ZAC WHITE · OPENAI HACKATHON · BUILD LOG 001
GenUI is a bounded protocol and native renderer for adaptive, interactive SwiftUI. One task. Twenty-six specialists. Every milestone tied to tests, commits, or live UI.
00 / ORIGIN STORY
I’m Zac White. At an earlier OpenAI Codex Hackathon ↗, I used GPT‑5.3‑Codex to build SwiftUIGen: an early attempt to turn model output into native SwiftUI.
It proved the shape of the idea, but stayed primitive. GenUI is the reboot: a protocol, renderer, app, corpus, and compiler built as one verifiable system.
Could GPT‑5.3‑Codex sketch the idea in a day?
NOWWhat can GPT‑5.6 Sol build when it can hold the plan, delegate, use the toolchain, and keep iterating?
The ambition stayed the same. The depth, evidence, and finish changed.
01 / CHRONOLOGY
Ten consequential turns, backed by task history, commits, tests, and live verification.
GenUI begins with one constraint: model-authored JSON must become real SwiftUI, with native layout, accessibility, state, and theming—not a web view or a generated image.
Codex builds the decoder, validator, renderer, diagnostics, adaptive layouts, and live JSON workbench. The first day ends with a usable loop from document to native preview.
Bounded specialist runs add the component registry, themes, state, controls, forms, resources, and repeated subtrees. Shared tests keep the pieces aligned as one model-facing contract.
Thirty canonical examples become executable documentation inside a searchable gallery and inspector. Screenshot and Simulator feedback then harden compact layouts, navigation, visibility, and a device-only stack overflow.
The executable catalog grows into a resumable generation pipeline with validation, render probes, quality gates, retries, and independent review. Smoke runs turn weak outputs into deterministic rejection rules.
Every stop becomes code: decorative bindings, fake repetition, duplicated constraints, stale values, image ratios, coverage skew, malformed batches, and hung judges all gain mechanical checks before the run can resume.
The guarded run finishes with 4,000 compile examples and 1,000 repair examples. Exact accounting, unique inputs, balanced coverage, checkpoint 50 review, and final verification all pass.
The corpus feeds a bounded generate_ui specialist with local validation, two repair attempts, Wavelike routing, and a 64-case evaluation harness. An MLX canary trains the longest 6,820-token record successfully, but uses swap—so no pilot starts.
Hugging Face publishes all 5,000 accepted examples as a public, ungated CC BY 4.0 dataset. The 4,253 / 457 / 290 splits match the verified local export; training remains a separate decision.
The iOS app streams assistant text, shows Thinking and Building interface states, routes output UI to the assistant and input UI to the composer, and repairs common path, wrapper, and truncation failures deterministically. A confirmed Clear action resets the thread. The tested hardening now lives on origin/main.
02 / THE AGENTIC DIFFERENCE
The differentiator is the loop: intent, delegation, tools, evidence, correction.
Long-horizon continuity
Architecture, feedback, screenshots, tests, and 119 commits remain connected across the original task and its continuations.
26 specialist runs
Specialists take narrow slices while the parent owns shared contracts, integration, and verification.
Remote, durable operation
While I golfed in Sonoma on my birthday weekend, Codex remotely supervised the synthetic-data run on my home Mac—checking tmux, handling review gates, fixing real stops, and resuming safely.
Proof before progress
Reviewer findings, test failures, and screenshots become gates. A milestone ships only when the artifact matches the claim.
THE MODEL IN THE LOOP
It preserved the architecture, coordinated specialists, operated the Apple toolchain, and kept working until the evidence matched the claim.
“The parent task moved between product, engineering, testing, and build operations without dropping the thread.”
OBSERVED IN THIS BUILD
03 / CURRENT SIGNAL
Only the current proof, decision, and next gate.
Hugging Face serves the verified 4,253 / 457 / 290 splits as a public, ungated CC BY 4.0 dataset.
PUBLIC · UNGATED · CC BY 4.0Assistant prose streams into a pending turn, interface generation exposes a clear progress state, and deterministic recovery handles common model-output failures. The tested commit is now on origin/main.
ORIGIN MAIN · 29/29 TESTS · STREAMING LIVEThe MLX canary is technically viable but misses the zero-swap gate. No local or paid pilot is running.
MLX FEASIBLE · ZERO-SWAP MISSED · NO PILOT RUNNINGHOW THIS RECORD STAYS HONEST
The “GenUI Implementation” continuations.
Commits, tests, and live artifacts.
Keep only consequential milestones.
Republish when the evidence changes.