIP_SEED=1 ./scripts/deploy-prod.sh — migrate ran, seed skipped, API image built (3m17s), both services redeployed. Also applied the staging CPU fix to prod: gcloud run services update verex-api-prod --no-cpu-throttling --min-instances 1 (revision verex-api-prod-00003-t5s) — prod's rev 00001 predated that discovery. Verified live: health OK, both PATCH routes respond 403 to non-operators, market pages render the 56px logo, domain serves the new revision.

Data gap found: prod has the 10 original binary markets and zero groups — its one-time seed predates the group feature (staging got its 21 group members via the jul-29 idempotent re-seed). Group features are live in code but have nothing to render. Fix requires re-running the real seed against the prod backbone — destructive (wipes prod Trade/PricePoint/Outcome/Market rows, ~15 min of Sepolia txs from the prod operator) — parked for jay's decision.

Seed script: bake in the prod-curated market logos

jay hand-picked logo URLs for all 10 binary markets on the live prod site (via the new edit feature); those exact URLs are now in seed.ts (SeedMarket.imageUrl, threaded through both the DB-only and on-chain create paths) so any future seed — including a prod group re-seed — reproduces them instead of losing them to the picsum fallback. Verified by diffing seed.ts against the prod API per slug: all 10 match.

Prod: featured market switched to the Coachella K-pop market (volume bump)

jay wanted kpop-headliner-coachella-2027 featured on prod. Featured is derived (highest volume), so per jay's call we set that market's volume to 6,000,000 directly in the prod DB (Cloud SQL proxy + Prisma; ETH was at 5,620,010) instead of adding a FEATURED_SLUG env override (drafted, then reverted — jay preferred the data route). Side effect, accepted: the market now displays $6M Vol and tops the Hot list. No code change.

Prod re-seed: groups created, empty books refilled — liquidity error resolved

jay hit "no liquidity at this price" on prod and prod had no groups, so we ran the real seed against the live prod backbone (jay-approved; DB wipe accepted). Result: 10 binaries re-created (existing conditions skipped, inventory topped up) + 3 groups / 21 members created fresh on Sepolia + MM ladders posted for all 31 books; curated logos survived via the seed change. Verified: books show 5×5 depth, and the previously-failing case (kr-world-cup Yes BUY) filled at 32¢ and settled on-chain (0x441b1a94…6317) — the settlement also exercises the new no-cpu-throttling worker. Root cause of the drained books: pre-fix CPU throttling froze the in-process re-quote worker between requests. The kpop featured volume bump was wiped by the reset; re-apply pending (permission gate).

Local re-seed + first-visit welcome overlay

jay asked for a local reset ("seed again" — a prod re-run was started by misreading and stopped harmlessly during idempotent operator setup, before the DB wipe). Local seed ran clean: 10 binaries + 3 groups, curated logos, fresh MM books on anvil. New WelcomeOverlay (root layout, localStorage-flagged) tells first-time visitors to trade with the Demo Wallets and resolve with the Operator Wallet — verified show/dismiss/stay- hidden locally; committed 446f103 and deployed app-only to verex-web-prod. Dev-server gotcha again: adding a new component file mid-run 404'd Next's own chunks → clean-.next restart.

Staging parked: down/up scripts, ~95% cost cut while idle

jay asked to stop paying for idle staging. New scripts/staging-down.sh / staging-up.sh (runbook §10 bullet added): down = verex-api scale-to-zero + cpu-throttling restored + verex-db stopped (~$40-65/mo → ~$1-2/mo, storage/images/ secrets kept); up reverses in ~2 min. Deliberately all-or-nothing — a throttled API with a live DB would drain the MM books (today's prod incident). Ran down: verex-db STOPPED, API min-instances 0, prod verified untouched.

Gotcha: next build while next dev is running corrupts .next

Running a production next build in packages/web while jay's dev server was up broke the dev server ("Cannot find module './vendor-chunks/tailwind-merge…'") — both write to the same .next directory. Fix: kill the dev server, rm -rf .next, restart. Avoid verifying web changes with pnpm build when a dev server is already running on the machine.