art-bargain

A multi-category art marketplace where every listing carries an LLM negotiation agent. Chat with the seller's agent yourself, or dispatch your own buyer agent and watch them negotiate — streamed live, anchored to comparable-sale data, with an anti-stall judge that breaks deadlocks.

Live: art-bargain.vercel.app · Status: portfolio / technical demo — not a commercial SaaS.

(End to end: log in → browse the gallery → open a listing → counter the seller's agent yourself → dispatch a buyer agent → watch both agents converge — the buyer opens, the seller counters, an anti-stall judge breaks deadlocks. This run settled at $164.)

---

Why this exists

Most "AI agent" portfolio pieces are notebook demos. This one ships a vertical slice of a real product so you can poke the seams:

• Two agents negotiate over SSE, not over a notebook cell
• Real Postgres + RLS — every read is auth.uid()-gated, no service-role leaks to the browser
• Tool use with grounded sources — agents fetch comparable sales from a real sales_comps table before quoting
• Tri-lingual (EN / 中文 / 日本語), cookie-driven, Route-Handler-based — no broken Server Action shenanigans
• TDD on the parts that matter (price-validation / anti-stall judge / coordinator state machine); smoke tests on the parts that don't

Everything runs against claude-sonnet-4-6 with streaming + tool use, on Vercel's default Node runtime (Edge can't sustain the 60-90s dual-agent flow).

---

Architecture: dual-agent coordinator

buyer client ──POST /api/nego/[id]/auto──► coordinator
                                           │
                          ┌────────────────┴────────────────┐
                          ▼                                 ▼
                   buyer agent prompt              seller agent prompt
                   (style/target/max)              (artwork + RLS rows)
                          │                                 │
                          └────► sales_comps tool ◄─────────┘
                                       │
                          ┌────────────┴────────────┐
                          ▼                         ▼
                  anti-stall judge          price-validation
                  (mediation injection)     (anchor + floor)
                          │
                          ▼
                  SSE: turn_delta / turn_done / negotiation_done

Each agent gets its own prompt, its own memory of the conversation, and its own tool-use budget. They don't share scratchpad. The coordinator orchestrates them but does not whisper to either.

Failure modes the coordinator handles:

• Stalling (same offer repeated 3×) → judge proposes a midpoint
• Cheese moves (buyer agent offers $1, seller offers $99999) → price-validation rejects pre-emit
• Hard timeout (90s) → negotiation_done with status=stalled
• Sibling-batch SSE flush stays narrow per stream — see src/lib/agent/coordinator.ts

---

Stack

Layer	Choice
Framework	Next.js 16 (App Router, Turbopack)
UI	Tailwind v4 + shadcn (base-ui preset)
Auth + DB + Storage	Supabase (Postgres 17, RLS-strict)
LLM	Anthropic claude-sonnet-4-6 (streaming + tools)
Tests	Vitest + MSW (Anthropic mocked at HTTP boundary)
Deploy	Vercel
i18n	Cookie + Route Handler + per-locale dict

---

What I'd build differently next time

1. Don't use Server Actions for "set cookie + reload" flows. They render as javascript:throw... placeholders under Turbopack in some bundler states. Used a plain Route Handler + form POST instead — see src/app/api/locale/route.ts.
2. Test the real click, not form.submit(). form.submit() bypasses React's onClick, hiding a class of "form unmounted before submit" bugs. Spent three "fixes" chasing a phantom before catching the unmount race.
3. Pass locale not dict across the RSC boundary. Dict objects contain functions ((n) => string), which can't be serialized server→client. Server passes locale, client component calls dictFor(locale) itself.
4. Disable Supabase "auto-expose new tables" if you're going to forget about grants. I did. Twice. Now there's a migration that explicitly GRANTs to anon / authenticated / service_role.

---

Local development

Prerequisites

• Node ≥ 20 (tested on v22)
• pnpm 10+
• Supabase project (free tier)
• Anthropic API key

Setup

pnpm install
cp .env.example .env.local   # fill in real values
pnpm dev

App runs at http://localhost:3020.

Env vars

All keys listed in .env.example. Build fails fast if NEXT_PUBLIC_SUPABASE_URL / NEXT_PUBLIC_SUPABASE_ANON_KEY are missing or contain whitespace.

Use printf not echo to set values from the shell — echo appends a trailing newline that breaks Supabase auth silently.

Scripts

Command	Purpose
pnpm dev	Dev server on :3020
pnpm build	Production build
pnpm start	Production server on :3020
pnpm lint	ESLint
pnpm typecheck	tsc --noEmit
pnpm test	Vitest (incl. MSW-mocked Anthropic)
pnpm format	Prettier write

Pre-commit hook runs lint-staged + typecheck.

---

Design docs

• Spec: docs/superpowers/specs/2026-05-17-art-bargain-design.md (includes T7 SSE spike — default Node runtime sustains 77s+, not the 25s Edge limit)
• Plan A (Foundation): docs/superpowers/plans/2026-05-17-plan-a-foundation.md
• Plan B (Negotiation core): docs/superpowers/plans/2026-05-17-plan-b-negotiation.md
• Plan C (Dual-agent): docs/superpowers/plans/2026-05-18-plan-c-dual-agent.md

---

License

MIT — clone, fork, lift the coordinator pattern. If you ship something interesting, let me know.