UI/UX designer seller: the model designs, the review pipeline is tooling

[maxy-player]

What this is

A runnable UI/UX designer seller package, not a persona file. A real model harness, driven
through maxplayer's own local driver, is handed a natural-language brief and authors the design
itself — then renders it, inspects its own screenshots, and audits its own accessibility.

This is the corrected round. The previous revision was rejected because its "designer" could only
accept a supplied candidate and rewrite it deterministically — it could not create a design. That
pipeline is retained here but demoted to tooling.

The gate

cd sellers/ui-ux-designer
cargo build -p maxplayer --features acp --release   # from repo root; required by check 3
npm ci
bash scripts/smoke.sh

Last verified run: 9 checks run, 0 failed, SMOKE GATE PASSED.

Expect several minutes — the gate spends a real model turn letting the agent design something.

The generative path (check 3)

briefs/nova-status-brief.md is prose, phone-first, with red/green-colour-blindness and
screen-reader constraints. No candidate implementation is supplied or present in the tree. The
model must author status.html, render it, read the tool output, and iterate to zero accessibility
violations at both viewports.

It runs through the actual maxplayer local driver — no relay, wallet, sats, or deployment. The
stock release refuses this path (maxplayer run requires rebuilding with the acp feature), so the
binary is built from source.

From the passing run (runs/generative-smoke/summary.json, verdict PASS, failures: []):

Evidence	Result
Driver	exit 0, job.execution_changed → completed
Deliverable	status.html authored by the model, ~14 KB, not byte-identical to anything shipped
Agent-invoked tools	shoot.mjs true, a11y.mjs true
Accessibility, desktop	0 violations, 15 passes, 63 rules, axe-core 4.11.4
Accessibility, mobile	0 violations, 15 passes, 63 rules, axe-core 4.11.4
Load proof	HTTP 200, title "Nova service status", 225 elements
Screenshots	desktop 122,207 B, mobile 291,881 B, both content-asserted

agent/design-run.mjs refuses to report success without evidence: the run must finish, the
deliverable must exist and be non-trivial, the driver log must show the agent invoking the
tools, and the agent's own axe JSON must show a real page load and zero violations at both
viewports. --verify-only re-checks finished evidence without spending another model turn.

Harness — probed, not assumed

Presence on PATH is not availability, so each adapter was driven over ACP stdio:

Adapter	Model answered	Verdict
claude-agent-acp	no	refused: errorKind: rate_limit, org monthly spend limit — reproduced twice, an hour apart
codex-acp	yes	used — gpt-5.6-sol, stopReason: end_turn, non-zero output tokens
claude-code-acp, cursor-agent, cursor-agent-acp, goose	—	absent

Detail in sellers/ui-ux-designer/evidence/RUNTIME.md.

Four negative controls, each of which must exit nonzero

• an empty page that axe scores as 0 violations is rejected by the load-proof guard;
• a 404 URL yields no screenshot;
• two identical images fail a required-change diff;
• a generative run whose status.html has been deleted is not reported as a pass.

A gate that cannot fail proves nothing.

A detector bug worth recording

The first generative run passed on the merits, yet the gate failed it, claiming the agent "never
rendered its design". The driver writes two streams: events.jsonl nests its type under
.payload and never carries tool calls, while the ACP tool calls and turn_ended arrive on driver
stdout. Reading only the first produced a confident false accusation against correct work. The
gate now reads both.

Limitations (not exhaustive — see report §7)

• No live relay, wallet, sats, mint, or seat. Seller registration, job-claim, delivery and
  payment remain entirely unexercised. This proves the agent and its tools, not that the seat earns.
• Output varies between runs; the gate is not deterministic and costs real model tokens.
• Single harness in practice — codex-acp alone. If that account hits its limit the check fails
  loudly rather than falling back to deterministic execution.
• No human designer has judged the output. Zero automated violations is a floor, not taste.
• Docker sandbox mode untested; the design identity "Plainsong" is proposed, not adopted.

Holds observed

No wallet, no sats, no paid service, no new subscription, no production deployment, no merge, no
tag, no push to main. Delivery is a branch on the fork plus this PR. Git identity was set
per-command; git config was never run.

[vercel]

Someone is attempting to deploy a commit to the MakePrisms Team on Vercel.

A member of the Team first needs to authorize it.