HIKYAKU

Inspiration

Most fitness apps ask for commitment before they give anything back. We wanted one that asks for five minutes. In Edo Japan a hikyaku was a courier who ran messages across the country on foot — a good frame for a short walk that matters.

What it does

You choose your time and energy, and HIKYAKU issues an order with a real target distance derived from those two choices. You walk it; the route rail lights checkpoint by checkpoint. On arrival you get a sealed record.

Then you log what you ate. We look the food up in Open Food Facts. Real food data is incomplete, so nutrients are often missing — instead of hiding that, we fill the gaps with GPT-5.6 and disclose the source of every single number: Open Food Facts, GPT-5.6, or a deterministic fallback. Nothing is presented as measured when it was estimated.

Judges do not have to walk. Judge Demo mode completes a full mission at a desk, and simulated distance is labelled as simulated everywhere it appears.

How we built it

React 19 + TypeScript + Vite on the front end, a Cloudflare Worker behind it. No database, no login, no sign-up.

• gpt-5.6-sol fills nutrition gaps and scores the meal, using strict Structured Outputs so responses always match our schema.
• gpt-5.6-terra writes the mission narrative, so the story changes with your choices instead of repeating a template.

Codex built it. Product code was written by Codex Sol running in parallel git worktrees, one lane per feature, merged through an integration branch. Codex also generated the competing UI design directions we chose from, produced art direction, and audited its own output. The commit history is the record.

Challenges we ran into

The honest ones. An early build scored a Kit Kat above tofu because the metric rewarded contribution without penalising excess. Nutrient references were applied at the wrong scale, so everything read as "too much". The core loop was briefly unreachable because tab routing shadowed the journey state machine. Each was found by driving a real browser and measuring, not by reading the code and assuming.

The hardest discipline was refusing to fabricate. No invented rivals, no fake leaderboards, no populations. Where a value is unknown the app prints an em-dash and says why.

Accomplishments that we're proud of

Per-nutrient provenance. A nutrition app that admits what it does not know is more useful than one that quietly guesses, and we could not find another entry doing this.

It also just runs: a fresh clone, one command, no API key required. If the endpoint is unreachable it falls back to local generation and says so on screen. 105 tests pass.

What we learned

Code-level review does not catch what a human sees. Every serious UI problem in this project was found by looking at a rendered screen or measuring element geometry — never by an audit that only read source.

What's next for HIKYAKU

Opening the locked flag grounds, real GPS runs beyond the demo path, and widening the provenance model so every derived number in the app carries its source the way nutrients already do.

Built With

• cloudflare-workers
• codex
• open-food-facts
• openai
• playwright
• react
• typescript
• vite
• vitest